Comparative Advantage

Stanford Undergraduate Economics Journal

Spring 2017

Volume 5

1

Note from the Editor

On behalf of the Comparative Advantage Editorial Board, I am honored to present the fifth

volume of the Stanford Undergraduate Economics Journal.

The journal has grown tremendously this year in its e↵orts to create an accessible platform

for readers to engage with our authors’ work. Our new website is designed to mirror the

layout of professional academic journals, with individual articles available on separate pages

and the journal displayed in its entirety in the Archives section. We additionally created

the Comparative Advantage Blog to make our publication more approachable for younger

readers. Shorter length submissions, literature reviews, and opinion pieces are published

directly on the website in this section. As we are constantly developing new ideas to

improve the journal, we appreciate any feedback our readers may have.

Although the outward aspect of the journal has changed, we have maintained our

dedication to rigorous research. This year, our publication contains 7 original research

papers on a diverse set of topics that our reviewers found thoughtful and compelling. In our

selection process, we emphasize both empirical analysis and theoretical foundations, and we

believe our final result will be valuable to the undergraduate economics community.

Finally, we would like to thank the Stanford Economics Association (SEA) and the

Stanford Economics Department for their continued support and the Stanford Institute for

Economic Policy Research (SIEPR) for collaborating with us on our new initiatives this

year. We are excited to resume our partnerships in 2017-2018 and hope you will join us.

Laura Zhang

2017 Editor-In-Chief

2

Editors and Sta↵

Editor-in-Chief

Laura Zhang

Production and Design Editors

Aakash Pattabi

Vidushi Singh

Associate Editors

Udai Baisiwala

Toren Fronsdal

Matthew Galloway

Maya Ganesan

Raymond Gilmartin

Spencer Guo

Barrett Medvec

Genevieve Selden

3

Contents

1 How Does Technology A↵ect Skill Demand? Technical Changesand Capital-Skill Complementarity in the 21st CenturyYifan Gong 5

2 Has Indonesia’s Growth Between 2007-2014 Been Pro-Poor?Evidence from the Indonesia Family Life SurveyAriza Gusti 14

3 Place and its Role in Venture Capital FundingLuke Heine 23

4 Temporary Assistance with Lasting E↵ects: A Report on Poli-cies of Self-Determination in Native AmericaRyan Mather 31

5 Does Intercropping Improve the Outcomes of Export Assis-tance Among Kenyan Smallholders?Noah Nieting 45

6 The Influence of Collectivism on Microfinance in SenegalCole Scanlon, Keaton Scanlon, and Teague Scanlon 56

7 Vocabulary as an indicator of creditworthiness: An analysis ofpublic loan dataJustin Wagers 66

1

4

How Does Technology Affect Skill Demand? Technical Changes andCapital-Skill Complementarity in the 21st Century

Yifan GongAdvisor: Professor Mario Solis-Garcia

Macalester College, Department of Economics

Abstract— This paper attempts to examine technology’simpact on the labor market through the lens of skilled labor.Technical changes in the late 20th century are skill-biased innature, because they are found to complement with skilledlabor who are adept at adopting new technologies. However,recent studies document a lower demand for high-skilled laborin the 21st century, compared with the late 20th century. Aretechnologies starting to substitute for human skills instead ofcomplementing them? Drawing on the wage share data from1975 to 2015 for 18 sectors in the United States, I find strongand robust evidence of complementary relationships betweentechnical changes and demand for skilled labor. Furthermore,my results suggest that technologies have become more skilled-biased, not less, in the 21st century.

I. INTRODUCTION

This paper aims to shed light on the the relationship be-tween technological changes, capital and skill demand in the21st century. It attempts to explore how recent technologicalchanges affect the demand for skilled labor, and how thatrelationship varies over time and across industries.

The concern over new technologies destroying jobs is nota new one. Numerous scholars have expressed concernsover the impact of recent technological changes on thelabor market. In his 2014 book The Second Machine Age,

Brynjolfsson argues that while the Industrial Revolution,or First Machine Age, is all about power systems to aug-ment human muscle, in the Second Machine Age we arebeginning to automate a lot more cognitive tasks, a lotmore of the control systems that determine what to usethat power for. According to Brynjolfsson, computerizedmachines nowadays are so smart and powerful that theywill start substituting for skilled human labor rather thancomplementing it.

Along these lines, Karabarbounis and Neiman (2013)study labor market data in 59 countries from 1975 to 2012and observe downward trends in the labor share for 42 ofthem. These findings lead to a widespread concern that therelationship between capital and skill is changing in the 21stcentury, as machines start replacing human labor at the topof the skill distribution. Do recent technological changeschallenge the capital-skill complementarity assumption?

There is also empirical evidence revealing a lower demandfor skilled labor within the last decade. Economists fre-quently characterize information and communication tech-nologies (ICT) in the 21st century as skill-biased in nature

because they favor skilled workers who are suitable toadopt new technologies over unskilled ones. However, somerecent studies find evidence against the complementaryrelationship between technologies and skilled labor. Beaudryet al. (2013), using data from the Outgoing Rotation GroupCurrent Population Survey Supplements for the years 1979-2011, document a decline in the demand for cognitive tasksand highly-educated workers from 2000 to 2010. Is thisreversion temporary, or does it signify a change in the skill-biased nature of technological changes?

This paper answers the two questions above with a datasetcovering the years 1975 to 2015. The rest of the paper isorganized as follows: Section II reviews studies relevant tothis subject. Section III explains economic of the capitaland labor market. Section IV describes empirical modelsestimated in the paper. Section V presents data and summarystatistics. Section VI summarizes empirical results. SectionVII concludes and points out areas for future studies.

II. LITERATURE REVIEW

A. A Capital-Skill Complementarity

The role of skill and education in a production functionwas tested first by Griliches almost 40 years ago (Griliches1969). Drawing on post-World War II data from U.S.manufacturing sectors, Griliches finds a positive relationshipbetween capital employment and skill demand. He formal-izes this phenomenon as capital-skill complementarity, a hy-pothesis stating that physical capital is more complementaryto skilled labor than to unskilled labor. Fallon and Layard(1975) confirm this hypothesis with data at both aggregateand sectoral levels.

However, capital and skilled labor have not always beencomplements. Studies drawing on data from the late 19thcentury reveal evidence to the opposite. Cain and Paterson(1986) examine the U.S. manufacturing sector from 1850 to1919 and find that physical capital complements with rawmaterials and substitutes for skilled labor. In the same vein,James and Skinner (1985) divide manufacturing sectors inthe 1850s into skilled and less skilled sectors. They findstrong complementary relationship in the skilled sector butrelative substitutability in the remaining sectors.

Summing up, while post-World War II data reveals acomplementary relationship between capital and skill, in the

5

late 19th century capital is found to substitute for skilledlabor at industry levels.

B. Skill-Biased Technical Change (SBTC)

1) Evidence for the United States: After capital-skill sub-stitutability in the 19th century and capital-skill complemen-tarity in the mid 20th century, the late 20th century is knownas a period with growing demand for skilled labor (i.e., skillupgrading). Most studies attribute the accelerated skilledupgrading in the late 20th century to skilled-biased technicalchange (SBTC). SBTC, also known as the technology-skillcomplementarity, is a shift in the production technology thatfavors skilled over unskilled labor by increasing its relativeproductivity and, therefore, its relative demand (Violante2008). Unlike the capital-skill complementarity hypothesis,the technology-skill hypothesis supports the complementar-ity between new capital (e.g., technology-embodied capital)and skilled labor.

The paper of Berman et al. (1994) is among the firststudies that examines SBTC empirically. Relying on datadrawn from the Annual Survey of Manufactures (ASM) inthe 1980s, Berman et al. identify an increasing share ofskilled labor in total employment within the 450 industriesin U.S. manufacturing. Through an econometric analysis thatrelates the shift in favor of skilled workers to production-labor-saving technical change, they confirm the SBTC hy-pothesis. They attribute the increasing wage share of skilledlabor in American manufacturing in the 1980s to the levelof investment in R&D and computers.

Autor et al. (1998) extend Berman et al.’s study by addingmore sectors over a longer period. They link educationalwage-bill share data with computer utilization records fromthe Current Population Survey for years 1960 to 1990. Theyfind a positive relationship between growth in computerusage and skill upgrading for 47 U.S. private industry sectorsstarting in the 1970s. Although the strong correlation is byno means a causal relationship, their findings are valuablein pointing out that the skill upgrading in the U.S. has beenconcentrated in the most computer-intensive sectors.

2) International Evidence: Empirical studies outside theUnited States reveal mixed findings. On the one hand,studies of OECD countries strengthen the SBTC hypoth-esis. Machin and Reenen (1998) study the changing wageshare and employment in seven OECD countries (UnitedStates, Denmark, France, Germany, Japan, Sweden, and theUnited Kingdom). All countries show a shift in relativelabor demand in favor of skilled labor and significantcomplementarity of capital with new technology. In thesame vein, Michaels et al. (2010) update the model bycategorizing labor into low, middle, and high educatedworkers. Using a panel dataset covering the U.S., Japan, andnine European countries from 1980-2004, they find strongcorrelation between the growth in ICT and the growth in thedemand for the most educated workers. They conclude thattechnological changes since the 1980s can account for upto 25% of the growth in the demand for college-educatedworkers.

On the other hand, countries in the Asia-Pacific region areless influenced by the diffusion of skill-biased technologies.Berman et al. (2003) study the manufacturing sector inIndia during the 1980s. They find that India does not showsignificant growth in the demand for skilled labor that iscommon to other high-income countries. They also findthat increased capital investment can explain very littleof the increased wage share of skilled labor in Indianmanufacturing sectors.

Summing up, countries outside the United States showinconclusive evidence regarding SBTC. Despite a rich bodyof literature on this topic, there remain few studies on theeffect of ICT and capital on the skill demand in the U.S. after2000. The next section describes a theoretical frameworkthat can be used to test the capital-skill and technology-skillcomplementarity hypotheses in the 21st century.

III. THEORY

At the industry level, the shift away from unskilled toskilled labor can happen between and within industries. Inthe former case, trade and immigration are likely to cause la-bor to shift away from less-educated and import-competingsectors. In the latter case, skill-biased technological changescould reduce the demand for unskilled labor and increasethe demand for skilled labor within an industry. To explorefactors that might explain within-industry changes in theskilled labor’s employment share, I start from a firm’s costfunction. Following the practice of Berman et al. (1994)and Autor et al. (1998), I assume heterogeneity of laborby categorizing it into skilled and unskilled labor groups.I also assume that the firm’s capital input is quasi-fixed. 1

Therefore, the firm’s variable cost function is

CV (Ws,Wu,K,Q) (1)

where Ws and Wu are the wage rates of skilled and unskilledlabor, K stands for quasi-fixed capital, and Q represents realoutput.

Drawing on Berman et al. (1994), Machin and VanReenan (1998), and Meschi et al. (2008), a translog func-tional form 2 of the variable cost implies

ln(CV ) = ↵0 +X

i=s,u

�i ln(Wi)

+X

i=s,u

X

j=s,u

�ij ln(Wi) ln(Wj) + �y ln(Y )

+X

i=s,u

�iy ln(Y ) + �k ln(K) +X

i=s,u

�ik ln(K)

(2)

where the � parameters denote the effect of factor prices offactor prices, output, and the capital stock over total variablecost. Following Shephard’s lemma, the cost-minimizingdemand for an input can be derived by differentiating the

1Quasi-fixed capital assumes the capital to be fixed in the short-run.2A translog functional form provides a second-order approximation to a

Cobb-Douglas production function and does not impose any restriction onthe substitutability of various inputs.

6

cost function with respect to the factor price. Therefore, theshare equation for skilled labor can be derived as

Sti = �0+�1 ln(Wsti

Wuti)+�2 ln(

Kti

Qti)+�3 lnQti+ ✏ti (3)

where t indexes year, i indexes industry, and ✏ti is the errorterm.

In equation (3), the sign of �1 depends on whether theelasticity of substitution between skilled and unskilled laboris larger than 1. Estimates of �2 indicate the relationshipbetween capital and skilled labor: capital and skill are com-plementary inputs if �2¿0 and substitutes if �2¡0. Estimatesof �3 show the relationship between growth in output andthe wage share of skilled labor.

To account for the impact of technologies, I augmentequation (3) by including a new variable TECHti torepresent technology-embodied capital stock in industry iand year t. The new equation becomes

Sti = �0 + �1 ln(Wsti

Wuti) + �2 ln(

Kti

Qti) + �3 lnQti

+�4TECHti + ✏ti

(4)

where estimates of �4 denote the relationship between tech-nologies and skill demand. The SBTC hypothesis suggeststhe sign of �4 to be positive.

IV. EMPIRICAL MODELS

In empirically estimating the skilled labor’s wage share,the wage share variable Wsti

Wutiis frequently removed from the

model because it is likely to be highly endogenous (Machinand Reenen 1998). Assuming (i) complete mobility of em-ployees across industries and (ii) that wage differentials arefully absorbed by industry dummy variables, I include fixedeffects to capture any unobserved heterogeneity betweenindustry that is time-invariant (Di). Equation (4) becomes

Sti = �0 + �1 ln(Kti

Qti) + �2 lnQti + �3TECHti

+✏ti +Di

(5)

Most studies in the early 1990s proxy the stock of tech-nology by the ratio of employees using computers at work.Of course, this is hardly a good measure of technologiesin the 21st century due to the variety of electronic devicesemployed in the work place. Later studies frequently useinvestment in research and development (R&D) instead.However, R&D is recorded separately from software pur-chases and is not the best variable to measure the technologystock either. In this paper, I choose to use the stock ofintellectual property products from the Bureau of EconomicAnalysis (BEA) that is comprised of R&D, software, andoriginals work to get a more complete account of firm’stechnology stock. Admittedly, this is not the most accuratemeasure of technology because it contains the stock ofentertainment, literary, and artistic originals at the industrylevel. However, because R&D and software data is notavailable for industries of interest separately, the stockof intellectual property products is the best measurement

available to proxy for an industry’s technology stock. Iadjust the variable (IPPti) by output and take the logtransformation for a consistent specification on the right-hand side of the equation, giving the final equation:

Sti = �0+�1 ln(Kti

Qti)+�2 lnQti+�3 ln(

IPPti

Qti)+✏ti+Di

(6)Table 1 summarizes estimates of �3 in two relevant studies.Despite the disparity in measurements of capital or selec-tions of industries, both papers document positive estimatesof �3 in the U.S. from 1960 to 1980. Both findings indicatethe complementary relationship between capital and skilledlabor and the presence of skill-biased technological change.Using a similar framework, this paper reexamines the valueof �3 in 18 U.S. industries from 1975 to 2015.

Table 1: Estimates of �3 in Relevant Studies

V. DATA AND SUMMARY STATISTICS

The data used in this paper comes from two sources.The first is the Current Population Survey March samplesfrom 1975 to 2015. It contains information on annual wageincome, weeks worked, and usual hours worked per week, aswell as demographical information regarding age, educationlevel, sex, and race for the nearly 8 million individualssurveyed. Following common practice in the field, I limit thedataset to employees within age range 16-64 and who areworking full time throughout the year (i.e., working for morethan 35 hours per week and more than 40 weeks per year).The second data source is the BEA, which provides 2-digitindustry level data on (1) real output, (2) stock of privateintellectual property products, and (3) stock of equipmentand structures.

A. Crosswalk between CPS and BEA

In this paper, I consider 18 private sectors that are mappedbetween the CPS and the BEA. The ind1990 variable in theCPS provides a set of industry codes from 1968 forward thatare consistent with the North American Industry Classifica-tion System (NAICS) used in the BEA datasets. (Refer tothe Appendix for the exact crosswalk between two sources.)

B. Wage Share

I categorize employees as either skilled or unskilled basedon their education background. The education backgroundindicates an individual’s level of expertise and is a goodproxy for the skill level. Those who have obtained abachelor’s or more advanced degree (master’s, professionalschool, doctorate degrees, etc.) are categorized as skilled

7

labor.3 The remaining employees are classified as unskilledlabor, whose education levels range from no degree to highschool diploma and associate’s degree. The wage sharevariable is then derived as the ratio between the sum ofthe skilled labor’s wage income and total wage income.Table 2 displays the skilled labor group’s average wage sharein 10-year intervals for the industries under analysis. Theaverage share grows from 19% in 1975-85 to 39% in 2005-15, an upward trend that is well-documented.4 Disparitiesremain in the changes of wage share for different sectors.For sectors such as finance and chemical products the wageshare ratio is always high (around 40%). Sectors such aselectrical products and paper products experience the largestpercentage growth, the ratio of which grows from 28% to62% and from 10% to 45%, respectively. The ratio remainslow for sectors such as wood and and plastics product acrossdecades.

Table 2: High Education Wage Share

C. Explanatory Variables

To proxy for technological change, I use the stock ofintellectual property products from the BEA. It is the bestmeasure of technology stock available at the industry level,as explained in Section IV. The output-adjusted sum ofequipment and structure stocks is used to proxy for physicalcapital. Real output is calculated as nominal output dividedby the price level, as documented in the BEA dataset. Astatistical summary of all variables used in the paper isreported in the Appendix.

VI. EMPIRICAL RESULTS

A. Benchmark Model

Table 3 reports a set of fixed-effects regressions coveringthe four time periods 1975-1985, 1985-1995, 1995-2005,and 2005-2015. It estimates the change in the skilled labor

3Would different definitions of skilled labor change my estimationresults? In this paper I follow the practice of Berman et al. (1994) andAutor et al. (1998) to group college graduates and beyond as skilled labor.

4The upward trend is also documented by Berman et al. (1994) andGoldin and Katz (1996).

share of wage bill on indicators of changes in physicalcapital, intellectual property products, and real output.

Model 1 includes only time dummies for time periods1985-1995, 1995-2005, and 2005-2015. Coefficients havepositive signs, which indicate a continued growth of skilledlabor share of wage bill through the early 21st century.Model 2 estimates the share equation (6) and shows asignificant and positive relationship between skill demandand capital stock. According to the model estimates, aone percent increase in physical capital will lead to a0.11% increase in skilled labor wage share, and one percentincrease in technology-embodied capital will lead to a0.02% increase. The positive coefficients support the skill-capital and skill-technology complementarity hypotheses.Incidentally, the three independent variables can collectivelyexplain more than 60% of the variations in skill demand.Model 3 in enhances equation (6) by interacting the intel-lectual property products stock with time dummy variables.Compared with the base period 1975-1985, technologicalchanges appear to be progressively skill biased in the 1990sand afterwards. The interaction term has a positive sign forall three periods and is largest in 2005-2015. This upwardtrend indicates a stronger relationship between technologicalchanges and demand for skilled labor in the 21st century.In model 4, I interact physical capital with time dummyvariables, and the interaction terms have positive thoughinsignificant coefficient estimates. The evidence suggestscapital-skill complementarity across all periods, and nosignificant changes in the complementary relationship in the21st century.

Summing up, through econometric analysis I find contin-uously growing demand for skilled labor in the past fourdecades. The wage share of skilled labor is significantlyand positively correlated with the stock of physical capitaland intellectual property products, supporting the capital-skill and technology-skill hypotheses. The complementaryrelationship between technology stocks and capital exhibitsan upward trend in the four periods studied.

B. Robustness Checks

Table 4 reports tests of robustness to alternative measuresof demand for skilled labor. I use the employment shareof skilled labor to proxy for skill demand, based on theassumption that wage differentials across industries can becontrolled by the fixed-effects estimator. Model 2 againreturns positive coefficient estimates. Similar to my resultsfrom Table 3, models 3 estimates the interaction terms to bepositive, and model 4 estimates them to be positive thoughinsignificant. Therefore, the technology-skill complemen-tarity, capital-skill complementarity, and a trend towards astronger complementarity relationship between technologiesand skilled labor are statistically significant and robust.

C. Industry-level Evidence

I next examine how the complementary relationships be-tween skill and capital vary across industries. I create indus-try dummy variables for all sectors and interact them with

8

Table 3: Changes in the Skilled Labor’s Wage Share

Note: T-statistics in parentheses (z-statistics for random effects model).***Significant at 0.01 level. **Significant at 0.05 level. *Significant at 0.1level.

the technology stock and physical capital stock separately.Model 1 in Table 5 presents estimates of the coefficient onthe stock of intellectual property products. The coefficientestimates are highest among electrical products manufactur-ing, motor vehicles manufacturing and machinery manufac-turing. These three sectors are also the ones that invest mostintensively on technologies.5 Model 2 includes interactionterms with physical capital: the coefficient estimates are thelargest among the same three sectors. On the other hand,retail and transportation industries have negative thoughinsignificant coefficient estimates, suggesting capital-skillsubstitutability. This finding indicates that the skill-biasedtechnology changes and capital-skill complementarity aremost obvious in capital-intensive sectors, whereas sectorsthat hold low stock of physical capital exhibit potentialcapital-skill substitutability.

Summing up, disparities remain in the relationships be-tween capital and skill in different sectors. On one hand,capital-intensive sectors (e.g., electric products manufactur-ing) show strong evidence in favor of SBTC and capital-skill complementarity. On the other hand, sectors that holdlow stock of physical capital (e.g., transportation) exhibitpotential capital-skill substitutability.

VII. CONCLUSION

My results show robust evidence of capital-skill andtechnology-skill complementarities across the 18 sectors

5See the Appendix for a plot of the technology intensity across industries.

Table 4: Changes in the Skilled Labor’s Employment Share

Note: T-statistics in parentheses (z-statistics for random effects model).***Significant at 0.01 level. **Significant at 0.05 level. *Significant at 0.1level.

analyzed. The complementary relationship is becomingstronger across decades and is strongest in the 21st century.Contrary to some claims that suggest the possibility of smartmachines replacing high-skilled labor, my econometric anal-ysis of the wage share ratio over last 40 years indicatesa continued trend for technological changes to favor, andcomplement with skilled labor. Disparities remain in themagnitude of the capitals effect on skilled labors wage sharefor different industries. When examining the coefficientsseparately for each sector, all of them exhibit technology-skill complementarity and most of them exhibit capital-skill complementarity. The complementary relationship isstrongest for capital-intensive sectors such as electricalproducts manufacturing, motor vehicles manufacturing, andmachinery manufacturing. On the other hand, less capital-intensive sectors such as retail and transportation suggestcapital-skill substitutability.

However, the positive and significant covariance betweencapital and technology stock wage share is rather mechanicalthan causal. My finding reveals that whatever factors thatcause industry-level technology and capital stock to increasein the past 40 years also lead to an increase in the skilllabors wage share. The models are subject to potentialendogenous bias as it is possible that increased supply, ratherthan demand, of highly skilled labor motivates companiesto invest more in technology-embodied capital. For futurestudies instrumental variables uncorrelated with the wageshare ratio (such as government spending on R&D) could

9

Table 5: Changes in the Skilled Labor’s Wage Share, byIndustries

Note: T-statistics in parentheses (z-statistics for random effects model).***Significant at 0.01 level. **Significant at 0.05 level. *Significant at 0.1level.

be used to correct the endogeneity bias.Another potential area for future studies is to redefine

skilled labor not based on education levels but on oc-cupational tasks performed at work. Autor et al. (2003)introduce a new methodology for analyzing changes in theskill demands: instead of using average educational levelsof workers as a proxy for skill demands, they draw a dis-tinction between skills and tasks, and argue that advances intechnologies first change the labor division between workersand machines, then task composition, and finally the demandfor different skills. Using data on task requirements from theDictionary of Occupational Titles (DOT) and the Census andCurrent Population Survey, the authors form a panel dataset

of occupational task inputs from 1960 to 1998. They find aconsistent increase in the demand for non-routine cognitivetasks (e.g., consulting, marketing, engineering), and non-routine manual tasks (e.g., driving cabs, cleaning buildings),and a decrease in routine cognitive and manual tasks (e.g.,clerical and bookkeeping jobs). They argue that the infor-mation and communication technologies function throughpredefined rules and algorithms, and therefore substitute pro-grammable routine tasks and complements non-routine tasksthat are beyond present programming capacities. Becauseoccupational datasets are not available for years after 2005,the framework can not be tested over a longer horizon. Inthe future, it remains of interest to examine the effect ofnew technologies on task composition and skill demand.

REFERENCES

[1] Autor, David H. and Brendan Price. (2013) The Changing TaskCompostition of the US Labor Market: An Update of Autor, Levyand Murnane (2003). MIT working paper, June 2013.

[2] Autor, D., L. F. Katz and A. B. Krueger. (1998) Computing Inequality:Have Computers Changed the Labor Market? Quarterly Journal ofEconomics. November, 113:4, pp. 1169-1213.

[3] Beaudry, Paul; David A. Green and Benjamin Sand. (2013) The GreatReversal in the Demand for Skill and Cognitive Tasks NBER WorkingPaper No. 18901.

[4] Berman, Eli; John Bound and Zvi Griliches. (1994) Changes inthe Demand for Skilled Labor within U.S. Manufacturing: Evidencefrom the Annual Survey of Manufacturers The Quarterly Journal ofEconomics. Vol. 109, No. 2 (May, 1994), pp. 367-397.

[5] Berman, Eli, Rohini Somanathan and Hong Tan. (2005) Is skill-biasedtechnological change here yet? Evidence from Indian manufacturingin the 1990 Policy Research Working Paper Series 3761, The WorldBank.

[6] Brynjolfsson, E. (2014) The Second Machine Age: Work, Progress,and Prosperity in a Time of Brilliant Technologies (First Edition.).New York: W. W. Norton & Company.

[7] Brown, Martin and Peter Phillips. (1986) Craft Labor and Mecha-nization in Nineteenth

[8] Century Canning Journal of Economic History 46: 743-756.[9] Brown, Randall S., and Laurits R. Christensen. (1981) Estimating

Elasticities of Substitution in a Model of Partial Static Equilibrium:An Application to U. S. Agriculture, 1947 to 1974 Modeling andMeasuring Natural Resource Substitution, Ernst R. Berndt and BarryC. Field, eds. (Cambridge, MA: MIT Press, 1981), pp. 209-29.

[10] Bound, John and George Johnson. (1992) Changes in the Structureof Wages in the 1980s: An Evaluation of Alternative ExplanationsAmerican Economic Review 83 (June 1992): 371-392.

[11] Cain, L., and Paterson, D. (1986) Biased Technical Change, Scale,and Factor Substitution in American Industry, 1850-1919 The Journalof Economic History,46(1), 153-164.

[12] Fallon, P. R., and Layard, P. R. (1975) Capital-Skill Complementarity,Income Distribution, and Output Accounting Journal of PoliticalEconomy,83(2), 279-302.

[13] Griliches, Zvi. (1969) Capital-Skill Complementarity Review of Eco-nomics and Statistics November 1969, 51(4), pp. 465-68.

[14] James, J., & Skinner, J. (1985) The Resolution of the Labor-ScarcityParadox The Journal of Economic History, 45(3), 513-540.

[15] Karabarbounis, Loukas and Neiman Brent. (2013) The Global Declineof the Labor Share NBER 19136.

[16] Machin, S., and Van Reenen, J. (1998) Technology and Changes inSkill Structure: Evidence from Seven OECD Countries The QuarterlyJournal of Economics,113(4), 1215-1244.

[17] Michaels, Guy; Ashwini Natraj, and John Van Reenen. (2010) HasICT Polarized Skill Demand? Evidence from Eleven Countries over25 Years NBER Working Paper No. 16138. June 2010.

[18] Violante, G. L. (2008) Skill-Biased Technical Change in The NewPalgrave Dictionary of Economics, S. N. Durlauf and L. E. Blume(eds.), Basingstoke, England: Palgrave Macmillan.

10

APPENDIX

A. Crosswalk between the CPS and the BEA

The table below demonstrates how I map industries between the CPS and the BEA datasets. The CPS dataset uses athree-digit coding system to store industry information in the ind1990 variable. The BEA uses North American IndustryClassification System (NAICS) and aggregates industry information to the two-digit level.

B. Summary Statistics

The table below is a statistical summary of the variables employed in the paper.

11

C. Technology Intensity

The figure above plots the log of the output adjusted intellectual property product stocks for each industry. The mosttechnology-intensive industries are chemical products manufacturing, electrical products manufacturing, motor vehiclesmanufacturing and machinery manufacturing. The least intensive ones are wood manufacturing, retail and transportation.

D. Capital Intensity

The figure above plots the log of the output adjusted physical capital stocks for each industry. The most capital-intensiveindustries are finance, retail and transportation industries. The least intensive ones are furniture and apparel sectors.

12

E. Aggregate Wage Share and Employment Share

The two figures above plot two different measures of the independent variables used in the paper: the wage share and theemployment share ratio. From 1975 to 2015, skilled labor is progressively taking up a larger percent share of total wagebill and employment.

13

Has Indonesia’s Growth Between 2007-2014 Been Pro-Poor? Evidencefrom the Indonesia Family Life Survey

Ariza Atifan GustiAdvisor: Dr. Paul Glewwe

University of Minnesota, Department of Economics

Abstract— A country’s economic growth is said to help thepoor and eradicate poverty if it is pro-poor, in that its impactsare broad-based, and benefit the poor in absolute terms. Thisresearch seeks to explore whether Indonesia’s sustained growthbetween 2007-2014 were pro-poor by examining a panel data ofhousehold survey results given by the Indonesian Family LifeSurvey. Furthermore, since measurement errors are plentifulespecially in household survey datasets, appropriate measureswill be taken to minimize the possible bias.

I. INTRODUCTION

There is no denying that the growth of an economycan lead to reductions in poverty, especially in developingnations. The Department for International Development ofthe UK strongly advocates economic growth for developingcountries, stating that it is the most potent tool in reducingpoverty and enhancing the quality of life in those countries(DFID, 2008). However, the extent to which economicgrowth can help the poor and eradicate poverty depends onhow broad-based the growth is. One recent notion to describegrowth that boosts the poor’s income and possible outcomes,is pro-poor growth. An economy’s growth is said to be pro-poor if and only if there are benefits reaped by the poor inabsolute terms, as indicated by an appropriate measure ofpoverty (Ravallion and Chen, 2003).

How pro-poor a country’s economic growth has beenis an increasingly popular topic for economists and otheracademics alike. This study will contribute to the literaturesurrounding pro-poor growth by investigating whether In-donesia’s recent economic growth has been pro-poor. Overthe last 15 years, Indonesia has experienced sustained eco-nomic growth. The average annual GDP per capita growthrate is 5.4%, leading to its inclusion as the only South-EastAsian country in the G20 (World Bank, 2016). However, thisrapid growth has not been enjoyed by households at all levelsof income. Inequality in Indonesia has been rising rapidly,as indicated by an increase in the Gini coefficient from 0.31points in 2000 to 0.43 in 2013 (ADB, 2015). Consumptionis also very unevenly distributed, with the richest 10% nowconsuming as much as the poorest 54% (World Bank, 2016).

This study tries to capture the extent to which this eco-nomic growth is pro-poor by drawing upon Glewwe andDang (2011), who analyzed Vietnam’s economic growth inthe 1990s. Following their approach, this study employs twomethods to examine whether Indonesia’s growth has beenpro-poor. The first method is a cross-sectional analysis of

household consumption that compares the mean of per capitaexpenditures of a given quintile of the population in twodifferent years. The second method compares, for a givenquintile, the same households’ mean per capita expendituresboth in the first and second year, regardless of the quintilesthose households are placed in the second year.

This article utilizes two of the most recent iterationsof the Indonesia Family Life Survey (IFLS): the IFLS4,which was conducted in 2007 and the IFLS5, which wasconducted in 2014. Since household surveys datasets areutilized, the main concern with the analysis is the presence ofsubstantial measurement error in household survey datasets,which would cause serious bias in the results. Thus, a largepart of this study involves trying to correct for measurementerror to minimize the resulting bias. This is achieved by usinginstrumental variables and simulating the joint distribution ofexpenditure levels at two points in time.

This paper proceeds as follow. Section 2 presents a liter-ature review of current theories of pro-poor growth. Section3 reviews the quantitative methodology underpinning thisstudy, with a strong focus on how to correct for measure-ment error. Section 4 presents the results of the analysesconducted. Section 5 concludes.

II. LITERATURE REVIEW

The notion of economic growth reducing poverty was firstdeveloped in the 1950s and 1960s, with the introductionof the trickle-down development concept. The trickle-downeffect revolved around the idea that the benefits of economicgrowth vertically flow from rich to poor (Kakwani andPernia, 2000). However, by turn of the century, this ideawas widely contested as growth that consistently favors therich which would instead result in a persistent increase ininequality between rich and poor (ADB 1999,6).

As a result, the concept of pro-poor growth has sincegained in popularity among economists. However, althoughpro-poor growth has been an increasingly popular topic ofdiscussion, there is not yet a widely-accepted definition ofpro-poor growth nor a framework to determine whether aneconomy’s growth is pro-poor. Ravallion and Chen (2003)deem pro-poor growth to be when the poor reap benefits ofgrowth in absolute terms. This absolute benefit results in anabsolute decrease in the level of poverty. However, manyview this definition as too loose since it pertains solely to

14

the poverty rate and ignores the socioeconomic gap betweenincome groups.

Considering the distribution of growth between the poorand non-poor, Kakwani and Pernia (2000) define pro-poorgrowth as inclusive economic growth that provides, propor-tionally, more benefit to the poor than to the rich. Theyalso argue that pro-poor growth is achieved by intentionallyfavoring the poor over the rich. Similarly, Grosse et al. (2008)state that growth is said to be pro-poor in the strongest sensewhen the poor’s income growth rates are strictly higher thanthe non-poor’s, which results in a decrease in inequality.

III. METHODOLOGY

As mentioned above, this study follows the approach ofGlewwe and Dang (2011) to analyze whether Indonesia’sgrowth has been pro-poor. This approach incorporates two in-dependent analytical frameworks. The first involves a cross-sectional analysis of the mean per capita expenditure of eachquintile in the first year with the mean per capita expenditureof households in that same quintile in the second year. Incontrast, the second takes advantage of the availability ofpanel data, and compares the mean per capita expenditureof the same households in each quintile over time regardlessof which quintile the households are in for the second year.In both frameworks, sample households were divided intofive quintiles in the first year according to their real percapita expenditure. Therefore, the first quintile representsthe poorest 20% of the population while the fifth quintilerepresents the richest 20% of the population. If we assumethat there is income mobility in that some households moveto different quintiles between the two periods, then we expectthe second method to generally produce results with highergrowth rates for the poorest quintile than the first method.

Both methods produce useful interpretations of pro-poorgrowth. The first method is beneficial in that it shows thedistribution of income in a country across quintiles andreflects the changes in inequality over time. On the otherhand, the second approach reveals the degree of mobility forthe poor to move into higher quintiles and therefore reflectsthe extent to which the growth of an economy can reduceinequality and eradicate poverty.

A. Data and Estimation Issues

The data utilized in this study were obtained from theIndonesia Family Life Survey, an ongoing longitudinal so-cioeconomic and health survey of a sample of householdsrepresentative of about 83% of the Indonesian population.Dating back to the first version in 1993, four more iterationsof the IFLS have been implemented, with the most recentcompleted in 2014. This study uses IFLS 4 and IFLS 5,conducted in 2007 and 2014, respectively. Every wave of thesurvey targets the original households/respondents initiallyinterviewed in IFLS 1, along with their split-offs. Split-offmembers are those family members who have moved fromthe original household, as identified in the previous wavesof survey, and are therefore counted in a new household. As

a result, the number of households interviewed grew from7,224 households in 1993 to 16,204 households in 2014.

The IFLS provides information on individuals, their fam-ilies, households, communities, and health and educationalfacilities. In assessing whether economic growth has a sub-stantial impact on household welfare, the two most importantvariables to analyze are income and consumption. In thisstudy, I have decided to use consumption/expenditure as themain variable of interest since data on consumption are likelyto be more accurate than income data. A possible reason forthe inaccuracies of income data is that respondents, hopingfor additional financial support from the government, tendto report lower incomes. Furthermore, with tendencies tosmooth consumption over time, expenditure data are alsolikely to be less volatile than income and therefore possessa stronger link with households’ overall welfare (Deaton,1997). This study employs a pre-existing consumption vari-able constructed by the IFLS, which aggregates all foodconsumption (including self-produced food), and almost allkinds of non-food consumption, including utilities, educa-tion, and rent.

One issue with longitudinal household survey data isthat it is difficult to keep constant the unit of observation(household) across time since household members are likelyto move out or new members could move in. However,the fact that the IFLS keeps track of the split-offs of theoriginal households allows one to keep the households assimilar as possible across time. This is done by adding thesplit-off household members in 2014 back to their originalhouseholds in 2007. To check for robustness of the overallresults, this study will conduct two separate analyses, onethat doesn’t add the split-off members back and one thatdoes (See Appendix).

A larger problem with household surveys is that they arevery likely to measure income and expenditure with error,which can result in serious biases, especially for panel dataanalysis. Unlike the resulting bias caused by measurementerror in the cross-sectional analysis, which is likely to besmall, bias arising in the panel data analysis is likely tobe very large and can significantly affect results (Glewwe,2007). The reason for this is that measurement error tends toput households in the wrong quintiles, and since householdsare followed over time, this skews the analysis. For example,in the first year a household might report expenditure lowerthan the true value and is therefore included in a lowerquintile than what it should have been. If the householdreports a value closer to its true value in the second yearand is included in the higher quintile, the analysis wouldsuggest upward mobility for the household. However, this isa misleading since the household has always been in thathigher quintile and there has been no upward mobility. As aresult, it is vital to account for measurement error to produceresults with minimal bias. The next section discusses howthis study minimizes such bias.

15

B. Correcting for Measurement Error

Correcting for measurement error bias is extremely diffi-cult. To attain an unbiased result free of measurement error,an ideal scenario would be to have the joint distributionof the true expenditure values in both years whereas theonly data available are the joint distribution of the observedvalues, which are reported with error. Given this situation,this paper attempts to account for measurement error bymaking inferences on the density, mean and variances ofthe true values and simulating a joint distribution of the truevalues of per capita expenditures in 2007 and 2014. Usingthis simulated distribution, it then calculates quintile-specificgrowth rates and compares them with the actual growth ratesobtained from the observed data.

To make inferences on the density of the true valuesrequires some derivations, which are explained in detailthis section. Assume that the relationship between the truevalues of expenditure in 2007 and the observed values ofexpenditure in 2007 is given by the following equation:

y1 = y

⇤1✏y1 ) ln(y1) = ln(y⇤1) + ln(ey1) (1)

In the above equation, y

⇤1 indicates the true values of

expenditure in 2007, y1indicates the observed values ofexpenditure in 2007 while ey1 is the random measurementerror. With the assumption that ln(ey1) is symmetricallydistributed and has a mean of 0, we can infer that themedians of ln(ey1) is also zero and median of ey1 is one.As can be observed from the model above, a multiplicativerandom measurement error framework is used instead ofan additive one. The reason for this is that an additivemeasurement error could potentially generate unreasonablenegative values of expenditure when there is a large negativemeasurement error. Furthermore, an additive measurementerror also suggests that error is unrelated to household expen-diture whereas a multiplicative measurement error impliesthat error is proportional to expenditure values, which isfound to be more likely.

One can also form a similar equation, presented in equa-tion (2) that shows the relationship between true values ofexpenditure in 2014 and the observed values of expenditurein 2014. The analogous equation is as follows:

y2 = y

⇤2✏y2 ) ln(y2) = ln(y⇤2) + ln(ey2) (2)

In equation (2), y2 denotes the observed expenditure valuesin 2014, y⇤2 denotes the true expenditure values in 2014 whileey2 denotes the random measurement error in the model. Likein equation (1), ln(ey2) is also assumed to be symmetricallydistributed with a mean of 0. Thus, ln(ey2) has a median ofzero and ey2 has a median of 1. A key assumption is thatthe random measurement errors, ey1 and ey2 , are both classicrandom measurement error. This means not only that they areuncorrelated with each other, but they are also uncorrelatedwith y

⇤1 and y

⇤2 .

With the assumptions of how measurement error relates tothe observed and true values firmly established, one can nowproceed to the framework to simulate the joint distributions.

Assume that the relationship between the true values ofexpenditure in 2014 and 2007 is determined by the followingequation:

ln(y⇤2) = ↵

⇤2 + �

⇤2 ln(y

⇤1) + u2 (3)

In this equation, ↵

⇤2 and �

⇤2 indicate a simple linear

relationship between the true values of expenditure in 2014and 2007, while u2 is a residual with a mean of zero.If one observed the true values of expenditures in 2007,Ordinary Least Squares (OLS) regression can be used toobtain unbiased estimates of both ↵

⇤2 and �

⇤2 However, since

y

⇤1 is measured with error, its observed value is correlated

(endogenous) with the residual term, u2, and thus OLSestimates of ↵

⇤2 and �

⇤2 will be biased and inconsistent.

To rectify this, one can use instrumental variables to runa 2SLS regression, which provides us with consistent es-timates of ↵

⇤2 and �

⇤2 . The difficulty is in finding suitable

instrumental variables which, by definition, are variables thatare correlated with the independent variable, in this case y

⇤1 ,

and uncorrelated with u2. A similar equation to equation(3) could also be formed by switching the independent anddependent variables, as shown in equation (4):

ln(y⇤1) = ↵

⇤1 + �

⇤1 ln(y

⇤2) + u1 (4)

As with equation (3), this equation displays the rela-tionship, indicated by ↵

⇤1 and �

⇤1 , between the true values

of expenditure in 2007 and 2014. In the equation above,u1 denotes the residual term, has a mean of zero, and isuncorrelated with y

⇤2 . Since the true value of expenditure

in 2014 is not observed, one can once again run a 2SLSregression and make use of instrumental variables to obtainconsistent estimates of ↵⇤

1 and �

⇤1 .

As previously mentioned, this study attempts to correct formeasurement error by simulating a joint distribution of thetrue expenditure values in 2007 and 2014. The simulation ofthe joint distribution will be done using either equation (3) orequation (4). Which equation we choose to adopt dependson which equation exhibits a more linear relationship. Tocheck for linearity, one can regress both equations usingtheir observed values while adding a squared term of theindependent variable as an additional exogenous variable. Tocheck for the linearity of equation (3), one can add ln(y⇤1)

2 tothe regression equation. The relationship is said to be linear ifthe squared term of the regression produces an insignificantcoefficient. Therefore, this study adopts the equation thatproduces a more insignificant quadratic term. After runningboth regressions, we find equation (3) better approximatedby a linear regression when analyzing the panel data withoutadding back the split-offs. We find equation (10) is moreappropriate when analyzing the panel data with adding backthe split-offs.

To simulate the joint distribution using either equation, it isnecessary to obtain estimates of the relevant components ineach equation. So, if one were to simulate using equation(3), estimates of ↵

⇤2, �

⇤2 , V ar(u2) and V ar[ln(y⇤1)] are

required. Furthermore, in addition to the assumption of alinear relationship between true expenditure values in both

16

years, several other assumptions must be made. Two keyassumptions that pertain to the classical linear model areexogeneity, i.e. E[u2|y⇤1 ] = 0, and homoscedasticity oferrors, i.e. E[u2

2|y⇤1 ] is constant. Another necessary assump-tion is that expenditure in both years follows a log-normaldistribution. This implies that ln(y⇤1) , ln(y⇤2) and u2 arenormally distributed. To test this assumption, one can plota kernel density of the observed expenditures in both yearsand compare it to a normal distribution with the same meanand variance. Figures 1 and 2 displays the density plotsfor observed expenditure in 2007 and 2014 respectively.Although they do not perfectly follow a normal distribution,this fit is still close. Therefore, it is not unreasonable to claimthat the expenditures follow a log-normal distribution.

Figure 1: Density Plot of log expenditure in 2007

Figure 2: Density Plot of log expenditure in 2014

As previously shown, �

⇤1 and �

⇤2 can be obtained by

running a 2SLS regression using instrumental variables. Inthis study, BMI (body mass index), constructed from thesurveys height and weight variables, is used as instrumentalvariable. It is not unreasonable to choose BMI as an instru-ment because it satisfies the two criterions of instruments.First, BMI is likely to be correlated with current expenditurelevels since people who consume more food are likely tobe heavier therefore households with heavier members arelikely to have higher expenditure values. On the other hand,an individuals current BMI is unlikely to be correlated with

the level of income in the other time period after conditioningon current income.

The constants of the regressions, ↵

⇤1 and ↵

⇤2, can be

estimated using the equations (4) and (3) respectively, byusing properties of expectations. Using equation (4) to solvefor ↵⇤

1 and taking expectation of equation (4) yields

↵

⇤1 = E[ln(y⇤1)]� �

⇤1E[ln(y⇤2)]� E(u1) (5)

↵

⇤1 = E[ln(y⇤1)]� �

⇤1E[ln(y⇤2)] (6)

Equation (6) makes use of the fact that E[ln(✏y1)] = 0and E(u1) = 0. As a result, an unbiased estimate of ↵⇤

1 canbe obtained using equation (6), where �

⇤1 is estimated using

2SLS regression. The same procedure is used to acquirean estimate of ↵

⇤2. Another component that needs to be

estimated is V ar(u2), which could be obtained by takingthe variance of equation (3). Taking the variance of equation(3) yields us the following equation.

V ar(u2) = V ar[ln(y⇤2)]� (�⇤2)

2V ar[ln(y⇤1)] (7)

Observing equation (7), it is clear that one needs to findestimates of V ar[ln(y⇤1)] and V ar[ln(y⇤2)]. This can be doneby the following equation, which is a standard equation forthe OLS estimates if a regression is run for equation (3)using the observed values.

�2 =Cov[ln(y1), ln(y2)]

V ar[ln(y1)]=

Cov[ln(y⇤1), ln(y⇤2)]

V ar[ln(y1)](8)

Equation (8) follows from the fact that adding uncorrelatedrandom measurement errors to each variable does not changethe covariance between the two variables. Furthermore, ifone were to run an OLS regression using the true values,one would get the following equation

�

⇤2 =

Cov[ln(y⇤1), ln(y⇤2)]

V ar[ln(y1)](9)

Taking the ratio of equation (8) and equation (9), providesan estimate of the variance of ln(y⇤1).

�

⇤2

�2=

V ar[ln(y1)]

V ar[ln(y⇤1)](10)

V ar[ln(y⇤1)] =�2

�

⇤2

V ar[ln(y1)] (11)

To obtain an estimate of V ar[ln(y⇤2)], assume proportionalmeasurement error, whereby the contribution of measurementerror to V ar[ln(y1)] is proportionally the same as the con-tribution of measurement error to V ar[ln(y2)]. Using thisassumption provides following derivations.

V ar[ln(y2)]

V ar[ln(y⇤2)]=

V ar[ln(y1)]

V ar[ln(y⇤1)]=

�

⇤2

�2(12)

V ar[ln(y⇤2)] =�2

�

⇤2

V ar[ln(y2)] (13)

With all of the necessary components derived, these com-ponents can then be plugged into either equation (3) or (4) to

17

simulate the joint distribution between the true expenditurevalues in 2007 and 2014. The following table summarizeshow to obtain estimates for the necessary components ofequation (3).

Table 1: A Summary of How to Obtain the Estimates of theComponents to Simulate Equation (3)

Estimate Equation↵

⇤2 = E[ln(y⇤2)]� �

⇤2E[ln(y⇤1)]

�

⇤2 2SLS Regression using IV

E[ln(y⇤1)] E[ln(y1)]

V ar(u2) V ar[ln(y⇤2)]� (�⇤2)

2V ar[ln(y⇤1)]

V ar[ln(y⇤1)]�2

�⇤2V ar[ln(y1)]

V ar[ln(y⇤2)]�2

�⇤2V ar[ln(y2)]

IV. RESULTS

To determine whether Indonesia’s economic growth be-tween 2007 and 2014 has been pro-poor, this section ap-plies the two analytical approaches discussed above to theIndonesia Family Life Survey. The results shown first arefor the cross-sectional analysis. Then, results are presentedfor the panel data analysis. The panel data results first showgrowth rates from an analysis that does not include the split-off household members. A similar analysis that includes thesplit-off household members is included in the Appendix.Lastly, we present growth rate results from a simulation ofthe joint distribution of true expenditure values.

A. Cross-Sectional Analysis

Table 2 shows the growth rates in expenditure between2007 and 2014 by quintiles using the cross-sectional methoddiscussed in Section 3. In this analysis, the unit of observa-tion is the household and the consumption expenditure valuesare expressed in real terms according to 2014 price levels.The fourth column in Table 2 contains the overall growthrate over seven years while the last column is the averageannual growth rate.

From the last row of Table 2, we see that the average realper capita expenditure increased from Rp. 735,276 in 2007to Rp. 1,009,687 in 2014, which amounts to a 4.63% averagegrowth rate per annum. This is in line with Indonesia’soverall real GDP growth rate reported by the NationalAccounts, which was estimated to be around 5%. Lookingat the results by quintiles, the first quintile experienced thelowest overall growth rate over seven years, with mean percapita expenditure rising from Rp. 247,689 in 2007 to Rp.335,571 in 2014. This amounts to an annual growth rate of4.43%. Compared with the other 4 quintiles, one can seethat the poorest 20% experienced the lowest growth rate.The third quintile has the highest growth rate, with meanper capita expenditure rising from Rp. 542,975 in 2007 toRp. 756,134 in 2014, which equals an overall growth rate of39.26% or 4.84% annually.

Table 2. A Summary of Growth Rates According to QuintilesUsing the Cross Section Method

Whether one would classify these growth rates as pro-poor depends on the definition one is willing to use. If pro-poor growth were defined using the definition of Ravallionand Chen (2003), Indonesia’s growth would be classified aspro-poor since the poor (indicated by the first quintile) havefared better in absolute terms and thus poverty has declined.However, using the definition of pro-poor growth providedby Kakwani and Pernia (2000), the fact that the poorestquintile does not experience the highest growth among thefive quintiles suggests that Indonesia’s economic growthbetween 2007 and 2014 has not been pro-poor. This supportsthe proposition that inequality in Indonesia has risen over theseven years, evident in the noted increase in Gini coefficient.

B. Panel Data Analysis without Correcting for Measurement

Error

Table 3 presents the growth rates in per capita expenditurebetween 2007 and 2014 across quintiles using the panel datamethod, which compares the same households over time. Thedata analyzed were constructed by using only the householdswho were found and interviewed in both years, and withoutadding back household members who have moved awayfrom the original household (split-off members). Once again,the unit of observation is the household and per capitaexpenditure figures are expressed in real terms using 2014prices.

Table 3 shows that the mean of per capita expenditureincreases from Rp. 788,929 in 2007 to Rp. 1,091,589 in2014, amounting to an average growth rate of 4.75% per year.This is slightly higher than the overall growth rate reportedin Table 2, which was 4.63%.

The next section of Table 3 shows the growth rates ofper capita expenditure when households are being followedbased on their per capita expenditure in 2007. Clearly,growth rates of per capita expenditure when households areranked based on 2007 per capita expenditure are dramaticallydifferent to the growth rates obtained from the cross-sectionanalysis. Unlike the cross-sectional analysis, which suggeststhat the poor fared the worst among the other quintiles,the panel data analysis shows that the poor experienced the

18

highest growth rate. The poorest quintile experienced anaverage growth rate of 11.85% per year, with expenditureincreasing from Rp. 245,110 in 2007 to Rp. 536,909 in2014. Furthermore, another major difference with the cross-sectional analysis is that growth rates of expenditure seemto be decreasing as one moves to higher quintiles. Forexample, the richest 20% experienced the lowest growth,and had a negative average annual growth rate of -0.08%.In terms of pro-poor growth classification, Table 3 suggeststhat Indonesia’s growth between 2007 and 2014 has beenpro-poor according to the requirements of both Ravallionand Chen (2003) and Kakwani and Pernia (2000). There are

Table 3. A Summary of Growth Rates According to QuintilesUsing the Panel Data Method Without Adding Back SplitoffHousehold Members

several reasons why Table 3’s results, which were obtainedfrom a panel data analysis is markedly different from Table2’s results, which were obtained from a cross-sectionalanalysis. The first reason is that the households included inthe panel data analysis are not representative of the sample ofthe entire population, which was used in the cross-sectionalanalysis. This is possible since in the construction of thepanel dataset for Table 3 used only those households thatwere found in both 2007 and 2014. However, the last 5 rowsof Table 3, which present a cross-sectional version of paneldata in which the mean of per capita expenditures of 2014was defined according to 2014 quintiles, show results similarto those of Table 2. The similarity of these two analysessuggests that panel attrition does not explain the differencesbetween the growth rates in Tables 2 and 3.

Furthermore, the result that poorer quintiles fare betterthan the richer quintiles is expected assuming that there isupward mobility. With upward mobility, some householdsthat were found to be in the first quintile in 2007 may end upin the second quintile in 2014 and thus contribute to a largergrowth rate. This differs from a cross-sectional analysis,

where the first quintile for 2014 includes only householdsthat are found in the first quintile in 2014. Furthermore, anincrease in expenditure in absolute terms for households inthe poorer quintile will contribute to a larger growth ratethan an increase in expenditure in absolute terms of the samemagnitude experienced by a richer quintile. Another possiblereason that can explain this difference in growth rates inTable 3, and one that will be the subject of the next section,is measurement error. As explained in Section 3, householdsurvey datasets often measure income and expenditure witherror, and this is likely to cause bias in the analysis. Thisbias problem is particularly severe for panel data analysisthat follows the same households over time, as we did above.

C. Simulation Correcting for Measurement Error

Table 4 presents simulated growth rates that have beencorrected for measurement error using equation (3) as dis-cussed in Section 3. The annual growth rate for the overallpopulation, 4.33%, is nearly equivalent to previously com-puted overall growth rates presented in Tables 2 and 3.This suggests, as predicted, that measurement error causeslittle or no bias when taking the mean of all households.Measurement error also does not cause substantial bias whentaking the mean of households across quintiles (withoutfollowing the same households over time) as corroboratedin the last five rows of Table 4. The last five rows of Table 4show a cross-sectional analysis using panel data in which onetakes the mean of per capita expenditure in 2014 calculatedaccording to 2014 quintiles. As one can see, the growth ratesacross quintiles is very similar to the growth rates presentedin Table 2 and thus indicate that measurement error does nothave a large effect on these cross-sectional analyses.

Table 4. A Summary of Simulated Growth Rates UsingEquation (3)

On the other hand, the top five rows of Table 4, whichshow the simulated growth rates of expenditure using panel

19

data analysis, differ greatly from the corresponding resultsfrom Tables 2 and 3. This implies that measurement errorcauses serious bias when one computes growth rates byfollowing the same households over time. Comparing Table3 and 4, one can see that measurement error, which wasnot accounted for in Table 3, overestimates growth rates forall quintiles except the fifth quintile and results in a widerdispersion of growth rates in Table 3. According to Table 4,the poorest quintile experienced an annual growth rate of percapita expenditure of 10.67%, more than 1 percentage pointlower than the growth rate reported in Table 3. Similar over-estimations of growth rates also occur with the second, thirdand fourth quintiles which were overestimated by 1.8, 1.2and 0.30 percentage points respectively. Another interestingobservation from Table 4 is that the poorest 20% performedbetter than the other quintiles and in fact, experienced agrowth rate of nearly 10 times higher than that of the top20%. This suggests that Indonesia’s growth between 2007and 2014 has been pro-poor and contradicts the previousnotion that inequality in Indonesia has been rising. However,it is unlikely that the poorest 20% did indeed perform 10times better than the top 20%, suggesting that there is stillmeasurement error that has not yet been corrected and thatthe methodology from Section 3 may only partially correctthis persistent problem with panel data analysis.

V. CONCLUSION

This study was conducted with the intention to contributeto the pro-poor growth literature that has been gaining pop-ularity recently. This study focuses on Indonesia’s economicgrowth between 2007 and 2014, which has been impressivelyhigh but may have adverse effects by widening inequality be-tween the poor and the rich. To fully grasp whether economicgrowth in Indonesia has been pro-poor, this study followsthe approach of Glewwe and Dang (2011), who examinedwhether Vietnam’s growth in the 1990s was pro-poor. Thisstudy uses two analytical approaches to determine whetherIndonesia’s growth has been pro-poor. The first, which isuseful in giving a picture of the distribution of income, com-pares the mean of per capita expenditure per given quintilein both 2007 and 2014. The second method, which utilizesa panel data of households, follows the same householdsover time and compares their per capita expenditures overtime. An important aspect of this research involves dealingwith measurement error, which is plentiful in householdsurveys, and causes serious biases when analyzing paneldata. To correct for measurement error, this paper simulatesa joint distribution of the true expenditure values in 2007and 2014 by making inferences on the joint density, meanand variances of the variables.

The results of our two analyses produce two varyingconclusions. Findings from the cross-sectional analysis showthat growth rates across quintiles are very similar, with thepoorest quintile experiencing a somewhat lower growth rate.This suggests not only that Indonesia’s growth has not beenpro-poor (using Kakwani and Pernia’s (2000) definition), but

also that the pattern of income distribution and inequalitybetween 2007 and 2014 has not changed. This analysis didnot correct for measurement error since studies have shownthat the resulting bias of cross-sectional analysis is likelysmall.

On the other hand, findings from the panel data analysisshow that Indonesia’s growth between 2007 and 2014 hasbeen pro-poor. In particular, analyses using both observedvalues and simulated growth rates indicate that the poor-est 20% experienced a higher growth than the other fourquintiles. This implies that Indonesia’s growth is likelyto be accompanied by upward mobility between quintiles.Furthermore, simulation of growth rates, which correct formeasurement error, demonstrates how large the bias thatmeasurement error can cause in panel data analyses. Findingsfrom the simulated growth rates show that measurementerror leads to overestimating of growth rates, and widensthe dispersion of growth rates among quintiles.

Although useful in providing a picture of how pro-poor Indonesia’s growth has been, the question of pro-poor growth in Indonesia can certainly not be answeredby this paper alone. Further research must be conductedto better understand what pro-poor growth entails and howto measure whether an economy’s growth has been pro-poor. Furthermore, since measurement errors are plentifulin household survey data, which are required to conduct thesecond methodology, more studies should also be dedicatedin trying to better correct for measurement error.

VI. ACKNOWLEDGEMENTS

I would like to thank Dr. Paul Glewwe of the Univer-sity of Minnesotas Applied Economics department for histremendous support and constant guidance throughout thisresearch project. I would also like to thank the UndergraduateResearch Opportunity Project (UROP) at the Univesity ofMinnesota for funding this research project.

REFERENCES

[1] Aji, P. (2015). Summary of Indonesias Poverty Analysis (ADB Paperson Indonesia No. 4). Manila, Philippines: Asian Development Bank.

[2] Deaton, A. (1997). The Analysis of Household Surveys: A Microe-conometric Approach to Development Policy. Baltimore, MD: JohnsHopkins University Press.

[3] Department for International Development. (2008). Growth BuildingJobs and Prosperity in Developing Countries.

[4] Glewwe, P. (2012). How Much of Observed Economic Mobility isMeasurement Error? IV Methods to Reduce Measurement Error Bias,with an Application to Vietnam. The World Bank Economic Review,26(2), 236264. https://doi.org/10.1093/wber/lhr040

[5] Glewwe, P., & Dang, H.-A. H. (2011). Was Vietnams economic growthin the 1990s pro-poor? An analysis of panel data from Vietnam.Economic Development and Cultural Change, 59(3), 583608.

[6] Grosse, M., Harttgen, K., & Klasen, S. (2008). Measuring Pro-Poor Growth in Non-Income Dimensions. World Development, 36(6),10211047. https://doi.org/10.1016/j.worlddev.2007.10.009

[7] Kakwani, N., Pernia, E. M., & others. (2000). What is pro-poorgrowth? Asian Development Review, 18(1), 116.

[8] Ravallion, M., & Shaohua, C. (2003). Measuring Pro-Poor Growth.Economic Letters, (78), 9399.

[9] World Bank. (2016). Indonesias Rising Divide. Jakarta, Indonesia: TheWorld Bank.

20

Table 5. A Summary of Growth Rates According toQuintiles Using the Panel Data Method Using Dataset

in which Splitoff Members are Added Back

VII. APPENDIX

21

Table 6. A Summary of Simulated Growth Rates usingEquation (3) Using Dataset in which Splitoff Members

are Added Back to Original Household

22

Place and its Role in Venture Capital Funding

Luke Heine

Harvard College, Department of Sociology and Computer Science

Abstract— How are city demographics correlated with theamount of venture capital they receive? The paper uses aunique dataset of 58,000 venture deals from 2000 2014from the CrunchBase dataset and census data from the sameperiod. Place and the Role of Venture Capital asserts venturecapital’s spatial dependency and uses statistical software to finda strong positive correlation between the amount of venturecapital funding and foreign, international, male professionalswithin a city, the gendering of venture capital, and the negativecorrelation of unskilled, foreign labor with funding.As venture capital travels along social ties, the paper suggeststhat foreign, international, and male professionals’ positivecorrelation may be due to these members having a wider andmore diverse social network, allowing the ability to conjurefunds. Moreover, the demographic may be a synonym forSassen’s International Class, allowing the study to dovetail witha broader set of research. Finally, the paper also provides amechanism to classify cities based off their venture capital ac-tivity. The implications of this study are a better understandingof the trends correlated with venture capital, a classificationsystem for cities, and a possible caveat to ’virtuous cycle’ theory.A supplement to the paper and to visualize implications forcities, we also created this d3 visualization visualizing thegeographic positioning and relationships of those 58,000 deals,providing communicable and interactive research.

I. LITERATURE REVIEW

From Athens to Florence to Silicon Valley, humanityhas always associated innovation with geography. Innovativeplaces are, by definition, regions where humans innovate.Vicinity to research universities, cultural disposition towardsrisk, and access to capital are all factors impacting an area’sinventiveness and ability to create (Florida 1996, Hambrecht1984).

A cornerstone of entrepreneurship, modern venture capitalarose from investment firms formerly specializing in rail-roads and traditional machines with the first firm specializingin investment into Boston’s textile industry (Florida 1996,Hambrecht 1984). Once a profession where men had adifficulty describing to their wives what they did,’ venturecapitalism now underscores the success of three of theworld’s five most valuable companies as firms have restruc-tured their need for upfront capital in hopes of rapid scaling(Florida 1996, Green and McNaughton 1987, Hambrecht1984).

With the perceived impact of venture capital on innova-tion rising, cities and governments are increasingly craftingeconomic policies to capture venture capital funding for theirown regions or fund their own. A 2001 National GovernorsAssociation report stated Venture capital is critical to grow-ing the new businesses that will drive the new economy’.

Finding ways to nurture the culture of entrepreneurs, andthe capital that feeds them, must be the top priority ofstates (Henry Chen a, et al. 2009). The National Associationof Seed and Venture Funds estimated that state venturecapital funds in 2008 totaled $2.3 billion; meanwhile, anincreasing share of the approximately $50 billion that statesspend on industrial incentive areas is going to venture-backedfirms (Henry Chen a, et al. 2009). Therefore, geographicallystudying venture capital is necessary and timely.

The theory behind incentivizing venture capital invest-ment, virtuous cycle theory,’ argues that easier funding forcompanies will result in additional organizations basingthemselves in a specific area, resulting in more opportunity,and the attraction of a highly-educated workforce (Dahl andSorenson 2010, Henry Chen a, et al. 2009, Khorsheed andAl-Fawzan 2014).

Historically, however, areas outside their contemporaryvirtuous cycles but able to connect with those existing havebeen most successful, showing greater geographic complex-ity than that presented solely by virtuous cycle theory (Engeland del-Palacio 2011, Hambrecht 1984).

For example, only a round of funding secured from NewYork based Fairchild Camera and Instruments by ArthurRock, a financial analyst at the Wall Street investmentfirm of Hayden Stone Arthur, for Robert Noyce, a defectorfrom Shockley Laboratories would set Santa Clara Valley–far outside the then current establishment–down the road tobecoming Silicon Valley (Silicon).

Additionally, in the 1970’s Dan Tolkowsky, a retired Israelimilitary officer, joined Discount Investment and flew toSilicon Valley to interest the young U.S. venture capitalindustry to invest in Israel, attracting some of the initialSilicon Valley investments in Israeli companies (Engel anddel-Palacio 2011). Though outside the funding centers of itstime, now Israel ranks third in number of companies listedon NASDAQ and has twice the venture capital investmentsas the whole of Europe (Engel and del-Palacio 2011).

Place matters, but clearly–when examining the historicallydetached nodes of Israel and Silicon Valley–it may matterless than ties to place and capital, providing hope and a pathforward to new tech areas without strong VC bases (Engeland del-Palacio 2011). Research validates. In a study ofover 3,132 investment decisions, personal ties from investorto company were found to be more important in terms ofwhether to invest than the prestige of other participating firmsin the round, with both direct and indirect connections havingimpact on venture capital decisions (Wuebker et al 2015).

23

The amount of investment dramatically impacts a city’sfunding structure, with a one standard deviation increase inthe number of venture capital offices in an area associatedwith an increase of venture capital investments in thatarea of 49.7% (Henry Chen a, et al. 2009). But gettingthe right investment matters. Perhaps more important thanthe monetary infusion foreign investment brings, high-statusinvestors bestow legitimacy that produces future investmentsbecause they are believed to be capable evaluators thataffiliate only with promising organizations (Petkova et al.2016). Therefore, foreign investment can legitimize behaviorwhich is then imitated by those with local power and capital,meaning that investments in smaller, newer cities outsidefunding centers can have dramatic, cascading effects (HenryChen a, et al. 2009, Petkova et al. 2016). Even in the case ofearly Silicon Valley, a New York investment started a wave ofdomestic activity directly contrasting other funding centersat the time. In the words of an active venture capitalist ofthe time:

Looking back, I am still amazed at how easy it was toraise money to start Hambrecht & Quist. My partner and Idecided to start our firm one evening in San Diego. We wrotea brief four-page business plan on the plane. The next day, wevisited four prominent San Francisco families that afternoonand by that evening we had raised a million dollars. I couldn’timagine doing that in New York, Boston, or Philadelphia.I know that my cousin, a Philadelphia banker, wouldn’thave made the loan. But here in California, our investorsare only one generation removed from the risk takers whohad created the capital in the first place. Their willingnessto take risks has its cultural roots in the pioneer traditionsof this state...As a result, in 1981, California venture fundsraised nearly three times the amount of capital raised bythose of any other state, and accounted for over one-third ofthe capital raised nationwide. California also accounted for36% of the new venture financings during 1981, over half ofwhich were located in Santa Clara County alone. (Hambrecht1984). As shown in Santa Clara Valley, previous investmentinto Fairchild Instruments by a high-power actor not onlyinfused capital but loosened capital from domestic investorswho then may make riskier choices, showing that wherefunding comes from matters and meaning that proportionaldomestic spending may offer a proxy to gauge a region’sVC stage (Silicon). And as tastes change for those withaccess to capital, we may be on the cusp of a venturecapital revolution. Increasingly transnational and globalized,those that have the power to invest are an increasinglycosmopolitan class, less bounded by place than ever before,presenting a huge opportunity for newer, less well-knowninnovation regions (Sassen 2000). With an increased will-ingness to inhabit and invest in places fluidly, increasinglycosmopolitan investors open opportunities for new citiesenabled by heightened globalization and telecommunicationto acquire capital outside their conventional centers (Sassen2000). Most encouraging for newer cities, coupled with theelite’s changing tastes, research states that geographicallydiverse investments also produce the highest investment re-

turns when compared to domestic investments in establishedcenters (Henry Chen a, et al. 2009).

Encompassing more than half the 1,000 venture capitalfirms listed in Pratt’s Guide to Private Equity and VentureCapital, the venture capital firms located in Boston, NYC,and SF outperform those in other cities, but–importantly–not from their domestic, local investments but by theirinvestments in different regions (Henry Chen a, et al. 2009).Though perhaps counter intuitively, why geographically di-verse firms outperform VC’s that invest in only local, blue-chip, virtuous cycle portfolios’ is thought to be due to ahigher barrier of entry, resulting in an extra layer of qualitycontrol and skepticism before investment (Henry Chen a, etal. 2009). And as firms continue to seek areas with largerreturns, being a new innovation region outside current centersmay not be a disadvantage but a huge opportunity.

Further necessitating the consideration of domestic in-vestment in funding regions, predictably diminishing returnson investments occur as more venture capital flows to aspecific area past a certain point (Henry Chen a, et al.2009). Therefore, an opportunity for newer innovation cen-ters seeking capital, older innovation centers look outsideestablished centers to invest, captured in a lower domesticVC investment percentage (Henry Chen a, et al. 2009).Therefore, the percentage of domestic investment may alsoindicate a center’s maturity by its investors’ actions ininvesting elsewhere (Henry Chen a, et al. 2009).

II. DATA

To test the international class’s impact and inspect fundingcenters’ nature, this paper uses data from CrunchBase.comand the 2010-2014 American Community Survey. Sourcedfrom a foremost venture capital database, the CrunchBasedataset has 500,000 companies and 86,000 investor roundsreported. Extracted via a public API on March 4th, 2014, theCrunchbase data set contains all the information precedingthe pull date with the study investigating 86,000 instancesof funding rounds and their information including InvestorLocation, the investor’s geographic location upon time of in-vestment; Round Amount, the amount of money transmittedbetween investor firm and capital; and Company Location,the region where a company is headquartered.

Allowing us to better understand the impact of ties to placein global venture capital, the impact of legitimization, andwhether an area is financially saturated, examining not justfunding amount but the origin of funding and asserting therole of foreign and domestic origin offers a more granularconsideration of place’s impact. By considering place, thedata also allows the assertion of the percentage of fundingcompanies receive from the city they are based in alongwith the amount of investment by domestic firms, useful forregions as they seek funding and for understanding ties’ rolein an ever expanding venture capital field.

The second data set is sourced from the 2010-2014American Community Survey via Social Explorer on April9th, grouped by Metropolitan Statistical Area for the top48 domestic VC metropolitan areas as identified by the

24

CrunchBase aggregated dataset and contains all availableACS statistics on demographics, socio-economic standings,gender, and residents’ birth country for a given area.

A. Data Treatment

Treating the CrunchBase data for use involved subsettingthe Investor Location, Round Amount, and Company Loca-tion entries from the complete data set and converting allstring entries for those columns to lower case, stripping anynon-alphanumeric characters from the entries and trimmingany trailing spaces. After the automated data conditioning,I then corrected near string matches to equalities to allowvalue summing by equivalent values–for example, manuallyconverting singapur’ to singapore’ in the data sheet. Dueto the scraping method, the CrunchBase data erroneouslyduplicated the investment round’s total aggregate dollar valuefor each individual investor’s round contribution. To correct, Igrouped data for where investment dates, unique company id,and investment series were equal, selected the first inputtedindividual round amount–which correctly identifies the totalnot individual round investment–counted the duplicate num-ber, and then replaced the individual investor amounts by thetotal amount raised divided by the duplicate amount.

After correcting the individual investment amounts, Icoded for when investment place and investment locationequaled each other–indicating an investment by a firm in itsown city–then aggregated total investments by region stringname to find an area’s total investment. I then repeated theprocess for only domestically coded instances to find the totalamount of domestic investment. To find the domestic investedpercentage, I divided the area’s total domestic aggregatedinvestmentsummed by the method describedby the region’stotal investment. To find the percentage of domesticallyraised funds, I substituted raised funds for invested andrepeated the process.

In order to test Sassen’s proposed international class’simpact and better understand how cities attract investment,I then merged the top 48 US VC metropolitans’ totalinvestment, total domestic investment, and the percentage ofdomestic investment with the respective ACS data to createa new data set including ACS and CrunchBase data. Partlydue to the San Francisco Bay Region, the aggregated totalfunding values behave exponentially, justifying a logarith-mic transformation on the response variable before linearmodeling. Transforming total invested dollars by log10, theregional total funding data displays less skew and behavesmore normally.

B. Data Methods

Testing Sassen’s international class’s impact and the rolesof ties on a region’s total VC funding is possible byregressing the traits she ascribes to the class on the nowlog-transformed data set and inspecting their impact. Bynamesake, she attributes rising internationality with the class,writing that not only the transmigration of capital...takesplace in this global grid but people. Regressing the ACS’s

foreign-born-population metrics on total investment, there-fore, provides a way to proxy the international class’s hy-pothesized positive impact on total and domestic investment.Sassen also notes rising inequality with internationalization,writing that when cities become more internationalized, theyattract both a transnational professional workforce, and poor,mostly migrant workers. As the two foreign classes Sassendescribes have disparate capital and means, a detectable splitshould exist in foreign residents’ impact on regional funding,meriting the use of an interaction variable. And as moreindividuals immigrate with dramatically different capital to aregion, the Gini coefficient should increase and be positivelycorrelated with total funding. Sassen additionally highlightswealth and class. Possessing disproportionate affluence, theACS’s percentage of those earning over $200,000 in a regionand the percentage of professionals in a region providesanother means to proxy the international class and its impacton cities by linearly modeling against total investment.

C. Data Analysis

1) Impact of Gini: Statistically testing Sassen’s proxies,linearly regressing the regional Gini coefficient in a leastsquares model as a predictor for total funding displays theGini coefficient being correlated with near positive signifi-cance in predicting a city’s total funding with a p-value above.05. The variable has limited descriptive power, though,with only 9.367 log10 dollars in total funding describing aperfectly equal to unequal transition. Though displaying thepositive significance in line with Sassen’s theory, the Ginicoefficient by itself is marginally descriptive.

2) Impact of Foreign Born: Validating Sassen, a city’sforeign-born population is positively correlated with a city’stotal funding though lacking descriptive power as a per-fectly domestic to international population change displays amarginal 5.23 log10 dollar increase, showing an only foreign-born model’s limitations. Capturing the Gini coefficient’sdescriptive power, adding the foreign-born percentage to themodel causes the Gini coefficient to lose near significance.1

3) Impact of Wealth: Indicating personal wealth’s positiveeffect, regressing a region’s percentage of individuals earningmore than $200,000 as an independent variable along withthe foreign-born percentage to predict total funding showsthat the wealth variable has a strong, statistical positiverelationship with a city’s total raised funding amount. A10% increase in a region’s population making over $200,000is correlated with a $5.50 increase in log10 funding orthe same impact that moving from a completely domesticto foreign-born city offers. Due to its descriptive power,when incorporating the wealth percentage the foreign-born-percentage variable moves from high significance to near

1 Using foreign population to approximate the international class has somelimitations as it excludes the American international class, a dispropor-tionately large international class subsection due to disproportionately highaffluence. Using the percentage of internationally born citizens, however,still has justification, while not a complete encapsulation of the class, as itdirectly relates to Sassen’s definition.

25

significance with much of the descriptive power of thepercentage of foreign born being captured in the percentageof those earning over $200,000.

4) Gendering of Venture Capital: When regressing themale percentage of professionals from all backgrounds in aregion as defined by the ACS versus total investment, there isa slight significant relationship between the percentage andtotal funding, illustrating male professionalization’s impor-tance. When regressing the percentage of female profession-als to total funding, however, the model has no significance,showcasing an outcome differing by gender.

5) Impact of Foreign Born Professional and Unprofes-sional labor: To assert the impact of foreign-born profes-sionals I introduced a variable interacting a region’s maleprofessionals and foreign-born to test Sassen’s internationalclass. When used to predict total funding, the interactionvariable displays a strong positive correlation with vastlymore descriptive power than either the foreign born or pro-fessional percentage alone. Directly inline with the Sassen’stheory, a 10% interaction variable increase is correlated witha $52.3 log10 increase in funding, illustrating the powerfulcorrelations that the professional, international class has on acity’s funding. The foreign born variable when not interactedwith professionals additionally displays a significant negativecoefficient, illustrating Sassen’s writing on immigrant differ-ences and capturing the human capital split in immigrants.

Variables Estimate Standard Error

% foreign -12.0025* (4.8422)

% male professional -15.8173’ (7.8716)

% foreign*% male professional 169.069*** (47.1770)

Constant 9.6866*** (.7984)

’ p < 0.1, * p < 0.05, ** p < 0.01, *** p < 0.001

6) Impact of Foreign-born Females and Professionals vs.Male: Repeating the interaction steps with female profes-sionals, despite having little descriptive value alone, thefemale professional interaction variable has stronger sig-nificance with total funding than the male professionalinteraction variable as indicated by the p-value, with a10% increase resulting in a $29.9 log10 increase. As withthe male, the foreign-percentage variable by itself displaysthe negative correlation predicted by Sassen’s human cap-ital split. Significant in predicting a region’s total fundingwhen interacted with foreign-born, the female professionalinteraction variable demonstrates the unique relationshipbetween internationality and professionalization as proposedby Sassen.

Internationalization and professionalization have signifi-cance, but comparing the model’s descriptive quality with aplace’s wealth–a critical component in virtuous cycle–meanscomparing the foreign-born male professional interactionvariable with a region’s percentage of those earning over$200,000 and foreign born. The foreign-born and wealthinteraction variable displays strong statistical significancefor predicting total funding with the un-interacted wealth

percentage variable remaining significant. Comparing thewealth-foreign interaction variable with the female-foreignand male-foreign interaction variables, high collinearity andshared variance exists between the three, underscoring thevariables’ similarities and descriptive nature.

Not just a means to an end, though, the foreign-bornand male professional model provides a more useful modelfor predicting total funding as it provides a theoreticalmechanism for how cities become wealthy while offeringa similar MSE, R-squared value, and a more stable q-qplot than the wealth-foreign interaction. The female foreign-born professional percentage, on the other hand, performsworse than the professional, foreign-born percentage in termsof q-q fit and predicting total funding, again showing thegendering of capital and variation between male and femaleprofessionals. While having wealth and being professionaland foreign born are highly correlated, the foreign-bornprofessional interaction variable offers higher value to citiesas it provides a means by which cities acquire wealth.

7) Comparing Wealth and Foreign-born, Professionals’sImpact: Expanding past the wealthy international demo-graphic, comparing the foreign-born, male, professional per-centage regression with the percentage of those earning$200,000 from all backgrounds when regressed on totalfunding, the foreign-born-professional interaction variablehas only marginally higher MSE and lower adjusted r-squared values while the male international-professional q-qplot displays similar fitting qualities and absorbs outliers.Both having validity, the foreign-born, male interaction vari-able explains similar variance with the percentage of thoseearning $200,000 from all backgrounds. Demonstrating theinternational class’s significance and forwarding a mecha-nism past just an area’s immediate wealth, the percentageof male foreign-born professionals provides an equally ormore descriptive model than the female-professional interac-tion model, the interaction between the percentage of thosemaking over $200,000 and foreign-born, and the percentageof those earning over $200,000 from all backgrounds, in-dicating both venture capital’s gendering and internationalprofessionalization’s nuanced significance in a city’s abilityto raise capital.

III. ANALYSIS IMPLICATIONS

Displaying the international class’s impact, foreign-born,male, professionals provide a reasonable predictor for totalfunding possibly through wider, international social net-works, important for funding as noted by Wuebker. Showingthe disproportionate effect the wealthy have on attractinginvestment, the median income of a city has little impact onoverall funding. And as professionalization involves buildingties to wealth and management, their social networks wouldhold high power for cities andas shown by the modelparticu-larly when globalized. Individuals earning ACS professionalstatus may also better describe traits that contribute tofunding companies than the wide range of how one canearn more than $200,000. Another possibility, the foreignprofessionals living in the United States may also comprise

26

an elite group of global professionals. In any case, theinteraction variable displays that a city’s ability to attract aprofessional, male, and international class is correlated withits ability to secure investment to a similarly high degree asthe percentage earning over $200,000 in a city.

So while a region’s immediate wealth is highly corre-lated with garnering investment so too is foreign profes-sionalization independent of wealth. Instead of a virtuouscircle, where all activity takes place within an establishedecosystem, the significance of foreign-born professionalssuggests a virtuous lattice, where the ability to attract tiesto the international classand the ties that come with themhassignificance along with the nodes themselves, a conceptfurther solidified by city classification and when visualizing

the data. 2

IV. PROPOSING A NEW METHOD TO CLASSIFY CITIES’ROLE IN VENTURE CAPITAL

A refinement of virtuous circle, and as detailed by Wue-bker, the ability to receive funding depends on ties as wellas place. How cities attract these connections and theirsignificance widely varies, and as some ties have as muchsignificance as immediate domestic wealth on the ability toraise funds, it merits creating a city classification system thatexpands past virtuous cycles strictly wealth-based focus andthat includes funding origins. Summing investment instancescoded for domestic regionwhen a firm invests in their owncityand dividing by the total amount invested by the region’sfirms yields the percentage of domestic investment fromVC’s. By repeating the process for the funding raised bycompanies, summing the amount of domestically 3 codedinstance by regionthe investments raised from domestic fun-dersand dividing by the total amount of money raised by theregion yields the domestic percentage of raised money. Forclarity, the two percentages, though seemingly similar, mea-sure completely different scenarios. For example, a regionwith a small VC scene may heavily invest in their own city,making one percentage high, but that funding may accountfor only a small percentage of the total money that companiesraise in the city. Dividing cities into four groups by a 25%domestic investment cutoff and a 25% domestically raisedcutoff results in four uniquely different city charactersself-sustaining centers, tech hubs, spatially mismatched centers,and financial center. See table in appendix and visualization.

A. Self Powered Cities

Self-powered centersBoston and San Franciscodisplayhigh proportional amounts of investment from investorsbased in the city along with high proportional amounts ofdomestically raised funding by companies. Here, the highpercentage of domestic financial activity is indicative oflocalized, geographic capital and also strong companies thatattract investment. In other words, self-powered cities haveboth strong finance and tech presences that complementone another. In addition, these cities are also some of theworld’s most diverse, providing local access to internationalnetworks. 4 Exemplifying virtuous cycle theory, self-powered

2 Importantly, Sassen never mentions a particular educational pedigreefor the international class, and when running correlations between totalfunding and the percentage of those receiving bachelor or doctorate degrees,the coefficients are negligible for the dataset. In Adam Grant’s bookOriginals, the Wharton professor notes that individuals will make riskierdecisions if their risk is minimized in other areas, a perspective repeated byan interviewed Anthos Capital associate. The investor noted that whetherthe firm’s invested founders had health care impacted their portfolio andcompany outcomes. In this data set, however, the percentage of health carecoverage has low predictive power on total investment most likely dueto health care’s availability in sectors not traditionally funded by venturecapital.

3 Domestic in this paper refers to an interaction within the same city whileforeign refers to an interaction outside the city.

4Cities that are conducive to immigration may also have political qualitiesthat are more favorable to fund raising and the International class.

27

cities have companies attracting high foreign investmentalong with high domestic investment. 5

B. Tech Hubs

Exhibiting a more cosmopolitan nature than explained byvirtuous circle theory, the second class of cities representssmaller, younger cities that lack a high domestic VC presencecompared to self-powered centers that disproportionatelyreceive funding from other cities. These cities, such as Austinand Seattle, represent a class of cities where domestic VC’schannel high amounts of investment into the region theyoccupy but an amount only accounting for a small proportionof the total received funding due to the high influx of foreigncapital. Methodologically, these cities’ investors invest over25% of total investment domestically with the total domesticamount of money raised by the city below 25%. As ex-plained by Henry Chen et al, these cities possess desirablebusinesses, accounting for the high inflow of capital, andas with a young Silicon Valley and Israel, the high degreesof foreign investment’s legitimization may also influence thehigh level of domestic investment. Due to these cities’ highreliance on foreign funding, ties to funding centers mattergreatly for these cities’ companies.

C. Spatially Mismatched

Exemplified by LA, Minneapolis, Shanghai, and Dallas’sdomestic activity, the third city type displays spatial mis-match between local VC and tech needs–either with VCcapabilities or interests being unable to match local compa-nies’ needs or local companies being unable to match the VCneeds. Methodologically, cities exhibiting spatial mismatchhave less than 25% total domestic investment along withsub 25% in total money raised domestically. In practice, thismay result from sub-tier local tech not fitting the portfolioneeds of highly-rated VC firms in primarily investor cities orfrom nascent local VC firms being unable to fulfill the needsof top-rated companies. As a result, firms invest elsewherealong with companies drawing their funds from other means.In these cities, ties and links to other cities, along withattracting the International class, may hold most importancedue to regions’ heavy reliance on both non-domestic fundingand investment.

Geographic positioning to other cities may also causespatial mismatch as one city’s virtuous cycle may be avicious cycle for others, making it more difficult for otherregions to build critical mass as neighboring regions cap-tures its funding. While economically efficient, allocation ofresources may not be desirable from the perspective of localgovernments and other cities that seek local employmentgrowth (Henry Chen a, et al. 2009). Considering their spatialrelation to one another, SF Bay and Austin may havedetrimental impact on Sacramento and Dallas’s ability to starttheir own centers.

5 While also relatively self-feeding, due to the strength of these cities, itshould also be noted that they also fund many other cities.

D. Finance Centers

The fourth category represents financial centers, citiesthat have high amounts of funding sourced by companiesfrom local investors but that accounts for only a smallpercentage of a city’s total investment, meaning that thesecities primarily function as funding centers. In other words,the city’s companies raise more than 25% of total fundingfrom within the city but the city’s venture capital firms investless than 25% of their portfolio domestically. Exemplifiedby New York City, London, and Paris, these cosmopolitancities contain high international class amounts and seeminglyhave more capital than startups, with most funding propor-tionately going to other cities. Though sharing many censuscharacteristics with self-sustaining cities, in financial centersonly a small portion of foreign investor money funnels intothe cities’ startupsperhaps also due to financial crowdingby domestic centers. In other words, financial centers areVC first, startup center second. And as shown with historicSilicon Valley and New York City, ties to these cities holdmuch value for fledgling venture capital cities. 6

V. IMPLICATIONS

Considering the percentage of domestic funding investedand raised provides a way to assess a city’s ties to others,better describing city natures and opening the possibility toassess cities beyond their immediate space but also theiractions. Showcasing the range of feasible funding structuresfor cities through domestic activity also shows virtuous circletheory’s limitations through the variety in which region’s ac-quire funding. Creating a city classification system providesadditional and transmissible insight for cities by movingpast the binary, one-sized virtuous cycle and is useful forcities when considering strategic partnerships and symbioticrelationships with other cities along with better governancefor their own. For example, in known finance cities city gov-ernance may transition slotted subsidies encouraging venturecapital to tech companies and universities.

Illuminating the significant effect foreign-born, male pro-fessionals have on a city’s total funding allows the opportu-nity to display the importance of ties to placenot just place it-self. Instead of a solely wealth-based narrative, international,male, professionals also have high importance in predictingtotal funding, possibly through highly socially integratedcareers, as shown by Wuebker et al. Through social tieshaving impact, by definition a city’s ability to raise capital isin part relational, indicating a necessary addition to virtuouscircle theory: place matters, but so too, and sometimes moreimportantly, does its ties to others. And by funding beingimpacted by more than just a region’s ability to fund itselfliberates cities that do not currently have a high venturecapital concentration.

6 Importantly, the classified city examples are sorted by total fundingand total money raised but classifications may vary based upon industry.Minneapolis, MN is classified as spatially mismatched by total funding, asan example, but is considered a self-sustaining city for biotech. In addition,as enabled by the visualization, city types can also differ by year, thoughandas shown by the visualizationthere’s high inertia in cities to do so.

28

VI. CONCLUSION

Venture capital is more spatially fluid and dynamic thandescribed by a solely place-based narrative as shown throughestablishing the international class’s significance and illus-trating the variety of city structures, therefore meriting amore flexible framework that includes ties to other places.Further research should examine how cities transition fromone classification to another–as exemplified by historic Sil-icon Valley exhibiting high Tech Hub characteristics be-fore becoming a self-sustaining powerhouse. Additionally,Sassen’s argumentation opens an interesting opportunity totrack how residents move from one city to another, poten-tially illuminating a tightly interconnected human and finan-cial marketplace, possibly showing how one can influencethe other. Further study should additionally delve deeper intoventure capital’s gendering and the theoretical and statisticalreasons why female professionalization has no significancewhen regressed with total funding.

Providing cities and governments with a better modelof how venture capital acts and who can conjure funding,including ties’ importance fosters greater understanding ofwho funds innovation and from where. And as Sassen’s inter-national class becomes increasingly cosmopolitan, a possiblerenascence of venture capital could result as male, foreign,professionals connect with more regions. In the 21st century,place matters but by showing internationalities significanceand cities’ interdependence, so too does connection to place,meriting varying city classifications, asserting that the maleinternational class has significance in determining a city’sfunding, and providing a flexible addition to the virtuouscircle paradigm.

REFERENCES

[1] Antoaneta P. Petkova, Violina P. Rindova, Anil K. Gupta, (2013) NoNews Is Bad News: Sensegiving Activities, Media Attention, and Ven-ture Capital Funding of New Technology Organizations. OrganizationScience 24(3):865-888.

[2] Chen, Henry, Paul Gompers, Anna Kovner, Josh Lerner. Buy Local?The Geography of Venture Capital. Journal of Urban Economics 67(2010): 90-102. 28 Mar. 2016.

[3] Dahl, Michael S and Olav Sorenson. The Migration of TechnicalWorkers. Journal of Urban Economics 67 (2010): 33-45. Web. 28 Mar.2016.

[4] Engel, Jerome S., and Itxaso Del-Palacio. ”Global Clusters of Innova-tion: The Case of Israel and Silicon Valley.” California ManagamentReview 52.2 (2011). Web.

[5] Florida, Richard, and Mark Samber. ”Capital and Creative Destruction:Venture Capital, Technological Change, and Economic Development.”(1994). Web. 28 Mar. 2016.

[6] Green, Milford B., and Rob B. McNaughton. ”Interurban Variation inVenture Capital Investment Characteristics.” Urban Studies 26 (1989):199-213. Web. 29 Mar. 2016.

[7] Hambrecht, William R. ”Venture Capital & the Growth of SiliconValley.” California Management Review 26.2 (1984). Web. 29 Mar.2016.

[8] Khirsheed, Mohammad S., and Mohammad A. Al-Fawzan. ”FosteringUniversityindustry Collaboration in Saudi Arabia through TechnologyInnovation Centers.” Innovation: Management, Policy & Practice 16.2(2014): 224-37.

[9] Sassen, S. New frontiers facing urban sociology at the Millennium.The British Journal of Sociology, 51: 143159. (2000). Web. 29 Mar.2016.

[10] ”Silicon Valley.” American Experience. PBS WGBH. Boston, MA,2013. Web.

[11] Wuebker, Robert, Nina Hampl, and Rolf Wustenhagen. The Strengthof Strong Ties in an Emerging Industry: Experimental Evidence ofthe Effects of Status Hierarchies and Personal Ties in Venture CapitalDecision Making. Strategic Entrepreneurship Journal 9 (2015): 167-187.

29

APPENDIX

30

Temporary Assistance with Lasting Effects: A Report on Policies of

Self-Determination in Native America

Ryan A. MatherAdvisor: Dr. Paul Glewwe

University of Minnesota, Department of Applied Economics

Abstract— Native American reservations are marked by

poverty rates that remain persistently above national averages

through generations, but a recent shift in policy toward greater

tribal control over previously federal and state-operated pro-

grams shows great promise in improving the situation. This

report seeks to better understand the economic impact of such

self-governance policy, first by examining the marginal effect

of having a tribe administer its own Temporary Assistance for

Needy Families (TANF) program, and second by examining

results from 75 household surveys that I conducted on the

Rosebud Sioux Reservation. I find that the implementation of a

TANF program by a reservation government produced a five-

percentage-point drop in the poverty rate above and beyond

any reduction in poverty that occurred in state-run programs

on Native reservations. Further, within tribal TANF programs,

there seem to be community gains associated with both ge-

ographical proximity and administrative proximity, that is,

having decisions made by a single local reservation government

as opposed to a consortium of reservation governments. The

survey shows similarly positive effects associated with proximity

and gives some reason to suggest that local programs would be

preferred to federal programs. Finally, I end the report with a

brief look at the Rosebud Reservation’s preferences for future

economic development programs as revealed in the survey.

I. INTRODUCTION

Present economic conditions on Native American reserva-tions are concerning and leave much to be hoped for, butthere is good reason to have hope. Concerning, because the2010 Census revealed that an estimated 23% of Native Amer-ican families were below the poverty line (U.S. Departmentof the Interior, 2014), and this poverty has brought with it ahost of other problems. Among major ethnic groups, NativeAmericans have the highest rates of depression, anxiety,and behavioral problems requiring treatment or counseling,and they are also the most likely to report difficulty inreceiving specialty care (Flores, 2013). They are victims ofviolent crime at double the rate of an average US citizen(Perry, 2004; Sarche & Spicer, 2008), and in 2000, one ofevery four suspects in federally prosecuted violent crimecases examined by government attorneys hailed from Nativereservations (Perry, 2004).

Particularly impoverished are reservations in SouthDakota, which has the highest rate of Native Americanpoverty in the nation. This report will place special em-phasis on one South Dakota reservation called Rosebud.2010 Census data reveal that the Rosebud Sioux Tribefaced unemployment rates of 50.8% (U.S. Department ofthe Interior, 2014) and that those who were working received

an average annual salary of $30,000 (United States CensusBureau, 2010). More recent census data placed Todd County,which is encompassed by the Rosebud Reservation, as thesecond poorest county in the nation based on a poverty rateof 47.4% (Release Highlights of 2014, 2015).

As difficult as the conditions on Native reservations are,however, recent progress is giving strong reason for hope. Af-ter a century of policies ostensibly designed to benefit NativeAmericans by forcefully assimilating them into the broaderUS culture, the late 1900s showed a gradual transition towardpolicies of self-determination, granting reservations and thetribes that govern them more freedom to chart their owneconomic futures. The results have far surpassed those ofpast programs (Kalt et al., 2008). In the 1990s, policies ofself-determination were correlated with an income growthon reservations three times that of the United States as awhole (Begay, Cornell, Jorgenson, & Kalt, 2007). This trendhas largely continued, leading Joseph Kalt of the HarvardProject on American Indian Development to say that for thefirst time in a century, the United States seems to have founda policy that works, and Indian nations are taking hold ofself-determination and making the most of it (Kalt et al.,2008, p. 112).

While this general positive trend is now well establishedin the literature, relatively little scholarly analysis has beendedicated to the specific implementation of these policies(King, 2007). Part of the problem is that Native Americanshave historically been a rather difficult ethnic group onwhich to collect data for several reasons. First, a highdegree of intermarriage has blurred the lines between Na-tive Americans and other ethnic groups, so that the sameindividual might primarily identify with one of two separateraces depending on the circumstance. Second, there is agreat degree of cultural difference between Native tribes,not unlike differences between Europeans who emigrated tothe United States. Third, Native Americans tend to spreadout widely across rural reservations and urban environmentsrather than concentrate in certain areas (Ericksen, 1996).

Census data is often the only type available, which isunfortunate. Not only does the census leave out importantdetails about tribes (U.S. Department of the Interior, 2014),but the Census Bureau has historically misrepresented NativeAmerican tribes through undercounting (Goodluck & WhiteHat, 2011). This lack of information limits the ability ofpolicymakers to evaluate the effects of self-determination

31

policies. In the case of employment programs like NativeEmployment Works (NEW), for example, official unemploy-ment rates can underestimate the problem by as much as 75percentage points by only counting those who are activelyseeking work as being unemployed. Similar data shortfallsfor Temporary Assistance to Needy Families (TANF) makeit difficult for states to determine whether Native families aregetting the welfare they need (Brown et al., 2001).

I will use recently released datasets to better understandthe effects that certain self-determination policies have hadon Native American reservations and to inform future self-government policy. First, the Akee Taylor 2014 dataset onNative American reservations is used, which I have supple-mented with variables pulled from government documents(United States Government Accountability Office, 2011;Fourth Annual Report to Congress, 2002) and Dippel’s 2014work on Forced Coexistence to assess the economic effects ofTANF programs operated by Native governments. For eachprogram that I examined, the US government provided anear equivalent program to reservations that did not electto administer their own, and so difference in differencesanalysis will provide a useful estimation of these programs’effects. Second, I will use survey data that I collected on theRosebud Native American Reservation to examine the impli-cations of possible self-government in a more local context,acknowledging that each Native American reservation has itsown respective culture and will experience policies of self-government differently.

In both sections, my hypothesis is that services providedby tribal institutions are preferred to, and achieve betterresults than, similar services provided by the US government.Also, I predict that tribal institutions in closer proximity tothe people they represent achieve greater results than thosethat are farther away. This hypothesis is largely confirmed:The implementation of a TANF program by a reservationgovernment seems to yield a five-percentage-point drop inthe poverty rate above and beyond any reduction in povertythat occurs under state-run TANF programs on reservations.Further, within tribal TANF programs, there seem to be com-munity gains associated with both geographical proximityand administrative proximity, that is, having decisions madeby a single local reservation government as opposed to aconsortium of such governments. The survey results alsolargely confirm my hypothesis by showing that people wholive closer to a tribal government are more likely to view itpositively and by giving some reason to suggest that localprograms would be preferred to federal programs. I end witha brief look at the Rosebud Reservation’s preferences forfuture economic development programs as revealed in thesurvey.

II. HISTORY

Self-governance on Native American reservations is nota new idea. It was the norm before Columbus’ journey,of course, but even during the colonization period, Nativeswere treated as belonging to separate nations (Strommer &Osborne, 2014). The trend continued for some time after

the United States became a nation, so that some 367 inter-governmental treaties were ratified with Native tribes duringthe first 90 years. Concessions made in those agreementsstill form much of the basis for the federal government’spresent obligations to Native tribes, as they were intendedto justify taking much of the land that the United Statesstill occupies (Warne & Frizzell, 2014). It was an officialshow of goodwill that belied an underlying racism towardNative Americans. In federal government documents, theywere recognized as civilized peoples, but private letters castthem as uncivilized savages (Strommer & Osborne, 2014).Nevertheless, the formal documents remained, and broadpromises to provide all proper care and protection typicalof those treaties created an official responsibility to seek thewell-being of Native tribes (Warne & Frizzell, 2014).

Unfortunately, this unspoken rule of self-governance on atribal level was soon altogether reversed by assimilationistpolicies intended to forcefully incorporate Natives into thebroader economic systems and cultural norms of the UnitedStates. Following the confinement of Native Americans ontoreservations, the first major attempt of the United Statesin ostensibly honoring their responsibilities to Native tribescame as the General Allotment Act of 1887, which providedfor the distribution of reservation land to individual NativeAmericans living on those lands (Washburn, 2006). Theidea was to make them into independent farmers, much likethe white settlers in rural areas. Whatever the government’sintentions, though, this and other assimilationist policiesthat characterized the era are now understood largely asfailed endeavors that, in most cases, contributed directly toreservation poverty (Kalt et al., 2008). The assimilationistera was brought to an end by the Meriam Report in 1928,which revealed conclusively that the General Allotment Acthad resulted in a large loss of land for Native Americans asoutside groups came in to buy up the best parcels (Washburn,2006).

Eventually federal policy shifted once again toward greaterself-determination, searching for a balance in what it meantfor tribes to be sovereign nations yet still under the authorityof the federal government. President Lyndon B. Johnsoncreated the Office of Economic Opportunity during his Waron Poverty, which all but ignored state and local governmentsin favor of partnering with grassroots community organiza-tions. Native tribes fit the bill, and so from 1965 to 1967,community action program grants to Native tribes went up by$16.5 million (Washburn, 2006). The trend continued underNixon, who is quoted as saying that Indians will get betterprograms and that public monies will be more effectivelyexpended if the people who are most affected by theseprograms are responsible for operating them (Strommer &Osborne, 2014, p. 17). One year after Nixon’s presidencyended, this positive sentiment finally found its way intolegislation with the passing of the Indian Self-Determinationand Education Assistance Act of 1975 (ISDEAA). TheISDEAA allowed Native tribes to contract with the Bureauof Indian Affairs (BIA) for federal funding to provide certainservices (Washburn, 2006). Even then, however, tribes were

32

held back from taking full advantage of the law becausetheir federal partners were hesitant to approve applicationsand thereby cede control (Strommer & Osborne, 2014).

Ironically, this unwillingness of certain agencies to fullycomply with ISDEAA helped tribes form an even strongercase to Congress that more sovereignty was needed (Strom-mer & Osborne, 2014). This time rebranded as Self-Governance, the new movement began with the Tribal Self-Demonstration Project Act of 1991. Under this act, 20 tribeswere selected to pilot a new program in which they couldnot only apply for federal funds toward certain programs,but would also have a great degree of freedom to tailorthose programs for their own individual needs (Tribal Self-Governance, 1991). John McCain, in a statement for theSenate Committee on Indian Affairs, would later affirm thatthis project has been a success and deserves to be establishedas a permanent option for all tribes (McCain, 1993, p. 1),a sentiment that held for the entire committee (Strommer& Osborne, 2014). As though making good on McCain’spromise, the 1994 Tribal Self-Governance Act (TSGA) didjust that, allowing all tribes to modify and administer pro-grams not just from the BIA, but from those offered byother federal agencies as well (King, 2007; Washburn, 2006;Strommer & Osborne, 2014).

A decade later, self-governance programs comprisedaround half the annual budget for both the BIA and theIndian Health Service (IHS) (Washburn, 2006). Since then,however, there have been no major developments in therelationship between tribes and the federal government. Thisis baffling both because the language used in the legislationis sufficiently broad to allow Native tribes more sovereignty(King, 2007) and because tribes now have an unprecedentedlevel of influence in Washington to further a policy thathas been shown successful (Washburn, 2006). The literaturegenerally predicts that self-governance is on the brink ofmore and greater change (Washburn, 2006; King, 2007;Kalt et al., 2008) and calls for deeper analysis on theimplementation of past programs to inform the possibilityof new ones (Washburn, 2006; King, 2007). This study aimsto take humble steps in that direction.

III. THE EFFECTS OF TRIBAL SELF-DETERMINATION

As tribes were gaining more freedom to implement theirown programs, the economic conditions on reservationsimproved tremendously. From 1970 to 2010, during whichtime real per capita incomes for the entire United States roseby 49%, those of Native Americans on reservations grew bya full 104% (Akee & Taylor, 2014). Changes in per capitaincome on reservations seem to be clustered around thepassing of key self-determination legislation, first increasingin the 1970s around the ISDEAA, then decreasing duringthe Reagan-era funding cuts, and finally increasing againover the 1990s around the TSGA, this time at a rate threetimes that of the national average (Vinje, 1996; Akee &Taylor, 2014; Begay et al., 2007; Kalt et al., 2008). Theshift has also been linked to an increase in privately ownedNative businesses, a number that more than doubled from

1992 to 2002 (Kalt et al., 2008). Unfortunately, the originaldisparities were great enough that even changes as significantas these made little dent in the average per capita incomedisparity between Native Americans on reservations and therest of the US population. Even if the current trends ofaccelerated growth were to continue, real per capita incomeson reservations would not intersect the averages of the UnitedStates until 2054 (Akee & Taylor, 2014).

Still, this is a remarkable movement in the right direction,and there are good reasons to believe that rising standards ofwelfare are more than just correlated with self-determinationpolicy. The first is the strength of the correlation even outsideof these particular policies. In its thirty years of researchacross a variety of tribal contexts, the Harvard Project onAmerican Indian Economic Development presents as itsprimary conclusion that self-determination is fundamentalto successful development on Native reservations, especiallywhen backed by effective Native government institutions(Ritsema, Dawson, Jorgensen, & Macdougall, 2015; Kaltet al., 2008). Second, this correlation arose even as federalspending on Native programs decreased in real terms. Ad-justing for inflation, all major federal expenditures on NativeAmericans, save those of the Indian Health Service (IHS),decreased from 1975 to 1996 (United States Senate, 1995).A report by the US Commission on Civil Rights over thisperiod found that the federal agencies primarily responsiblefor Native welfare were not only failing to meet their specialobligations to Native groups, but also providing services oflesser quality than those offered to the general Americanpublic (United States Commission on Civil Rights, 2003).This lower funding is contrasted against relatively higherfederal funding before 1975, during which time there are nosuccessful cases where federal planning and management hasproduced sustained economic development in Indian Country(Kalt, 1996, p. 4).

There are also good theoretical reasons to believe that self-determination would improve outcomes, both in terms ofcultural resonance and more pragmatic efficiency. It is wellestablished in social contract theory that groups of peopletend to work together best when their governance is carriedout according to mutually agreed upon principles (Cornell& Kalt, 1995). In the Native context, numerous studiessuggest that the extent to which a reservation’s currentsystem of government aligns with its cultural norms priorto the intervention of the United States has a positive andstatistically significant effect on its current standards of well-being (Cornell & Kalt, 2000; Akee, Jorgenson, & Sunde,2015; Dippel, 2014). By this reasoning, it seems natural thatthe level of government closest to individual Natives wouldbest represent their interests and thereby achieve greateroutcomes, especially considering how much tribal groupswant sovereignty. Freedom to chart their own futures isthe most common self-reported goal of Native nations inthe context of economic development (Kalt et al., 2008),and all but one of the tribal respondents to a GovernmentAccountability Office survey listed the flexibility to addresslocal cultural issues as a major benefit of administering

33

their own TANF programs (Temporary Assistance for Needy,2000).

With a greater understanding of, and closer proximity to,Native clients also comes efficiency gains for tribal groupsadministering previously federal programs. In surveys, tribaladministrators universally felt that they had a better un-derstanding of their tribe’s unique needs than state TANFworkers did (Hillabrant & Rhoades, 2003) and that the abilityto shape local TANF programs to fit those needs constituteda major benefit (Temporary Assistance for Needy, 2000).Locally driven programs can uniquely fit an area in ways thatcookie-cutter programs from the federal government cannot,allowing for a spectrum of programs ranging from computerskills training for elders on the Mille Lacs Reservation toencouraging subsistence hunting in the Yukon Flats (Begayet al., 2007). Proximity also allows tribal members greateraccess to tribal programs than their counterparts operatedby the state, and greater ability to coordinate with otherlocal welfare efforts in employment, child care, and the like(Hillabrant & Rhoades, 2003; Brown et al., 2001).

Equally true is that proximity begets greater accountabil-ity for tribal groups, which comes with both benefits andchallenges. Unlike the BIA, which is accountable mainly tothe federal government, tribal governments and consortia areforced to answer directly to their constituents. This greateraccountability is a benefit to self-determination policies, butwith it comes the reality that if tribal groups fail, there islittle federal safety net to catch them (Kalt et al., 2008).As the BIA gradually released a portion of its control, theability of reservations to govern themselves was put to thetest, and they diverged from each other as certain reservationswere better equipped to take on this task than others (Dippel,2014; Kalt et al., 2008). This difference may very well havebeen greater were it not the case that reservations neededto self-select into taking on programs from the government;a reservation that knew itself to be unprepared might havebeen content to stay with the state program.

On the surface, this suggests a mixed approach, where onlycertain tribal governments should be given increased control.Recent research, however, has again linked much of thisdivergence in tribal performance to how well a government’sstructure and ideology align with the cultural values of itsconstituents. Cornell and Kalt (1995) studied how well theconstitutions imposed on tribes by the United States fit witheach tribe’s previous systems of government by examiningfour characteristicsstructure, scope, location, and sourceinterms of their effect on modern unemployment and per capitaincome. Rosebud, for example, came from the Sioux Tribewhose original governmental characteristics were almostexactly opposite those of the imposed constitution in eachof the four categories. The study found a clear negativeeffect associated with the imposed constitution being a poorcultural fit, but an effect that was only allowed to becomeprominent when recent self-determination policies allowedtribal governments a real say in how federal dollars shouldbe spent. Cornell and Kalt then made the bold conclusionthat a match between extra-constitutional cultural norms and

formal institutions is a necessary condition for existence ofan economy based on real production [as] opposed to federaltransfers (1995, p. 406). A subsequent study over 67 tribesgave their results further weight (Cornell & Kalt, 2000), andwork by Akee, Jorgenson, and Sunde (2015) also showedsimilar results.

Aside from a tribal government’s constitution, another fac-tor which lay dormant until the advent of self-determinationpolicy was whether the United States had forcefully joinedpolitically distinct peoples onto a single reservation. Whenthe federal government was forming reservations, it decidedto group the Native Americans based on their tribe. Manyof these tribes, while being interconnected by languageand family structures, had not previously been politicallyunified. Dippel (2014) conducted a study to test the impactof this forced coexistence on current per capita incomes onreservations and found that forced coexistence was associatedwith a 30% lower per capita income in the year 2000.Strikingly, most of this difference was created only once self-determination policies came into effect. Another regressionwithin Dippel’s paper helps explain the reason by showingthat governments on reservations with forced coexistencehad three times as many incidents of internal conflict overthat same period as measured by the number of newspaperreports.

To the extent that negative tribal outcomes can be thusexplained, the situation seems to call for still greatersovereignty of a kind that helps tribes better express them-selves at a fundamental and constitutional level, especiallygiven the overarching trend of success for policies of self-determination on Native reservations. Constitutional reforminitiatives like this are gaining traction in reservations acrossthe United States, spurred in part by success stories ofreforms in the Mescalero Apache Tribe, the Mississippi Bandof Choctaw, the Confederated Salish and Kootenai Tribesof the Flathead Reservation, and others (Kalt et al., 2008).This along with more federally-funded capacity building willlikely help tribes manage their new responsibilities (Hill-abrant & Rhoades, 2003). Nevertheless, the larger positivetrends can disguise what is happening on a program-by-program level, and more information is needed to ascertainthe effects of each independently. The next section of thispaper seeks to do this for tribal TANF programs.

IV. TRIBAL TANF

When the Personal Responsibility and Work OpportunityReconciliation Act of 1996 (PRWORA) replaced the Aidto Families with Dependent Children (AFDC) program withTANF, it kept with the spirit of the TSGA by allowingfor Native tribes to operate their own TANF programs.Should a tribe apply to do so and be approved, the tribeis provided a yearly block grant to cover the expenses ofoperating the program and has the responsibility to do sofor the people within its jurisdiction (U.S. Department ofHealth, 2000). Any reservation that does not administerits own program, either because it did not apply or wasnot approved, is covered under its respective state program

34

(Brown et al., 2001). Both states and tribes were given aconsiderable amount of flexibility in operating their ownprograms. Toward the end of promoting work, responsibility,self-sufficiency, and strong families, states and tribes are freeto allocate their funding according to their own eligibilitylevels and service types (U.S. Department of Health, 2000).

TANF provides an ideal case by which to study therelative impact of state policy versus tribal policy under self-determination for several reasons: First, all reservations areprovided TANF services under PRWORA. This creates cleartreatment and non-treatment groups, with some reservationsreceiving TANF from their own tribal governments and somefrom the state government. Of course, there is the possibilityof self-selection bias for tribes electing to operate their ownprograms, but if factors can be found that help to explainthe bias, a still fuller picture will be created as to why somereservations prosper while others struggle. Second, fundingfor tribal administration of TANF was held constant at thelevel provided to that tribe under AFDC in 1994, and didnot increase relative to the states (Brown et al., 2001). Infact, as will be shown below, tribal services received lowerfunding than the state system. Third, the ways tribal and stateprograms could differ from one another are precisely thoserelevant to the question at hand. Namely, given a certainallocation of resources, are programs operated by tribes orby states able to achieve greater results on reservations?

V. METHODS

This ideal testing scenario is met, unfortunately, by a lessthan ideal dataset on which to do the testing. Specifically,there are several factors that should be controlled for whichthe following Ordinary Least Squares (OLS) models arenot able to incorporate. First, it would be useful to havea plausible proxy for reservation motivation, that is, theextent to which a government is trying to take advantage ofnew opportunities under self-determination. One such proxyfor this case might be to include a dummy variable forall tribes who applied to operate a TANF program even ifthat application was not approved. For example, Rosebudand six other Sioux tribes from its area jointly applied tooperate their respective TANF programs as a consortia oftribes (Humphrey, 1997), but funding decisions by the stateof South Dakota led to it not being approved (Brown et al.,2001). Data listing which other reservations might be in thissame situation would be relevant for two reasons: First, suchtribes would provide a better comparison for the treatmentgroup wherein a tribal TANF program was implementedbecause, in both cases, the desire to implement a programwas present, and the only change was the ability to do so.There may be some bias here as well in that the rejection ofapplications is also non-random, but nevertheless, such datawould provide a step in the right direction. Second, suchtribes may have found other ways to help needy familiesin their area that tribes with no desire did not, and theimpact of those possible initiatives would also be relevant toquestions about the effectiveness of tribally led poverty-reliefprograms. Unfortunately, I was not able to find any kind of

reliable data across all reservations that listed whether theydid or did not want to implement a program.

Another limitation on the dataset I am using is that itdoes not include information on the exact level of fundingavailable to reservations relative to the level of fundingavailable to states. Such funding is often obtained frommultiple levels, including local, state, and federal coffers,and so I was unable to get reliable numbers across bothstates and reservations for the time periods this study covers.Nevertheless, because reservation programs were generallyunderfunded compared to state programs, this lack of in-formation should, if anything, bias the regression resultsin favor of state programs. Funding levels for tribal TANFprograms were based upon AFDC funding for that tribe in1994, which, in turn, was determined by census data thatsystematically undercounted the number of Native familiesin need. Furthermore, unlike the states, which already had thenecessary infrastructure for TANF administration in place,tribes were starting fresh programs and had no additionalfunding to cover their start-up costs (Brown et al., 2001;Hillabrant & Rhoades, 2003). While states could receiveperformance bonuses, tribes could not (US Department ofHealth and Human Services, 2000), and while states servedan even distribution of TANF clients, tribal caseloads weredominated by clients classified as hard-to-serve (Brown etal., 2001). Thus, because the data do not reflect this fundingdisparity, they naturally make tribal programs seem worserelative to state programs, and will thereby underestimatethe positive effects per dollar spent of TANF programsadministered by reservation governments.

For each of the OLS regressions below, I began by mergingAkee and Taylor’s Native American Databook (2014) withDippel’s data on forced coexistence (2014). 1 Between thesetwo datasets, most of the indicators for reservations areavailable at three dates: 1990, 2000, and 2010. Then, usinggovernment documents (United States Government Account-ability Office, 2011; Fourth Annual Report to Congress,2002), I entered dummy variables based on whether eachreservation had elected to provide a particular service itself,or receive that same service from the federal government.TTANF2000, for example, is a dummy variable that takeson a value equal to 1 if a reservation decided to offer TANFon or before the year 2000, and 0 if they did not. Finally,I reduced the dataset to only those reservations that hadmore than 10 families in 1990, 2000, and 2010. This actioneliminated reservations that were too small for questions ofunemployment to take on any real significance and those thatwere subsumed into another between 1990 and 2010. Theresult was a sample of 163 reservations with data recordedover three dates each, for a total of 489 observations.

For the OLS regressions involving TANF services, specif-ically, I made further modifications to the dataset. First, Ientered dummy variables to capture whether a tribe admin-

1 To clarify, I only included reservations occupied by Native Americanson the mainland of the United States. Akee and Taylor’s dataset alsoincluded measures for Alaskan Natives, which I removed for the purposesof my regressions.

35

istered its own TANF program in the years 1997, 1998, 2000,2001, 2006, or 2010, where the earliest possible time theycould have begun doing so was 1997. Data for the otherindicators only existed for the years 1990, 2000, and 2010,so to correct for this I only allowed each dummy variable toaffect the regression in years at or after it had been put inplace. The TTANF2000 dummy variable, for example, whichindicates whether a reservation was administering their ownTANF program in the year 2000, was set to zero in 1990 evenif a tribe would later begin administering a TANF programin 2000. TTANF2000 provides the most useful point ofcomparison for this dataset because each reservation involvedcan be traced from a decade before this treatment to a decadeafter, and so it will be the main variable of interest in theregressions that follow. To remove interference from TTANFprograms added later, I removed from the dataset those 2010values of any reservation that added a new program after2000 but before 2010. Next, I included an additional dummyvariable called TTANFConsortia to account for whetherthe tribal TANF programs operated in the year 2000 wereadministered by a single tribe, or as a grouping of multipletribes in a tribal consortium. The resulting dataset had a totalof 463 observations. A fuller description of these and othervariables used in this section is provided in Appendix 1. Theregression equation for difference in differences analysis is:

PFamPovi,t =�0 + µt + PFamPovi,1990

+ �TTANF2000i,t +

JX

j=1

�jxj,i,t + ✏i,t

Where PFamPov is the percentage of families in povertyindexed by reservation i and time t = 2000, 2010 condi-tional on a constant �0, the set of year fixed effects µt,PFamPovi,1990, which remains constant at reservation ispoverty rate in 1990, a dummy variable TTANF2000 thatequals 1 for reservation i if that reservation was imple-menting its own TANF program by the year 2000, and anerror term. Beginning with that baseline model, J otherreservation controls xj,i,t are added in subsequent models tofurther establish the effects of TTANF2000i,t. To be clear,the regression coefficient associated with TTANF2000i,t

will not measure the impact of the TANF program per se.All reservations participated in some version of the TANFprogram, and so the effects of that program as a wholeare difficult to separate from the general trends of povertyreduction that occurred over this time period. Instead, the �

coefficient measures the expected benefit or loss associatedwith a reservation government administering the program asopposed to a state government. This is an important clari-fication because it sidesteps a larger debate on the efficacyof TANF as a whole, which, while largely applauded, hasalso been criticized for inadequately transitioning recipientsoff welfare (Bavier, 2001; Katz, 2012). Even if it were thecase that TANF as a whole had a negative effect, a positivecoefficient associated with TTANF2000 would show thattribal implementation yields positive results relative to stateimplementation.

Before running any regressions, it useful to examine a fewbasic statistics of the data. As noted earlier, the percentageof families in poverty across reservations decreased from1990 to 2010, a trend illustrated graphically in Figure 1.A second conclusion that appears immediately is that thecorrelation between TTANF2000 and PFamPov in theyear 1990 is positive at around 0.1186. This shows thatpoorer reservations in 1990 were more likely to take on theirown TANF programs when that option became available,which is exactly what would be expected if reservationswere effectively using self-determination opportunities toprovide programs consistent with their community’s uniqueneeds. TTANF2000 is also positively correlated with theunemployment rate in 1990, which underscores the need tocontrol for these original values in the regressions to isolatethe effects of TTANF .

Fig. 1: Percentage of American Indian Families in Povertyby Reservation

Table 1 presents the regression results of six differ-ent variations on the reservation controls portion. Overall,TTANF2000 is found to have a clear downward effect onthe percentage of families in poverty equivalent to aboutfive percentage points, and one that generally becomes onlystronger as more reservation controls are added. This is asignificant effect considering how in 1990, the average familypoverty rate across Native American reservations was 36.59percent. For a reservation to administer its own TTANFprogram, then, took approximately a seventh off its originalfamily poverty rate in addition to the general downwardtrend in poverty that occurred over the next two decades.Multiplying .05 by the mean number of families on a NativeAmerican Reservation in 1990 puts the results in humanterms: An average of 26 additional families were raised outof poverty on each reservation that administered its ownTANF program.

The first model does not include the TTANF2000 vari-able, and instead provides a slightly more rigorous de-scription of the downward trend in family poverty ratesestablished by Figure 1. Adding TTANF2000 into Model2 shows that it has a downward effect on the family povertyrate with a p value of .107, which is just above typical

36

standards for statistical significance. That p value decreasesto .006 by Model 3 with the addition of several key internalreservation controls. First, Unemployment1990 is added,providing a constant measurement of each reservation i’sunemployment rate in 1990. It is particularly importantto control for unemployment because TANF recipients arerequired to either hold a job or complete employment-relatedactivities as defined by the agency administering their bene-fits. The comparative lack of available jobs on reservations,then, makes this a large obstacle for tribal programs. Second,PHighSchoolDiploma measures the percentage of the Nativepopulation ages 25 or older in reservation i during year t witha high school diploma or higher degree. This is included forreasons similar to those used Unemployment1990 becausea worker’s ability to pull his or her family out of povertyis in part determined by the human capital that they canleverage for that purpose. Finally, the log of the NativeAmerican population in reservation i in year t is includedto account first for how larger reservations are typicallypoorer, and second for the added administrative challengeof administering a TANF program over a larger reservation.2 Controlling for these terms, then, helps to isolate theeffect of TANF programs and reveals that those operated byreservation governments have significantly greater benefits.

Once tribal TANF programs have been associated witha downward shift in family poverty rates, we are in aposition to test the second portion of my hypothesisthattribal institutions in closer proximity to the people theyrepresent achieve greater results than those that are fartheraway. There are at least two relevant ways of interpretingproximity to answer that question within the given dataset,one of which being physical proximity, that is, the geographicdistance between the site administering TANF and TANFrecipients. If physical proximity were important to the suc-cess of a tribal TANF program, then it would be expectedthat reservations farther from the next major city wouldexperience greater gains from operating their own TANFprograms than reservations closer to those cities where thestate government might have more influence. This is testedby two separate models above, the first being Model 4. Init, logdist 00 is included to measure the natural log of thedistance in miles between a reservation and the next majorcity measured in the year 2000. The result is statisticallysignificant at .01, and describes the same upward effect onfamily poverty rates that the theory predicts. Model 5 isless conclusive, which includes an interaction term betweenlogdist 00 and TTANF2000 to test the specific relationshipbetween reservation TANF programs and that reservationsdistance from a city. The sign of the interaction term isnegative as the theory would suggest, but to compensatethe TTANF2000 variable now predicts a positive changein the poverty rate, and neither variable takes on statisticalsignificance. Nevertheless, these regressions together give

2 It may seem here that a dummy variable should also be included tocontrol for whether or not a particular reservation i was operating a casinoin the year 2000, but I have excluded that variable from this analysis becauseit does not take on statistical significance in any of the regression models.

strong evidence that proximity to external population centersis important for the family poverty rate, and somewhatless precise evidence that more isolated reservations achievegreater gains for operating their own TANF programs.

Another relevant way to interpret proximity is admin-istrative proximity, or the degree to which a program islocally run. Among tribal programs, some are operated bya single reservation government, while others are jointlyoperated by a consortium of such governments. Using model4 as a baseline, the relative effect of having a tribal con-sortium implement a program can be found by adding thedummy variable TTANFConsortia that only equals 1 whenreservation i administered its TANF program as part of aconsortia in the year 2000. To clarify, all nonzero valuesin TTANFConsortia are also nonzero in TTANF2000,and so TTANFConsortia will only be significant if thereis some benefit or harm in consortia programs that cannotbe explained by an analysis of tribal TANF programs as awhole. The results, shown as Model 6, are that a consortiumprogram is associated with a .034 increase in family poverty,while TTANF2000 moves down to -.058 to compensate andretains a .01 significance. Because the TANF programs undertribal consortia are included in TTANF2000, this meansthat tribal consortia programs are still associated with abouta two-percentage-point reduction in family poverty levels. Aprogram administered by a single reservation government,however, is expected to reduce poverty rates by 5.8 per-centage points. There is, therefore, a clear positive benefitassociated with having a single tribal government administerits own TANF programeven greater than that of Tribal TANFprograms generally.

VI. SURVEY METHODS

Having argued for the effectiveness of TANF programsconducted with Self-Determination on Native Americanreservations, I now turn to a more local analysis of the Rose-bud Sioux Reservation to inform future self-determinationpolicy. The data used are from a survey that I conductedbetween July and August in 2016. In preparing this survey,there were several unique challenges presented by the reser-vation environment. The ordinary method of conducting asurvey like this would involve randomly selecting a sampleof home addresses from some central list, but such a methodis unfeasible on the Rosebud Reservation for a few reasons:First, it is a predominantly rural area, with some isolatedcommunities as far as a three-hour drive from the centralcapital. Second, once at those communities, it is not com-monly accepted for outsiders to visit individual homes. Onthe few occasions where I had opportunity to visit individualhomes on the reservation, I was often greeted by guard dogsand No Trespassing signs. Third, and finally, a centralizedlist of addresses from which to pull a sample does not exist.The closest thing would be the 911 directory, but even ifthat were to be made available to a visiting researcher likemyself, many addresses amount to the red house five milesdown Highway 83 and cannot be meaningfully sorted.

37

Instead of surveying at individual houses, then, I used aform of cluster sampling. Most of the eighteen communitieson the Rosebud Reservation have both monthly communitymeetings and an annual pow-wow. By calling the tribalcouncil member overseeing that community, I could some-times obtain permission to attend one of those meetings asa guest to administer surveys. This method removes muchof the challenges associated with surveying at individualhomes but does bring with it the risk of possibly double-counting some households. Due to the sensitive nature ofsome financial information on the survey, no informationwas collected on the surveys that could personally identifya respondent. The accompanying consent form did includea signature line where respondents listed their names, butagain, privacy demanded that a list of such names not betaken to each new community for cross-referencing purposes.Therefore, it would have been possible for one member ofa household to take a survey in the Corn Creek community,for example, and a second member of that same householdto enter the same information on a survey in the Missioncommunity. While this possibility cannot be ruled out, theclustering method was still used because the chances of sucha double-counting occurring are probably very low. I neverwent to a meeting in the same community twice, and therewere no incentives associated with the survey that mightmove a household to try and take the survey more than once.Further, having now completed my review of the surveys, itdoes not appear that any survey contains the same householdinformation as another.

In sum, I administered 75 household surveys representinga total of 279 respondents across seven of the Rosebudcommunities: Corn Creek, Milks Camp, Mission, Rosebud,Soldier Creek, St. Francis, and Upper Cut Meat. Before look-ing at the research questions this survey attempts to answer, itis useful to briefly describe some of the demographics of thesample. A full 52% of households surveyed were receivingfood stamps, and among such households, food stampsaccounted for an average of 77% of the household’s monthlyfood budget. It was a predominantly young sample, with 42%of household members being children under the age of 18,and only 6% over the age of 60. Finally, and significantfor this study, 17% of households were receiving TANFbenefits from the state government. A fuller description ofthe variables used in the regressions that follow is deferredto Appendix 2.

VII. SURVEY RESULTS

Because the Rosebud Sioux Tribe is more limited thanother tribes in terms of the self-determination policies ithas taken over at a tribal level, this survey sought mainlyto investigate the possible effects of expanding such pro-gramshow they would be received and what areas they oughtto focus on. In determining how they would be received, eachhousehold was asked to compare their experiences with theRosebud Sioux Tribal Council against those with state andfederal programs acting in the area, both in terms of theservices they received from each level of government and

the extent to which they believe each level of governmentto be looking out for their best interest. Perhaps it comesas no surprise that government is unpopular at all levels,as summarized in Figure 2, and that a respondent’s opinionof whether a governing group is looking out for their bestinterest exhibits a high positive correlation with their opinionof said government’s services: .812 in the case of theRosebud tribal government, and .740 in the case of the USgovernment.

Comparing average responses on this five-point scale,respondents ranked services from the Rosebud Sioux tribalgovernment slightly lower on average than services from theUS government, but considered the Rosebud Sioux tribalgovernment to be looking out for their best interest slightlymore than the US government. Taken together, these twofacts seem to suggest that respondents would welcome ashift in programs from the federal or state to the tribal level,where intentions better reflect those of the people, but wherethere is currently a relative lack of meaningful services.Alternatively, this difference in the respondents’ relativerating of Rosebud tribal services could also be explainednot as a problem with the quality of current services butrather the quantity. Under this second interpretation, I wouldargue that because the intentions of the Rosebud Sioux tribalgovernment are ranked higher in either case, the answer stillinvolves placing more programs under tribal oversight inRosebud. That shift, nevertheless, should be made in tandemwith capacity-building efforts so that the intentions of theRosebud tribal government are better conveyed in the results.

As can be seen in Figure 2, however, there is very littledifference in average responses between questions relatingto the Rosebud Sioux Tribe and parallel questions regardingthe US government. In fact, none of these differences is sta-tistically significant in the data, and even if such a differencecould be shown conclusively, respondents’ rankings of theirlocal government are not so positive as my initial hypothesiswould have predicted. In the aggregate, respondents to thissurvey did not seem to have a strong preference towardeither their local tribal government or the US governmentbut clearly would give low rankings to both. This, in part,confirms the results of Cornell and Kalt’s (1995) studymentioned earlier, which connected a poor match betweenthe Rosebud Sioux Tribe’s original governmental structureand the constitutions imposed on them by the United Statesto a downward effect on present economic outcomes. Forthem, the answer seems to be still greater sovereignty of akind that comes alongside tribes like Rosebud and helps themto restructure their own constitutions (Kalt et al., 2008).

Before moving to examine what possible federal programsthe Rosebud Sioux tribal government might seek to operatefor themselves, a secondary prediction of my hypothesiswas that proximity matters when it comes to a person’sexperience of governments, where those closer to the placewhere decisions are made and programs are administered aremore likely to have a better experience with government. Thecommunities that I visited to administer these surveys aregeographically spread far enough apart to meaningfully test

38

this by adding a variable that controls for the driving distanceto Rosebud, the capital of the Rosebud Sioux Reservation. Asimple regression for each respondent i measures the effectof this distance variable on the respondent’s opinion of howwell the Rosebud Sioux tribal government looks out for theirbest interest, while controlling for J other rankings x ofopinions and services on the same five-point scale:

RosebudSiouxTribeBestInteresti =

�0 + � log (DistanceToRosebudi) +

JX

j=1

�jxj,i + ✏i

The full results of this regression are shown in Table 2. Formodel 1, I control only for the respondent’s opinion of theservices provided by the Rosebud Sioux Tribe. As predicted,the physical distance between a respondent’s community andRosebud has a negative effect on their opinion of the RosebudSioux Tribe’s intentions statistically significant at .1. Whatis more significant for the purposes of this paper, however,is the respondent’s opinion of the Rosebud Sioux tribalgovernment relative to their opinion of the US government.Therefore, to control for this in model 2, I also includecontrols for the respondent’s opinion of whether the UnitedStates is looking out for their best interest and providingservices that improve their quality of life. Here, the down-ward effect of a respondent’s distance to Rosebud becomesstronger and more statistically significant. Thus, while thereis some uncertainty over which level of government thepeople of Rosebud view more favorably, there is strongerevidence that proximity does matter.

VIII. COMMUNITY PREFERENCES FOR FUTUREPROGRAMS

Should the Rosebud Reservation decide to take over moreprograms from the US government, the next step will beto determine which programs to focus on. To that end, thesecond portion of the survey included the question: Howimportant is it that new programs are created in each ofthe following areas on the Rosebud Reservation? followedby a scale on which to rank several categories of programswhich can currently be administered either by the federal orlocal government. The results of this question are presentedin Figure 3. Again, it is perhaps unsurprising that very fewrespondents ranked any of these programs as Not Importantor Not Very Important, considering that they are all goodthings described in words with positive connotations.

Popular vote is by no means an economically rigorousway of choosing economic development programs, but it is asignificant one in the context of tribal self-determination. Toreference an argument made earlier in this paper, programswhich are initiated locally and fit the culture of a givenarea generally tend to do better than cookie-cutter initiativesimposed by the federal government. Insofar as this dataaccurately represents the preferences of Rosebud, then, thereis reason to believe that programs like these should beinvested in more heavily. Furthermore, while this sort ofsurvey does not take into account the relative costs of these

different programs, it does provide one means of aggregatingcommunity preferences. It goes without saying that all surveyrespondents were rational agents making decisions basedupon their own preferences, which in the aggregate becomelocal indifference curves, that is, local values.

The two program areas ranked by respondents as mosthighly importantwork skills and housingare areas where theRosebud Sioux tribal government has already taken someinitiative in designing local programs. This is encouragingbecause I have claimed that this type of community respon-siveness should be an advantage of local tribal governments.Still, both areas demand more investment. For work skills,the Family Support Act of 1988 allowed reservations tooperate their own Job Opportunities and Basic Skills Training(JOBS) programs in 1990 much like reservations wouldlater be able to administer TANF, and Rosebud was oneof the tribes to do so. A General Accounting Office reportundertaken two years later in 1992, however, said that thepositive effects of the JOBS program would likely be limitedfor Rosebud because there simply were not jobs enough forthe labor force, no matter how skilled (Delfico et. al, 18). Alocally sponsored survey taken near that time in 1985 showedthat there were only 1,406 full-time positions among a laborforce of 7,241 people, and a mere 214 of these positions werein the private sector (Hargreaves & Chang, 1989). Eventually,the JOBS program became the NEW program, and theRosebud Sioux Tribe has continued to operate this programto present day (Grants amount allocated, 2015). Still, thesame problems associated with few job openings remain.Augmenting this program with other programs designed tocreate local businesses and jobs, which was the third mostdesired policy by respondents, will therefore be crucial.

The second most desired programs were in the categoryof housing. Rosebud’s remote and rural location strainslocal infrastructure in the provision of energy, water, androads to individual homes, and among those homes thatare adequately supplied, overcrowding can be a challenge.Studies prepared for the Rosebud Economic DevelopmentCorporation (REDCO) show that there is an immediate needfor over 500 homes in the area (Keya Wakpala, 2016). This isa large challenge, but in responding to this challenge REDCOand the broader Rosebud Sioux Tribe have exemplified thepotential of locally driven development initiatives. Commu-nity meetings and local surveys were used to identify keyvalues that local people wanted to have represented in futurehousing developments, which were in turn used as a basis forthe Keya Wakpala WAAGEYAPI Initiative, a plan to developsome 590 acres of tribally owned land into a residentialarea. In 2015, the plan was recognized with an internationalSocial Economic Environmental Design (SEED) award, andREDCO is currently in the process of putting it in place(REDCO Lakota, 2015).

IX. CONCLUSION

Self-Governance programs of the 1990s came with, andare generally accepted to have been a major factor in,

39

improved living conditions across Native American reser-vations. Using these results to determine the future of self-governance policy, however, demands a more nuanced un-derstanding of what specifically worked and what did notwithin the broad reforms of the TSGA and subsequent acts.The contribution of this study has been to focus in one suchfacet of self-governance and provide the humble beginningsof an analysis on whether there are benefits associated with areservation government administering its own TANF benefits.In answer, I found that there are, in fact, benefits on the orderof an additional five-percentage-point drop in the povertyrate beyond any poverty reduction that occurs in state-runTANF programs. Further, these relative benefits are expectedto be even greater the closer a tribal TANF program is tothe individuals receiving TANF benefits, both in terms ofphysical and administrative proximity. Similar research willbe needed to assess the relative costs and benefits associatedwith other facets of self-governance policy toward the endof a more complete picture that can be used in informingfuture government policy.

Should the government continue giving tribal entitiesgreater sovereignty over their portion of federal dollars,the beliefs and preferences of individual communities willbecome even more important than they already are in termsof shaping local economic outcomes. The survey portion ofthis report focused on the beliefs of one such community,the Rosebud Sioux Reservation, and described both theiropinions on the different levels of government as well as theirpreferences for future economic development programs. Thisresult does not constitute an official vote by the RosebudSioux Tribe, nor does it necessarily prescribe what shouldbe done, but instead provides an initial view of what maybe done as the Rosebud Sioux Tribe continues to defineitself apart from programs imposed by state and federalgovernments.

Just like my analysis of tribal TANF programs, this surveywill also need to be supplemented with further work definingthe local preferences of the Rosebud Sioux Tribe, but nogroup is as well-positioned to do that as the Rosebud SiouxTribe itself and its tribally-chartered partners like REDCO. Ifnothing else, the message of this study is not only that suchgroups can be trusted with chartering the economic futuresof Native American reservations, but, when supported withthe proper capacity-building efforts, are often the optimalgroups to do so. The present economic conditions on NativeAmerican reservations are truly concerning and leave muchto be hoped for, but tribal entities like these give strongreason for hope.

X. ACKNOWLEDGEMENTS

I would like to thank Dr. Paul Glewwe of the University ofMinnesota’s Applied Economics department for his adviceand mentorship over the course of this project, as well asDr. Richard Todd of the Minneapolis Federal Reserve andDr. Darin Mather of Crown College for their commentson the drafts of this paper. Special thanks also to WizipanLittle Elk, CEO of the Rosebud Economic Development

Corporation (REDCO), and all the REDCO staff for sograciously welcoming me this summer onto the RosebudReservation and providing countless insights through every-day conversation. This project would not have been possiblewithout Glen Marshall of the Lakota Journey, who providedme with both housing and a critical local perspective. I wouldfurther like to thank William Kindle, current president of theRosebud Sioux Tribe, for his approval of the survey that wasadministered at certain community meetings and pow-wows.This project was supported by the University of Minnesota’sUndergraduate Research Opportunities Program.

REFERENCES

[1] Akee, R. K. Q., and Taylor, J. B. (2014, May). Socialand economic change on American Indian reservations: Adatabook of the US censuses and the American CommunitySurvey 19902010. [PDF file]. Sarasota, FL: The Taylor PolicyGroup. Retrieved from https://static1.squarespace.

com/static/52557b58e4b0d4767401ce95/t/

5379756ce4b095f55e75c77b/1400468844624/

AkeeTaylorUSDatabook2014-05-15.pdf

[2] Akee, R., Jorgenson, M., and Sunde, U. (2015, November). Criticaljunctures and economic development evidence from the adoption ofconstitutions among American Indian nations. Journal of ComparativeEconomics, 43(4), 844861. http://dx.doi.org/10.1016/j.jce.2015.08.004

[3] Bavier, R. (2001, July). Welfare reform data from the Surveyof Income and Program Participation [PDF file]. Monthly LaborReview, 124(7), 1324. Retrieved from http://www.bls.gov/

opub/mlr/2001/07/art2full.pdf

[4] Begay, M. A., Cornell S., Jorgenson M., and Kalt, J. P. (2007).Development, governance, culture: What are they and what do theyhave to do with nation building? [PDF file]. Rebuilding Native Nations:Strategies for Governance and Development. Retrieved from http:

//isites.harvard.edu/fs/docs/icb.topic524968.

files/RNB_NNB-SGD_Dev_Gov_Cult_09-06.pdf

[5] Brown, E. F., Whitaker, L. S., Springwater, M., Cornell, S., Jorgenson,M., Hale, M., and Nagle, A. (2001, November). Welfare, work, andAmerican Indians: The impact of welfare reform. Report to theNational Congress of American Indians. Retrieved from https:

//buder.wustl.edu/research/Documents/Welfare,

%20work,%20and%20american%20indians-the%

20impact%20of%20welfare%20reform.pdf

[6] Cornell, S., and Kalt, J. P. (1995, July). Where does economic develop-ment really come from? Constitutional rule among the contemporarySioux and Apache [PDF file]. Economic Inquiry, 33(3), 402426.Retrieved from http://www.sfu.ca/

˜

allen/indian.pdf

[7] Cornell, S., and Kalt, J. P. (2000). Where’s the glue? Institu-tional and cultural foundations of American Indian economic de-velopment [PDF file]. Journal of Socio-Economics, 29, 443470.Retrieved from https://www.innovations.harvard.edu/

sites/default/files/cornell_kalt_2000.pdf

[8] Delfico, J. F., Bentsen, L., Packwood, B., Rostenkowski, D., andArcher, B. (1992, March). Welfare to work: Effectiveness of TribalJOBS programs unknown [PDF file]. Retrieved from http://www.

gao.gov/assets/80/78318.pdf

[9] Dippel, C. (2014, November). Forced coexistence and economic devel-opment: Evidence from Native American reservations. Econometrica,82(6), 21312165. doi:10.3982/ECTA11423

[10] Dippel, C. (2014, May). Forcedcoexistance webfinal.dta [Datafile]. Retrieved from http://www.anderson.ucla.edu/

faculty/christian.dippel/forcedcoexistence_

webfinal.dta

[11] Ericksen, E. P. (1996). Problems in sampling the Native Americanand Alaska Native populations. In G. D. Sandefur, R. R. Rindfuss,and B. Cohen (Eds.), Changing numbers, changing needs: AmericanIndian demography and public health (pp. 113129). Washington,DC: National Academy Press. Retrieved from https://www.nap.

edu/read/5355/chapter/9

40

[12] Flores, G., and Lin, H. (2013). Trends in racial/ethnic disparities inmedical and oral health, access to care, and use of services in USchildren: Has anything changed over the years? International Journalfor Equity in Health, 12, 10. doi:10.1186/1475-9276-12-10

[13] Fourth Annual Report to Congress. (2002, April). Administration forChildren & Families Archives, US Department of Health & HumanServices. Retrieved fromhttp://archive.acf.hhs.gov/

programs/ofa/data-reports/ar2001/indexar.htm

[14] Goodluck, C., and White Hat, M. (2011, September). Reclaim-ing Futures initiative from a Sicangu Lakota tribal perspective:Lessons shared. Children and Youth Services Review, 33(Supplement1), S29S33. http://dx.doi.org/10.1016/j.childyouth.2011.06.010

[15] Grant amounts allocated to American Indian entities administering afederal TANF program, a federal NEW program, and a federalTribal TANF Child Welfare Coordination Grant program.(2015). Rep. Office of Family Assistance. Retrieved fromhttps://www.acf.hhs.gov/sites/default/files/

ofa/tanf_new_and_ttcw_grant_amounts_fy_2015.pdf

[16] Hargreaves, M. B., and Chang, H. N.-L. (1989, April). Evaluatingthe impact of federal welfare reform legislation on Indian country:A case study of the Rosebud Sioux Reservation [PDF file] (Rep. no.PRS 89-2). Harvard Project on American Indian Economic Develop-ment. Retrieved from http://hpaied.org/sites/default/

files/publications/PRS89-2.pdf

[17] Hillabrant, W., and Rhoades, M. B. (2003, August). Operating TANF:Opportunities and challenges for tribes and tribal consortia. Reportsubmitted to United States Department of Health and Human Services,Office of the Assistant Secretary for Planning and Evaluation. Re-trieved from https://aspe.hhs.gov/report/operating-tanf-opportunities-and-challenges-tribes-and-tribal-consortia

[18] Humphrey, K. (1997, May). Seven Sioux tribes favor joining forceson TANF. Indian Country Today [Oneida]. Retrieved from ProQuest.

[19] Kalt, J. P. (1996, September). Statement of Professor Joseph P.Kalt before the United States Senate Committee on Indian Affairs[PDF file]. Harvard Project on American Indian Development. Har-vard University. Retrieved from http://hpaied.org/sites/

default/files/publications/senatfnl.pdf

[20] Kalt, J. P., Cornell, S., Curtis, C. E. A., Henson, E. C., Grant, K. W.,Jorgensen, M. R., Lee, A. J., and Taylor, J. B. (2008). The state of thenative nations: Conditions under U.S. policies of self-determination.New York, NY: Oxford University Press.

[21] Katz, S. (2012, July). TANF’s 15th anniversary and the great recession:Are low-income mothers celebrating upward economic mobility?Sociology Compass, 6(8), 657670. doi:10.1111/j.1751-9020.2012.00479.x

[22] Keya Wakpala Green Development. (2016). REDCO. Retrieved fromhttp://sicangucorp.com/community-development/

keya-wakpala-green-development

[23] King, M. A. (2007). Co-management or contracting? Agreements be-tween Native American tribes and the U.S. National Park Service pur-suant to the 1994 Tribal Self-Governance Act. Harvard EnvironmentalLaw Review, 31(2), 475530. Retrieved from http://www.law.

harvard.edu/students/orgs/elr/vol31_2/king.pdf

[24] McCain, J. (1993, October). Implementation on the tribal Self-governance Demonstration Project (S. Doc. 103-441). Hearing beforethe Committee on Indian Affairs U.S. Senate, 103rd Congress, Firstsession. Washington, DC: U.S. Government Printing Office.

[25] Perry, S. W. (2004, December). American Indians and crime: ABJS statistical profile for 19922002. The Bureau of Justice Statis-tics. Retrieved from http://www.bjs.gov/content/pub/

pdf/aic02.pdf

[26] REDCO Lakota Nation building project wins internationaldesign award. (2015, January 24). Lakota Voice. Retrieved fromhttp://www.lakotavoice.com/2015/01/24/redco-lakota-nation-building-project-wins-international-design-award/

[27] Release highlights of 2014. (2015, December). Small area

income and poverty estimates. U.S. Census Bureau. Retrievedfrom http://www.census.gov/did/www/saipe/data/

highlights/2014.html

[28] Ritsema, R., Dawson, J., Jorgensen, M., and Macdougall, B. (2015).Steering our own ship? An assessment of self-determination and self-governance for community development in Nunavut. The NorthernReview, 41, 157180. Retrieved from http://journals.sfu.ca/

nr/index.php/nr/article/view/474/510

[29] Sarche, M., and Spicer, P. (2008). Poverty and health disparities forAmerican Indian and Alaska Native children: Current knowledge andfuture prospects. Annals of the New York Academy of Sciences, 1136,126136. doi:10.1196/annals.1425.017

[30] Strommer, G. D., and Osborne, S. D. (2014). The history, status, andfuture of tribal self-governance under the Indian Self-determinationand Education Assistance Act. American Indian Law Review, 39(1),375. Retrieved from http://digitalcommons.law.ou.edu/

cgi/viewcontent.cgi?article=1001&context=ailr

[31] Temporary Assistance for Needy Families (TANF) Program:Third annual report to Congress [PDF file]. (2000 August).U.S. Department of Health and Human Services. Retrievedfrom http://archive.acf.hhs.gov/programs/ofa/

data-reports/annual3/annual3.pdf

[32] Tribal Self-Governance Demonstration Project Act (H.R. 3394). (1991,December). Bill passed 102nd Congress, First session. Washington,DC: U.S. Government Printing Office.

[33] U.S. Department of Health and Human Services, Administrationfor Children and Families. (2000, September). Temporary As-sistance for Needy Families (TANF). Almanac of Policy Issues.Retrieved from http://www.policyalmanac.org/social_

welfare/archive/tanf.shtml

[34] United States Senate. (1995, May). Report of the Committee on theBudget United States Senate to accompany S. Con. Res. 13 togetherwith additional and minority views (S. Rept. 10482). Report to the104th Congress, First session. Retrieved from ftp://ftp.loc.

gov/pub/thomas/cp104/sr082.txt

[35] U.S. Department of the Interior, Office of the Assistant SecretaryIndianAffairs. (2014, January). 2013 American Indian Population and LaborForce Report [PDF file]. Retrieved from http://www.bia.gov/

cs/groups/public/documents/text/idc1-024782.pdf

[36] United States Government Accountability Office. (2011, September).Temporary Assistance for Needy Families: HHS needs to improveguidance and monitoring of tribal programs (GAO-11-758). Reportto the ranking member, Committee on Natural Resources, Houseof Representatives. Retrieved from http://www.gao.gov/new.

items/d11758.pdf

[37] United States Census Bureau. (2010). Rosebud Sioux Tribe’s 2010statistical profile. South Dakota Department of Tribal Relations. Re-trieved from http://www.sdtribalrelations.com/new/

tribalstatprofiles/rststatprofile2011.pdf

[38] United States Commission on Civil Rights. (2003, July). A quietcrisis: Federal funding and unmet needs in Indian country [PDFfile]. Retrieved from http://www.usccr.gov/pubs/na0703/

na0204.pdf

[39] Vinje, David L. (1996). Native American Economic Developmenton Selected Reservations: A Comparative Analysis. The AmericanJournal of Economics and Sociology, 55(4). 427-442. JSTOR. Re-trieved from http://www.jstor.org.ezp2.lib.umn.edu/

stable/pdf/3487617.pdf

[40] Warne, D., and Frizzell, L. B. (2014, June). American Indianhealth policy: Historical trends and contemporary issues. Amer-ican Journal of Public Health, 104(Supplement 3), S263S267.doi:10.2105/AJPH.2013.301682

[41] Washburn, K. K. (2006, May). Indian law at a crossroads: Tribalself-determination at the crossroads. Connecticut Law Review, 38,777796. Retrieved from http://www.lexisnexis.com.ezp2.

lib.umn.edu/hottopics/lnacademic/

41

APPENDIX

TABLE I: Examining the Effects of Tribal TANF Programson Family Poverty Rates

(1) (2) (3) (4) (5) (6)

Intercept �0 .110*** .110*** .077´ .041 .030 .030

Year µt -.011 -.010 -.026 -.002 -.001 -.001

PFamPovi,1990 .439*** .449*** .272*** -260*** .266*** .270***

TTANF2000 -.027 -.044** -.049** .038 -.058**

PUnemployed1990 .173* .095 .058 .081

PHighschoolDiplomai,t -.024*** -.254*** -.261** -.258***

log(Population)i,t .023*** .022*** .022*** .022***

logdist 00 .022** .024** .023**

TTANF2000 ⇤ logdist 00 -.024

TTANFConsortia .034

R2 .125 .356 .379 .4362 .608 .457’ p < 0.1, * p < 0.05, ** p < 0.01, *** p < 0.001

Fig. 2: Views of Federal and Local Government

TABLE II: Relative Views of Government Controllingfor Distance to the Capital City

Variables (1) (2)

Intercept .839*** .011*

Log(DistancetoRosebud) -.103 ´ -.111*

RosebudSiouxTribeServices .804*** .747***

USBestInterest .030

USServices .126

R2 .125 .356’ p < 0.1, * p < 0.05, ** p < 0.01, *** p < 0.001

42

Fig. 3: Importance of Economic Development Program

Tribal TTANF Variable Summary

Mean/Percentage of Tribesand (Standard Deviation) across

Variables Description 1990, 2000, and 2010PFamPov This will be the regressand for all regressions on

TANF. It gives the percentage of Native Americanfamilies for reservation in poverty.

.3015 (.1393)

Intercept/µt µt is a dummy variable that takes on a value of 1 inthe year 2010. By this structuring, the constant termcan be interpreted as the baseline constant effect inthe year 2000.

TTANF2000 A dummy variable that takes a value of 1 if atribal group was administering the TANF programfor reservation in the year 2000, a 0 if one was not.This variable is set to zero for all reservations in1990 because at that point, the effects of a programadministered in 2000 would not have been felt.

14.11%

PUnemployed90 The unemployment rate in 1990 for reservationgiven by the number of Native Americans unem-ployed divided by the number of Native Americansin the labor force.

.2064 (.0924)

PHighschoolDiploma The percentage of Native Americans aged 25 or olderon reservation that have either a high school diplomaor a college degree, given by the sum of NativeAmericans who have obtained either a high schooldiploma or college degree by age 25 divided by theNative population ages 25 and older. This variablecan change from year to year in the regression for agiven reservation.

.4041 (.0936)

log(AIANPopulation) The natural log of the sample population on reser-vation as reported by US Censuses. This variable isalso allowed to change from year to year for eachreservation.

6.778 (1.367)

logdist 00 The natural log of the distance between the borderof reservation and the nearest major city, defined asany city with 50,000 people or more.

3.611 (1.014)

PFamPov90 The percentage of families in reservation that werein poverty in the year 1990. This is held constantthrough all years in the regression.

.3659 (.1470)

TTANFConsortia A dummy variable that equals 1 when a consortiumof tribes together administered the TANF servicesto some reservation , and 0 if a reservation electedto administer TANF services on its own or simplyreceived TANF services from the state government.

2.808%

43

Survey Summary

Variable Description Mean and (Standard Deviation)RosebudSiouxTribeBestInterest Respondents were asked to what extent they agreed

or disagreed with the statement The Rosebud SiouxTribal Council is looking out for my best interestwith the options strongly disagree, disagree, indiffer-ent, agree, and strongly agree. These answers werecoded 1-5 such that an answer of strongly agreewould equal 5 and an answer of strongly disagreewould equal 1. This is the regressand for the surveyregressions.

2.5733 (1.21)

Log(DistanceToRosebud) Gives the natural log of the distance in miles alongrecognized roads between a particular communityand Rosebud, the capital of the Rosebud Reservation.In some cases, this may have overestimated thedriving distance between two communities becausethere are many backroads on the Rosebud Reser-vation that cut through private property but are,nevertheless, widely used. For surveys administeredwithin Rosebud, that distance was zero, and so zerois used for those values of Log(Distance to Rosebud).The next closest community surveyed was 7.2 milesaway, and so Log(Distance to Rosebud) never takeson negative values.

2.606 (1.419)

RosebudSiouxTribeServices Respondents were asked to what extent they agreedor disagreed with the statement The Rosebud SiouxTribal Council provides services that improve myquality of life with the options strongly disagree, dis-agree, indifferent, agree, and strongly agree. Theseanswers were coded 1-5 such that an answer ofstrongly agree would equal 5 and an answer ofstrongly disagree would equal 1.

2.493 (1.234)

USServices Respondents were asked to what extent they agreedor disagreed with the statement The United Statesgovernment provides services that improve my qual-ity of life with the options strongly disagree, dis-agree, indifferent, agree, and strongly agree. Theseanswers were coded 1-5 such that an answer ofstrongly agree would equal 5 and an answer ofstrongly disagree would equal 1.

2.547 (1.166)

USBestInterest Respondents were asked to what extent they agreedor disagreed with the statement The United Statesgovernment is looking out for my best interest withthe options strongly disagree, disagree, indifferent,agree, and strongly agree. These answers were coded1-5 such that an answer of strongly agree wouldequal 5 and an answer of strongly disagree wouldequal 1.

2.460 (1.149)

44

Does Intercropping Improve the Outcomes of Export Assistance AmongKenyan Smallholders?

Noah NietingAdvisor: Amy Damon

Macalester College, Department of Economics

Abstract— Commodity export markets provide opportunitiesfor gains from trade in developing economies, but smallholdingfarmers remain least likely to access these markets due to un-scalable factors, risks of monocropping, meeting internationalregulations, and volatile cash crop prices. This paper usesrandomized control trial data from western Kenya that gavesmallholders access to export assistance and credit services, thusremedying issues of price dispersion and imperfect export andcredit markets found in the smallholder literature. An ecologicaladaptation of production possibilities theory hypothesizes thatintercropper income should be equal to or greater than thatof monocroppers, ceteris paribus. A series of difference-in-difference empirical approaches finds that program participa-tion interacted with intercropping has a positive but statisticallyinsignificant effect on non-wage agricultural income per acre.Results confirm the expected sign from theory and show thatresearchers and development practitioners should not forgointercropping given small scales.

I. INTRODUCTION

Commodity export markets provide opportunities for gainsfrom trade and income generation among the world’s poorestfarmers. These smallholding farmers, however, remain leastlikely to access these markets due to unscalable factors,risks of monocropping, meeting international regulations,and volatile cash crop prices. Traditional smallholder agricul-tural practices such as intercropping, the practice of growingmultiple crops in a field, can mitigate risk by means of cropdiversification and soil management. It remains to be seenwhether the benefits of intercropping translate to success inexporting when smallholders finally access global markets.

Using randomized control trial (RCT) data, this studyasks: Does intercropping improve the outcomes of exportassistance among Kenyan smallholders? While incomes werepreviously found by Ashraf, Gine, & Karlan (2009) toincrease from export assistance at a statistically insignif-icant level, perhaps farmers who practiced intercroppingexperienced more positive results. Results found a positivebut statistically insignificant relationship between non-wageagricultural income per acre and RCT participation interactedwith intercropping.

Section II surveys the literature, finding that the relation-ship between exporting and intercropping remains contested.Section III describes the RCT implementation and summarystatistics. Section IV sets up the theoretical framework forthe study, drawing from a synthesis between economics andecology developed in Ranganathan, Fafchamps, & Walker(1991), and Section V applies this to an empirical framework.

Empirical results in Section VI show positive but statisticallyinsignificant effects of RCT participation interacted with in-tercropping on income per acre. While a few data limitationshandicapped the study, results encourage further study anddevelopment practice linking intercropping and exportationas a means to boost incomes without scaling land and capitalinputs.

II. LITERATURE REVIEW

According to economic and ecological literature, inter-cropping contributes positively to input and overall produc-tivity. A number of studies demonstrate that intercroppingmaize and beans contributes to higher soil productivity andquality by keeping and replacing nitrogen among plant roots(Altieri, 1999; Hodtke, de Almeida, & Kopke, 2016; Peter& Runge-Metzger, 1994). Other studies explore the positiverelationship between intercropping and labor productivity(Lai et al., 2012; Norman, 1977; Okigbo & Greenland, 1977).Alene & Hassan (2003) note how productivity increases insoil and labor can interact to improve overall efficiency useof inputs with intercropping. Also, studies suggest that yieldsand overall farm productivity increase when smallholdersintercrop (Alene & Hassan, 2003; Arslan et al., 2015; Laiet al., 2012). Relatedly, yields have lower variability, whichreduces production risk (Achoja et al, 2012; Arslan et al.,2015; Dorsey, 1999; Just & Candler, 1985; Okigbo, 1978;Rodriguez, Rejesus, & Aragon, 2007). Studies consistentlyfind evidence that smallholder farmers, much like thosefeatured in this study, are risk averse. (Carter, 1997; Schaefer,1992).

Additionally, a large number of studies show that inter-cropping has substantial effects on output market outcomes.Allison-Oguru, Igben, & Chukwuigwe (2006) note howNigerian smallholders who intercropped experienced higherrevenues than those who did not. Numerous studies find thatintercropping increases profitability via some combination ofincreasing yields, improving input efficiency, and decreasingrisk and income sensitivity to crop prices (De Groote et al.,2012; Dorsey, 1999; Lai et al., 2012; Manda et al., 2016; Pe-ter & Runge-Metzger, 1994; Rodriguez, Rejesus, & Aragon,2007; Saka, 2015). In sum, efficiency and productivity gainsof intercropping often lead to increased agricultural profitsin domestic markets.

Consensus on intercropping outcomes divides, however,in the context of exporting. Two major studies relate inter-

45

cropping to income generation by means of export. Bothcoincidentally took place in Kenya. Dorsey (1999) findsthat small Kenyan farms experienced higher income perhectare by having both high rates of crop diversification andspecialization. The results of Dorsey (1999) at first seemcontradictory, but they stem from intercropping producingsubstantial amounts of crops for export (specialization) aswell as other crops for household and domestic market con-sumption (diversification). These results are consistent withthe predominating notion in the literature that intercroppingis more productive and profitable than monocropping forsmallholders.

On the other hand, Minot & Ngigi (2004) discover thatFrench bean monocropping in Kenya resulted in highersmallholder yields and incomes than maize-bean intercrop-ping. Minot & Ngigi (2004) note, however, that their resultsdo not incorporate the implicit labor costs of householdlabor. Given that crops like French beans are more labor-intensive than intercropping with maize, they may haveunderestimated production costs. Despite these results, Minot& Ngigi (2004) do not completely discount intercroppingand conclude that more perfect markets, (homogenous pricedispersion, reduced transaction costs and credit constraints,and improved information and access to exporters) remainthe crucial components to export success.

The results in the literature on the complementarity ofintercropping and exporting are inconclusive, but Minot &Ngigi (2004) specifically note the importance of more perfectmarkets. The RCT data used in this study, and explained inthe following section, include an intervention that relaxedinformation and credit constraints for smallholder exporters.This intervention allows study of the counterfactual desiredin Minot & Ngigi (2004), asking: Does intercropping im-prove the outcomes of export assistance among smallholders?

III. DATA DESCRIPTION

Ideal data to assess this question would allow regress-ing the income generated from export cash crops on thequantity of crop sold. The data would be able to dis-tinguish between crops sold from intercropped fields andthose from monocropped fields, all while controlling forrelated production endowments and other technology. AnRCT in Ashraf, Gine, & Karlan (2009; 2015) in collaborationwith the Abdul Latif Jameel Poverty Action Lab at theMassachusetts Institute of Technology provides much ofthese data, available through Harvard University’s Dataverse.Their study introduced export assistance and credit servicesto farmers in the Kirinyaga district of Kenya through a non-governmental organization called DrumNet. The baselinesurvey was conducted in April 2004 with the follow-upsurvey in May 2005.

To participate in DrumNet, farmers needed to belong to aself-help group (SHG, also known as a farmers’ association),express interest in exporting cash crops through their SHG,have some irrigated land, and at least a week’s worth ofwages, or approximately 10 USD, in DrumNet’s TransactionInsurance Fund (TIF). Participants in the treatment group

first embarked on a four week Good Agricultural Practicescourse to ensure crop quality suitable for export to the EU.This was followed by placement into joint liability groupsof 5 for each farmer chosen for treatment who met theTIF requirement. Farmers could take on loans up to fourtimes the size of their TIF collateral. DrumNet distributedseeds (primarily French beans and baby corn) and otherinputs shortly after group placement. It later negotiated priceswith farmers after harvest and forwarded quality produce toexporters. Loan payments and TIF savings were deducted byDrumNet from gross sale of produce, and remainders werecredited to farmers in private savings accounts. The offer ofthe Good Agricultural Practices course, production inputs,and the facilitation of sending crops to exporters comprisedthe “export assistance” treatment. The offer of TIF savingsaccounts and loans comprised the “credit” treatment.

Farmers’ entrance into the control group of no DrumNetservices and the two treatment groups, DrumNet serviceswith credit and without, was randomized. Randomizationminimizes biases of self-selection because treatment is as-signed instead of taken up voluntarily. With minimizedbiases, inter- and intra- group characteristics are likely toresemble each other excepting treatment and eliminate mostproblems of endogeneity. Preliminary analysis by Ashraf,Gine, & Karlan (2009) found no significant characteristicdifferences in the data within and among control and treat-ment groups.

Further attention is necessary to account for differencesbetween monocroppers and intercroppers. A summary ofvariables of interest is featured in Table 1. The total numberof observations is 1,109 members across the baseline andfollow-up periods. The variables of interest are non-wageagricultural income per acre (dependent variable), a dummyequal to 1 if farmers practice any form of intercroppingFrench beans or baby corn (maize) and 0 otherwise, and alist of controlled variables that affect production and incomeincluding household size, field size, machine or animaluse, pesticides use, land quality, and distance from sale.Intercropped field size is also included to show how muchtotal production uses intercropping.

The mean of the dependent variable (non-wage agriculturalincome per acre) is 25,994.6 Kenyan shillings (225.10 USD)with a standard deviation of 58,859.22 shillings. This signalsa skew rightward resulting from a few outliers of very largeincome. There remains a slight rightward skew as well inthe intercropping dummy, where the mean value is 0.77 witha standard deviation of 0.42. Since the mean is not 0.5, itshould be noted that the number of intercropped cash cropplots is larger than those that are monocropped. Intercroppedfield size is on average approximately 40 percent of totalfield size, showing modest level of diversification of all cropswithout taking into account the other 60 percent of remainingland.

Machine or animal use, land quality, and pesticide use aremore evenly distributed with means closer to 0.5. Householdsize, field size, and distance from sale are non-normallydistributed and skewed to the right. Histograms for these

46

variables are featured in Figures 1, 2, and 3, respectively.Three observations were deleted in household size due tosuspected input error. Attention will be paid in the nextsection regarding the impact of non-normally distributed datafound in household size, field size, and distance from sale.

A more refined method to assess the validity of thesegroups (monocroppers and intercroppers) is through a two-sided t-test, which determines whether the means of thegroups for variables are statistically significantly differentfrom each other. The t-test results are available in Table2. Three independent variables, machine or animal use,household size, and land quality, point to a statistically sig-nificant difference in the distributions of the variables. Thesedifference are significant at the p0.1, p0.05, and p0.01levels, respectively, meaning there is little likelihood that thetrue means are equal to each other. For further analysis, re-gressions should take these household characteristic variablesinto account to remedy the incomplete randomization amongintercroppers and monocroppers. Finally, it is worthy ofnote that no statistically significant difference exists betweenthe means of the groups regarding the dependent variable.There remains, however, a larger distribution of outcomesamong intercroppers, which may portend heterogeneity ofoutcomes attributable to intercropping interacting with RCTparticipation.

IV. THEORETICAL FRAMEWORK

Economic theory provides a framework with which toassess the effectiveness of intercropping in terms of yieldsand income generation. Ranganathan, Fafchamps, & Walker(1991) adapt the ecological concept of the land equivalencyratio to production possibilities theory. After a brief intro-duction to the concept of the land equivalency ratio, a modelfor production efficiency between two crops can be used todetermine whether intercropping produces more, equally, orfewer yields than monocropping with implications for trade.

The land equivalency ratio (LER) measures the yields ofintercropping against those of monocropping. In essence, theratio expresses how much more or less a given plot of landcan produce of crops by intercropping over monocropping.If the LER is greater than 1, then intercropping producesgreater yields than monocropping, ceteris paribus. If the LERequals one, intercropping yields equal monocropping yields,and an LER less than 1 means intercropping produces feweryields than monocropping. The LER is given by (1) and isthe summation of the ratios for each crop i of intercrop yieldsover monocrop, “pure stand,” yields.

LER =MX

i=1

Y ield of Intercrop

Y ield of purestand

(1)

As Ranganathan, Fafchamps, & Walker (1991) note, thevalues of the LER correspond to the coefficients appliedto a standard Cobb-Douglas production possibilities frontier(PPF) as shown in Figure 4. The PPF demonstrates allpossible output bundles between two crops, x and y, whileholding constant resources and technology. The PPF is

always downward sloping because resources and technologymust be diverted from one product to another in order toproduce more of the latter.

An LER of 1 is equivalent to monocropping and a direct1:1 tradeoff between crops. It is thereby represented bycurve A. Curve A is straight because there is no change inopportunity cost by switching crops when the LER equals 1.An LER greater than 1 means intercropped species comple-ment each other to increase yields, which corresponds to aconcave curve like B. Its concavity comes from the varyingopportunity costs of having disproportionate amounts of onecrop, which decreases the complementarity of intercrop pro-duction. Finally, an LER less than 1 means crops compete forresources, and so their mixture decreases overall productionand creates a convex curve like C. The convexity of curveC arises from the high opportunity costs of expanding pro-duction of a second, competitive crop. Thus, if intercroppinguses complementary crop species that support each other’smineral needs, productivity for a given level of a crop shouldbe greater under intercropping than other methods.

The PPF model posits that the income generated fromcomplementary intercropping will always be equal to orgreater than that of monocropping. Figure 5 imposes an arbi-trary price ratio, represented by the dotted lines, to establishthree optimal points given by the points where the negativeprice ratios are equal to the marginal rate of transformationbetween crops. All optimal points for monocropping arebelow or equal to the output of complementary intercropping,meaning the latter always meets or exceeds the income ofthe former.

Producers accept world crop prices upon the eliminationof trade barriers. If the world price rests above the domesticprice, then the country exports the crop. According to classi-cal Ricardian trade theory, producers derive this comparativeadvantage from technology that allows them to produce at alower price. It is possible intercropping may serve as sucha technology and establish a price advantage. This is likelythe case with Kenyan smallholders switching to horticulturecrops bound for wealthy European markets who pay a globalpremium for a product that uses globally cheaper inputs.

This result depends on whether the assumption of constantresources and technology remains intact. It is expected thatendowments of land, labor, and capital as well as non-intercropping technologies, such as pesticides, will differacross individuals and affect output rates. Data descriptionfrom the previous section found this to be true betweenmonocropping and intercropping groups regarding householdsize, machine and animal use, and land quality. Nonetheless,a proper model specification can account for these issues.

V. EMPIRICAL FRAMEWORK

The theoretical understanding of intercropping grantedby the PPF translates rather directly to a specification foran outcomes like non-wage agricultural income per acre, adescaled measurement isolating income from crop sales:

47

yit = �0 + �1Tit + �2Ait + �3ICit + �4TitAit+�5TitICit + �6AitICit + �7TitAitICit + �Vit + ✏it

(2)where y equals income generated for time t and household

i. T is a binary variable equal to 1 if household i belongsto any treatment group and 0 otherwise. A is a binaryvariable that equals 1 if the time period t is the follow-up, post-program period. IC is also a binary variable equalto 1 if household i intercropped cash crops. Present arethe first (single variable), second (two variables), and third(all three binary variables) orders necessary to test theinteraction of treatment placement, time, and intercropping.V is a vector of control variables for household i at timet, including household size, machine or animal use, soilquality, pesticide use, field size, and distance from sale.These variables proxy endowments, technologies, and costsin hopes of both recovering the broken assumptions of theory,accounting for statistically significantly different means, andisolating the interaction effect of RCT program participationand intercropping. Finally, ✏it is the error term with meanzero.

The sign of �7 tests the hypothesis that intercroppingparticipants experienced improved program outcomes overtheir monocropping peers by interacting time, treatment, andintercropping. If positive, the coincidence of intercroppingand RCT program participation contributed positively toincome generation. If negative, the converse is true. Thecoefficients �1, �2, and �4 are expected to be positive as theyshould mirror the results of Ashraf, Gine, & Karlan (2009).The coefficient �4 interacting time and treatment could benegative, however, because its result could be reconstructedonce interacted with the likely positive effect of intercroppingon income. The coefficient �3, the effect of intercropping, isexpected to be positive as it relates to outcomes like yieldsand income. The signs for the intermediary coefficients of �5

and �6, are ambiguous, but their interpretation is not usefuland is thus ignored. Additionally, generated LER estimatesfrom the data using (1) will validate or cast doubt upon theconnection between the data and the theory established inthis section. The LER can discern the validity of the theoryas it applies to the data and regression results, but is unlikelyto cast doubt upon the results of the regression themselves.It will prove a helpful tool in conceptualizing the outcomespresented in the regressions.

VI. EMPIRICAL RESULTS

Before any presentation or analysis of econometric results,a glance at the calculated LER should determine expectationsfor regression results as well as the adequacy of productionpossibilities theory in explaining intercropping productiontrends in the data. In the context of the data used, the LERmeasures average yields of intercropped over monocroppedmaize, beans, French beans, and coffee. Data on yields forother crops were unavailable but represent only 12 percent ofcultivated area in the dataset, so the estimation should retainmost of its precision.

The LER estimate for the data equals 3.12, which meansthat the crops in question have a complementary relationshipin the data and that, on average, 212 percent more land isrequired of monocroppers to match the yields of intercrop-pers. With regards to production possibilities theory, the PPFis concave and should establish that intercroppers earn moreper acre from growing the same crops as monocroppers, allelse equal.

The results for an initial regression of equation [2] arefeatured in column (1) of Table 3. It defines RCT treatmentsimply as placement into either treatment group, otherwiseknown as intent-to-treat (ITT). Ashraf, Gine, & Karlan(2009) warned against use of other treatment definitions,such as treatment on the treated (TOT), as they wouldlikely overlook spillover effects by measuring only RCTcompliance. The regression uses robust standard errors toaccount for issues of heteroskedasticity.

The coefficients for treatment, after, and intercropping areall positive, as predicted by theory. Intercropping interest-ingly has a large effect that is statistically significant atthe p0.1 level. The treatment and after interaction alsomaintains its positive but statistically insignificant coefficientfrom Ashraf, Gine, & Karlan (2009), featured in column(5). The coefficients for the other two second order interac-tions are negative with that of treatment and intercroppingstatistically significant at the p0.1 level. These signs donot match predictions, but their interpretive value remainsminimal. The third order interaction term interacting after,treatment, and intercropping is positive and insignificant atthe p0.1 level, confirming only in sign the prediction ofproduction possibilities theory and the LER derived from thedata. The triple interaction also has the greatest magnitudeof all coefficients excepting the constant.

Most control variables match the prediction of theory andprevious studies. Household size has a negative sign, whichresults from the rising dependency ratios of larger households(McCulloch & Otas, 2002; Muriithi & Matz, 2015). Farmsize and income per acre are inversely related and significantat the p0.01 level, which results from decreasing returnsof scale among labor-intensive smallholders (Von Braun,Hotchkiss, & Immink, 1989). Related to this finding isthat of the inverse relationship between income per acreand machine or animal use, which stems from their useon larger, though less efficient, smallholdings. Land qualitycontributes positively to income as a component of totalfactor productivity (Witcover, Vosti, Carpentier, & de AraujoGomes, 2006). The case is the same for pesticide use, whichis positive and statistically significant at the p0.1 level.Distance from sale does not match theory’s prediction oftransaction costs, but results remain statistically insignificantat the p0.1 level. This likely stems from outliers that,because input error is not suspected, do not warrant exclusionfrom analysis and will be corrected for in another regressionusing a logarithmic transformation.

The sign and magnitude of these results remain when usingclustered errors at the SHG level, featured in column (2).Clustered errors recognize unobserved heterogeneity at the

48

group level as a tradeoff for not recognizing unobservedheterogeneity among individuals. Concerning the data athand, farmers received treatment at the SHG level, whichposits an appropriate identifier for difference across groupsas it corresponds to both treatment designation and likelyavenues for spillover effects. Significance remains the sameexcept for the treatment and intercropping interaction, whichno longer has significance, and the triple interaction term,which becomes significant at the p0.01 level. Clusterederrors, however, do not account as much as robust standarderrors for heteroskedastic error biases that affect statisticalinference.

A logarithmic transformation of the data is featured incolumn (3) to account for nonnormality caused by outliersand skewed distributions. The signs and general magnitudeof previous results remain excepting the after-treatment inter-action, machine and animal use, and distance from market.The interaction of after and treatment turns negative butstatistically insignificant, which falls within the analysis ofAshraf, Gine, & Karlan (2009) in that the program was notfound to have a statistically significant effect on incomeper acre. Machine and animal use turns positive but statis-tically insignificant, reflecting theory in that capital deep-ening should increase smallholder productivity and income;decreasing returns to scale may have been overestimated.Finally, distance from market turns negative and statisticallysignificant at the p0.01 level. This result matches theory ontransaction costs and demonstrates the overall improvementof the logarithmic transformation on the previous regressionin matching theoretical predictions.

The triple interaction remains positive but loses its sta-tistical significance at the p0.1 level. The interaction ofprogram participation and intercropping thus had a positiveimpact on non-wage agricultural income per acre, but it is notstatistically discernable from zero. These results have robust-ness at the SHG level as shown in column (4). These resultsagain resemble those of Ashraf, Gine, & Karlan (2009)in that the program had no statistically significant impacton income per acre. It also relates back to the statisticallyinsignificant difference between dependent variable outcomemeans found in the t-tests of Section III.

These final results reflect the direction predicted by theory,but not the magnitude. While the LER value of 3.12 showssoils are conducive to intercropping, the LER may haveoverestimated the effect of intercropping on income. A highLER could have resulted from pesticide use, which wouldinflate the perceived effect of intercropping on yields andincome per acre. Sadly, the LER does not decompose toaccount for the effect of pesticides in these data and thereforecannot be re-estimated.

VII. CONCLUSION

This study asked whether intercropping improved the out-comes of export assistance among Kenyan smallholders. The

use of RCT data and its intervention of export assistance andcredit services corrected for imperfect price dispersion andmarkets that plagued Minot & Ngigi (2004). Results reflectedthe expectations of the production possibilities theory derivedin Ranganathan, Fafchamps, & Walker (1991) with non-wageagricultural income per acre increasing with intercropping.The results were not statistically significant, however, whenaccounting for heteroskedasticity and nonnormality. Resultshave a sense of internal validity in that they reflect the overallprogram conclusions of Ashraf, Gine, & Karlan (2009). Itis possible that data overestimated the LER by neglectingpesticides.

A number of limitations exist for this study and its impli-cations. First, results do not take into account any adoptionor abandonment of intercropping. The resulting bias is notlikely to be significant, however, because intercropping asa traditional practice is not one to be quickly abandoned,nor did DrumNet ever encourage monocropping. Second,this study does not separately assess both treatment groups,which may bias results downward if credit services had animpact on income masked by grouping treatments together.Also, the external validity requires substantial attention tolocal environmental conditions since the benefits of inter-cropping rely heavily on local soils and crops. Finally, cashcrop analysis was undertaken at the household level, the onlylevel available for dependent variable outcomes. Future studyof crop level outcomes would more directly measure howintercropping affects income generated.

Nonetheless, my results show that intercropping farmersare shown to be just as, if not more, capable of incomegeneration as monocroppers when both export. The theorybehind intercropping and the positive, though statisticallyinsignificant, effect intercropping had on exporter incomewarrant further attention from researchers and policymakers.Future research could better substantiate the interaction ofintercropping and exporting by better randomizing acrossagricultural practices and by studying crop level outcomes.It could also pay closer attention to scale in data andtheory because the benefits of intercropping experience largedecreasing returns to scale, a dynamic only tangentiallymentioned in this study.

Ashraf, Gine, & Karlan (2009) explain the difficultiessmallholders have in meeting EU quality regulations, butnothing in their description of such regulations preventssmallholding intercroppers from coordinating with organi-zations like DrumNet to capture gains from trade. Thus, thefailure of this study to reject intercropping as a sustainable,income-generating agricultural practice in a global marketmeans that policymakers and development practitioners canstill see intercropping as a potential, though small, stopgapto food insecurity, environmental degradation, and ruralpoverty.

49

VIII. ACKNOWLEDGEMENTS

I would like to first acknowledge the mentorship of Profes-sor Amy Damon, particularly in putting me on the right pathto select the topic and data for this paper. Second, JessicaTimmerman reviewed this paper and offered comments andcorrections of immense worth. She made it more readable,coherent, and relevant. Ibrahima Dieye also offered someedits to improve its clarity. I would like to mention thebountiful data, ease of access, and ease of use that goeswith using J-PAL papers and datasets. I appreciate how theirdiligence and attention to detail permitted me to undergo thisproject. Finally, I would like to thank the support network ofmy peers in the classroom and those who happen to inhabitthe Macalester Econometrics Lab. We would not have knownwhat we were doing, nor had as much fun doing it, had itnot been for each other.

REFERENCES

[1] Achoja, F. O., Idoge, D. E., Ukwuaba, S. I., & Esowhode, A. E.(2012). “Determinants of export-led cassava production intensificationamong small-holder farmers in Delta State, Nigeria.” Asian Journal ofAgriculture and Rural Development 2(2):142-148.

[2] Alene, A. D., & Hassan, R. M. (2003). “Total factor productivity andresource use efficiency of alternative cropping systems in two agro-climatic zones in Eastern Ethiopia.” Agricultural Economics Review4(2):32-46.

[3] Allison-Oguru, E., Igben, M. S., & Chukwuigwe, E. C. (2006).“Revenue maximising combination of crop enterprises in Bayelsa Stateof Nigeria: A linear programming application.” Indian Journal ofAgricultural Economics 61(4): 667-676.

[4] Altieri, M. A. (1999). “Applying agroecology to enhance the produc-tivity of peasant farming systems in Latin America.” Environment,Development and Sustainability 1(3):197-217.

[5] Arslan, A., McCarthy, N., Lipper, L., Asfaw, S., Cattaneo, A., &Kokwe, M. (2015). “Climate smart agriculture? Assessing the adap-tation implications in Zambia.” American Journal of AgriculturalEconomics 66(3):753-780.

[6] Ashraf, N., Gine, X., & Karlan, D. (2009). “Finding Missing Markets(and a Disturbing Epilogue): Evidence from an Export Crop Adoptionand Marketing Intervention in Kenya.” American Journal of Agricul-tural Economics 91(4):973-90.

[7] Ashraf, N., Gine, X., & Karlan, D. (2015). “Finding Missing Markets(and a disturbing epilogue): Evidence from an Export Crop Adoptionand Marketing Intervention in Kenya,” doi:10.7910/DVN/PES3RD,Harvard Dataverse, V1.https://dataverse.harvard.edu/dataset.xhtml?per

sistentId=doi%3A10.7910/DVN/PES3RD

[8] Carter, M. R. (1997). “Environment, technology, and the social artic-ulation of risk in West African agriculture.” Economic Developmentand Cultural Change 45(3):557-590.

[9] De Groote, H., Vanlauwe, B., Rutto, E., Odhiambo, G. D., Kanampiu,F., & Khan, Z. R. (2010). “Economic analysis of different optionsin integrated pest and soil fertility management in maize systemsof Western Kenya.” American Journal of Agricultural Economics41(5):471-482.

[10] Dorsey, B. (1999). “Agricultural intensification, diversification, andcommercial production among smallholder coffee growers in centralKenya.” Economic Geography 75(2): 178-195.

[11] Hodtke, M., de Almeida, D. L., & Kopke, U. (2016). “Intercropping ofmaize and pulses: an evaluation of organic cropping systems.” OrganicAgriculture 6(1):1-17.

[12] Just, E. R. & Candler, W. (1985). “Production functions and rationalityof mixed cropping.” European Review of Agricultural Economics12(3):207-31.

[13] Lai, C., Chan, C., Halbrendt, J., Shariq, L., Roul, P., Idol, T.,Chittanrajan, R., & Evensen, C. (2012). “Comparative economic andgender, labor analysis of conservation agriculture practices in tribalvillages in India.” International Food and Agribusiness ManagementReview 15(1):73-86.

[14] Manda, J., Alene, A. D., Gardebroek, C., Kassie, M., & Tembo, G.(2016). “Adoption and Impacts of Sustainable Agricultural Practiceson Maize Yields and Incomes: Evidence from Rural Zambia.” Amer-ican Journal of Agricultural Economics 67(1):130-153.

[15] McCulloch, N., & Ota, M. (2002). “Export horticulture and povertyin Kenya.” Brighton, UK: Institute of Development Studies, pp. 1-33.

[16] Minot, N., & Ngigi, M. (2004). “Are horticultural exports a replicablesuccess story?: Evidence from Kenya and Cote d’Ivoire.” Washington,DC: International Food Policy Research Institute, pp. 1-113.

[17] Muriithi, B. W., & Matz, J. A. (2015). “Welfare effects of vegetablecommercialization: Evidence from smallholder producers in Kenya.”Food Policy 50:80-91.

[18] Norman, D. W. (1977). “Economic rationality of traditional Hausa dry-land farmers in north of Nigeria.” In R. D. Stevens, ed. Tradition andDynamics in Small-Farm Agriculture Ames IA: Iowa State UniversityPress, pp. 63-91.

[19] Okigbo, B.N. (1978). “Cropping systems and related research inAfrica.” Association for the Advancement of Agricultural Science inAfrica: Ibadan, Nigeria.

[20] Okigbo, B. N. & Greenland, D. (1977). “Intercropping systems intropical Africa.” In R. I. Papendick, P. A. Sanchez & G. B. Triplett,eds. Multiple Cropping. Madison WI: American Society of Agronomy,pp. 63-101.

[21] Peter, G. & Runge-Metzger, A. (1994). “Monocropping, Intercroppingor Crop Rotation? An Economic Case Study from the West AfricanGuinea Savannah with Special Reference to Risk.” Agricultural Sys-tems 45(2):123-143.

[22] Ranganathan, R., Fafchamps, M., & Walker, T. (1991). “EvaluatingBiological Productivity in Intercropping Systems with ProductionPossibility Curves.” Agriculture Systems 36(2):137-157.

[23] Rodriguez, D. G. P., Rejesus, R. M., & Aragon, C. T. (2007). “Impactsof an agricultural development program for poor coconut producersin the Philippines: An approach using panel data and propensityscore matching techniques.” Journal of Agricultural and ResourceEconomics 32(3):534-557.

[24] Saka, J. O. (2015). “Productivity and income potentials of intercropcombinations among food crop farmers in Southwestern Nigeria.”African Journal of Agricultural Research 10(52):4730-4737.

[25] Schaefer, K. C. (1992). “A portfolio model for evaluating risk ineconomic development projects, with an application to agriculture inNiger.” American Journal of Agricultural Economics 43(3):412-423.

[26] Von Braun, J., Hotchkiss, D. & Immink, M. (1989). “Non-traditionalexport crops in Guatemala: Effects on production, income, and nutri-tion.” Washington, DC: International Food Policy Research Institute,pp. 1-98.

[27] Witcover, J., Vosti, S. A., Carpentier, C. L., & de Araujo Gomes,T.C. (2006). “Impacts of soil quality differences on deforestation, useof cleared land, and farm income.” Environment and DevelopmentEconomics 11(3):343-370.

50

APPENDIX

51

52

53

54

55

The Influence of Collectivism on Microfinance in Senegal

Cole ScanlonHarvard University

Keaton ScanlonWarren Wilson College

Teague ScanlonPomona College

Advisors: Andrei Shleifer and Susan Kask

Abstract— Microfinance, despite its mixed results in economicliterature, continues to proliferate in many developing countries(Rooyen et al., 2012). This research project investigates therelationship between collectivism and microfinance. It analyzesthe question: how does a collectivist culture and its normsinfluence the ways in which borrowers spend loaned fundsand interact with microfinance institutions? We generate atheoretical model for how norms of informal redistributionaffect borrowing decisions and use a robust dataset of all ofthe loans facilitated by Kiva, a global microfinance institution,to compare microloan borrowing in countries with differentcultures of collectivism. A case study of Senegal, a culturallycollectivist country, includes surveys and detailed interviewsof individuals and microfinance institutions (MFIs). We findthat the strong social networks associated with collectivism arewell-adapted to the structures of many MFIs. However, wealso uncover that some of the collectivist social norms, such asnorms of informal redistribution, can deter individuals fromusing microfinance.

I. INTRODUCTION

“Mi anda si mido faala heebude prłt. No saati fii si hidajoogi kalis bwee en Sngal. Si hida joogi, a footike joonudeyeembe beng bwee. Woono jaatigibe maa waala benguremaa buuri faala heebude kalis. Si ko dun, haray mi wowatarootude kalise.”

“I don’t know if I truly want a loan. It’s hard if you havea lot of money in Senegal. If you do, you have to share it.Maybe your families and friends need the money more. Ifthat’s the case, then I can’t pay back the loan.”

- Mamadou D*, subsistence farmer in Thiabekaare, Senegal,January 2017.

There is wide-ranging research on microfinance in de-veloping countries, yet less is understood about the effectsof cultural norms on how individuals and groups interactwith microfinance institutions (MFIs). Strong social networksthat typically exist in collectivist cultures have been foundto play an important role in the diffusion of microfinance(Banerjee et al., 2013). Countries in sub-Saharan Africa,which have been characterized as having high or moderatelevels of collectivistic values, also have well-documentednorms of redistribution whereby individuals frequently trans-fer a substantial share of their income to members of theirsocial networks, i.e. members of the household or extendedfamily, friends, and neighbors (Baland et al., 2016; di Falcoand Bulte, 2011). This informal redistribution shapes thesocial and economic habits of individuals; people makeresource-allocation choices accounting not only for their

own socioeconomic status but also for that of the membersof their social networks (Platteau, 2000, 2006, 2014). Inseeking to understand the role of collectivism and strongsocial networks in microfinance, economic research hasfocused primarily on the risk-sharing dimension of informalredistribution in economies where people are structurallyvulnerable to income shocks, but have limited access tofinancial markets and to formal redistribution systems (Coxand Fafchamps, 2007 ).

Recent economic literature, however, has increasingly fo-cused on the potential adverse effects of collectivist culturalnorms. Norms of informal redistribution have been foundto distort economic decisions related to investment and en-trepreneurialism (Grimm et al., 2013; Hadness et al., 2013).Such norms have also been found to result in strategies toescape the pressure to redistribute, for instance, by favoringnon-easily-sharable assets, hastening some expenses, andhiding income sources and easily-shared resources (Balandet al., 2011; Boltz et al., 2015; di Falco and Bulte, 2011;Goldberg, 2013).

In this paper, we aim to investigate the influence ofcollectivism on microloan borrowing decisions. A simpleeconomic model is used to detail how norms of informalredistribution, in the form of sharing profits from a microloanproject, could reduce borrowing. A cross-country comparisonusing a complete dataset of the 1,049,576 loans facilitatedfrom 2007 to 2016 by Kiva, a global microfinance insti-tution, is used to uncover whether, in practice, borrowersfrom countries with different levels of collectivism interactwith microfinance differently. We then detail a case studyof microfinance in Senegal, which has been defined as acollectivist culture with well-documented collectivist culturalnorms, including informal redistribution, to better understandthe mechanisms by which collectivism influences borrowingdecisions (Senegal - Geert Hofstede, 2017; Boltz, 2015).Surveys and ethnographic interviews of lending institutions,small business owners, and individuals in Senegal wereconducted.

Through the cross-country analysis, we find evidencesuggesting that collectivist cultures tend to borrow smallerloans and tend to have shorter loan terms. Through theresearch trip in Senegal, including surveys and detailedinterviews with 103 Senegalese natives, we uncover howcollectivist cultures are well-suited for the structure of groupborrowing that is used by microfinance to distribute risk.On the other hand, we also discovered some of the waysin which collectivist cultural norms can deter entrepreneurs

56

from borrowing money or starting ventures, because of thenorms of informal redistribution.

Our research adds to the few ethnographic accounts onmicrofinance that engage with the processes of borrowingrather than the institutional outcomes driving the indus-try (Brett, 2006; Perry, 2006; Gurin, 2006; Duffy-Tumasz,2005). Our account highlights the importance of culturalcontext, not only in structuring of microfinance products suchas microloans, but also in spreading the concept of financialservices to isolated regions and villages.

II. METHODOLOGY

In this study, a combination of economic theory, cross-country data analysis, surveys, and interviews were used.The surveys and interviews were conducted in the Dakarand Kedougou regions of Senegal from December 2016to January 2017 with participants ranging from lendinginstitutions, to small business owners, to individuals.

A. Theoretical FrameworkFollowing from the Stiglitz (1990) model of peer moni-

toring in a competitive credit market, we create a theoreticaleconomic model to uncover how informal norms of redistri-bution - specifically profit sharing from successful microloanprojects - could influence borrowing decisions.

B. Empirical StrategyThe cross-country data used for this research includes

94 countries and was aggregated from three sources. First,data on all loans facilitated from 2007 to 2016 by Kiva,a global microfinance institution, were downloaded usingthe Kiva API (Kivatools, 2017). A total of 1,049,576 loansfacilitated by Kiva were included. Our data second source isGeert Hofstede’s Index of Individualism, which was mergedwith the Kiva dataset on the country variable (Index ofIndividualism, 2014). The index of individualism measuresthe “degree of interdependence a society maintains among itsmembers [which is]...related to whether people’s self-imagesare defined in terms of ‘I’ or ‘We’. In individualist societies,people are supposed to look after themselves and theirdistinct family only [while] in collectivist societies peoplebelong to ‘in groups’ that take care of them in exchangefor loyalty” (Senegal - Geert Hofstede, 2017). The range ofscores on Hofstede’s Index of Individualism observed in thedataset includes a least individualistic, minimum score of 6(Guatemala), a most individualistic, maximum score of 91(United States), and a mean score of 27. Senegal received ascore of 25, which is a relatively low score. The final datasource was GDP per capita, PPP (current international $),which was downloaded from the World Bank Dataset andwas used as a control in the cross-country analysis (WorldBank Databank, 2016).

In investigating the differences in how individualist andcollectivist cultures interact with microfinance, cross-countrycomparisons were used. The average loan size and loan termwere used as metrics to compare countries with differentcultures of collectivism (as measured by Hofstede’s Index ofIndividualism).

C. Ethnographic Strategy

To investigate the influence of collectivist cultural normson microfinance, a detailed ethnographic study was con-ducted in Senegal from December 2016 to January 2017.Senegal is considered a relatively collectivist society and, inGeert Hofstede’s Index, is described in detail:

“A low score of 25 in this dimension means that Senegalis considered a collectivistic society. This is evident in aclose, long-term commitment to the member ‘group’, be thata family, extended family, or extended relationships. Loyaltyin a collectivist culture is paramount and overrides mostother societal rules and regulations. The society fosters strongrelationships where everyone takes responsibility for fellowmembers of their group. In collectivist societies: offenceleads to shame and the loss of face, employer/employeerelationships are perceived in moral terms (like a familylink), hiring and promotion decisions take account of theemployee’s in-group and management is the management ofgroups (Senegal - Geert Hofstede, 2017).”

During the research trip, 103 individuals were interviewedfrom a range of professions, backgrounds, ages, genders,and geographical locations within Senegal. The locations ofthe research include Dakar, the urban capital of Senegal,Kedougou, a rural city in the south-east of Senegal, and fourrural villages in the Kedougou region of Senegal (Dindefelo,Thiabekaare, Segou, and Pellel).

Fig. 1: Map of Senegal

While in the Dakar and Kedougou regions, surveys andextensive ethnographic interviews were used to investigatepeople’s attitudes towards microfinance and to understandthe influence of collectivist cultural norms on people’s in-teractions with microfinance institutions (MFIs) in Senegal.The survey (included in Appendix A), includes 6 questionsrelated to people’s attitudes and interactions with MFIs. Theextensive interviews immediately followed the surveys andwere used to provide greater detail and substance to theanswers recorded from the surveys.

57

Various precautions were taken to mitigate bias in answers,to both the survey and interview questions, that wouldplausibly emerge given that the researchers are foreignersfrom the United States. The Kedougou region, which hasa population of approximately 152,134, has a relativelyminimal foreigner presence, and was selected because ofthe extensive social network that one of the researchers,Keaton Scanlon, had previously developed in this region(Senegal Census Data, 2013). Keaton spent four months inthe Dakar region and a cumulative year and a half living inthe Kedougou region of Senegal, primarily in the village ofThiabekaare, but also in Dindefelo, Segou, Pellel, and thetown of Kedougou. In addition to becoming proficient in thelocal language of Pular, Keaton developed a strong socialnetwork in both regions. Many of the people in Keaton’sestablished social networks in the Dakar and Kedougou re-gions were surveyed and interviewed, because of the relativereliability and expected honesty of their answers. To expandbeyond Keaton’s network in getting honest attitudes andanswers, Seydou Diallo, a bilingual native to the Kedougouregion, utilized his network established from his 20 yearsliving in the region, and assisted in administering surveysand conducting interviews. All interviews were conducted inPular, the local language in which Keaton is fluent. In Dakar,where the majority of people speak French and Wolof, peoplewho spoke Pular were sought out, and even provided someaid translating responses from proximate participants.

In addition to interviewing 103 individuals in the Dakarand the Kedougou regions, employees at various micro-finance institutions in the regions were also interviewed.Employees of Credit Mutuel du Senegal, a microfinanceinstitution, as well as Orange Money, a mobile moneyservice, were interviewed to get a greater understandingof microfinance in Senegal from the business providers’perspective. These interviews were conducted in Wolof bySeydou Diallo, and then translated into English.

III. BASIC THEORETICAL FRAMEWORK

Our theoretical economic model is directly specified at theborrower level and sheds light on how norms of informalredistribution in collectivist cultures could affect borrowingdecisions. Our analysis in this paper follows from Stiglitz(1990), who constructs a model of peer monitoring in acompetitive credit market.

Assume that all individuals have various projects whichthey can undertake, such that project i, if successful, returnsYi(L) when undertaken at scale L (measured in CFA francof expenditure). If project i fails, its returns are zero. Theprobability of success for each project is pi. To account fornorms of informal redistribution in collectivist cultures, let �be the proportion of the profits that are shared if a project issuccessful, such that 0 � 1. We assume that � is largerin countries with higher levels of cultural collectivism. If theproject is successful a proportion of profits from the project,�Yi(L), is shared and (1+r)L, the loan amount plus interest,is paid back to the MFI where the rate of interest r is fixed.

If the project is unsuccessful, then nothing is shared but, still,(1 + r)L is paid back to the MFI.

Thus, the expected consumption tomorrow from undertak-ing project i, with rate of interest r, is:

C = �piYi(L)� pi(1 + r)L+ (1� pi)(�(1 + r)L)

For the borrower presented with the expected consumptiontomorrow from undertaking project i represented above, wehave the following maximization problem:

max

L�0�piYi(L)� pi(1 + r)L+ (1� pi)(�(1 + r)L)

The slope of the indifference curve if the individualundertakes project i is:

@C

@L= �piY

0i (L)� (1 + r) s.t.

@2C

@L2< 0

Thus, in the borrower’s maximization problem, we are leftwith the following comparative static:

�piY0i (L) = (1 + r)

The indifference curve, @C@L , represents the marginal con-

sumption of scale, where scale is measured in CFA francsof expenditure. It is evident in the comparative static, wherethe rate of interest, r, and the probability of success, pi,are fixed, as the proportion of the profits that are shared, �,increases, the marginal returns of scale, Y 0

i (L), decreases.This suggests that, in practice, a borrower in a culturewith norms of informal redistribution would have a lowerpropensity to take on projects with larger scale compared toa equivalent borrower not in a culture with norms of informalredistribution. Given that there is an opportunity cost ofundertaking a new project, such as forgone consumptionof goods that are not easily sharable, the profit sharingproportion, �, decreases the propensity of a borrower to takeout a loan in the first place (i.e. given the lower marginalutility of scale, utility could be maximized when L = 0 andall wealth is invested into consumption). We investigate thesedynamics through cross-country analysis of microloan dataand through ethnographic research strategies in Senegal.

IV. EMPIRICAL RESULTS: THE RELATIONSHIP BETWEENCOLLECTIVISM AND MICROFINANCE

In analyzing the relationship between collectivist cultureand microfinance, we investigate whether, in accordance withthe theoretical model, there is an association between acountry’s level of collectivism and its average microloan sizeand term length.

A. Loan Amount ComparisonAcross all of Kiva’s 1,049,576 loans across 94 countries

between 2007 and 2016, the maximum loan facilitated byKiva is $100,000 and the minimum loan is $25. The meanloan facilitated by Kiva is $842.09. Based on the theoreticaleconomic model defined earlier, we’d expect countries withhigher levels of collectivism to have lower average loansizes. There is a clear positive association between average

58

loan size and level of collectivism on the country-level (seeFigure 2).

Fig. 2: Individualism vs. AverageLoan Amount

Since countries with higher levels of development typi-cally have higher levels of individualism - which is notedin Hofstede (2001) and supported in our data (see Fig-ure 3) - we use level of development as a control. Across-country regression of average microloan size on levelof individualism, controlling for level of development asmeasured by the logistic of GDP per capita, supports thehypothesis derived from the model that there is a relationshipbetween collectivism and loan size (see Table 1), which isstatistically significant. A 10-point increase in a country’slevel of individualism (measured on a 100-point scale), isassociated with a $295.50 increase in the country’s averagemicroloan size.

Fig. 3: Individualism vs. GDP percapita

B. Loan Term ComparisonThe theoretical economic model suggests that informal

redistribution reduces the marginal returns of scale. With alonger duration loan, there are more opportunities to shareprofits with others. For instance, if a Senegalese farmer took

out a loan to buy chicken for resale, given a longer loanduration, there would be more opportunities for people toask the farmer for chickens or for profits from the business.Therefore, it is plausible that norms in collectivist culture,such as norms of informal redistribution, influences both thesize and the duration of loans. Across the Kiva dataset, theminimum loan term facilitated by Kiva is 1 month and themaximum loan term is 195 months (16.25 years). The meanloan term facilitated by Kiva is 13 months. There is a clearpositive association between countries with individualisticcultures and countries with longer average loan terms. Thiscould suggest that borrowers in collectivist cultures takeon more short-term projects compared to individualisticcultures.

Fig. 4: Individualism vs. AverageLoan Duration

In a regression of the average loan term on the levelof individualism across countries, again controlling for thelogistic of GDP per capita, there is a significant relationshipbetween individualism and the average microloan term suchthat a 10 point increase in a country’s level of individualism(measured on a 100-point scale) is associated with a 2.53month increase in the average microloan term.

TABLE I: Regression Results

Dependent Individualism Log GDP Constant R2

Variable (Hofstede Index) per capita N

Loan Amount 29.55*** 392.72*** -2938.40*** 0.41(USD, $) (8.59) (149.24) (1278.77) 41

Loan Term 0.25*** 1.08 -1.23 0.25(months) (0.08) (1.40) (12.07) 46

Note: * p < 0.1, ** p < 0.05, *** p < 0.01.

V. THE INFLUENCE OF COLLECTIVISM ONMICROFINANCE: A CASE STUDY OF MICROFINANCE IN

SENEGAL

In an effort to provide a more detailed account of therelationship between collectivism and microfinance, weconducted detailed surveys and ethnographic interviews inSenegal, which has been defined as a collectivist culture

59

with well-documented collectivist cultural norms includingnorms of informal redistribution (Senegal - Geert Hofstede,2017; Boltz, 2015).

A. Current Landscape of Microfinance in SenegalMicrofinance institutions in Senegal, as in many other

developing countries, are on the rise. In 2015, there were1,168 microfinance institutions in Senegal, some of themost popular being the Alliance of Credit and Savings forProduction (ACEP), Senegalese Mutual Credit (CMS), andthe Partnership for Mobilizing Savings and Credit in Senegal(PAMECAS) (International Monetary Fund, 2016). In paral-lel to the increase in the number of microfinance branches,the number of accounts has been rising substantially since2004.

Fig. 5: MFIs and CommercialBank Branches in Senegal

Fig. 6: MFIs and CommercialBank Depositors in Senegal

While the growth of microfinance in Senegal is sharedby many countries in Sub-Saharan Africa, Senegal’s mi-crofinance landscape has some unique characteristics. Forinstance, the dominance of Islam in Senegal, where 94%of the population is Muslim, influences the structure ofinformal microfinance systems (Mahmud, 2013). A blogpost by Kiva describes in detail that many informal lendingstructures in Senegal do not operate with interest due toIslamic Law, since Islamic scholars claim that “charginginterest on loans is usurious and a violation of Islamic law”(Islamic microfinance, 2012). To accommodate Islamic Law,while maintaining its core business model, Kiva charges aservice fee instead of interest in some cases.

B. Structure of Microfinance in SenegalTo gain insight into the structure of microfinance insti-

tutions in Senegal and the impact of Senegal’s collectivist

culture on these structures, bank managers from the Ke-dougou branch of Credit Mutuel du Senegal and a director ofoperations at the Kedougou branch of Orange Money wereinterviewed.

C. MicrocreditCredit Mutuel du Senegal is one of the largest microfi-

nance institutions in Senegal and, at the end of 2014, hadmore than 125,000 active borrowers (FCCMS, 2015). In theKiva dataset, the mean loan amount in Senegal from 2007to 2016 was $1415.77 and the average loan term was 11.05months. In accord with the data from Kiva, the bank managercited average loan sizes of 15,000 CFA to 1,000,000 CFA(about 1600 USD) (Badji E*, January 12, 2017). While theaverage number of borrowers per loan for all of Kiva’s loanswas 1.97, the average number of borrowers per loan for the10,865 loans to borrowers in Senegal is 6.05 people.

In accord with the analysis of Kiva’s data on loansfacilitated in Senegal, the bank manager from Credit Mutueldu Senegal in Kedougou stated that the most common type ofloan system is that of a large group. The advantage of sucha structure is that the benefits and burden of the loan aredistributed. A group loan also makes the loan less risky forthe banks, since the borrowers keep each other accountable.According to the bank manager, “It is better to loan to agroup because it is less risky. A lot of women’s groups andbig families borrow money from us.” (Badji E*, January 12,2017).

The well-documented collectivist social norms and strongsocial networks in Senegal are plausibly conducive for sucharrangements. Badji E*, the bank manager, outlined that, ifpeople in the group are unable to payback their loans, theyare given a few months of additional time and, if still unableto pay, then other people in the borrower’s group are askedto contribute on the behalf of the defaulting customer. Next,the defaulted borrowers’ family members are visited by thebank in an attempt to acquire the owed funds, and finally,the chief of the village is approached. Upon this step, anagreement is made between the village chief and the bankas to how the customer should best account for their default(such as with the taking of collateral). Since social networksare so strong, it is common, according to Badji E*, for thebank to get money for the loan: ”The social shame of notabiding by one’s word is often enough to get someone topay back their debt. However, friends and family often willhelp as well” (Badji E*, January 12, 2017).

Microloan borrowers in Senegal also cited benefits ofgroup borrowing in expanding the scope of who can par-ticipate in microfinance. Ibrahima*, a college student inKedougou, said that “To get a loan, you have to show thebank that you have something, like money, or a motorcycle,or cows. I don’t have enough.” (Ibrahima A*, January 5,2017). In groups, however, such burdens are distributed.In the case of Credit Mutuel du Senegal in Kedougou,the bank manager stated that loan amounts are determinedby past credit history with the bank, including consistencyand size of past deposits. Since only two of the borrowers

60

are required to have their identification cards photocopied -along with having savings accounts of at least 15,000 CFA(approximately 24 USD) - often a head of household or awealthier person in the village join a loan group on behalfof other people in their social network.

While the group-borrowing system we observed in Senegalis a common structure for microloans and seems to benefitbanks and borrowers, responses during the ethnographicinterviews shed light on significant drawbacks as well. Mostprominently, there is pressure for all parties within a borrow-ing group to only use money for low-risk investments. BadjiE*, the bank manager, stated that “different interest rates aregiven to loan recipients from different industries. We chargea lower fee to buy and sell produce at the local market, buta higher fee to buy and sell livestock. Livestock is morerisky” (Badji E*, January 12, 2017). In comparing uses ofthe 10,865 microloans facilitated by Kiva in Senegal between2007 and 2016, there is a clear skew towards relatively low-risk investments. Some of the most popular activities forwhich loans through Kiva were borrowed include retail (3102loans), food production/sales (979 loans), fruit & vegetables(578 loans), livestock (449 loans), animal sales (376 loans),food market (365 loans), and cloth and dress making supplies(352 loans).

In this way, there appears to be a tension between micro-credit and true entrepreneurship, which is usually associatedwith taking risks. Unlike entrepreneurialism in the UnitedStates, for instance, with very few venture capital fundedstartups being successful, MFI rules in Senegal are set up notto tolerate any failure. However, as mentioned in a chapteron microfinance in Poor Economics: A Radical Rethinking ofthe Way to Fight Global Poverty, “it is a necessary byproductof the rules that have allowed microcredit to lend to a largenumber of poor people at low interest rates...microfinancegives its clients every incentive to play it safe, so it is notwell suited to discover who has an appetite for risk taking”(Banerjee et al., 2011, p.177).

Fig. 7: Microloans in Senegal bysector

D. MicrocreditMicrosavings institutions, like microcredit institutions, fill

a void unattended by traditional banks. Azi T*, the Directorof Operations at Orange Money, a mobile money service,explained that Orange Money does not offer loans or creditopportunities, but instead provides clients with the ability tosafely store savings which fits into the overarching categoryof microfinance.

As is the case with microloans, financial requirements de-ter many from using the services. According to the Directorof Operations at Orange Money in Kedougou, “putting inmoney is completely free, but withdrawing money comeswith a small fee” (Azi T*, January 8, 2017). This feenaturally makes it undesirable to withdraw small amountsof money frequently, and adds extra costs when one could,potentially, just physically hold onto their money. Many ofthose interviewed stated that they avoid savings accounts,even through platforms such as mobile money, because of thewithdrawal fees. Mamadou Jang G*, a construction worker inKedougou said, “I don’t use Orange Money because the feeswould take away my money” (Mamadou Jang G*, January9, 2017). When asked where he puts his money, MamadouJang G* said, “I hide it. I put it in my hut and lock it whenI go away or take it with me. It’s not good, though. It canbe dangerous carrying around all of my savings” (MamadouJang G*, January 9, 2017). In addition, some cited the fees toopening a savings account as a barrier. Adama H* cited a feeof 10,000 CFA to open an account (approximately 16 USD)that, although not astronomical by standards in Kedougou,is enough to cause some like Adama to avoid opening anaccount.

During interviews in Senegal, the benefits of collectivistcultural norms in the context of micro savings becameapparent. A few people interviewed described how, giventhe withdrawal fees, groups have merged together under oneaccount. Coumba E*, for instance, described how norms incollectivist culture have adapted to the structure of microsav-ings in Senegal: “I know many people who share one accountand all withdraw money if they need it. Some people don’thave savings accounts because of the fees. But other peopleshare the fees by using one account” (Coumba E*, January11, 2017).

E. Attitudes towards Microfinance Institutions (MFIs)During interviews, there were noticeable differences in

awareness of microfinance institutions between those in-terviewed in the Dakar region and those interviewed inthe Kedougou region. The Senegalese capital of Dakar isan urban center for many MFIs, while Kedougou is thepoorest region of the country with a lower population density(152,134 people in Kedougou, versus the 1.056 millionpeople living in Dakar according to Senegal Census Data,2013).

All people interviewed in the Dakar region were familiarwith microfinance. When asked about their thoughts aboutmicrofinance, most gave answers reflecting a sentiment sim-ilar to Issa J*’s answer: “I think microcredit and savings are

61

good. They help people start business and save for family andfor school” (Issa J*, December 30, 2017). All of the peopleinterviewed in Dakar seemed to have a positive view towardsmicrofinance, even though less than half used at least one ofthe services provided by microfinance institutions.

While most people interviewed in the Kedougou regionknew what microfinance was, multiple people interviewedindicated that they were extremely hesitant or afraid of suchinstitutions. Most apparently, there was a consistent fearthat one would be put in jail if they weren’t able to paytheir loan. The origin of such a fear could potentially besourced to the structure of microloans. To overcome theexpensive administrative costs involved in lending to thepoor, including due diligence on borrowers, MFIs divergefrom traditional banks and informal moneylenders in howloan contracts are enforced. In addition to threatening tocut off all future lending to anyone who defaults outright,MFIs have a track record of removing almost all flexibility ofloan repayments (Banerjee et al., 2011, p.167). The stringentnature of loan contracts with many MFIs is also unsurprisinggiven the politics of microfinance: low default rates are oftenused as the proof-of-concept for microfinance institutions.The “portfolio at risk” (loans that may default, but willnot all) was less than 4% in South Asia and no more than7% in most Latin American and African countries in 2009(Microfinance Information eXchange, 2017). According tothe Kiva website, the repayment rate on all of their loans is97.3% (The risks of lending, 2016).

However, in an interview with MFIs in Senegal, the actualenforcement of default cases was not excessively stringent.The bank manager at Credit Mutuel du Senegal, Badji E*,clarified the actual protocol of loan contract enforcement.According to Badji E* if a borrower did not pay, the bankwould firstly follow up with the borrower, followed by theborrower’s family and the village chief and elders. Onlywhen the borrower’s network has been fully explored are thepolice involved, initiating a process of repossession in whichlivestock or homes are taken over and sold by the bank.People are never arrested, however, as putting the customersin jail is not regarded as a productive way to recover missingfunds.

Interestingly, when many of the village residents wereinterviewed regarding their attitudes towards loans, peopleseemed scared of the banks. One strategy that seems tokeep MFI defaults low, whether used intentionally or not, isfear. Many of the people interviewed mentioned their fear ofmicrofinance institutions. Awa B* from the village of Segouin the Kedougou region stated that: “I’m scared of the loans.The police will come if [they] can’t repay the loan.” (AwaB*, January 5, 2017). With a similar sentiment towards MFIs,Ibrahima G*, a subsistence farmer from Pellel, claimed that“Banks are scary. They will take everything you have andput you in jail if something bad happens and you can’t paythem back in time.” When asked what would happen if onewas able to avoid the banks, the interviewee said, “They willfind your family and take their things, and might put themin jail too. You can’t hide from the police” (Ibrahima G*,

January 11, 2017).While there is clearly a large degree of misinformation

between the microfinance institutions and the potential bor-rowers, it seems to be understood by the institutions. BadjiE*, the bank manager, said that “the people in the villages arescared of us. They think banks are evil and that we want totake all of their things. They don’t understand how bankingworks” (Baji E*, January 12, 2017). From our conversationsin the larger town of Kedougou with people who have familyin the surrounding villages, we discovered that when no onein a village has ever borrowed from a bank, it is rare thatbank loans will be sought out. However, once some in thevillage have successfully taken out microloans, people aremore likely to pursue similar opportunities themselves.

F. The Influence of Collectivism on Borrowing Decisions

Groups are an integral part of microloans - not onlyfor institutions that often use group-borrowing structures- but also in how individuals interact with microfinanceinstitutions. While interviewing individuals in the Kedougouregion, it became apparent that strong social networks influ-ence borrowing decisions.

The case study of Mamadou B*, a man of thirty twoyears born and raised in Thiabekaare, Senegal, sheds light onhow collectivist culture norms can influence borrowing andspending decisions. He is the son of the chief of the village,which may have brought some heightened social status, buteconomically he grew up much like the rest of the village- as a subsistence farmer. His family grows corn, peanutsand occasionally cotton and sells their excess at the end ofharvest time. He is the father of two children, aged threeand one, and has a wife, who, like almost the entirety of thewomen in the village, is a homemaker.

Throughout the time Keaton has known Mamadou B*,which has been about four years as of December 2016, he hasalways expressed a strong work ethic and, by observation,is perhaps one of the more entrepreneurial young people inthe village. He has been unique in that he has been hired totake on a part time job with a non-profit called EcoSenegalwhile also managing his own vegetable garden. He said thatthe garden barely pays for the work and time put into it,yet he expressed dreams of one day making the garden largeenough that he can hire younger boys to work, thus allowinghim more time to focus on selling what he grows. If he hadfunding, he says that he would buy a motor pump, whichwould allow him to expand his garden (right now his luggingwatering cans to and from the river is inefficient enough thatit limits the amount of produce he can grow). He stated thatlast year he had enough money to purchase a pump, however,both his wife and he fell sick. The hospital fees meant thathe could no longer afford a pump.

Mamadou B* gave multiple explicit indications of howcollectivist social norms affect his borrowing and spendingactivities. After being asked whether he’d consider taking outa loan, Mamadou B* stated his worry:

62

“If people find out you have money, they all ask for it. It goesaway quickly to things that are most important. Families arehuge here. If someone gets sick or needs the money morethan me, I’d give it to them. I’m scared that if I get a loan, Iwill share it and then be unable to pay it back” (MamadouB*, January 7, 2017).

With funds from Stride 4 Senegal, a nonprofit organizationdedicated to Senegalese development, we had the capacityto facilitate an interest-free loan to Mamadou B*. Beforeeven discussing the possibility of a loan, we spoke at lengthabout the pros and cons of informal loaning structures withhim. He said he wouldn’t want to borrow money unless heknew he could repay it, so would want to think “mootya!”(hard!) before taking out an actual loan. “Yeembe no sekudenolugol sata, konno nolugol no sati bwee” he told me, whichtranslates to, “People think loans are easy, but loans arevery difficult.” He spoke at length about problems he hasseen in the past with aid organizations just “giving’, andhe expressed that he has seen firsthand that when peopleare given ‘free’ money they may not put in much work tosustain a project. When asked if people sometimes asked himfor money because he has a garden and he said that it doesoccasionally happen.

Mamadou decided that he did want a microloan so thathe could buy a gas water pump for his garden. In describingthe benefits of such an investment, he said: “I don’t mindwatering the garden, but it takes a lot of time. It’s hardto water it two times a day and I cannot expand thegarden anymore with my other job in Dindefelo. If I havea water pump, I could save time and expand the garden.”(Mamadou B*, January 7, 2017). The conditions of the loan(in Appendix B), describe the structure and payment planof the interest-free loan for 200,000 CFA (the equivalent ofabout $330).

While Mamadou’s garden and, now, his own gas pumpstatute him in a better economic position than the majorityof the village, his initial concern for how his loan would fitinto the context of his large social network was shared bymany. Seydou L*, when asked what he would do if he gota large loan, stated that “I’d share it with my family. Peopleshare in Senegal” (Seydou L*, January 2, 2017). While moreresearch needs to be done in this regard, there was certainly acommon theme of informal redistribution, not only of money,but also of other items such as crops, food, and clothes.

VI. CONCLUSION

As more microfinance organizations are being foundedand more branches are being built, even in remote ruralareas like the Kedougou region of Senegal, more researchis needed to understand the influence of culture on mi-crofinance. Naturally, there are effects in both directions:microfinance institutions can affect culture and culture canaffect microfinance institutions. This research project aims touncover how collectivist culture - and the norms associatedwith that culture - influence microfinance.

We construct a theoretical economic model to demonstratehow norms of informal redistribution common in collectivistcultures can reduce borrowing. Through a cross-countrycomparison of loans facilitated by Kiva from 2007 to 2016,we find that collectivist cultures tend to borrow less moneyfor a shorter duration, of which a greater proportion isspent on sectors such as food, livestock, and agriculturethan their more individualistic counterparts. Importantly, ourresearch also involved three weeks of ethnographic fieldworkin Senegal to uncover some of the ways in which collectivistculture influences people’s attitudes towards, and interac-tions with, microfinance. We find that collectivist culture iswell-suited for the loan structures of many MFIs, such asgroup-borrowing. However, we also found, mainly throughethnographic interviews with individuals, that collectivistscultural standards such as norms of redistribution makepeople hesitant to borrow money.

Future research should explore how culture influencesmicrofinance in other cultural contexts, or better uncover theways in which collectivist culture affects how people interactwith institutions (MFIs being one of them). Additionally,research should be done on the specific loaning structuresthat should be different depending on cultural context. Ourhope is that our research - along with future research -not only raises interesting questions but sheds light on thelives and attitudes of those affected by development efforts,whether related to microfinance or not.

VII. ACKNOWLEDGEMENTS

We gratefully acknowledge financial support from theKenneth I. Juster Fellowship and logistical support fromClare Putnam. We also thank Andrei Shleifer and Susan Kaskfor significant advising, suggestions, and general supportwith the research project. Hassana Diallo and Seidou Dialloprovided in-country assistance and translated interviews.

63

REFERENCES

[1] A*, Ibrahima*. Interview by author. January 5, 2017.[2] B*, Alerji Moussa*. Interview by author. January 6, 2017.[3] B*, Awa*. Interview by author. January 5, 2017.[4] B*, Badji*. Interview by author. January 4, 2017.[5] B*, Dianabo*. Interview by author. January 3, 2017.[6] B*, Mamadou*. Interview by author. January 7, 2017.[7] Baland, Jean-Marie, Catherine Guirkinger, and Charlotte Mali, “Pre-

tending to Be Poor: Borrowing to Escape Forced Solidarity inCameroon,” Economic Development and Cultural Change, 2011, 60(1).

[8] Baland, J., Bonjean, I., Guirkinger, C., and Ziparo, R. (2016).The economic consequences of mutual help in extendedfamilies. Journal of Development Economics, 123, 38-56.doi:10.1016/j.jdeveco.2016.07.004

[9] Banerjee, A. V., & Duflo, E. (2011). Poor economics: a radicalrethinking of the way to fight global poverty. New York: PublicAffairs.

[10] Banerjee, A., Chandrasekhar, A. G., Duflo, E., & Jackson, M. O.(2013). The diffusion of microfinance. Science, 341 (6144), 1236498.

[11] Boltz, Marie, “Sharing the Burden or the Burden of Sharing: Intra-family Redistribution and Personal Savings in Senegal,” TechnicalReport 2015.

[12] Boltz, M., Marazyan, K., & Villar, P. (2015). Income hiding and infor-mal redistribution: A lab-in-the-field experiment in Senegal. Mimeo.

[13] Brett, John A (2006). “We Sacrifice and Eat Less: The StructuralComplexities of Microfinance Participation.” Human Organization.Washington. Vol. 65, Iss. 1; pg. 8, 12 pgs.

[14] B*, Mamadou*. Interview by author. January 7, 2017.[15] Cox, Donald and Marcel Fafchamps, “Chapter 58 Extended Family

and Kinship Networks: Economic Insights and Evolutionary Direc-tions,” in T. Paul Schultz and John A. Strauss, eds., Vol. 4 of Handbookof Development Economics, Elsevier, 2007, pp. 3711 3784.

[16] di Falco, Salvatore and Erwin Bulte, “A Dark Side of Social Capital?Kinship, Consumption, and Savings,” Journal of Development Studies,June 2011, 47 (8), 1128?1151.

[17] Duffy-Tumasz, A. (2005). Credit and Co-Wives: exploring empower-ment in Senegal. CUREJ-College Undergraduate Research ElectronicJournal, 2.

[18] D*, Ciatatou*. Interview by author. January 3, 2017[19] D*, Teneng*. Interview by author. January 3, 2017[20] D*, Hassana. Interview by author. December 31, 2016.[21] E*, Badji*. Interview by author. January 12, 2017.[22] E*, Coumba*. Interview by author. January 11, 2017.[23] FCCMS, Fdration des Caisses du Crdit Mutuel du Sngal. (2015).

Retrieved February 02, 2017, from https://www.oikocredit.coop/what-we-do/partners/partner-detail/11022/fccms-federation-des-caisses-du-credit-mutuel-du-senegal

[24] G*, Mamajang*. Interview by author. January 8, 2017.[25] G*, Ibrahima*. Interview by author. January 11, 2017.[26] Grimm, Michael, Renate Hartwig, and Jann Lay, “Does Forced Sol-

idarity Hamper Investment in Small and Micro Enterprises”, IZADiscussion Papers 7229, Institute for the Study of Labor (IZA)February 2013.

[27] Goldberg, Jessica, “The lesser of two evils: The roles of social pressureand impatience in consumption decisions”, Department of Economics,University of Michigan., December (mimeograph), 2013.

[28] Gurin, Isabelle. (2006) “Women and Money: Lessons from Senegal.”Development and Change 27(3), 549-67. Oxford: Blackwell Publish-ing.

[29] H*, Adama*. Interview by author. December 30, 2016.

[30] Hadness, Myriam, Bjrn Vollan, and Michael Kosfeld, “The Dark Sideof Solidarity,” Technical Report 2013.

[31] Hofstede, Geert. (2001), Culture’s Consequences: Comparing Values,Behaviors, Institutions and Organizations Across Nations. 2nd Edition,Thousand Oaks CA: Sage Publications.

[32] I*, Samba*. Interview by author. January 12, 2017.[33] Index of Individualism. (2014). All loans. Retrieved from

https://geert-hofstede.com/countries.html

[34] International Monetary Fund. (2016). Financial Access Survey. Re-trieved from http://data.imf.org/

[35] Islamic microfinance: How is it different? (2012). Retrieved Febru-ary 18, 2017, from https://www.kiva.org/blog/kiva/2012/05/01/kivas-approach-to-lending-and-islamic.html

[36] J*, Issa*. Interview by author. December 30, 2016.[37] Kivatools. (2017). All loans. Retrieved from

http://kivatools.com/

[38] L*, Seydou*. Interview by author. January 2, 2017.[39] Mahmud, S. S. (2013). Sharia or Shura: contending approaches to

Muslim politics in Nigeria and Senegal. Lanham, MD: LexingtonBooks.

[40] Microfinance Information eXchange (2017) Retrieved January 21,2017, fromhttp://www.mixmarket.org

[41] Perry, Donna. (2006) “Microcredit and women moneylenders: Theshifting terrain of credit in rural Senegal.” Human Organization.,Washington: Vol. 61, Iss. 1; pg. 30, 11 pgs.

[42] Platteau, Jean-Philippe, Institutions, social norms, and economic de-velopment, Vol. 1, Routledge, 2000.

[43] Platteau, Jean-Philippe, “Chapter 12 Solidarity Norms and Institutionsin Village Societies: Static and Dynamic Considerations,” in Serge-Christophe Kolm and Jean Mercier Ythier, eds., Foundations, Vol. 1of Handbook of the Economics of Giving, Altruism and Reciprocity,F, Elsevier, 2006.

[44] Platteau, Jean-Philippe, pp. 819 “ 886. , ”Redistributive Pressures inSub-Saharan Africa: Causes, Consequences, and Coping Strategies,?in Emmanuel Akyeampong, Robert H. Bates, Nathan Nunn, andJames Robinson, eds., Africa’s Development in Historical Perspective,Cambridge University Press, 2014, pp. 153?207. Cambridge BooksOnline.

[45] Rooyen, C. V., Stewart, R., & Wet, T. D. (2012). The Im-pact of Microfinance in Sub-Saharan Africa: A Systematic Re-view of the Evidence. World Development, 40(11), 2249-2262.doi:10.1016/j.worlddev.2012.03.012

[46] Senegal - Geert Hofstede. (2017). All loans. Retrieved fromhttps://geert-hofstede.com/senegal.html

[47] Senegal Census Data (2013). Retrieved fromhttp://senegal.opendataforafrica.org/

[48] Stiglitz, J. E. (1990). Peer Monitoring and Credit Markets. The WorldBank Economic Review,4(3), 351-366. doi:10.1093/wber/4.3.351

[49] T*, Azi*. Interview by author. January 8, 2017.[50] T*, Pende*. Interview by author. January 3, 2017.[51] The risks of lending. (2016). Retrieved February 03, 2017, from

https://www.kiva.org/about/due-diligence/risk

[52] World Bank Databank. (2016). GDP per capita,PPP (current international $). Retrieved fromhttp://data.worldbank.org/indicator/NY.GDP.PCAP.PP.CD

[53] X*, Awa*. Interview by author. January 11, 2017.

* Last name hidden and first names changed to preserveanonymity of subjects.

64

APPENDIX A: ETHNOGRAPHIC RESEARCH QUESTION**

1) Si goo‘to joonee maa kalis seeda, ko hondung sortata? Imagine we gave you a small amount of money. What wouldyou spend it on?

2) Si goo’to joonee maa kalis bwee, ko hondung sortada? Imagine we gave you a large amount of money. What wouldyou spend it on?

3) Si goo’to joonee maa kalis seeda fii boniface maa, ko hondung sortata? Imagine we gave you a small amount of moneyfor business. What would you spend it on?

4) Si goo’to joonee maa kalis bwee fii boniface maa, ko hondung sortata? Imagine we gave you a large amount of moneyfor business. What would you spend it on?

5) Ko hondung wooni relation maa con der banque? Esque hida joogi compte? Esque hari a yihee ka der banque fiinawalagol kalis? What is your relationship with banks? For instance, do you have a savings account or have you everborrowed money from a bank?

** These questions were carefully translated and asked in Pular by a local, native speaker.

APPENDIX B: MAMADOU B*’S MICROLOAN CONTACT (TRANSLATED INTO ENGLISH)

Dindefelo, date : 08/01/2017Name: Mamadou B*Address: ****ID Number: ****Telephone number: *****

Objective: A loan applicationI am a gardener, and do not have the assets to purchase what I wish for my business. I hope after my plea, you willbe able to assist me in my wishes. I have a plot of land 40 meters by 40 meters and wish to irrigate and expandmy vegetable growing operation. With a 200,000cfa loan, I will be able to purchase motor pump. The work is asfollows: January-February: Preparation of land; March-April: Planting; May-September: Watering and tending to plants;October-November: Harvest/vegetable sales.

Signed: Mamadou B*Dindefelo, date: 01/08/2017

ContractI, Mamadou B* (birthdate, village of residence and identification number provided), commit to having received 200,000CFA,to be used for the purchase of a water pump, from the hands of Keaton Scanlon (Batouli Ba). Signed: Mamadou B*.

Dindefelo, date: 01/08/2017

Plan for payment of billI, Mamadou B*, born *****, village of residence ****, ID number *****, has been paid the full amount of 200,000cfawithout the charging of interest, with Ibrahima T* as a witness, on the 8th of January, 2017. The year of repayment willbegin March 2017, with a 100,000cfa installment to be paid back by the 31st of August, 2017. The second installment of100,000cfa will be paid back by the 31st February, 2018. These installments will be paid into the hands of Ibrahima T*,on behalf of Batouli B*.

Signed: Keaton Scanlon, Ibrahima T*, Mamadou B*

65

Vocabulary as an indicator of creditworthiness:

An analysis of public loan data

Justin Wagers

University of Puget Sound, Department of Economics

Abstract— The purpose of this research is to determine

the usefulness of a borrower’s vocabulary in determining

his/her creditworthiness. The analysis takes a word-frequency

approach to 36,055 loans from the peer-to-peer lending platform

Lending Club, and evaluates text submitted by borrowers to

improve the prediction of whether they will pay back their

loan through a nave-Bayes classifier model. Vocabulary, when

paired with traditional creditworthiness measures, is found to

significantly improve the prediction accuracy of a borrower’s

creditworthiness as compared to the accuracy of traditional

credit measures alone.

I. INTRODUCTION

The common credit score has become a default indicatorwhen it comes to determining how trustworthy a borrowermight be. The financial industry is quite dependent on thecredit score: banks, credit card companies, and mortgagelenders all use it as a major tool when determining whether acustomer will receive a loan. However, some evidence showsthat the credit score might not be the precise tool that it ismade out to be; credit scoring has had particular accuracyproblems with underrepresented groups. The industry isalways looking for tools to improve knowledge about aborrower before lending to them; there is a continual searchfor ways in which to improve the imperfect credit score. Thissearch for improvement has led to the very recent analysisof a borrower’s vocabulary as a potential indicator of theircreditworthiness. Often borrowers submit a description oftheir purpose when applying for a loan, but this text hashardly been utilized by lending entities. However, scholarlyliterature in psychology suggests that there is power inusing vocabulary as a predictor of future behavior. Thus,the vocabulary of borrowers, in the form of a typed loandescription, might provide additional accuracy to the creditscore in estimating the probability of default for a given loan.

The peer-to-peer lending industry, by making all of its loandata publicly available, presents an opportunity to analyzethe links between text submitted by borrowers and defaultprobability. Investors on these sites have taken advantage ofthe publicly available loan data to improve the returns ontheir portfolios primarily by taking into account additionalnumerical information on borrowers (inquiries, delinquen-cies, income, etc.). But only recently have they begun to seethe text descriptions as a useful indicator of creditworthiness.Peer-to-peer investors have just begun to dig for patterns inthe loan descriptions, looking at simple characteristics suchas description length. However, a much more in-depth anal-ysis is needed to determine if there existed any relationships

between the actual content of these descriptions and defaultrates of borrowers. Previous research utilizing non-linearregression established a correlation between individual wordusage and default rate for a set of 17 words of high frequencyin loan descriptions, some indicating a higher chance ofdefault and some indicating a lower chance (Wagers 2016).So, given this evidence of correlation for individual words,the next question is whether it is possible to use thesewords collectively to improve the current estimate of defaultprobability as determined by credit score.

An attempt to answer this question requires a predictivemodel that can smoothly incorporate natural language inconjunction with numerical variables. A Bayesian statisticalapproach offers this ability, and is commonly used in the fieldof natural language analysis in general. Bayesian statistics,as opposed to the more commonly used frequentist statis-tics, is advantageous for its ability to incorporate a priorbelief in the construction of an accurate predictive model.I use a nave Bayes classifier predictive model, perhaps themost commonly used model for natural language analysis,to assess the ability to improve a prediction of defaultprobability given a borrower’s loan description. This linkbetween vocabulary and creditworthiness would be excitingbecause it could support the importance of non-numericalvariables for financial prediction purposes: by constructinga model that quantifies previously overlooked data, financialpredictions can become more precise. In particular, lenderscould use patterns in vocabulary to improve knowledge onlenders beyond the numerical credit score.

II. BACKGROUND ON PEER-TO-PEER LENDING

The dataset used in this study is publicly available datafrom the Lending Club platform. Lending Club is thebiggest peer-to-peer lending platform today, with nearly $16billion in loans funded since their start in 2007 (LendingClub Statistics, 2015). Lending Club and other peer-to-peerlending platforms have eliminated the need for a bank asa middle man, providing individual borrowers with lowerinterest rates and individual lenders higher returns than theywould receive by parking their money in a savings account(Athwal, 2014). Borrowers on Lending Club are essentiallycrowd-funded by investors in $25 increments, and investorsin turn receive interest from the borrowers. Lending Clubevaluates a borrower’s credit information including debt-to-income ratio, number of bankruptcies, and credit score.These variables give Lending Club some insight into a

66

borrower’s trustworthiness, which allows them to assign aninterest rate to each loan. In addition to this information,borrowers submit a short description of why they are takingout a loan; these descriptions will serve as the substance forthis analysis.

III. LITERATURE REVIEW

This area of research is especially important given thatboth peer-to-peer lending and natural language analysis areon the rise. The US peer-to-peer lending market generated$6.6 billion in loans in 2015, up an enormous 128% from2014 (Bakkler 2016). As the market grows, so will the poolof investors, and these investors will be looking for newstrategies to take advantage of the large quantities of dataavailable on these platforms. After they have evaluated allof the numerical characteristics of borrowers, some will turn(and already have turned) to other ways to increase theirreturns; analysis of natural language might provide the nextbest way to infer information about a borrower. To under-stand how natural language could improve the estimationof creditworthiness on peer-to-peer sites, it is important toevaluate the credit score and its effectiveness, how verbal andwritten language analysis has been conducted in psychologyand economics, past investing strategies on peer-to-peerlending, and the advantages of a Bayesian statistical modelwith natural language.

The Fair Isaac Corporation started working on a standard-ized measure of creditworthiness in 1954, and began useof the FICO score in 1989; a remarkably similar formulais now used by all three major credit-reporting agenciesin the U.S. (Trainor, 2015). The systematic reliance onthe FICO score cannot be understated: credit score helpsdetermine insurance rates, employment options, protectionagainst fraud, and the ability to borrow money in any formfor consumers (Hamm, 2014). From a lender’s standpoint,credit score is the critical measure of creditworthiness; manyhave a concrete credit score threshold to determine who canborrow certain amounts of money. However, credit scoreis not a 100% accurate reflection of true creditworthiness.Mester (1997) emphasizes that there are some cases wherepeople high credit scores will default, and vice versa. Mesteralso points out a number of flaws with credit scoring, inparticular how it inaccurately estimates the creditworthinessof underrepresented groups (Mester 1997). Borrowers in thepeer-to-peer lending industry aren’t forced to rely on thissometimes flawed measure of creditworthiness; the indus-try has differentiated itself in that all loan data has beenmade available to the public, allowing investors to analyzeinformation about the borrowers themselves to determinewhat characteristics lead to a trustworthy borrower. Amongthis information is a description submitted by borrowers.Thus, an opportunity has presented itself to quantify the linksbetween natural language, in the form of loan descriptions,and creditworthiness, in the form of default rates on theseloans.

The study of language and its effect on behavior are fre-quently discussed in social psychology. Many of the original

theories linking language and behavior are credited to Ken-neth Pike’s 1954 Language in Relation to a Unified Theoryof the Structure of Human Behavior, Pike theorized thatlanguage and behavior were too ingrained within each otherto consider them separately; The activity of man constitutes astructural whole with language in a behavioral compartmentinsulated in character, content, and organization from otherbehavior (Pike 1954). This theory that combines languageand behavior into one should alone be reason enough toinvestigate the prediction power of language. However, whilePike was a forefront theorist on language and behavior,he left the door open for more empirical work on theinterworkings between the two. Researchers have continuallykept Pike’s theory in mind while investigating more concreterelationships within language analysis. Lera Boroditsky, aprofessor of Cognitive Science at UCSD, has provided evi-dence in various studies for the effect of language on visualperception, risk taking and the way people perceive events.Most notably, Boroditsky has thoroughly investigated corre-lations between language and time perception. She providesstrong evidence that speakers of different languages perceivetime in different terms: English speakers think in length oftime, while Mandarin speakers think of time vertically, whichcan influence whether and how speakers of each languageplan for the future (Boroditsky 2009). Her observations arenotable because they indicate the ability to correlate languagewith future behavior. This finding is further supported byeconomist Keith Chen’s determination that language can bean indicator of economic behavior, namely savings rates ofindividuals. Chen found that speakers of languages withobligatory future-time reference had higher savings ratesthan speakers of languages that didn’t force them to reflecton time scale when speaking (Chen 2013). Chen’s resultsare especially important because they suggest that languagecan in some ways be a predictor of economic behavior,which is particularly relevant to this study. However, bothBoroditsky and Chen focus strictly on verbal language, whileI am primarily concerned about the written language thatborrowers submit in their loan applications.

Luckily, written language analysis has shown similar sig-nificance. One 2010 study showed that written word choicehas proven useful in determining personal information aboutindividuals, including one’s opinions on controversial issues.Klebanov et. al. (2010) conducted a word frequency analysisof opinion-based text from abortion debates, death penaltyblogs, and film reviews on Bitter Lemons to investigatethe extent to which vocabulary was not only a matterof topic but was reflective of an individual’s perspectiveon an issue. The researchers found that use of a smallnumber of keywords could signal an individual’s opinion ona controversial issue. Written language has also been usedto determine an individual’s level of expertise: Chujo andUtiyama (2005) constructed an extensive list of vocabularyalong with the level of specialization that each word signaled.With this indicator, they found that an individual’s writtenvocabulary could be evaluated to determine their level ofspecialization or education in a certain field. Klebanov and

67

Chujo & Utiyama’s research provides strong evidence thatwritten language can be used to predict future behavior.Subsequently, there is reason to believe that loan descriptionscould reveal information about a borrower in the peer-to-peerlending setting.

There has been widespread interest in investing strate-gies on peer-to-peer lending sites and analyzing the toolsinvestors have at their disposal, the most noteworthy being aborrower’s credit score. In a 2009 analysis of 194,033 listingson Prosper Marketplace, researchers concluded that investorscould only infer 33% of the difference in creditworthinessbetween two similar loans. They observed, The credit scoreprovides an estimate of the true default probability, but it isonly based on a subset of predictors. Despite this limitation,the credit score is the best available measure of the ex-antedefault probability (Iver et. al., 2009). With the informationthat credit score was not the ultimate indicator of credit-worthiness, peer-to-peer investors began searching for otherways to improve the overall return on their portfolio. Thefirst analyst to consider descriptions as a potential indicatorof creditworthiness was Peter Renton. Renton ran an analysisof description length versus default rate, and determined thatloans with descriptions containing more than 2000 charactershad a default rate approximately three times the average:14.8% versus 4.6% (Renton, 2012). Although Renton haslaid the groundwork for using descriptions as an indicator oflikelihood for a borrower to default, his analysis was veryshallow in that it did not take into account the content of thedescriptions.

Previous research utilizing non-linear regression con-structed a model that evaluated the correlation between wordused by borrowers and default rates. When controlling for allother numerical variables, the words need, help, bills, andthank were found to be positively correlated with default,while credit, loan, and consolidate were negatively correlatedwith default (Wagers 2016). The most strongly correlatedwords indicated upwards of a .05% increased chance ofdefault. While the appearance of a single word is indicativeof a miniscule change in default risk, the collective effects ofa borrower’s entire vocabulary use could be a much strongerindicator.

Most, if not all research in the area of language, behavior,and peer-to-peer lending has taken a frequentist approach;however, the computer science field suggests a Bayesian ap-proach may be more useful when analyzing natural language.A notable characteristic of Bayesian statistics, as opposedto a frequentist approach, is that it allows one to approacha question with an estimation, or prior belief about somedata before any evidence is presented (Gelman 2002). In thelast decade, the computer science branch of statistical lan-guage modeling (SLM) has been leaning towards Bayesianapproaches. Rosenfeld (2000) gives a synopsis of whatlanguage analysis tools had been used in the previous decade,but then makes a strong case for Bayesian analysis as beingthe best way to approach natural language. Rosenfeld citeshuman biases as muddying the statistical process, saying: abetter solution might be to encode such knowledge as a prior

in a Bayesian updating scheme (Rosenfeld 2000). Rosenfeldsees the concept of the prior, specifically, to be the factorthat gives Bayesian a leg up on other statistical approaches.More specifically, scholars have recognized the ability of thenave Bayes classifier as being simple yet accurate. Friedman(1997), in a comparison of numerous classification methods,concluded that, One of the most effective classifiers, in thesense that its predictive performance is competitive withstate-of-the-art classifiers, is the so-called naive Bayesianclassifier. After being established as an accurate classifier,McCallum and Nigam (1998) endorsed the use of the naveBayes model specifically for text classification, citing thatit was the most popular and consistently accurate classifiermodel. It is clear that the Bayesian approach works wellwith natural language, and will be the logical next stepin providing further evidence linking natural language tocreditworthiness in peer-to-peer lending.

IV. HYPOTHESIS

The intention of this paper is to determine the extentto which a borrower’s vocabulary can improve upon theestimation of their probability of default. My hypothesis isas follows:

Vocabulary will be a predictor of creditworthiness and will

be able to improve upon the prediction accuracy of the credit

score.

V. DATA

The data itself consists of 36,055 loans from July 2007 toDecember 2011. Each loan is categorized as either ChargedOff or Fully Paid. Notes that are still current have beenexcluded from the analysis. A loan is charged off when thereis no longer a reasonable expectation of future paymentsand typically occurs when a loan is no later than 150 dayspast due (Lending Club, 2016). The data includes 66 piecesof information about each loan including information aboutthe borrower (credit history, employment, income, location),the loan itself (amount, purpose, description submitted byborrower), and information that Lending Club has logged forthe loan (interest rate, issue date, installment, loan status, pastpayments). The main interest here is of course the descriptionsubmitted by a borrower describing why they are takingout this loan, and the specific words they use within thatdescription that relate back to the hypotheses.

VI. MODEL

To test this hypothesis, I employ a nave Bayes classifiermodel. The nave Bayes gives us the probability of a classifier(i.e. the probability of default) given observed characteristics(predictors) about a certain instance (i.e. a given loan). Themodel estimates the probability of a loan defaulting with amulti-predictor rendition of Bayes’ theorem 1, the formulafor which is given by:

p(d|L) = p(c1|d) ⇤ p(c2|d) ⇤ · · · ⇤ p(cn|d)

1Bayes Theorem: p(cj |d) = p(d|cj)p(cj)p(d)

68

Table 1: Descriptive Statistics

Dummy variables for date, length of loan, income verification, location,length of employment, and loan purpose omitted from table. All words aredummy variables: 1 if yes, 0 if no

Where p(d|L) = probability of loan L defaulting given itscharacteristics c1, c2, . . . , cn, and the probability of its char-acteristics being in a defaulted loan. For example, p(c1|d)gives the probability of characteristic one (c1) being in adefaulted loan; perhaps the probability of the word needappearing in a defaulted loan as opposed to a non-defaultedloan.

To evaluate the additional predictive ability that vocabularybrings on top of traditional credit measures, I comparethe prediction accuracy of two nave Bayes models: onethat takes vocabulary into account and one that does not.The non-vocabulary model predicts loan default based ontraditional credit measures, primarily contained in a givenloan’s interest rate. A loan’s interest rate is primarily basedon a borrower’s credit score and thus reflects the probabilityof default for that borrower. Wagers (2016) showed thatinterest rate is primarily based on a borrower’s FICO score,and also somewhat dependent on loan amount, income, andinquiries, which are not factors included in the FICO scorebut may affect risk on a loan. In terms of the composition ofa FICO score, the algorithm is kept secret, but most believethat it is based upon the ratio of debt to available credit,which is in most cases a direct function of income. Thescore is then adjusted for payment history, number of recentcredit applications, and negative events such as bankruptcy orforeclosure, as well as changes in income caused by changesin employment or family status (Arya et. al., 2011). Thisnon-vocabulary model also uses a handful of other predictorsincluding: date, location, length of loan, length of borrower’semployment, prime rate at the time the loan was issued,whether the borrower’s income was verified, and the purposeof the loan.

The model that contains vocabulary includes the same set

of predictors as in the non-vocabulary model, complimentedby a set of words of significance. Words of significanceincluded in the model are split into two categories: positivewords that are believed to decrease the probability of defaultand negative words that are expected to increase the probabil-ity of default. The sets mirror the words of significance usedby Wagers (2016), and were chosen based on the hypothesisthat they reflected either financial awareness (positive) orfinancial desperation (negative). The words were chosen inadvance, before any observation of correlation with defaultrates. The sets are limited to avoid intercorrelation betweenwords. The positive words in the model include: credit, loan,rate, interest, payment, finance, consolidation, consolidate,financial, while the negative words in the model includehelp, need, bills, you, thank, please, and problem. The modelevaluates whether or not a borrower used each of these wordsin their loan description as an additional predictor towardswhether a borrower will default or not. The construction ofthis second model allows for its direct comparison to thefirst model; any change in predictive ability can be attributeddirectly to the inclusion of vocabulary in the model.

Each of the models are trained on 80% of the 36,055observations; on this portion of the data the models observethe probability of each characteristic being exhibited by adefaulted loan. They each take into account each of thevariables mentioned above: the average interest rate for adefaulted loan, the location, etc, while the vocabulary modeltakes into account how likely each word is to appear in a loanthat has been defaulted on. The models themselves provideconditional probabilities, essentially descriptive statistics, foreach of its respective predictors including word frequency infully paid and defaulted loans. The primary interest is usingthese models for prediction on the remaining 20% of thedata. The models, having learned from training data, makepredictions for whether each individual borrower defaultedor not on the test data. These predictions are then comparedwith the real-world outcomes for the same borrowers to de-termine the models’ accuracy. However, splitting the datasetinto 80% training data and 20% testing data means thatthe accuracy of the models for prediction will be somewhatdependent on how this split occurs. Thus, the partition into80% training, 20% test data sets is repeated 10 times foreach model, and results of the iterations averaged to providea more accurate estimate of the predictive accuracy of eachmodel. 10-fold cross validation is the minimum number ofrepetitions as recommended by Kohavi (1995) to provide areasonable estimate of model accuracy.

It is necessary to address the fact that the nave Bayesmodel makes the nave assumption of conditional indepen-dence: that each feature c1, c2, . . . , cn is independent of everyother feature. It is this assumption that allows the modelto achieve such simplicity, in that it allows each of thecharacteristics to be learned separately by the model. In thismodel, the implication is that all the characteristics of a loanare assumed to be independent of each other. Clearly, this isnot the case in reality; loan characteristics are often relatedto other loan characteristics, and words in particular may

69

be conditionally dependent on each other due to context.Scholars have widely recognized that the nave Bayes modelcontinues to perform very accurately even when breakingthis assumption of conditional independence; consequentiallyit has been widely employed by academics in a variety offields. McCallum and Nigam (1998) discuss this assumption:While this assumption is clearly false in most real-worldtasks, naive Bayes often performs classification very well thefunction approximation can still be poor while classificationaccuracy remains high. In this paper I fully acknowledgethe model breaking the conditional independence assump-tion, with the knowledge that it can still give an accurateprediction of loan default probability.

VII. RESULTS

We can first use the models to evaluate conditional prob-abilities for the words of significance for defaulted and fullypaid loans, respectively. These probabilities are shown inTables 2 and 3.

The conditional probabilities generally suggest that wordsthought to signify financial awareness appear more fre-quently in fully paid loan descriptions, while words thoughtto signify financial desperation appear more frequently inloan descriptions in which the loan was eventually defaultedon. Particularly strong differences in frequency are shownin the word rate (found in 4% more fully paid loans thandefaulted loans) and the word bills (found in 4% more de-faulted loans than defaulted loans). These probabilities givea good insight into the links between vocabulary and defaulttendencies; however, comparing the predictive abilities ofthe Vocabulary-Included and Non-Vocabulary models willprovide the true significance of these links. Below, tablesare shown of the predictive results for 10 iterations of boththe Vocabulary-Included and the Non-Vocabulary models.

Table 4 shows that the mean prediction accuracy ofthe Non-Vocabulary model after 10-fold cross validation is79.7101%, with a similar median. A small standard errorsignifies the reliability of this mean. The individual iterationsrange from 78.9459% to 80.2913% accuracy; the details ofeach iteration can be found in Appendix Table A1. Thesensitivity of the model is high, signifying that it is relativelygood at determining when someone is going to pay his orher loan in full, while the specificity is low, signifying themodel misses a high proportion of loans that were defaultedon in reality. Clearly there are some imperfections in themodel, but this is not concerning given that I am conductinga model comparison, so both models will have the sameimperfections.

Table 5 shows that the mean prediction accuracy of theVocabulary-Included model after 10-fold cross validation is80.1914%, with a similar median. Again, a small standarderror signifies the reliability of this mean. The individualiterations range from 79.6949% to 80.5825% accuracy; thedetails of each iteration can be found in Appendix Table A2.The sensitivity and specificity of this model are relativelysimilar to those of the non-vocabulary model, reassuring usthat the two models will share the same imperfections.

In a comparison of the two models, the Vocabulary-Included model predicts with a mean accuracy that is .4813%higher and a median accuracy that is .4715% higher thanthe accuracy of the Non-Vocabulary model. As displayedin Figure 1, the standard errors for the two models don’toverlap, indicating that the differences in the means arestatistically significant.

VIII. DISCUSSION

The results have significant implications in terms of theoriginal hypothesis that taking into account a borrower’svocabulary can improve the prediction of whether a borrowerwill default on a loan. The significance of the differencebetween the predictive abilities of the Vocabulary-Includedmodel and the Non-Vocabulary model provides clear evi-dence for the hypothesis, suggesting that vocabulary did, infact, improve the prediction of whether a borrower woulddefault on a loan. The magnitude of the difference, ap-proximately .48%, seems minuscule at first glance but issurprisingly large for a metric going unused in the financialindustry. While this might not be the exact amount by whichvocabulary improves the prediction of default, it signifiesthat there is potential for the use of vocabulary to improvethe accuracy of currently used creditworthiness measures,such as the common credit score. Below are the confusionmatrices of the two models and magnified to a sample sizeof 10,000 loans.

The confusion matrices imply that for a 10,000-loansample, the Vocabulary-Included model will predict 9 moredefaults than the Non-Vocabulary model, 819 to 810, and39 more Fully Paid loans, 7161 to 7200, for a 48-loanimprovement in loan accuracy overall. Although the samenumber of loans default in each case (1,834), an investorusing the Vocabulary-Included model would be able to turna more profitable investment from these loans, particularlybecause they would be able to foresee 9 more defaults thata traditional investor.

IX. INTERPRETATION

The somewhat striking significance also brings up a ques-tion of whether vocabulary is acting as a proxy for anothervariable that is not included in the model: education. Onecould see this as a plausible explanation: with more yearsof education, a borrower is more likely to be consideringthe financial repercussions of their loan, and thus usingmore words like consolidate, credit, etc. Although no dataexists on education levels for borrowers on Lending Club,many studies have proven education to be highly correlatedwith income (Porter, 2014). By controlling for income andevery other available factor that lending club provides aboutborrowers, the model aims to eliminate the impact of aborrower’s level of income to the best extent possible.Additionally, because the model evaluates only the frequencyof certain words and not grammar, sentence structure, etc.,education is less likely to be a confounding factor in thesignificance of vocabulary as a predictor of creditworthiness.

70

Table 2: Conditional probabilities for positive words of significance given default status

Table 3: Conditional probabilities for negative words of significance

Table 4: Results for the Non-Vocabulary model

Table 5: Results for the Vocabulary-Included model

Figure 1: Prediction accuracy comparison for Vocabulary-Includedand Non-Vocabulary models.

71

Table 6: Confusion matrix of Non-Vocabulary predictivemodel for theoretical sample size of 10,000 loans

Table 7: Confusion matrix of Vocabulary-Included predictivemodel for a theoretical sample size of 10,000 loans

While there does seem to be a bright future for the use ofvocabulary as a measure of creditworthiness, it also brings upan ethical question. If more weight is placed on vocabularyand similar metrics, it might have a disproportionate effecton low-income communities or certain racial groups. UnderTitle VII of the Civil Rights Act of 1964, employers areforbidden from using a racially neutral employment practicethat has an unjustified adverse impact on members of aprotected class. Thus, if vocabulary were adopted as a stan-dard measure of creditworthiness in the financial industry,steps would have to be taken to ensure that its use did nothave a disparate impact on any race, gender, etc. Kidderand Rosner (2002) showed how easy it was for even simplewording to have a disparate impact on certain ethnic groups,providing evidence that phrasing of SAT questions has adisproportionate impact on African-Americans. As morelinguistic patterns are developed to predict creditworthiness,there will be more potential for some of these patterns to

have a disparate impact on a protected class.It is clear that we have not yet found the boundary of

the predictive power of vocabulary. The significant resultsof this study provide momentum for future research on thepredictive power of vocabulary, particularly in the financialindustry. One direction this research could go would be tostart analyzing more complexities of the vocabulary usedby borrowers. This could include analysis of full phrases ormisspellings as potential predictors of default. This researchcould also be extended to datasets from other peer-to-peerlending platforms and hopefully to a dataset from a largerfinancial institution, if privacy policies do not interfere.

These results may have implications beyond their use toanalyze loan descriptions as well. If the usage of a few keywords can be used as a signal of financial desperation, it isimaginable that similar patterns could be found within socialmedia pages. In the future it may be possible to predict anindividual’s financial reliability by evaluating the vocabularyhe or she uses on Facebook and elsewhere in the digitalworld. In fact, FICO, the leading company in credit scoring,is currently working on a way to assign a credit score topeople who have been financially off the grid, meaning theyhave no credit history, by evaluating an individual’s Facebookpage (Selyukh, 2016). It has not been released whether FICOwill be using users’ vocabulary as a part of this evaluation,but from the results found in this paper, vocabulary patternswithin these pages could be a useful tool. It seems that weare just beginning to scratch the surface when it comes tothe predictive power of vocabulary.

X. ACKNOWLEDGEMENTS

Thanks to Garrett Milam for the continual support on thispaper, as well as to Geoff Considine, America Chambers,and Lisa Johnson.

72

REFERENCES

[1] Arya, E., Arya, S., Wichman, C., & Wichman, C. (2013). Anatomy ofthe Credit Score. Journal of Economic Behavior & Organization,95,175185.

[2] Athwal, N. (2014). The Disappearance Of Peer-To-Peer Lending. Re-trieved from: http://www.forbes.com/sites/groupthink/2014/10/14/the-disappearance-of-peer-to-peer-lending/

[3] Bakker, E. (2016, April). Peer-to-peer lending markets: The leadingcountries for alternative finance and the next high-growth markets.Retrieved from http://www.businessinsider.com/peer-to-peer-lending-markets-the-leading-countries-for-alternative-finance-and-the-next-high-growth-markets-2016-4-26

[4] Boroditsky, L. (2009). How Does Our Lan-guage Shape the Way We Think? Retrieved from:https://www.edge.org/conversation/lera boroditsky-how-does-our-language-shape-the-way-we-think

[5] Chen, M. K. (2013). The effect of language on economic behavior:Evidence from savings rates, health behaviors, and retirement assets.The American Economic Review, 103(2), 690-731.

[6] Chujo, K., & Utiyama, M. (2006). Selecting level-specific specializedvocabulary using statistical measures. System, 34(2), 255269.

[7] Civil Rights Act of 1964, Pub.L. 88-352, 78 Stat. 241 (1964).[8] Freeman, J. (2012). The relationship between lower intelligence,

crime and custodial outcomes: a brief literary review of a vulnerablegroup.Vulnerable Groups & Inclusion, 3(0).

[9] Friedman, N., Geiger, D., & Goldszmidt, M. (1997). Bayesian networkclassifiers. Machine learning, 29(2-3), 131-163.

[10] Gelman, A. (2002). Prior distribution. Encyclopedia of environmetrics.[11] Iyer, R., Khwaja, A. I., Luttmer, E. F. P., & Shue, K. (2009). Screening

in New Credit Markets: Can Individual Lenders Infer BorrowerCreditworthiness in Peer-to-Peer Lending? (SSRN Scholarly Paper No.ID 1570115). Rochester, NY: Social Science Research Network.

[12] Kidder, W. C., & Rosner, J. (2002). How the SAT Creates Built-in-Headwinds: An Educational and Legal Analysis of Disparate Impact.Santa Clara L. Rev., 43, 131.

[13] Klebanov, B., Beigman, E., and Diermeier, D. Vocabulary choice as anindicator of perspective. In ACL Short Papers, pages 253257, Strouds-burg, PA, USA, 2010. Association for Computational Linguistics.

[14] Kohavi, Ron (1995). ”A study of cross-validation and bootstrapfor accuracy estimation and model selection”. Proceedings of theFourteenth International Joint Conference on Artificial Intelligence.San Mateo, CA: Morgan Kaufmann. 2(12): 11371143

[15] Lending Club Investor Articles (2016). Retrieved from:http://kb.lendingclub.com/investor/articles/Investor/What-happens-when-a-loan-is-charged-off

[16] Lending Club Statistics. (2015). Retrieved from:https://www.lendingclub.com/info/download-data.action

[17] McCallum, A., & Nigam, K. (1998, July). A comparison of eventmodels for naive bayes text classification. In AAAI-98 workshop onlearning for text categorization (Vol. 752, pp. 41-48).

[18] Mester, L. J. (1997). What’s the point of credit scoring?. Businessreview, 3, 3-16.

[19] Pike, K. (1954). Language in relation to a unified theory of thestructure of human behavior. The Hague, Netherlands: Mouton & Co.

[20] Porter, E. (2014). A Simple Equation: More Education= More Income. The New York Times. Retrieved fromhttp://www.nytimes.com/2014/09/11/business/economy/a-simple-equation-more-education-more-income.html

[21] Renton, P. (2012). Loan Descriptions - Can They BeHelpful When Choosing Loans? Part 1. Retrieved from:http://www.lendacademy.com/lending-club-loan-descriptions-1/

[22] Rosenberg, J., & Tunney, R. J. (2008). Human Vocabulary Use asDisplay. Evolutionary Psychology, 6(3), 538549.

[23] Rosenfeld, R. (2000). Two Decades of Statistical Language Modeling:Where Do We Go From Here? Carnegie Mellon University ResearchShowcase.

[24] Trainor, S. (2015, July 16). The Long, Twisted History of Your CreditScore. Time. Retrieved from: http://time.com/3961676/history-credit-scores/

[25] Selyukh, A. (2016). Could Your Social Media Footprint Step On YourCredit History? Retrieved from: http://www.npr.org/sections/thetwo-way/2015/11/04/454237651/could-your-social-media-footprint-step-on-your-credit-history

[26] Wagers, J. (2016). Vocabulary as an indicator of creditworthiness: Ananalysis of public loan data. Working paper.

73

APPENDIX

Table A1: Cross-Validation for 10 iterations of the Non-Vocabulary model

Table A2: Cross-Validation for 10 iterations of the Vocabulary-Included model

74