Education, risk preference and wages

Education, Risk Preference and Wages

Sarah Brown and Karl Taylor

Department of Economics, University of Sheffield, 9 Mappin Street,

Sheffield S1 4DT United Kingdom [email protected], Tel: +44 (0) 114 222 3404

[email protected], Tel: +44 (0) 114 222 3420

Abstract: We explore the effect of risk preference on the educational attainment and wages of a sample of individuals drawn from the U.S. Panel Study of Income Dynamics (PSID). Using a sequence of questions from the 1996 PSID, we are able to construct measures of risk aversion and risk tolerance allowing us to explore the implications of interpersonal differences in risk preference for educational attainment. Our empirical findings suggest that risk preference has a significant influence on human capital accumulation, with the degree of risk aversion (tolerance) being inversely (positively) associated with educational attainment. In addition, our findings suggest that risk preference is a valid instrument for education in a standard Mincerian earnings function. Key Words: Human Capital; Risk Aversion; Risk Preference; Wages

JEL Classification: J24; J30

Acknowledgements: We are grateful to the Leverhulme Trust for providing financial support, grant reference F/00212/J. We are also grateful to the Institute for Social Research, University of Michigan for supplying the Panel Study of Income Dynamics 1968 to 2001. We are especially grateful to Bina Prajapati for providing excellent research assistance and to seminar participants at the Universities of Aberdeen and Sheffield for valuable comments. The normal disclaimer applies.

January 2007

mailto:[email protected]

mailto:[email protected]

2

I Introduction and Background

Given the uncertainty surrounding returns to investments in human capital, it is not surprising

that the risk preference of individuals has played a key role in the theory of human capital

accumulation.1 By definition, any investment in human capital can be considered risky, since

the return is unknown and uncertain. For example, Palacios-Huerta (2003) finds that, due to the

degree of risk associated with human capital investments, the actual gains from higher

education per unit of risk in the U.S. are in the region of 5 to 20 per cent higher than that from

risky financial assets.2 Furthermore, it is not clear how one can reduce the degree of risk

associated with human capital investments. As pointed out by Shaw (1996), the standard

approach to reducing risk in financial investment, namely diversification, is often not available

in the context of human capital. Typically, an individual holds one job with his/her human

capital investments tailored accordingly. Hence, given the risk associated with returns to

human capital investments, as well as difficulties with the diversification of such investments,

the risk preference of individuals plays an important role in the decision to acquire human

capital.

Given the obvious problems in measuring individuals’ risk preferences, it is not

surprising that attitudes towards risk have attracted limited attention in the empirical literature.

In some empirical models of human capital accumulation, a parameter of constant risk aversion

has been included,3 but such an approach does not allow variation in risk preferences across

individuals to play a role in the investment decision-making process. Belzil and Hansen (2004),

for example, estimate a dynamic programming model of schooling decisions where the degree

of risk aversion is inferred from school decisions. In this model, individuals are assumed to be

heterogeneous with respect to ability yet homogenous with respect to the degree of risk

aversion. One important exception in the literature is Shaw (1996) who jointly models

1 See, for example, Johnson (1978), Levhari and Weiss (1974) and Gibbons and Murphy (1992). 2 Some attempts have been made to adjust the returns to schooling for individual risk by estimating Mincerian wage equations allowing for random coefficients, which yields dispersion (i.e. risk) in the returns to schooling by assigning individual specific returns, Harmon et al. (2003a). 3 Such studies include Brown and Rosen (1987), Moore (1987) and Murphy and Topel (1987).

3

investment in risky human capital and financial wealth allowing for interpersonal differences

in risk preference. Shaw (1996) presents a theoretical framework that predicts an inverse

relationship between an individual’s degree of risk aversion and investment in risky human

capital. The model is based on a portfolio allocation model extended to incorporate an

individual’s decision to invest in risky human capital. Since human capital accumulation is

modeled as a standard investment process, the less risk averse individuals are predicted to

invest in relatively high levels of education. The empirical analysis suggests that risk

preference affects the returns to human capital, although the relationship between risk

preference and educational attainment is not directly explored. Brown and Taylor (2005) find

supporting evidence using British panel data. In a similar vein, Brunello (2002) presents a

theoretical framework which predicts that risk aversion affects educational choice via the

marginal utility of schooling. The theoretical framework predicts that selected years of

schooling decrease when absolute risk aversion increases. The empirical findings, which are

based on Italian household survey data, support the theoretical priors. This result has received

further recent empirical support from Guiso and Paiello (2005).

In a similar vein, Barsky et al. (1997) present measures of preference parameters

relating to risk tolerance, time preference and inter-temporal substitution based on the U.S.

Health and Retirement Study. The authors explore how risk preference varies across individual

characteristics and report a ‘U’ shaped relationship between years of education completed and

risk tolerance. Individuals with twelve years of schooling were found to be the least risk

tolerant, whilst individuals with more than sixteen years of schooling were found to have

greater than average risk tolerance. In the multivariate regression analysis, however, the

findings suggest that years of schooling are not associated with risk preference. It should be

acknowledged that intuitively one might argue that a risk averse individual may have an

incentive to invest heavily in human capital in order to safe-guard his/her future. Belzil and

Hansen (2004) find that a counterfactual increase in risk aversion increases educational

attainment, i.e. human capital accumulation.

4

In sum, the relationship between risk preference and educational attainment has

attracted attention in both the empirical and the theoretical literature on human capital

accumulation. According to such arguments, educational attainment (i.e. human capital

accumulation) is influenced by risk preference. A related interesting line of enquiry relates to

whether risk preference influences the returns to human capital investment. Human capital

theory regards educational attainment as the outcome of an investment decision-making

process with augmented lifetime earnings being the key benefit of investment in education. If

risk preferences influence educational attainment, which in turn influences earnings, it may be

the case that omitting risk preference in an educational attainment model may bias any

estimates of the influence of education on earnings, i.e. the returns to education. Our paper

contributes to this area – specifically exploring the relationship between risk preference, human

capital accumulation and wages.

II Data and Methodology

The obvious problem with exploring the relationship between human capital and risk

preference from an empirical perspective lies in locating a suitable measure of risk preference.

For this purpose, we exploit the Panel Study of Income Dynamics (PSID), which is a

representative panel of individuals ongoing since 1968 conducted at the Institute for Social

Research, University of Michigan. The PSID 1996 Survey includes a Risk Aversion Section

which contains detailed information on individuals’ attitudes towards risk. The Risk Aversion

Section contains five questions related to hypothetical gambles with respect to lifetime income.

To be specific, all heads of household were asked the following question (M1):

Suppose you had a job that guaranteed you income for life equal to your current total income.

And that job was (your/your family’s) only source of income. Then you are given the

opportunity to take a new, and equally good, job with a 50-50 chance that it will double your

income and spending power. But there is a 50-50 chance that it will cut your income and

5

spending power by a third. Would you take the new job?4 The individuals who answered ‘yes’

to this question, were then asked (M2): Now, suppose the chances were 50-50 that the new job

would double your (family) income, and 50-50 that it would cut it in half. Would you still take

the job? Those individuals who answered ‘yes’ to this question were then asked (M5): Now,

suppose that the chances were 50-50 that the new job would double your (family) income, and

50-50 that it would cut it by 75%. Would you still take the new job? Individuals who answered

‘no’ to Question M1 were asked (M3): Now, suppose the chances were 50-50 that the new job

would double your (family) income, and 50-50 that it would cut it by 20 percent. Then would

you take the job? Those individuals who replied ‘no’ were asked (M4): Now, suppose that the

chances were 50-50 that the new job would double your (family) income, and 50-50 that it

would cut it by 10 percent. Then would you take the new job?

We used the responses to this series of questions to create a six point risk aversion

index, iRA as follows (the percentage of individuals in each category are also shown below):

%82.29%36.19%90.15%57.15%42.13%93.5

4&3&154&3&14

3&132&12

5&2&115&2&10

======

====

======

=

NoMNoMNoMifYesMNoMNoMif

YesMNoMifNoMYesMif

NoMYesMYesMifYesMYesMYesMif

RAi

Thus, the index is increasing in risk aversion such that if an individual rejects all the

hypothetical gambles offered, the risk aversion index takes the highest value of 5, whilst if the

individual accepts all gambles offered the risk aversion index takes the value of zero. It is

interesting to note the low (high) percentage of respondents with the lowest (highest) value of

the risk aversion index. Intermediate cases lie in between these two extreme values such that

individuals are ranked according to their reluctance to accept the hypothetical gambles. The

series of questions, thus, enables us to place individuals into one of six categories of risk

4 As Luoh and Stafford (2005) point out it is important to acknowledge that the question states that the new job will be ‘equally as good’ such that there is no difference in the non monetary characteristics of the jobs. Without such a qualification, individuals may be less willing to accept the gamble if there are non monetary attachments to their current job (Barsky et al., 1997).

6

aversion. Furthermore, as stated by Barsky et al. (1997), p.540: ‘the categories can be ranked

by risk aversion without having to assume a particular form for the utility function.’

A further measure of risk preference is available from the PSID 1996 survey based

on Questions M1 to M5 above. Based on Barsky et al. (1997), a measure of risk tolerance

( iRT ) is available in the PSID where the answers to Questions M1 to M5 have been converted

into a single quantitative index of risk tolerance which has been corrected for measurement

error.5 Thus, iRT being a measure of risk tolerance is inversely related to risk aversion. 6

Our sample is restricted to those heads of household in full-time employment in

1996 aged between 18 and 65, yielding a total of 4,246 observations. We explore the

relationship between risk preference and human capital accumulation by modeling education,

ie , as a function of risk preference:

( ) 1, iiii rfe ε+= X (1) where ir denotes the measure of risk preference and iX represents a set of additional

explanatory variables, which draws on Wilson et al. (2005) and includes: age; gender;

ethnicity; the mothers’ marital status when the respondent was born; whether the respondent

lived with his/her parents until age 16; whether the mother worked when the respondent was

growing up; fathers’ occupational status when the respondent was growing up; number of

siblings; whether the respondent was the first born; whether the mother was born outside of the

U.S.; the educational attainment of both parents; an index of per student school expenditure in

the county of residence when the respondent was growing up; type of religion; whether the

family was poor when the respondent was growing up and controls for the region of birth.

5 A detailed explanation of the conversion procedure is given by Luoh and Stafford (2005), which is summarized here. The risk tolerance data are taken from the last column of Table 1 in Barsky et al. (1997). Assume a utility

function, ( ) ( )( ) qcqcU /11/11/1 −−= , that q is log-normally distributed and ( )qG ln= . We observe *G which lies in one of the categories determined by the hypothetical gamble questions. The product of each individual’s probability of being in a particular category yields the likelihood function. Maximizing the likelihood function and computing the expected means conditional on being in a particular category yields q (Luoh and Stafford, 2005). 6 It should be re-iterated that our measures of risk preference are based on hypothetical rather than actual behavior. In Section III, we explore an alternative measure of risk preference based on actual behavior.

7

In order to ascertain the robustness of our findings, we analyze both measures of risk

preference, iRA and iRT . Similarly, we explore two measures of education – an index denoting

the highest educational attainment of the head of household ( ie ) and the number of years of

completed schooling by the head of household ( is ). The highest educational attainment

variable is a five point index where 0 denotes less than high school completed, i.e. less than

grade 12; 1 denotes high school completed; 2 denotes that the individual went to college but

did not graduate; 3 denotes that the individual graduated from college; and, finally, 4 denotes

that the individual completed some postgraduate work. The number of years of completed

schooling is a continuous variable with a minimum (maximum) of 8 (17) years of schooling.

We estimate equation (1) as an ordered probit model when measuring education by the index

denoting highest educational attainment of the head of household ( ie ) given the inherent

ordering of the index. When measuring education using the number of years of schooling

completed by the head of household ( is ), equation (1) is estimated by ordinary least squares

(OLS).

Table 1 presents a correlation matrix; between iRA and iRT ; between ie and is ; and

between the risk preference and educational attainment measures. The degree of correlation

between the two measures of risk preference acts in accordance with a priori expectations, i.e.

the measure of risk tolerance is inversely related to the index of risk aversion. Moreover, the

strong inverse relationship is significant at the one per cent level. Similarly, the correlation

between ie and is is positive and significant at the one per cent level. Finally, the measures of

education are inversely associated with risk aversion and positively associated with risk

tolerance. Such relationships between the key variables are in accordance with the empirical

findings of Brown and Taylor (2005), Brunello (2002), Guiso and Paiella (2005) and Shaw

(1996).

A potential problem with our measures of risk preference is that, as argued by

Brunello (2002), educational choice depends upon risk attitudes at the time of the choice rather

8

than risk preference observed in 1996. Hence, the measures of risk preference are only

meaningful if risk preferences are time invariant or if the influence of the time variant

component of risk preference is small in terms of magnitude. In order to explore such issues,

researchers have analyzed the relationship between age and proxies for risk preference such as

the propensity to hold risky financial assets for individuals’ risk attitudes. Based upon U.S.

data, Haliassos and Bertaut (1995), for example, found the impact of an individual’s age on the

decision to hold risky stocks to be statistically insignificant. More recently, Guiso et al. (2003)

have also found the share of assets held in risky stocks to be invariant with respect to age

effects in a number of countries. Such findings suggest that risk preferences do not vary with

age.

We follow this literature and model our measures of risk preference ( ir ) conditional

upon: age; whether the individual’s home is owned outright; gender; marital status; whether the

individual is in full-time employment; occupation; industrial and regional dummy variables –

all contained in the vector Z and measured at 1996. We also control for household wealth , y,

which includes dividends, interest payments, and all other unearned income, as this may

influence the measures of risk preference, especially if individuals misinterpret ‘income’ in the

above questions M1-M5 to include wealth. Thus, we estimate the following equation as an

ordered probit regression:

( ) 2ln, iiii ygr ε+= Z (2) We then re-estimate our educational attainment equations replacing ir with it’s predicted

value, ir , derived from equation (2), in order to ascertain whether a relationship remains

between educational attainment and risk preference controlling for age and wealth effects.

Finally, we explore how educational attainment influences hourly wages, once we

allow risk preferences to influence human capital accumulation. Following Brunello (2002),

we explore the validity of risk preference as an over-identifying instrument for education in an

earnings function. The standard approach to ascertain the impact of education on wages

9

involves estimating a Mincerian wage equation whereby log wages ( wln ) are specified as a

function of personal controls, iH , and education (i.e. ie or is ):

3ln iiii ew εφλ ++= H (3)

The return to education φ is inconsistent if the error terms from the education equation (1) and

the wage equation (3) are correlated, i.e. 031

≠ii εεσ . Thus, in order to derive consistent

estimates, one must identify a variable which influences education but not wages, Card (1999).

If risk preference is insignificant when included in the wage equation, but significant in

explaining educational attainment, this would concur with the findings of Brunello (2002) and

provide further evidence that risk preference is a valid instrument for education attainment.

Moreover, if risk preference turns out to be a valid instrument, then using predicted education

( ie or is ), derived from equation (1), yields a more precise estimate of the returns to education.

Summary statistics of all of the variables used in our empirical analysis are presented in Table

2.

III Results

Risk Preference and Education

In Table 3 the results of estimating equation (1) are summarized for both measures of

education: Panels A and B presents the results for the highest education attainment index ( ie )

whilst Panel C presents the results for years of completed schooling ( is ). Due to the ordered

nature of the highest education attainment index ( ie ), we present the marginal effects of risk

preference on the probability of having each level of education from no education, i.e. less than

high school where index, ie , equals zero, through to having completed some postgraduate

work, where ie , equals four. For each measure of education, we estimate two specifications:

specification 1 includes iRA in the set of explanatory variables whilst specification 2 includes

iRT in the set of explanatory variables. The set of explanatory variables in each of these

10

models is as shown in Table 4, which gives the full estimation results of the educational

attainment equations where risk preference is measured by iRA .7

It is apparent that, for our sample of 4,246 individuals, risk preference exerts a

statistically significant impact on both measures of education. The marginal effects show that

risk preference as measured by the risk aversion index ( iRA ) increases the probability of

having no education (Table 3 Panel A). Indeed, a one standard deviation rise in the risk

aversion index increases the probability of having no education by 1.5%.8 Similarly, at the

opposite end of the educational attainment hierarchy, a one standard deviation increase in risk

aversion decreases the probability of having completed postgraduate study by 1.8%. Consistent

results are found with the alternative measure of educational attainment, shown in Table Panel

C, where increasing risk aversion has a negative impact on the number of years of completed

schooling ( is ). A one point move up the risk aversion index decreases the number of years of

completed schooling by 3%.9 We also consider the effect of risk tolerance ( iRT ) upon both

measures of education. As expected, risk tolerance decreases the probability of having no

education (Table 3 Panel B) and correspondingly increases the number of years of schooling

(Table 3 Panel C). For example, a one standard deviation increase in risk tolerance decreases

the probability of having no education by 1.3% and increases the probability of having

completed some postgraduate study by 1.6%. Noticeably, risk aversion (tolerance) has a

monotonic impact upon educational attainment. In general, our findings accord with those of

Brunello (2002) and Guiso and Paiella (2005) in that risk aversion is inversely associated with

educational attainment.

7 In accordance with the existing literature we find that educational attainment is increasing in: age; fathers’ occupation; mothers’ educational attainment; religion; school expenditure; whether the individual is male and ethnicity. Factors which significantly decrease educational attainment are whether the mother was single when the respondent was a child and whether his/her father had high school education only. 8 These calculations are based on the mean sample characteristics of individuals. For example, the 1.5% effect is calculated by multiplying the marginal effect, 0.00935, by the standard deviation of the risk aversion index, 1.6057. 9 This is calculated as a risk preference elasticity such that( ) ( )s r r s∂ ∂ × , where ( )ˆ s rφ = ∂ ∂ and r and s denote the mean values of r and s respectively.

11

Risk Preference and Time Invariance

As mentioned in Section II, exploring the relationship between the 1996 risk preference

measures and educational attainment may be problematic as human capital investments may

have been made some time ago, for instance, when the individual was at high school. The

inclusion of the 1996 risk preference measures in the educational attainment equation may be

appropriate if risk preferences do not vary over time or if the human capital investments were

made in 1996. The extent of the time variance issue may depend on the gap between the age of

the individual in 1996 and the age of the individual when the human capital investment was

undertaken. This potential problem was pointed out by Brunello (2002), but was not explicitly

addressed in his empirical analysis of Italian survey data. We explore this issue in three ways.

Firstly, we explore the significance and magnitude of age effects in the risk preference

equation. Secondly, we estimate the educational attainment equations by age cohorts in order

to ascertain the robustness of the influence of risk preference on education across time. Clearly,

for the youngest cohort, the gap between 1996 and the age of the human capital accumulation,

will, on average, be the smallest. Thirdly, we exploit information relating to risk preferences

from previous waves of the PSID.

The results from estimating the two risk preference equations are presented in

Tables 5A and 5B. It is apparent from the estimated risk preference equations, i.e. equation (2),

that iRA and iRT are both influenced by age which suggests that risk preferences are not time

invariant. However, the marginal effects of the quadratic in age (see Table 5B) are relatively

small for both measures of risk preference at the extreme values of the two measures and are

only statistically significant at the 10 per cent level for iRA and insignificant for iRT .

We replicate our analysis of Table 3 by estimating equation (1) based upon

predicted risk preference, ir estimated from equation (2), for both measures of education:

Panels A and B present the results for the highest education attainment index ( ie ) whilst Panel

C presents the results for years of completed schooling ( is ). The summary results shown in

12

Table 6 concur with those of Table 3, where risk preference is treated exogenously, in that

there is a statistically significant impact of both measures of risk preference upon education.

For example, a one standard deviation rise in the predicted risk aversion index increases the

probability of having no education by 1.7%. Similarly, at the opposite end of the educational

attainment hierarchy, a one standard deviation increase in predicted risk aversion decreases the

probability of having completed postgraduate study by 2.1%. Hence, our findings suggest that

the correlation between risk preference and education reported in Table 3 is not capturing, for

example, an unobserved wealth or age effect.

To further explore this issue, we estimate educational attainment equations by age

cohorts, i.e. by individuals born in the following decades: 1930s; 1940s; 1950s; 1960s and

1970s. We test whether the influence of risk preference on educational attainment in each

cohort (i.e. sub-sample) is significantly different from the estimated coefficient on risk

preference in the education equation estimated over the entire sample. The results presented in

Table 7 reveal that the null hypothesis that the influence of risk preference on education is the

same as the 1996 effect of risk preference on education cannot, in general, be rejected across

successive cohorts.10 The findings that the estimated coefficients on the risk preference

measures do not vary across the age cohorts provide further support for the time invariance of

the influence of risk preference on educational attainment.

Finally, over the period 1969 to 1972 an index of risk avoidance is available in four

waves of the PSID. This measure of risk avoidance is derived from questions relating to factors

such as the head of household’s seat belt usage, smoking behavior and purchases of medical

insurance and car insurance, i.e. actual rather than hypothetical behavior. It is possible that

individuals are in the sample between 1 to 4 times during the period 1969 to 1972. Hence, we

take an average of the risk avoidance index over a maximum of four years as our early measure

of risk preference, oiRA . There are 647 individuals who were heads of households in both 1996

10 Tests that the estimated coefficients between successive cohorts, i.e. 1930=1940; 1940=1950 etc., are equal cannot be rejected.

13

and over the period 1969 to 1972. These individuals are aged between 41 and 65. Hence, for

these individuals we can compare the influence of the risk preference measure reported in 1996

on educational attainment with that of the risk preference measure reported over 1969 to

1972.11 Indeed if risk preference is largely time invariant then we would expect oiRA to have a

similar effect upon education with the same direction of impact as iRA . It should be explicitly

acknowledged that iRA and oiRA do differ in terms of the underlying survey questions being

based on hypothetical behavior in the case of iRA and actual behavior in the case of oiRA .

Despite such differences, however, for this sub-sample of individuals, the correlation between

the risk aversion index of 1996 and risk avoidance index of the earlier period is 0.0797 which

is statistically significant at the 5 per cent level. Thus, the two risk preference variables,

although constructed from survey responses given two decades apart, are positively related

suggesting time invariance of risk preferences.

To further investigate the relationship between risk preference and human capital,

we estimate equation (1) based on the sub-sample of 647 individuals who were present in both

the 1996 PSID and at least one year in the 1969 to 1972 PSID. The results are shown in Table

8 for both measures of education: Panels A and B presents the results for the highest education

attainment index ( ie ) for each risk measure, and Panel C presents the results for years of

completed schooling ( is ). In Panels A and B the marginal effects are reported across each of

the education categories, along with the percentage impact of a one standard deviation increase

in risk. For this sub-sample of individuals, risk preference measured at 1996 and risk

preference measured over 1969 to 1972 both have a statistically significant impact on each

category of educational attainment and upon the number of years of completed schooling.

Noticeably, the impact of risk preference on educational attainment is larger for the earlier

period relative to 1996: 6.49% versus 2.95% for the probability of having no education

11 The risk avoidance index is increasing in risk aversion and so a priori we would expect it to be positively correlated with the 1996 measure of risk aversion.

14

( 0=ie ); 7.19% versus 3.28% for the probability of having completed some postgraduate work

( 4=ie ); and 17.08% versus 6.95% for years of completed schooling. Such findings are not

surprising as one might predict that the majority of educational attainment would have been

achieved closer to the early time period and, hence, its impact upon human capital

accumulation is predictably greater than the influence of risk preference in 1996 on educational

attainment.12 It is apparent that, year on year, the differential between the magnitudes of the

risk preference impacts is relatively small.

Education, Risk Preference and Wages

A large literature exists which has attempted to identify variables which influence education

but not wages in order to ascertain a valid over-identifying instrument for educational

attainment in earnings functions (see Card, 1999). Typically, the number of years of schooling

has been the measure of educational attainment adopted in this literature with the finding that

the impact of is upon wages based upon IV estimation is larger than that found using OLS,

where is is treated as exogenous, Card (1999). Following Brunello (2002), we explore the

validity of risk preference as an over-identifying instrument for education in a wage equation –

see equation (3). Table 9 Panel A shows the results of estimating standard Mincerian wage

equations where education is treated as exogenous. The findings accord with the previous

literature in that both measures of education are positively associated with wages.

To investigate whether risk preference is a valid over-identifying instrument for

education, we initially estimate equation (3) including the risk preference measure, iRA , in the

wage equation. From the results shown in Table 9 Panel B, it is apparent that the estimated

12 We also verify this by including the two risk preference measures, one from 1996 and one from averaging the reported measures over 1969 to 1972, in the educational attainment equation. The equality of the coefficients is always rejected at the 1 per cent level across both measures of education.

15

coefficient on the risk preference measure is statistically insignificant. Such findings endorse

the use of risk preference as an over-identifying instrument for education.13

Hence, we include predicted education measures, ie or is , conditional on risk

preference, in the earnings function without controlling for risk preference in the earnings

function. The results are shown in the first two columns of Table 10, where Panel A relates to

the highest educational attainment index and Panel B relates to completed years of schooling.

Results based upon the latter ( is ) indicate a 13% return to education under IV estimation

(when either risk preference measure is used as an instrument) compared to a 11% return based

upon OLS, which is consistent with findings in the literature, see Card (1999) and Harmon et

al. (2003b). Clearly, predicted education, based upon either measure of risk preference,

increases wages. Moreover, across the two definitions of education and the two measures of

risk preference, the impact of predicted education is significantly different to that of exogenous

education at the 1 per cent level, i.e. using risk preference as an instrument for educational

attainment has a significant impact on the magnitude of the estimated returns to education.

The final two columns of Table 10 present estimates based upon the effect of

predicted education upon wages, where the former is modeled conditional upon predicted risk

preference. It is potentially important to control for such effects as risk preference may vary

with household wealth, and wealth may be correlated with wages. However, after controlling

for such wealth effects, the positive influence of education upon wages remains. Hence, our

findings are robust to controlling for such effects.

IV Conclusions

This paper has focused on the relationship between human capital, risk preference and wages

using individual level U.S. data drawn from the PSID. Our empirical findings indicate that risk

preference has a statistically significant impact upon educational attainment. This result is

13 For brevity, we do not present the results presented in Table 9 Panel B for the alternative measure of risk preference, risk tolerance. The effect of risk tolerance in the earnings function is also statistically insignificant with an estimated coefficient (t statistic) of -0.0062 (0.13) and -0.0617 (1.45) for ie and is respectively.

16

robust across different measures of education and different measures of risk preference.

Specifically, greater levels of risk aversion decrease educational attainment. This is consistent

with the theoretical literature and the limited amount of empirical evidence in this area, such

as, Barsky et al. (1997), and Guiso and Paiella (2005). In addition, in accordance with Brunello

(2002), risk preference is found to be a valid instrument for education in a wage equation.

References

Barsky, R. B., F. T. Juster, M. S. Kimball, and Shapiro, M. D. (1997). ‘Preference Parameters

and Behavioural Heterogeneity: An Experimental Approach in the Health and

Retirement Study.’ Quarterly Journal of Economics, 112, 537-79.

Belzil, C. and Hansen, J. (2004). ‘Earnings Dispersion, Risk Aversion and Education.’

Research in Labor Economics, 23, 335-58.

Brown, J. and Rosen, S. (1987). ‘Taxation, Wage Variation, and Job Choice.’ Journal of Labor

Economics, 5, 430-51.

Brown, S. and Taylor, K. (2005). ‘Wage Growth, Human Capital and Financial Investment.’

The Manchester School, 73, 686-708.

Brunello, G. (2002). ‘Absolute Risk Aversion and the Returns to Education.’ Economics of

Education Review, 21, 635-40.

Card, D. (1999). ‘The Causal Relationship of Education on Earnings.’ In Handbook of Labor

Economics – Volume 3, Amsterdam, North Holland, 1801-63.

Gibbons, R. and Murphy, K.J. (1992). ‘Optimal Incentive Contracts in the Presence of Career

Concerns: Theory and Evidence.’ Journal of Political Economy, 100, 468-505.

Guiso, L., Haliassos, M. and Jappelli, T. (2003). ‘Equity Culture: Theory and Cross-country

Evidence.’ Economic Policy, 18, 123–70.

Guiso, L. and Paiella, M. (2005). ‘The Role of Risk Aversion in Predicting Individual

Behaviour.’ Temi di discussione (Economic working papers), Bank of Italy, Economic

Research Department, Number 546.

17

Haliassos, M. and Bertaut, C. (1995). ‘Why Do So Few Hold Stocks?’ Economic Journal, 105,

1110–29.

Harmon, C., Hogan, V. and Walker, I. (2003a). ‘Dispersion in the Returns to Education.’

Labour Economics, 10, 205-14.

Harmon, C., Oosterbeek, H. and Walker, I. (2003b). ‘The Returns to Education:

Microeconomics.’ Journal of Economic Surveys, 17, 115-55.

Johnson, W.R. (1978). ‘A Theory of Job Shopping.’ Quarterly Journal of Economics, 92, 261-

78.

Levhari, D. and Weiss, Y. (1974). ‘The Effect of Risk on the Investment in Human Capital.’

American Economic Review, 61, 950-63.

Luoh, M-C. and Stafford, F. (2005). ‘Estimating Risk Tolerance from the 1996 PSID.’

http://psidonline.isr.umich.edu/data/Documentation/Cbks/Supp/rt.html

Moore, M. (1995). ‘Unions, Employment Risks, and Market Provision of Employment Risk

Differentials.’ Journal of Risk and Uncertainty, 10(1), 57-70.

Murphy, K.M. and Topel, R. (1987). ‘Unemployment Risk and Earnings: Testing for

Equalizing Wage Differences in the Labor Market.’ In Unemployment and the Structure

of Labour Markets, New York, NY, Basil Blackwell, 103-40.

Palacios-Huerta, I. (2003). ‘An Empirical Analysis of the Risk Properties of Human Capital

Returns.’ American Economic Review, 93, 948-64.

Shaw, K. L. (1996). ‘An Empirical Analysis of Risk Aversion and Income Growth.’ Journal of

Labor Economics, 14(4), 626-53.

Wilson, K., Wolfe, B. and Haveman, R. (2005). ‘The Role of Expectations in Adolescent

Schooling Choices: Do Youths Respond to Economic Incentives?’ Economic Inquiry,

43(3), 467-92.

http://psidonline.isr.umich.edu/data/Documentation/Cbks/Supp/rt.html

Table 1: Correlation Matrix for Risk Preference and Educational Attainment

iRA iRT ie is

iRA 1.0000

iRT -0.9533 p=[0.000] 1.0000

ie -0.0554 p=[0.003] 0.0482 p=[0.001] 1.000

is -0.1062 p=[0.000] 0.0838 p=[0.000] 0.6326 p=[0.000] 1.000

Table 2: Summary Statistics

MEAN STD. DEV MAX MIN Highest educational attainment index ie 2.0607 1.1877 4 0 Number of years of completed schooling is 13.1213 2.2284 17 8 Risk Avoidance index iRA 3.1877 1.6057 5 0 Risk Tolerance iRT 0.2831 0.1592 0.57 0.15 Age 38.2230 10.0547 65 18 Male 0.7094 0.4541 1 0 White 0.0445 0.2063 1 0 Male×White 0.0249 0.1560 1 0 Mother single when child 0.0544 0.2268 1 0 Mother widow when child 0.0040 0.0632 1 0 Mother separated or divorced when child 0.0172 0.1300 1 0 Lived with parents until 16 0.6908 0.4622 1 0 Mother worked when child 0.3314 0.4708 1 0 Father professional or managerial 0.0824 0.2751 1 0 Father self employed 0.0438 0.2047 1 0 Father clerical or crafts 0.2075 0.4056 1 0 Father manual 0.1679 0.3738 1 0 Number of siblings 3.0391 2.6138 10 0 Firstborn 0.1992 0.3995 1 0 Mother born outside US 0.3022 0.4593 1 0 Mother high school education 0.7518 0.4320 1 0 Mother college education 0.1566 0.3635 1 0 Father high school education 0.7244 0.4468 1 0 Father college education 0.1606 0.3672 1 0 School expenditure in the county index 3.5318 3.3680 9 0 Roman Catholic 0.2310 0.4215 1 0 Jewish 0.0174 0.1309 1 0 Christian 0.5605 0.4964 1 0 Other religion 0.0224 0.1479 1 0 Family was poor when growing up 0.4086 0.4916 1 0 Log dividend payments 0.7610 2.2598 13.8155 0 Log interest payments 1.5459 2.8340 11.5129 0 Log all other payments 1.0772 2.6574 14.9469 0 Log hourly wage 2.4872 0.5647 5.0073 0.6539 Own home 0.5561 0.4969 1 0 Full-time employee 0.9479 0.2222 1 0 Married 0.5384 0.4989 1 0 Union member 0.1849 0.3882 1 0 Experience 20.5205 10.3397 53 0 OBSREVATIONS 4,246

Table 3: The Effect of Risk Preference on Educational Attainment

PANEL A: HIGHEST EDUCATION ATTAINMENT INDEX ( ie ) AND RISK AVERSION

LESS THAN HIGH

SCHOOL COMPLETED ( 0=ie )

HIGH SCHOOL COMPLETED ( 1=ie )

WENT TO COLLEGE DID NOT GRADUATE

( 2=ie )

GRADUATED FROM COLLEGE ( 3=ie )

SOME POSTGRADUATE STUDY ( 4=ie )

M.E. TSTAT M.E. TSTAT M.E. TSTAT M.E. TSTAT M.E. TSTAT Risk Aversion Index iRA 0.00935 (4.21) 0.00609 (4.19) 0.00196 (3.81) -0.00608 (4.16) -0.01132 (4.22) Chi Squared (77) 487.75 p=[0.000] Pseudo R Squared 0.0393

PANEL B: HIGHEST EDUCATION ATTAINMENT INDEX ( ie ) AND RISK TOLERANCE

LESS THAN HIGH




( 2=ie )



M.E. TSTAT M.E. TSTAT M.E. TSTAT M.E. TSTAT M.E. TSTAT Risk Tolerance iRT -0.08187 (3.68) -0.05319 (3.67) -0.01718 (3.39) 0.05321 (3.65) 0.09902 (3.68) Chi Squared (77) 484.41 p=[0.000] Pseudo R Squared 0.0389

PANEL C: YEARS OF COMPLETED SCHOOLING INDEX ( is )

COEF TSTAT COEF TSTAT Risk Aversion Index iRA -0.12852 (6.00)

Risk Tolerance iRT 1.00964 (4.68)

Adjusted R Squared 0.1000 0.0951

OBSERVATIONS 4,246

Notes: (i) Control variables are as shown in Table 4; (ii) M.E. denotes marginal effects; (iii) The results shown in Panels A and B are estimated from an ordered probit model, those in Panel C are from OLS estimation.

Table 4: The Determinants of Educational Attainment

HIGHEST EDUCATION

ATTAINMENT INDEX ie YEARS OF COMPLETED SCHOOLING INDEX is

COEF TSTAT COEF TSTAT Age 0.0378 (3.48) 0.1552 (6.95) Age Squared -0.0004 (2.58) -0.0018 (6.45) Male 0.1066 (2.65) 0.3041 (3.69) White 0.4342 (4.04) 0.7934 (3.01) Male×White -0.1572 (1.16) -0.4002 (1.26) Mother single when child -0.1848 (2.55) -0.4358 (3.21) Mother widow when child -0.0650 (0.34) -0.2972 (0.55) Mother separated or divorced when child -0.0248 (0.21) -0.4789 (2.05) Lived with parents until 16 0.1167 (3.04) 0.0494 (0.65) Mother worked when child 0.0457 (1.27) 0.1977 (2.76) Father professional or managerial 0.5478 (8.39) 1.5545 (12.92) Father self employed 0.4041 (5.08) 1.1876 (7.28) Father clerical or crafts 0.1061 (2.49) 0.3393 (3.98) Father manual 0.0353 (0.76) 0.1160 (1.26) Number of siblings -0.0069 (1.00) -0.0032 (0.24) Firstborn 0.0522 (1.29) 0.0013 (0.02) Mother born outside US -0.0046 (0.11) 0.0445 (0.55) Mother high school education 0.2724 (3.69) 0.2334 (2.75) Mother college education 0.3112 (3.72) 0.1990 (2.23) Father high school education -0.1722 (2.68) -0.0696 (0.57) Father college education -0.0779 (1.02) 0.1022 (0.67) School expenditure in the county 0.0116 (2.21) 0.0258 (2.48) Roman Catholic 0.2523 (4.34) 0.0545 (0.47) Jewish 0.3161 (2.67) 0.1273 (0.48) Christian 0.2611 (4.98) 0.0490 (0.47) Other religion 0.4124 (3.91) 0.1787 (0.77) Family was poor when growing up 0.1372 (3.81) 0.0234 (0.67) Risk Aversion Index iRA -0.0445 (4.23) -0.1285 (6.00)

Other Controls Regional Dummies (49) Chi Squared (77) 487.75 p=[0.000] – Pseudo R Squared 0.0393 – Adjusted R Squared – 0.1000 OBSREVATIONS 4,246

Notes: ie is estimated as an ordered probit model and is is estimated by OLS.

Table 5A: The Determinants of Risk Preference

RISK AVOIDANCE iRA RISK TOLERANCE iRT COEF TSTAT COEF TSTAT Age -0.0205 (1.73) 0.0174 (1.38) Age Squared 0.0005 (3.15) -0.0004 (2.68) Male -0.1850 (3.45) 0.2146 (3.75) White -0.0514 (0.64) 0.0250 (0.28) Married 0.1201 (2.56) -0.1681 (3.36) Full-time employee 0.0326 (0.40) -0.0645 (0.78) Own home 0.1139 (2.83) -0.1234 (2.89) Log dividends payments -0.0208 (2.52) 0.0242 (2.75) Log interest payments -0.0222 (3.46) 0.0186 (2.68) Log all other payments -0.0114 (1.76) 0.0097 (1.39) Other Controls Occupation, Industry and Regional Dummies (26) Chi Squared (36) 232.21 p=[0.000] 211.45 p=[0.000] Pseudo R Squared 0.0164 0.0203 OBSREVATIONS 4,246

Table 5B: The Determinants of Risk Preference – Marginal Effects of Age

PANEL A: Risk Aversion Index iRA LOWEST ( )0=iRA HIGHEST ( )5=iRA

M.E. TSTAT M.E. TSTAT

Age 0.0022 (1.73) -0.0070 (1.72)

Age Squared -0.0005 (3.13) 0.0002 (3.14)

PANEL B: Risk Tolerance iRT LOWEST ( )15.0=iRT HIGHEST ( )57.0=iRT

M.E. TSTAT M.E. TSTAT

Age -0.0069 (1.38) 0.0005 (1.38)

Age Squared 0.0002 (2.68) -0.0001 (2.69)

Notes: (i) Control variables are as in Table 5A; (ii) M.E. denotes marginal effects.

Table 6: The Effect of Predicted Risk Preference on Educational Attainment

PANEL A: HIGHEST EDUCATION ATTAINMENT INDEX ( ie ) AND RISK AVERSION

LESS THAN HIGH SCHOOL

COMPLETED ( 0=ie )

HIGH SCHOOL COMPLETED

( 1=ie )

WENT TO COLLEGE DID NOT

GRADUATE ( 2=ie )



M.E. TSTAT M.E. TSTAT M.E. TSTAT M.E. TSTAT M.E. TSTAT Predicted Risk Aversion Index ¶

iRA 0.01664 (5.21) 0.01088 (5.13) 0.00350 (4.55) -0.01083 (5.17) -0.02019 (5.18) Chi Squared (77) 492.85 p=[0.000] Pseudo R Squared 0.0398

PANEL B: HIGHEST EDUCATION ATTAINMENT INDEX ( ie ) AND RISK TOLERANCE

LESS THAN HIGH SCHOOL

COMPLETED ( 0=ie )

HIGH SCHOOL COMPLETED

( 1=ie )

WENT TO COLLEGE DID NOT

GRADUATE ( 2=ie )



M.E. TSTAT M.E. TSTAT M.E. TSTAT M.E. TSTAT M.E. TSTAT Predicted Risk Tolerance ¶

iRT -0.13665 (3.23) -0.08876 (3.20) -0.02863 (3.01) 0.08895 (3.21) 0.16509 (3.22) Chi Squared (77) 472.97 p=[0.000] Pseudo R Squared 0.0387


COEF TSTAT COEF TSTAT

Predicted Risk Aversion Index ¶iRA -0.18474 (6.01)

Predicted Risk Tolerance ¶iRT 1.55467 (4.02)


OBSERVATIONS 4,246 Notes: (i) Control variables are as shown in Table 4; (ii) M.E. denotes marginal effects; (iii) The results shown in Panels A and B are estimated from an ordered probit model, those in Panel C are from OLS estimation.

Table 7: The Influence of Risk Preference on Education Attainment across Age Cohorts

DECADE OF BIRTH

COEFFICIENT NULL HYPOTHESIS 1930 1940 1950 1960 1970

Dependent Variable = ie , Risk Preference = iRA

= 1996 effect

-0.0445

-0.1434

p=[0.0651]

-0.0958

p=[0.0592]

-0.0248

p=[0.2703]

-0.0166

p=[0.1295]

-0.0687

p=0.3807]

Dependent Variable = ie , Risk Preference = iRT

= 1996 effect

0.3894

1.7016

p=[0.0284]

0.8813

p=[0.0792]

0.1790

p=[0.2420]

0.1021

p=[0.1201]

0.7039

p=0.2364]

Dependent Variable = is , Risk Preference = iRA

= 1996 effect

-0.1285

-0.3436

p=[0.1113]

-0.3120

p=0.0750]

-0.1121

p=[0.6648]

-0.1123

p=[0.8585]

-0.1328

p=[0.9324]

Dependent Variable = is , Risk Preference = iRT

= 1996 effect

1.0096

3.5549

p=[0.1145]

2.7086

p=[0.0104]

0.7412

p=[0.4847]

0.9730

p=[0.9172]

1.2108

p=[0.6829]

OBSERVATIONS

(% SAMPLE)

199

(4.69%)

693

(16.32%)

1,426

(33.58%)

1,321

(31.11%)

607

(14.30%)

Notes: (i) Controls are as in Table 4; (ii) ie is estimated as an ordered probit model and is is estimated by OLS.

Table 8: The Effect of Risk Preference across Time on Educational Attainment

PANEL A: HIGHEST EDUCATION ATTAINMENT INDEX ( ie ) AND RISK AVERSION (1996)

LESS THAN HIGH




( 2=ie )



M.E. TSTAT M.E. TSTAT M.E. TSTAT M.E. TSTAT M.E. TSTAT Risk Aversion Index iRA 0.01949 (3.46) 0.01409 (3.33) 0.00691 (2.92) -0.01881 (3.35) -0.02167 (3.48) IMPACT 2.95% 2.13% 1.04% -2.84% -3.28% Chi Squared (35) 120.62 p=[0.000] Pseudo R Squared 0.0601

PANEL B: HIGHEST EDUCATION ATTAINMENT INDEX ( ie ) AND RISK AVOIDANCE INDEX (1969-72)

LESS THAN HIGH




( 2=ie )



M.E. TSTAT M.E. TSTAT M.E. TSTAT M.E. TSTAT M.E. TSTAT Risk Avoidance Index o

iRA 0.04662 (6.75) 0.03638 (6.14) 0.01888 (4.41) -0.05021 (6.21) -0.05167 (6.95) IMPACT 6.49% 5.06% 2.63% -6.99% -7.19% Chi Squared (35) 166.40 p=[0.000] Pseudo R Squared 0.0827


COEF TSTAT IMPACT COEF TSTAT IMPACT Risk Aversion Index iRA -0.25661 (4.04) 6.95%

Risk Avoidance Index oiRA -0.72379 (10.67) 17.08%


OBSERVATIONS 647

Notes: (i) Control variables are as shown in Table 4; (ii) M.E. denotes marginal effects relating to the probability of having no education; (iii) The results shown in Panels A and B are estimated from an ordered probit model, those in Panel C are from OLS estimation; (iv) in Panels A and B the “IMPACT” figures are derived by multiplying the marginal effect by the standard deviation of the relevant risk preference measure; (v) in Panel C the figures under the column “IMPACT” are derived from ( ) ( )s r r s∂ ∂ × , i.e. risk preference elasticity, where ( )ˆ s rφ = ∂ ∂ and r and s denote the mean values of r and s respectively.

Table 9: Wage Determination and Risk Preference

HIGHEST EDUCATION

ATTAINMENT INDEX ie YEARS OF COMPLETED SCHOOLING INDEX is

COEF TSTAT COEF TSTAT PANEL A Intercept 1.5532 (22.14) 0.2779 (3.49) Male 0.2535 (14.57) 0.2050 (12.74) Experience 0.0347 (14.42) 0.0329 (14.16) Experience Squared -0.0006 (11.67) -0.0005 (9.61) White -0.0724 (2.43) -0.0512 (1.80) Union member 0.1101 (6.65) 0.1050 (6.80) Full-time employee 0.2885 (7.96) 0.2686 (7.73) Highest education attainment index ie 0.1295 (21.13) – –

Years of completed schooling is – – 0.1128 (32.24)


PANEL B

Intercept 1.5613 (21.86) 0.2626 (3.25) Male 0.2532 (14.56) 0.2053 (12.77) Experience 0.0347 (14.43) 0.0328 (14.14) Experience Squared -0.0006 (11.64) -0.0005 (9.62) White -0.0726 (2.43) -0.0510 (1.80) Union member 0.1106 (6.67) 0.1042 (6.73) Full-time employee 0.2887 (7.97) 0.2682 (7.70) Highest education attainment index ie 0.1293 (21.11) – –

Years of completed schooling is – – 0.1131 (32.35) Risk Preference – iRA -0.0030 (0.67) 0.0046 (1.09)

Adjusted R Squared 0.3656 0.4577 Other Controls Occupation, industry and regional dummies OBSREVATIONS 4,246

Table 10: The Effect of Predicted Education on Wage Determination

PANEL A: PREDICTED HIGHEST EDUCATION ATTAINMENT INDEX ie

COEF TSTAT COEF TSTAT COEF TSTAT COEF TSTAT

ie based on iRA 0.03062 (3.88)

ie based on iRT 0.02955 (3.68)

ie based on predicted iRA – ¶iRA 0.03389 (4.29)

ie based on predicted iRT – ¶iRT 0.03597 (4.54)

Adjusted R Squared 0.3006 0.3004 0.3010 0.3013

PANEL B: PREDICTED YEARS OF COMPLETED SCHOOLING INDEX is

COEF TSTAT COEF TSTAT COEF TSTAT COEF TSTAT

is based on iRA 0.12741 (11.18)

is based on iRT 0.12884 (11.00)

is based on predicted iRA – ¶iRA 0.13417 (11.21)

is based on predicted iRT – ¶iRT 0.13287 (10.90)

Adjusted R Squared 0.3207 0.3204 0.3220 0.3211

OBSERVATIONS 4,246

Notes: With the exception of the measures of education, control variables are as shown in Table 9 Panel A.

Documents

Education, risk preference and wages