IdC

Sh

Qu

1.

ChhifaloPedescbefa20sy20(SHu(CZh

Economics and Human Biology 11 (2013) 474–487

A

Art

Re

Re

Ac

Av

Ke

Inf

De

Fam

Ch

Gr

*

15

htt

entifying the intergenerational effects of the 1959–1961hinese Great Leap Forward Famine on infant mortality

ige Song *

eens College & CUNY Institute for Demographic Research, City University of New York, 65-30 Kissena Blvd., Queens, NY 11367, USA

Introduction

The 1959–1961 Great Leap Forward Famine (GLFF) inina was one of the most tragic events in modern human

story. Based on commonly accepted estimates, themine caused 30 million deaths and 33 million fertilitysses during the three-year period (Ashton et al., 1984;ng, 1987; Yao, 1999). Recent research in the fields ofmography, economics, public health, and medicaliences suggests that the impact of the GLFF has extendedyond the three famine years. People born during themine suffered from increased risk of obesity (Luo et al.,06), schizophrenia (Song et al., 2009), metabolicndrome (Li et al., 2011b), and hypertension (Li et al.,11a; Huang et al., 2010b) as well as impaired fecundity

ong, 2013), reduced adult height (Chen and Zhou, 2007;ang et al., 2010b), and poor labor market outcomes

hen and Zhou, 2007), among others (Song, 2010; Mu andang, 2011; Almond et al., 2010). A number of studies

show that prenatal exposure to the GLFF may even have an‘‘intergenerational’’ effect and influence the health andwell-being of children of the famine cohort (Huang et al.,2010a; Kim et al., 2012; Fung and Ha, 2010).

The current study aims to contribute to this rapidlyexpanding literature by focusing on the relationshipbetween mothers’ prenatal exposure to the GLFF and theinfant mortality risk of their children. Using data from the2001 National Family Planning and Reproductive HealthSurvey, a large, nationally representative sample surveyconducted in China, and data from the 1982 ChinesePopulation Census, I conducted simple cohort difference(SCD) analysis, difference-in-differences (DID) analysis,and conditional cohort difference (CCD) analysis to isolatethe effects of developmental plasticity, developmentdisruption, and selection simultaneously. To the best ofmy knowledge, no such effort has been made before.

The remainder of this article proceeds as follows. First, Ibriefly review the relevant literature that connectswomen’s prenatal famine exposure to the infant mortalityrisk of their children, from the ‘‘fetal origins’’ hypothesis tolife history regulation theory and the developmentalorigins of health and disease (DOHaD) framework. I then

R T I C L E I N F O

icle history:

ceived 27 March 2013

ceived in revised form 26 August 2013

cepted 26 August 2013

ailable online 12 September 2013

ywords:

ant mortality

velopmental origins of health and disease

ine

ina

eat Leap Forward

A B S T R A C T

Using the 1959–1961 Chinese Great Leap Forward Famine as a natural experiment, this

study examines the relationship between mothers’ prenatal exposure to acute

malnutrition and their children’s infant mortality risk. According to the results, the

effect of mothers’ prenatal famine exposure status on children’s infant mortality risk

depends on the level of famine severity. In regions of low famine severity, mothers’

prenatal famine exposure significantly reduces children’s infant mortality, whereas in

regions of high famine severity, such prenatal exposure increases children’s infant

mortality although the effect is not statistically significant. Such a curvilinear relationship

between mothers’ prenatal malnutrition status and their children’s infant mortality risk is

more complicated than the linear relationship predicted by the original fetal origins

hypothesis but is consistent with the more recent developmental origins of health and

disease theory.

� 2013 Elsevier B.V. All rights reserved.

Tel.: þ1 718 997 2822; fax: þ1 718 997 2820.

E-mail addresses: shige.song@qc.cuny.edu, songsg@gmail.com.

Contents lists available at ScienceDirect

Economics and Human Biology

jou r nal h o mep age: h t t p: / /w ww.els evier . co m/lo c ate /eh b

70-677X/$ – see front matter � 2013 Elsevier B.V. All rights reserved.

p://dx.doi.org/10.1016/j.ehb.2013.08.001

inCsd

2

ethmeeedeTpoeo(Bc‘‘thhaa

2

o

inlaolowaplags2owrthuinRehmeteRCM

e

S. Song / Economics and Human Biology 11 (2013) 474–487 475

troduce the empirical context – the 1959–1961 GLFF inhina – and describe the data, variables, and analytictrategy. Finally, I present the key statistical results andiscuss their demographic significance and implications.

. Conceptual framework

The following three distinctive types of long-termffects of prenatal famine exposure have been proposed ine literature: a developmental disruption effect, develop-ental plasticity effect, and selection effect. The disruption

ffect, also known as the ‘‘fetal origins’’ effect, ‘‘scarringffect’’, or ‘‘debilitation effect’’, states that prenatal faminexposure leads to an increase in disease risk (and thus aecrease in health status and life expectancy) (Razzaquet al., 1990; Chen and Zhou, 2007; Huang et al., 2010b).he possibility that prenatal famine exposure has aositive, or plasticity, effect on health has been mentionednly recently (Gluckman and Hanson, 2006; Gluckmant al., 2005; Painter et al., 2008), based on biological theoryf developmental plasticity and life history regulationateson et al., 2004; Stearns, 1992). A selection effect

aused by differential population attrition may alsoreduce’’ the disease risk of the famine cohort. Unlike

e individual-level biological effects discussed above,owever, the selection effect operates at the cohort-levelnd must be controlled for when estimating the individu-l-level biological effects.

.1. Fetal origins, life history regulation, and developmental

rigins: a critical review

Barker and colleagues played a pivotal role in promot-g the idea that prenatal conditions may have a long-sting health consequences by articulating the ‘‘fetal

rigins’’ hypothesis and providing some early epidemio-gical evidence of an inverse relationship between birtheight and the risk of chronic disease later in life (Barker

nd Osmond, 1986; Barker, 1995, 1992). The notion thatrenatal exposure to adverse conditions may have a long-sting and nontrivial effect on adult health has been

radually accepted by health researchers as well as socialcientists, especially economists (Almond and Currie,011; Doyle et al., 2009; Osmani and Sen, 2003). Criticsf the fetal origins hypothesis claimed that the argumentas too broad and vague to be rigorously tested and

efuted. In addition, the lack of experimental results ande inherent weaknesses of the observational study design

sed in these studies made it difficult to make causalference (Joseph et al., 1996; Paneth and Susser, 1995;

asmussen, 2001; Tu et al., 2005; Lucas et al., 1999; Huxleyt al., 2002). As a response to these criticisms, researchersave recently employed a famine-based natural experi-ental approach to isolate the causal effect of prenatal

xposure to acute malnutrition on adult disease risk and tost the fetal origins hypothesis (Roseboom et al., 2001;

avelli et al., 1998; Painter et al., 2005; Song et al., 2009;hen and Zhou, 2007; Sotomayor, 2013; Almond andazumder, 2011).

Lumey and Stein (1997) tested the potential long-termffect of prenatal malnutrition on reproductive outcomes

using the 1944–1945 Dutch famine as a natural experi-ment. Their results showed that prenatal famine exposuredid not significantly affect women’s age at menarche, theproportion of women without children, women’s age atfirst delivery, or the number of children born. However,they found an excess of stillbirth and perinatal deathamong children of the famine-born women, providingsome support to the fetal origins hypothesis. In a morerecent study, Barker and colleagues reported that prenatalexposure to the 1944–1945 Dutch famine led to anenhanced female reproductive function, as evidenced bymore children born, more twin births, earlier onset ofchildbirth, and lower likelihood of childlessness (Painter etal., 2008). The authors utilized developmental plasticityand life history regulation theories to explain these results:fertility and body maintenance are mutually balanced; ifprenatal famine exposure led to a reduced survivalopportunity (as predicted by the fetal origins hypothesis),an enhanced reproductive success was likely to beobserved.

The DOHaD framework provides an opportunity toexplain these inconsistent empirical findings. Gluckman etal. (2005) presented a schematic map of the nonlinearrelationship between prenatal nutrition and adult fitness.They acknowledged the critical importance of the prenatalnutritional condition in determining health and reproduc-tive success later in life and argued that both overly poorand overly rich nutritional conditions can negativelyinfluence prenatal development. In contrast, it is themoderately low prenatal nutritional condition that is morelikely to achieve optimal fitness. The key to this process isthe predictive adaptive response mechanism of develop-mental plasticity during the prenatal stage, which predictsthe future (postnatal) environment based on the current(prenatal) environment and then regulates the life historystrategy of the unborn child to rebalance between survivaland reproductive efforts based on such predictions (Gluck-man et al., 2005; Gluckman and Hanson, 2006; Bateson etal., 2004). In the case of famine, prenatal malnutritionsignals the unborn child that it is about to enter a resource-poor world in which a compromised survival (e.g., poorhealth, reduced longevity) may be unavoidable. In thissituation, adopting a life history strategy that optimizesreproductive success, as Painter et al. (2008) suggested,may be evolutionarily advantageous. However, when thefamine-induced malnutrition becomes too severe for thefetus to respond adaptively, normal prenatal developmentwill be disrupted, which leads to pathological develop-mental outcomes that do not have adaptive values. Thisleads to the hypothesis that prenatal exposure to low levelof famine-induced malnutrition may lead to improvedreproductive outcomes, whereas exposure to severemalnutrition may lead to deteriorated outcomes.

2.2. Famine, natural experiment, and selection

From a study design point of view, famine-basednatural experimental studies have important advantagesover observational studies that use birth weight as a proxyfor prenatal conditions. Birth weight is an imperfectmeasure of prenatal nutritional condition. It may be

joenmtiosunizathSebesohegrthex

thnoaninsufro‘‘wdifa

3.

Plthidtoge

3.1

mbelivannearprththmcetraeafohopu20exincothdire

S. Song / Economics and Human Biology 11 (2013) 474–487476

intly determined with adult outcomes by genetic andvironmental factors, many of which cannot be directlyeasured. Without randomization, the estimated rela-nship between birth weight and adult outcome isbject to endogeneity bias. Famine provides an opportu-ty to tackle the endogeneity problem without randomi-tion. First, the occurrence of famine is typically beyonde control of most individuals or individual families.cond, although the influence of famine may differtween individuals due to the buffering mechanism ofcial institutions, there are cases in which individualterogeneity in the ability to cope with famine waseatly reduced by war and social disruption. Any famineat meets these criteria can be treated as a naturalperiment.The use of famine as natural experiment poses additional

reats to validity that randomized experiments typically dot face. Famine is often associated with increased mortalityd reduced fertility. If the famine-induced mortalitycrease or fertility reduction is not random (i.e., if thervivors of the famine cohort are significantly differentm those in the non-famine cohorts because the famineeeded out’’ the frail members of the famine cohort), a

rect cohort comparison between the famine and non-mine cohorts may be misleading.

The current study

Combining information from the 2001 National Familyanning and Reproductive Health Survey (NFPRHS) ande 1982 Chinese Population Census, this study aims toentify the intergenerational effects of prenatal exposure

the GLFF on the risk of infant mortality among the nextneration.

. The 1959–1961 Chinese Great Leap Forward Famine

The GLFF occurred at the climax of the collectivizationovement (Lin and Yang, 2000). By the time the faminegan, all land, factories, shops, machines, tools, andestock were collectivized. All individuals, both urband rural, worked for the state for food rations and othercessities under the supervision of local cadres. In ruraleas, ‘‘People’s Communes’’ were established as theimary administrative unit. Communes were much largeran traditional villages and typically consisted of tens ofousands of people. They were run centrally (in a semi-ilitary fashion) by cadres. Most communes establishedntrally managed dining halls as alternatives to theditional household kitchen. To ‘‘encourage’’ people to

t in the dining hall, commune-run militia confiscatedod and cooking utensils from individual rural house-lds. People who failed to relinquish their grain wereblicly criticized, humiliated, and punished (Dikotter,10; Becker, 1996). The People’s Communes represent antreme form of social control the state exerted on ruraldividuals that was rarely seen in modern history. Themmune controlled all the food in the rural area. Whene famine hit and the food supplied by the communening hall decreased, rural residents did not have foodserves to supplement their diet. Even worse, rural

residents could not resort to the traditional famine reliefmechanisms (family, neighbors, friends, relatives, andkinsmen) either because they were also under the tightcontrol of the commune system and did not have foodreserves. Migration, the last resort for survival in famine-infested areas, was strictly prohibited by local governmentand enforced by local police and militia.

These characteristics are important for the purpose ofthis study. First, the collectivization and communalizationgreatly reduced the level of population heterogeneity withrespect to the ability to cope with the famine-induced foodshortage, making the GLFF a viable choice for a naturalexperiment. Second, the ban on migration and thecommune’s tight control over the rural populationsseverely restricted their ability to leave the communefor food. This further reduced the population heterogeneitywith respect to the ability to cope with the famine.Heightened social control and greatly reduced socialinequality constitutes a rare combination. Along withthe drastic regional variation in famine severity, asreported by Lin and Yang (2000) and Peng (1987), theymakes the GLFF a particularly revealing case for thepurpose of assessing the long-term consequences ofprenatal exposure to malnutrition.

To some extent, what happened in China during theGLFF shared some similarities to what happened inUkraine during the 1933 famine (Davies and Wheatcroft,2004), presumably because of the resemblance in thepolitical system between China under Mao and the SovietUnion under Stalin. However, compared to the Ukrainianfamine, the GLFF lasted much longer, influenced a muchlarger population living in a much wider geographic area,and caused a much higher excess mortality. In addition, themajority of Chinese who were born around the time of theGLFF are still alive, which makes it possible to carry outnew data collections to study various kinds of long-termeffects of famine exposure. This is an opportunity that hasforever been lost for the Ukrainian famine as thegeneration of Ukrainians who were born around thefamine have gradually died out.

The GLFF occurred to a population that long sufferedfrom chronic malnutrition. Poor nutritional conditionpersisted in China throughout the 1960s, 1970s, and thefirst half of the 1980s, long after the end of the GLFF (Lardy,1984; Chai, 1992). This is in sharp contrast to the 1944–1945 Dutch Famine, which occurred to an otherwise well-nourished Western population. The GLFF and the Dutchfamine offer interesting contrast because the formerrepresents the case where the prenatal and postnatalenvironments match reasonably well whereas the latterrepresents the case where the prenatal environmentgrossly mismatches the postnatal environment.

3.2. Simple cohort difference, difference-in-differences, and

conditional cohort difference analysis

The validity of the SCD method rests on the ‘‘zero cohortdifference’’ assumption: in the counterfactual absence ofthe famine influence, there is no cohort difference in theoutcome of interest between the famine and non-faminecohorts. The presence of a non-zero counterfactual cohort

dothccatifacinthstudthti

inpbscuApmCawethfobcafa

thteodmartaansmceusrti

re

p

c

S. Song / Economics and Human Biology 11 (2013) 474–487 477

ifference can bias SCD results. This non-zero differenceccurs because events other than famine may also affecte outcome of interest. If this is the case, the DID method

an be used to partial out the influence of the non-zeroounterfactual cohort difference. The DID method rests on

somewhat weaker ‘‘constant cohort difference’’ assump-on: the counterfactual cohort difference between themine and non-famine cohorts, zero or not, does not

hange between the regions that were under the faminefluence and those that were not. It is important to noteat neither assumption can be assessed solely based on

tatistical evidence because they both involve counterfac-al information that is inherently unobservable. Well-

esigned ad hoc tests are informative, but the validity ofese assumptions should be assessed mainly on substan-

ve grounds.Frisbie (2005) identified a long list of risk factors of

fant mortality. Most of the risk factors that canotentially create a cohort difference in infant mortalityetween the famine and adjacent non-famine cohorts,uch as mother’s education, ethnicity, mother’s age athildbirth, birth order, sex of the child, and prenatal carese, can be directly controlled for in the current research.1

mong the factors that cannot be directly controlled for,overty, malnutrition, health insurance, and maternalorbidity are highly correlated with urban–rural status in

hina, which can be controlled for. Furthermore, smokingnd binge drinking behaviors are rare among Chineseomen and the pattern has been stable over time (Yang

t al., 1999; Cochrane et al., 2003). Maternal birth weight ise only important factor that cannot be directly controlledr in the present study. However, this is not a limitation

ecause the difference in infant mortality induced by theohort difference in maternal birth weight between faminend non-famine cohorts is part of the intergenerationalmine effect that the present study aims to estimate.

In addition, the literature on infant mortality suggestsat, with the introduction of highly effective life-savingchnology and services such as prenatal care, the pattern

f infant mortality has been increasingly dominated by theifferential access to such technology and services (Gort-aker and Wise, 1997; Frisbie et al., 2004; Wise, 2003). In

sense, prenatal care is an ‘‘equalizer’’ that can drasticallyeduce preexisting population differences in infant mor-lity by effectively lowering the infant mortality risk

mong people who have access, regardless of theirutritional conditions, socioeconomic position, healthtatus, and health behavior. In a recent study of infantortality in China, Song and Burgard (2011) showed that

ontrolling for prenatal care utilization status completelyliminates the direct effects of mother’s education andrban–rural status, the two most important indicators ofocioeconomic position in China, on infant mortality. Thisesult further supports the zero cohort difference assump-on.

With the zero cohort difference assumption, the choicebetween SCD, DID, and CCD becomes a statistical one. BothDID and CCD are based on statistical models withinteractions between measures of famine exposure andfamine severity, whereas SCD is based on statistical modelswithout such interactions. The two models can becompared based on model fit statistics. Extracted fromthe same statistical model, CCD and DID representdifferent ways of interpreting the same statistical results,both leading to important findings that are complementa-ry to each other.

3.3. Determining famine exposure status

Table 1 summarizes the prenatal and early-life famineexposure status of women conceived in different years andmonths. The pre-famine cohort includes women conceivedin 1955–1957; the famine cohort includes womenconceived between August and December of 1958, in1959–1960, and in January of 1961; the first post-faminecohort includes women conceived in 1962–1964 and thesecond post-famine cohort includes women conceived in1965–1967. Women conceived in the first seven months of1958 and the last eleven months of 1961 were excludedfrom the analysis because their prenatal famine exposurestatus was less clear than that of the other cohorts.

This cohort classification attempted to balance thefollowing two competing goals: making the contrastbetween the famine and the non-famine cohorts, especial-ly between the famine and post-famine cohorts, as clear aspossible while allocating a sufficient number of cases to thefamine cohort to ensure that the statistical analysis hadsufficient power. Sensitivity tests (available upon request)showed that the main results were not sensitive to thechanges in the famine cohort definition.

Two post-famine cohorts were included in the analysisto provide an ad hoc test of the constant (for DID analysis)or zero (for SCD and CCD analysis) cohort differenceassumption. For the assumptions to hold, no significantcohort differences in infant mortality between the firstpost-famine and second post-famine cohorts are expected.

3.4. Measuring regional variations in famine severity: a

multiple imputation approach

Because direct measures of famine severity are notavailable, two types of indirect measures have been used inprevious research. The first is a period measure of famine-induced mortality increase for all cohorts during the threefamine years (Chen and Zhou, 2007; Mu and Zhang, 2011;Luo et al., 2006; Almond et al., 2010; Fung and Ha, 2010),whereas the second is a cohort measure of all types offamine-induced cohort attrition that occurred to the threefamine cohorts (Huang et al., 2010b; Meng and Qian,2009). One problem with the mortality-based measure isthat there is no official estimate of the famine-inducedexcess mortality and existing estimates vary greatly. Thus,it is not clear whether and to what extent the province-level mortality estimates provided by Lin and Yang (2000),which are the most widely used province-level mortalityestimates, are subject to measurement error. In contrast,

1 Table A.2 in the online supporting materials reports sensitivity test

sults that do not have mother’s education and prenatal care use, the two

otentially endogenous variables, in the model. Excluding these two

ontrol variables has little impact on the key results.

thpuwco

imtiothfothfethorkninthfase

Ch20thoc20ni19Thapgeco10fa1927Figm

sizcocosizwimmseexth

Ta

Tim

T

Y

1

1

1

1

S. Song / Economics and Human Biology 11 (2013) 474–487478

e cohort size-based famine severity measure utilizes theblic-use samples of the Chinese Population Census,

hich are known to have high quality and good populationverage (Lavely et al., 1990).The cohort size-based famine severity index also hasportant advantages in handling selection effect. Selec-n effect is caused by (differential) cohort attrition, and

e cohort size-based famine severity index captures allrms of famine-induced cohort attrition that occurred toe famine cohort, including reduced pregnancy, increasedtal loss, and increased infant and early child mortality. Ife cohort size-based famine severity index indicates zero close to zero cohort attrition in certain regions but weow, a priori, that people in these regions were indeed

fluenced by the famine (Yang, 2012; Dikotter, 2010),en these regions give us an opportunity to estimate themine effect with zero or close to zero influence oflection.Computed from the 1% public use sample of the 1982inese Population Census (Minnesota Population Center,10), Fig. 1 shows the trend in birth cohort size arounde time of the famine. A sudden drop in cohort sizecurred between 1958 and 1959, from approximately0,000 to approximately 140,000, indicating the begin-ng of the GLFF. Cohort size remained unchanged between59 and 1960, indicating the continuation of the famine.ere was an additional drop in cohort size in 1961, fromproximately 140,000 to approximately 100,000, sug-sting that the famine worsened in this period. Finally,hort size began to increase in 1962, from approximately0,000 to approximately 160,000, suggesting that the

mine gradually ended. Suddenly, the cohort size rose in63 from approximately 160,000 to approximately0,000, indicating that the famine had completely ended.. 1 suggests that changes in cohort size can be a sensitive

easure of famine severity.In the current study, I use the observed famine cohorte and the ‘‘expected’’ size of the famine cohort under theunterfactual conditions of no famine influence tonstruct the famine severity index. I treat the expectedes of the famine cohorts under non-famine conditions,

hich are inherently unobservable, as missing data andpute them, from a set of cross-sectional time series,

ultiple times. For each of the imputed data, the famineverity measure is calculated as the ratio between thepected and observed famine cohort size. I then merge

statistical analysis on each of the merged data, andcombine all the results based on Rubin’s rule of combina-tion to obtain the final results (Rubin, 1987).2

This new famine severity index is substantivelyintuitive. It captures the difference between the observed(with famine impact) and expected (without famineimpact) size of the same cohort. Because the imputationis based on time series data, the imputed famine cohortsize can reflect the influences of short- and long-termpopulation dynamics. By excluding certain borderlinecohorts from the imputation (e.g., cohorts immediatelyfollowing the famine), it is possible to produce differentsets of imputed values to assess the robustness of theresults via sensitivity analysis. More importantly, themultiple imputation-based workflow (Rubin, 1987) pro-vides an elegant solution to the measurement errorproblem (Blackwell et al., 2011; Cole et al., 2006).

4. Research design

4.1. Data and variables

I used data from the 2001 National Family Planning andReproductive Health Survey (NFPRHS), a nationally repre-sentative sample survey conducted by the State FamilyPlanning Commission of China. The survey collectedinformation from 39,586 women (29,512 rural and10,074 urban) aged 15–49 living in family households inall 31 provincial administrative units in China through astratified multistage sampling technique. All selectedwomen were asked to provide their complete pregnancyand birth history, including the year and month ofpregnancy termination and the outcome of each pregnan-cy. For each live birth, the survey also collected informa-tion on the mother’s use of prenatal care, the newborn’ssex, and the newborn’s survival over the first year of life.The response rate of the NFPRHS is 98.3%.

The underreporting of birth and infant deaths has beena problem for most population surveys in China, presum-ably because of the strict family planning policy. TheNFPRHS implemented various methods to reduce under-reporting. For example, the sampling plan was designed tobe representative only at the national level so that local

ble 1

e of conception and exposure to the 1959–1961 Great Leap Forward Famine.

ime of conception Prenatal famine Early-life famine Cohort

ear Month Exposure status Exposure status

955–1957 1–12 No Yes Pre-Famine Cohort

1958 1–7 No to Some Yes –

1958 8–12 Some to Full Yes Famine Cohort

959–1960 1–12 Full Yes Famine Cohort

1961 1 Some to Full Yes Famine Cohort

1961 2–12 No to Some Some –

962–1964 1–12 No No 1st Post-Famine Cohort

965–1967 1–12 No No 2nd Post-Famine Cohort

2

Details of Rubin’s rule of combination can be found in the online

pporting materials.

ese imputed data back into the main data set, conduct su

fatrpZfrinarcyte

thathloainvasmemc3

th

S. Song / Economics and Human Biology 11 (2013) 474–487 479

mily planning officials could not be held responsible ifue fertility level in their areas were higher thanreviously reported (Chen et al., 2007). In a recent study,hang and Zhao (2006) compared fertility rates derivedom Chinese household registration data, censuses,tercensal surveys, retrospective fertility surveys, and

nnual surveys of population change and showed thatetrospective fertility surveys (including that used in theurrent study) and annual surveys of population changeielded the best possible enumeration and highly consis-nt estimates.

Infant mortality was constructed from the followingree questions: (1) What is the outcome (live birth,

bortion, miscarriage, or stillbirth) of this pregnancy? (2) Ifis is a live birth, is the child still alive? (3) If the child is nonger alive, what was his/her age at the time of death? The

nalytic sample includes only live births (boys or girls), andfant mortality was constructed as a dichotomous

ariable for which 1 = infant is no longer alive and thege of death is no more than one year old and 0 = infant istill alive or infant is no longer alive but the age of death isore than one year old. Women’s prenatal famine

xposure status was derived from their time (year andonth) of conception.3 This variable has the following four

ategories: 1 = pre-famine cohort, 2 = famine cohort, = first post-famine cohort, and 4 = second post-famine

cohort. The precise definition of these three cohorts can befound in Table 1. The famine severity measure wasconstructed from other data sources and then merged intothe main data set for statistical analysis, followingprocedures described in the next subsection (see onlinesupporting materials for a discussion of the possiblemismatch between province of birth and province ofresidence in 2001).

I also included the following three types of controlvariables: (1) women’s characteristics (type of residence,education, and ethnicity), (2) infants’ characteristics (birthorder, sex, and mother’s age at childbirth), and (3) theutilization of prenatal care.

Type of residence was a binary variable that indicatedwhether the respondent lived in a rural or urban location.Some women may have changed their type of residencebetween birth and the time of the interview. However, dueto the household registration system that has been inplace since the late 1950s, only a highly selected minority(e.g., college graduates) could have crossed the urban–rural boundary, and the move was nearly always fromrural to urban areas (Wu and Treiman, 2004). I conductedan additional sensitivity analysis using only the subsam-ple of rural women to ensure that the key findings werenot influenced by rural-to-urban migration. Women’syears of schooling was measured as a continuous variable.It is an important socioeconomic indicator that has beenwidely used in demographic research in developingcountries where other socioeconomic indicators, suchas household income or occupation, are not reliable or

Fig. 1. Cohort size in China, 1950–1970.

3 Time of conception was generated by subtracting nine months from

e year and month of birth of the child.

avva0 =

subilev1

animvaefva(FbeexinroI cwut

co19wa

sa

Ta

De

M

E

T

Y

A

I

B

G

P

I

N

N

Sou

4

pre

fam

S. Song / Economics and Human Biology 11 (2013) 474–487480

ailable (Desai and Alva, 1998). Ethnicity was a binaryriable, coded such that 1 = Han ethnic majority and

non-Han ethnic minorities.The first child-level covariate was birth order, mea-

red as an ordinal variable, with 1 = first birth, 2 = secondrth, and 3 = third and higher-order birth. A second child-el variable was the sex of the newborn, coded such that

= male birth and 0 = female birth. Because male infantsd higher-order births face higher mortality, it isportant to control for both its temporal and regionalriations to obtain an unbiased estimate of the faminefect. Mother’s age at childbirth was the third child-levelriable. It is an important determinant of infant mortalityrisbie, 2005) and thus must be controlled for. Finally,cause previous studies have shown that the rapidpansion in the prenatal care provision and utilization

China during the past decades has played an importantle in reducing infant mortality (Song and Burgard, 2011),ontrolled for prenatal care use to ensure that the resultsere not driven by the secular increase in prenatal careilization.I restricted the study sample to 14,218 women

nceived between 1955 and 1957, between August58 and January 1961, and between 1962 and 1967,

ho ever had at least one live birth. These women reportedtotal of 27,276 live births.4

Table 2 reports the descriptive statistics for the selectedmple.

4.2. Constructing the famine severity index through multiple

imputation

To create the famine severity index, I first extractedinformation on the size of the ith birth cohort in the kthprovince, Nik, from the 1982 census data, where i = 1951,1952, . . ., 1970. This yielded 20 cross-sectional (province-level) time series on cohort size. I then recoded theobserved cohort size of the three famine cohorts, N1959,k,N1960,k, and N1961,k, into missing values. Next, I imputedthese ‘‘missing values’’ under the assumptions that, withineach province, cohort size changes smoothly between1951 and 1970 to obtain the expected counterfactualcohort sizes N�1959;k, N�1960;k, and N�1961;k.5 The imputationutilized the algorithm optimized for cross-sectional timeseries data (Honaker and King, 2010), as implemented inopen source software Amelia II (Honaker et al., 2011).Finally, the famine severity index for the kth province, Sk,was computed as the ratio between the sum of theexpected famine cohort sizes and the sum of the observedfamine cohort sizes:

Sk ¼N�1959;k þ N�1960;k þ N�1961;k

N1959;k þ N1960;k þ N1961;k(1)

The missing data imputation is inherently stochasticbecause each imputed value comes from an independentdraw from the target multivariate normal posteriordistribution. Statistical results based on any one of these

ble 2

scriptive statistics of the analytic sample.

Mother’s cohort

Pre-famine Famine 1st post-famine 2nd post-famine

other information

thnicity

Ethnic majority (%) 90.48 89.76 89.53 89.32

Ethnic minority (%) 9.52 10.24 10.47 10.68

ype of residence

Urban (%) 17.75 21.39 18.18 16.84

Rural (%) 82.25 78.61 81.82 83.16

ears of schooling 5.03 6.46 6.96 6.66

(4.38) (4.37) (3.87) (3.70)

ge at childbirth 26.24 25.61 24.74 24.35

(3.79) (3.51) (3.28) (3.31)

nfant information

irth order

First birth (%) 46.27 50.60 52.94 57.50

Second birth (%) 33.10 32.30 33.74 33.18

Third and higher-order (%) 20.63 17.09 13.32 9.32

ender

Male birth (%) 51.85 52.18 53.53 54.49

Female birth (%) 48.15 47.82 46.47 45.51

renatal care use (%) 33.94 45.38 52.92 56.57

nfant death (%) 1.72 0.98 1.56 1.39

of mothers 3309 1966 4587 4356

of infants 7151 3885 8664 7576

rce: Author’s calculations based on the 2001 survey data.

The percentage of women who never had live births is 1.02% for the 5 The smooth, but otherwise unspecified, time trend was approximated

-famine cohort, 1.41% for the famine cohort, 1.11% for the first post-

ine cohort, and 1.28% for the second post-famine cohort.

by polynomials in the imputation model because polynomials tend to

outperform other smoothers with this kind of interpolation.

imaacr

4

s

wwim(eh‘‘rcGmetgtqtli

T

P

S

S. Song / Economics and Human Biology 11 (2013) 474–487 481

puted values underestimate the uncertainty involvednd is subject to measurement error. The recommendedpproach is to impute the missing value multiple times,onduct analysis on each imputed data, and combine theesults following Rubin’s rule (Rubin, 1987).

The new famine severity index is reported in Table 3.

.3. Generalized estimating equation and statistical

imulation

The NFPRHS data are multilevel in the sense that eachoman may have experienced multiple childbirths, all ofhich are present in the data. This data structure violatesportant assumptions of generalized linear models

GLM) and can lead to biased results. Generalizedstimating equation (GEE) is an extension of GLM thatandles clustered data structures by introducing aworking correlation’’ for the observed responses, whichepresents an a priori expectation about the within-luster variation (Zeger and Liang, 1986). Coefficients inEE have the same intuitive population average orarginal effect interpretation as those in GLM. Point

stimates of GEE coefficients are robust to misspecifica-ion of the working correlation structure, although aood working correlation can improve the efficiency ofhe model. Model selection for GEE can be based onuasilikelihood information criterion (QIC), a variation ofhe Akaike’s information criterion (AIC) for quasi-

independent variables as well as working correlations.More information on AIC and QIC can be found in onlinesupporting materials.

Let Cik be the birth cohort of the ith woman born in thekth region. Let Sk be the famine severity index of the kthregion (i.e., province). The probability of infant death of thejth birth of the ith woman born in the kth region can bemodeled as a GEE logistic regression

logPjik

1 � Pjik

� �¼ a þ b1X þ b21Cik1 þ b22Cik2 þ b24Cik4

þ b3logðSkÞ þ b41Cik1logðSkÞ

þ b42Cik2logðSkÞ þ b44Cik4logðSkÞ (2)

where X represents a vector of woman- and child-levelcontrol variables, Cik1 denotes the pre-famine cohort, Cik2

denotes the famine cohort, and Cik4 denotes the secondpost-famine cohort (Cik3, the first post-famine cohort, isused as the reference category). Cik1 log(Sk), Cik2 log(Sk),and Cik4 log(Sk) denotes the interactions between thepre-famine, famine, and second post-famine cohorts andthe logged famine severity index, accordingly.6

Statistical simulation provides a convenient way toextract quantities of interest from complicated statistical

able 3

rovince-level famine severity index via multiple imputation (M = 10).

Imputed values Mean

m = 1 m = 2 m = 3 m = 4 m = 5 m = 6 m = 7 m = 8 m = 9 m = 10

Beijing 1.06 1.32 1.06 1.15 1.24 1.16 1.24 1.06 1.27 1.34 1.19

Tianjin 1.51 1.40 1.34 1.44 1.49 1.47 1.38 1.65 1.35 1.12 1.42

Hebei 1.58 1.63 1.48 1.53 1.59 1.65 1.5 1.49 1.58 1.67 1.57

Shanxi 1.38 1.35 1.31 1.46 1.22 1.27 1.38 1.34 1.35 1.31 1.34

Neimenggu 1.28 1.35 1.12 1.29 1.12 1.14 1.29 1.28 1.32 1.25 1.24

Liaoning 1.55 1.37 1.53 1.40 1.45 1.38 1.50 1.49 1.44 1.37 1.45

Jilin 1.39 1.19 1.20 1.32 1.33 1.19 1.12 1.28 1.25 1.35 1.26

Heilongjiang 1.27 1.18 1.25 1.25 1.19 1.26 1.34 1.24 1.37 1.23 1.26

Shanghai 1.26 1.06 1.09 1.03 1.25 1.15 1.09 1.02 1.09 1.34 1.14

Jiangsu 1.83 1.56 1.68 1.80 1.73 1.77 1.75 1.86 1.87 1.47 1.73

Zhejiang 1.54 1.46 1.64 1.61 1.46 1.49 1.63 1.48 1.59 1.57 1.55

Anhui 2.86 3.13 3.06 3.16 2.90 3.28 3.34 3.18 2.84 3.07 3.08

Fujian 1.37 1.33 1.75 1.44 1.55 1.41 1.60 1.46 1.79 1.57 1.53

Jiangxi 1.51 1.47 1.40 1.44 1.43 1.36 1.37 1.40 1.51 1.35 1.42

Shandong 1.75 1.69 1.69 1.84 1.87 2.27 2.02 1.78 1.95 1.96 1.88

Henan 2.20 2.28 2.13 2.13 2.18 2.04 2.08 2.26 1.99 2.05 2.13

Hubei 1.61 1.76 1.84 1.66 1.75 1.68 1.79 1.78 1.60 1.67 1.71

Hunan 2.06 2.20 2.16 2.08 2.11 2.01 2.18 2.22 2.17 2.19 2.14

Guangdong 1.56 1.53 1.59 1.53 1.61 1.49 1.58 1.54 1.48 1.47 1.54

Guangxi 1.88 1.90 1.72 1.82 1.75 1.97 1.53 1.72 1.74 1.83 1.79

Sichuan 2.56 2.40 2.56 2.53 2.58 2.44 2.48 2.41 2.52 2.65 2.51

Guizhou 2.21 2.09 1.89 2.03 2.52 2.00 1.76 1.95 2.08 1.88 2.04

Yunnan 1.74 1.88 2.09 1.75 1.58 1.83 1.72 1.81 1.83 1.84 1.81

Xizang 1.18 1.25 .946 1.21 1.11 1.60 1.48 1.20 1.08 .993 1.21

Shaanxi 1.31 1.32 1.30 1.33 1.32 1.29 1.48 1.33 1.22 1.51 1.34

Gansu 2.15 2.16 2.1 2.03 2.05 1.93 2.02 2.00 1.78 1.7 1.99

Qinghai 1.82 1.72 1.79 2.2 2.05 2.25 1.62 1.97 1.58 1.93 1.89

Ningxia 1.84 2.02 2.09 1.75 1.92 1.65 1.82 1.84 1.33 1.92 1.82

Xinjiang 1.25 1.30 1.41 1.46 1.28 1.33 1.42 1.52 1.21 1.41 1.36

ource: The 1% public use sample of the 1982 Chinese Population Census.

kelihood models (Pan, 2001). QIC can be used to select

mthdipr

DD

wfacoasar

Ta

Re

Y

H

U

A

B

P

M

B

F

C

Q

E

*

*

*

6

co

ch

S. Song / Economics and Human Biology 11 (2013) 474–487482

odels (King et al., 2000). For example, the DID estimate ofe famine effect can be obtained in the form of the secondfference between two first differences in the predictedobabilities, obtained via simulation

Pðy ¼ 1jSk; Cik; XÞ ¼ ½Pðy ¼ 1jSk þ 1 S:D:; Cik

¼ 1; 2; 4; XÞ � Pðy ¼ 1jSk; Cik

¼ 1; 2; 4; XÞ� � ½Pðy ¼ 1jSk þ 1 S:D:; Cik

¼ 3; XÞ � Pðy ¼ 1jSk; Cik ¼ 3; XÞ� (3)

here Cik = 1 for the pre-famine cohort, Cik = 2 for themine cohort, and Cik = 4 for the second post-faminehort, while Cik = 3, the first post-famine cohort, is treated

the reference category against which the other cohortse compared. The famine severity index, Sk, is a

continuous variable. Its effect was taken as the changein the predicted probability of infant mortality producedby a one standard deviation (�0.44) increase in the famineseverity index. The values of the control variables X werefixed at meaningful levels (i.e., mean for continuousvariables and mode for categorical variables) in simulatingthe quantities of interest.

Similarly, the CCD estimate of the famine effect on infantmortality can be obtained in the form of cohort differencesevaluated at each different level of the famine severityindex:

DPðy ¼ 1jSk; Cik; XÞ ¼ Pðy ¼ 1jCik

¼ 1; 2; 4; Sk; XÞ � Pððy ¼ 1jCik ¼ 3; Sk; XÞ (4)

Details of statistical simulation are provided in theonline supporting materials.

ble 4

sults from GEE logistic regression models of infant mortality in China using one of the ten imputed data sets, N = 27, 276.

Model 1 Model 2 Model 3 Model 4

ears of schooling 0.96** 0.96** 0.96* 0.96*

[0.93, 0.99] [0.93, 0.98] [0.93, 0.99] [0.93, 0.99]

an ethnic majority 0.53*** 0.52*** 0.52*** 0.52***

[0.39, 0.71] [0.40, 0.69] [0.38, 0.72] [0.39, 0.71]

rban residence 0.94 0.96 0.96 0.96

[0.65, 1.41] [0.69, 1.34] [0.65, 1.42] [0.65, 1.42]

ge at childbirth 0.91*** 0.91*** 0.91*** 0.91***

[0.88, 0.95] [0.87, 0.95] [0.87, 0.95] [0.88, 0.95]

irth order

– First birth – – – –

– Second birth 1.24 1.24 1.24 1.24

[0.97, 1.57] [0.97, 1.59] [1.00, 1.54] [0.98, 1.58]

– Third and higher order 2.18*** 2.20*** 2.19*** 2.19***

[1.59, 3.02] [1.57, 3.09] [1.57, 3.06] [1.65, 2.91]

renatal care 0.75* 0.76* 0.75* 0.75*

[0.57, 0.98] [0.58, 0.99] [0.58, 0.97] [0.57, 0.99]

ale birth 0.93 0.93 0.93 0.93

[0.75, 1.17] [0.78, 1.12] [0.73, 1.19] [0.76, 1.15]

irth cohort

– Pre-famine cohort 1.06 1.17 1.14

[0.77, 1.44] [0.61, 2.24] [0.59, 2.20]

– Famine cohort 0.64* 0.20* 0.20**

[0.44, 0.93] [0.08, 0.52] [0.07, 0.58]

– 1st post-famine cohort – – –

– 2nd post-famine cohort 0.90 0.91 0.89

[0.61, 1.32] [0.32, 2.59] [0.38, 2.06]

amine severity (log-transformed) 1.19 1.04

[0.74, 1.90] [0.52, 2.09]

ohort and famine severity interaction

– Pre-famine cohort � Famine severity 0.83 0.87

[0.30, 2.35] [0.28, 2.66]

– Famine cohort � Famine severity 7.52* 7.82**

[1.43, 39.68] [1.59, 38.39]

– 1st post-famine cohort � Famine Severity – –

– 2nd post-famine cohort � Famine Severity 0.98 1.02

[0.20, 4.94] [0.27, 3.84]

IC 3683.2 3681.9 3681.1 3679.2

xponentiated coefficients; 95% confidence intervals in brackets.

p < 0.05.

* p < 0.01.

** p < 0.001.

Based on QIC, the most commonly used ‘‘independent’’ working

rrelation provided the best fit. The main results are not sensitive to the

oice of working correlation.

5

5

mdrmesimQ

inab3infatm(4cinpfaaer

ednc

T

K

S. Song / Economics and Human Biology 11 (2013) 474–487 483

. Results

.1. Model selection

Table 4 reports results from four GEE logistic regressionodels of infant mortality based on one of the ten imputed

ata set. Model 1, which only included control variables,epresented the baseline model against which all other

odels were compared. Model 2 included the faminexposure status and famine severity index. The QIC

howed that adding these two variables significantlyproved the model fit, as evidenced by a decrease in

IC from 3683.2 to 3681.9. By further including theteraction term between the famine exposure status

nd famine severity index, Model 3 evidenced an evenetter fit, as indicated by a further drop in QIC to681.1. By constraining the effect of famine severity onfant mortality among the reference (first post-mine) cohort to be zero (i.e., no famine effect among

he post-famine cohort), Model 4 achieved the bestodel fit among all four models considered

QIC = 3679.2). The comparison between Models 3 and provides an important ad hoc test of the validity of theohort size-based famine severity index. If offspringfant mortality increases with famine severity in the

ost-famine cohort who were not exposed to themine, the measure of famine severity used in the

nalysis may have captured residual regional differ-nces that are not related to famine. Fortunately, as theesults show, this is not the case here.

Based on Model 4, the best-fitting model, mother’sducation, ethnic majority status, and urban residenceecreased infant mortality, although the latter effect wasot statistically significant. Prenatal care use also de-

childbirth at an older maternal age significantly reducedinfant mortality, whereas third or higher-order birthsfaced significantly higher infant mortality than either thefirst or second birth. Male births showed a slightly lowerinfant mortality risk than female births; however, thedifference was not statistically significant. These resultsare in agreement with other studies of infant mortality inChina (Song and Burgard, 2011; Chen et al., 2007).

Model 1 does not include the imputed famineseverity index, its results do not change betweendifferent imputed data sets. In contrast, Models 2–4include the imputed famine severity index, which varybetween data sets due to the stochastic nature of themultiple imputation algorithm. Therefore, the estimatedcoefficients in these models may also vary between datasets. The results reported in Table 4 and the other ninesets of results, which are not reported here, are equallyvalid.7 The best way to utilize these model results is tocombine all of the results based on different imputeddata into a single set of results following Rubin’s rule ofcombination, as depicted by Eqs. (A.1)–(A.4) in onlinesupporting materials.

The first column of Table 5, under the heading of ‘‘mainanalysis’’, reports selected key coefficients from suchcombined results. Despite some minor numerical differ-ences, the combined results are highly consistent with theresults from Model 4 in Table 4. Columns 2–5 of Table 5report results from four sensitivity tests. Detailed discus-sion of these sensitivity test results can be found in SectionE in the online supporting materials.

able 5

ey coefficients from the main and sensitivity analyses on infant mortality in China: combined GEE logistic regression results from ten imputed data sets.

Main analysis Sensitivity analysis

Rural sample Urban sample First birth only Prenatal care user

Birth cohort

– Pre-famine cohort 1.12 1.07 1.55 0.97 0.86

[0.58, 2.16] [0.51, 2.24] [0.37, 6.54] [0.45, 2.09] [0.30, 2.46]

– Famine cohort 0.24** 0.24** 0.21 0.18* 0.20

[0.09, 0.64] [0.09, 0.67] [0.02, 2.84] [0.04, 0.74] [0.04, 1.04]

– 1st post-famine cohort – – – – –

– 2nd post-famine cohort 0.92 1.00 0.54 0.53 0.67

[0.38, 2.21] [0.39, 2.60] [0.06, 5.08] [0.19, 1.53] [0.23, 2.01]

Famine severity index (log-transformed) – – – – –

Cohort and famine severity interaction

– Pre-famine cohort � Severity 0.89 0.97 0.34 1.60 1.68

[0.31, 2.60] [0.31, 3.12] [0.01, 8.36] [0.49, 5.24] [0.28, 10.01]

– Famine cohort � Severity 5.65* 5.47* 9.32 11.83* 10.05

[1.23, 25.93] [1.17, 25.63] [0.09, 977.31] [1.43, 97.99] [0.69, 146.34]

– 1st Post-famine cohort � Severity – – – – –

– 2nd Post-famine cohort � Severity 0.97 0.84 2.69 2.00 1.54

[0.26, 3.66] [0.20, 3.54] [0.08, 91.68] [0.43, 9.21] [0.28, 8.62]

Observations 27,276 22,325 4951 14,218 13,462

Exponentiated coefficients; 95% confidence intervals in brackets.

* p < 0.05.

** p < 0.01.

reased infant mortality. Regarding biological factors,

5.2

fa

effamdipochstaarco

inrisinducofafu

Fig

fam

co

Th

sec

7

co

Mo

S. Song / Economics and Human Biology 11 (2013) 474–487484

. Main finding I: intergenerational effect of prenatal

mine exposure

Fig. 2 shows the simulated DID estimate of the faminefect on infant mortality, evaluated at each level of themine severity index, as depicted in Eq. (3). The left,iddle, and right panels of Fig. 2 represent the cohortfference between the pre-famine, famine, and secondst-famine cohort and the first post-famine cohort in theanges in infant mortality produced by an increase of onendard deviation in the famine severity index. The gray

eas surrounding each line represent the associated 95%nfidence intervals.Based on DID estimates reported in Fig. 2, an increase

famine severity produced an increase in the relativek (i.e., against the reference cohort) of children’s

fant mortality among women who were conceivedring the famine but not among women who werenceived either before or after the famine. The lack of amine effect among the pre-famine cohort deservesrther discussion. As shown in Table 1, both the pre-

famine and famine cohorts experienced early-life famineexposure but only the famine cohort experiencedprenatal famine exposure. Significant famine effectwas only observed among the famine and not amongthe pre-famine cohort, suggesting that the prenatalperiod is most critical for the development of femalereproductive function. The comparison between the firstand second post-famine cohorts provides an importantad hoc test of the constant cohort difference assumption.The absence of cohort difference between these twopost-famine cohorts suggests that this assumption is notviolated.

5.3. Main finding II: developmental plasticity, developmental

disruption, and selection effect among the famine cohort

Fig. 3 shows the simulated CCD estimate of the famineeffect on children’s infant mortality risk, evaluated at eachlevel of the famine severity index, as depicted in Eq. (4). Atany level of the famine severity index, children of neitherthe pre-famine nor the second post-famine cohortdisplayed significantly different level of infant mortalityfrom children of the first post-famine cohort (the referencecategory). In fact, there was virtually no relationshipbetween the cohort difference in children’s infant mortali-ty and famine severity among these two cohorts. In

. 2. Difference-in-differences estimates (i.e., cohort differences in the changes in infant mortality risk induced by one standard deviation’s increase in the

ine severity index) and associated 95% confidence intervals of the famine effect on the infant mortality risk of the next generation. Note: The pre-famine

hort includes those who were conceived in 1955–1957. The famine cohort includes those who were conceived between August 1958 and January 1961.

e first post-famine cohort, which was used as the reference group and thus not shown in the figure, includes those who were conceived in 1962–1964. The

ond post-famine cohort includes those who were conceived in 1965–1967.

Table A.1 in the online supporting materials reports model

mparison results for all ten imputed data sets. In all these data sets,

del 4 represents the best model.

ctycecthdws4thpcfac

sspAfadAimsKine

F

g

A

w

S. Song / Economics and Human Biology 11 (2013) 474–487 485

ontrast, the cohort difference in children’s infant mortali- between the famine and first post-famine cohorts

hanged significantly with famine severity. At the very lownd of the famine severity index, children of the famineohort showed a significantly lower infant mortality than

ose of the first post-famine cohort. This cohort differenceiminished rapidly as the famine severity increased, andas no longer statistically significant when the famine

everity measure reached a level of 1.8 (i.e., approximately4% of the original cohort were lost during the famine). Ate very high end of the famine severity index, the cohort

attern in infant mortality was reversed and the famineohort showed higher infant mortality than the first post-mine cohort, although this difference was not statisti-

ally significant.Because of the unique characteristics of the famine

everity index used in the present study, the CCD resultshown in Fig. 3 can help disentangle developmentallasticity, developmental disruption, and selection effects.t the lowest end of the famine severity index, where themine-induced cohort attrition is close to zero, by

efinition, the level of selection effect is also close to zero.s documented by earlier research, during the GLFF, the

pact of famine-induced malnutrition on daily life wasignificant even in low severity locations (Becker, 1996;ane, 1988; Brown, 2011; Yang, 2012). After all, famine-duced malnutrition had to reach a certain level before

xcess mortality and fertility reduction began to occur

(Bongaarts and Cain, 1981). With the absence of a selectioneffect, the significantly reduced infant mortality riskamong children of the famine cohort can only be attributedto the effect of developmental plasticity. The sensitivitytest results reported in Table 5 show that such a cohortpattern in infant mortality at the very low end of thefamine severity index was robust across different sub-samples. As the level of famine severity index increases,the following three events may occur simultaneously: theplasticity effect diminishes, the disruption effect increases,and so does the selection effect. At the very high end of thefamine severity index, the estimated cohort pattern ininfant mortality mainly depends on the relative strength ofthe disruption effect, plasticity effect, and selection effect.The fact that the famine cohort had a higher infantmortality than the post-famine cohort suggests that thedisruption effect played a more prominent role than boththe plasticity and selection effects in influencing the infantmortality of the next generation in these areas. The lack ofcohort difference between the first and second post-famine cohorts suggests that the zero cohort differenceassumption is not violated.

6. Conclusion

Using the 1959–1961 Chinese GLFF as a naturalexperiment, this study identified two different kinds ofintergenerational effects of mothers’ prenatal famine

ig. 3. Conditional cohort difference estimates and associated 95% confidence intervals of the famine effect on the infant mortality risk of the next

eneration. Note: The pre-famine cohort includes those who were conceived in 1955–1957. The famine cohort includes those who were conceived between

ugust 1958 and January 1961. The first post-famine cohort, which was used as the reference group and thus not shown in the figure, includes those who

ere conceived in 1962–1964. The second post-famine cohort includes those who were conceived in 1965–1967.

exchexrisfacodidetain

mimantimnivireobFomsta

suChnaIt

toinwGLdofasemprfacashcomscexthim

incoreAdstualithcomfuonstu

S. Song / Economics and Human Biology 11 (2013) 474–487486

posure on the risk of infant mortality among theirildren. In regions of low famine severity, prenatal famineposure significantly reduced children’s infant mortalityk, whereas in regions of high famine severity, prenatal

mine exposure increased such risk. These results arensistent with the DOHaD prediction regarding thestinction between the developmental plasticity andvelopmental disruption effects of prenatal environmen-l insult and the crucial role of the severity of such insult determining the final outcomes.

The strengths of this study include its natural experi-ental design, the DID and CCD methods, and the multipleputation-based famine severity index. The unique sociald political atmosphere of Chinese society around thee of the famine, including the collectivization, commu-

zation, and heightened social control, made the GLFF aable natural experiment. The presence of significantgional variations in famine severity makes it possible totain a more fine-grained estimate of the famine effect.rmulating the famine severity index construction as aissing data problem leads directly to an eleganttistical solution, multiple imputation.The natural experimental nature of the study design

ggests that, although this study was based on theinese population, the insights gained are causal inture and can be generalized to other social contexts.may be argued that the Chinese GLFF is unique due

its unparalleled severity, which is unlikely to occur even the most impoverished parts of the modernorld, and thus the lessons learned from the ChineseFF are of little practical relevance. This argumentes not stand up to scrutiny because it ignores the

ct that, as Table 3 shows, whereas very high famineverity was observed in some provinces, very low oredium famine severity was also observed in otherovinces. It is the presence of such a wide spectrum ofmine severity that makes the GLFF a uniquely suitable testse to reveal the otherwise deeply hidden causal relation-ips. The causal knowledge gained from such extremenditions can then be used to inform and improve policyaking and clinical practice in the same way that theientific knowledge obtained from physics or chemistryperiments, which usually represent extreme conditionsat are not found in ordinary conditions, can be used toprove our life.One limitation of this study is that the infant mortality

formation used was based on retrospective informationllected in a cross-sectional survey. Recall bias and under-porting are known to be associated with this type of data.ditionally, the probability of being selected into thedy sample was dependent on whether a woman was

ve at the time of the interview. It is reassuring to noteat the level of infant mortality reported here ismparable to that based on other data sources (Develop-ent Data Group, 2011). The results reported here must berther verified and corroborated by future studies based

independent data sources (preferably based on differentdy designs) for a more definitive conclusion.

Appendix A. Supplementary data

Supplementary data associated with this article can befound, in the online version, at doi:http://dx.doi.org/10.1016/j.ehb.2013.08.001.

References

Almond, D., Currie, J., 2011. Killing me softly: the fetal origins hypothesis.The Journal of Economic Perspectives 25, 153–172.

Almond, D., Edlund, L., Li, H., Zhang, J., 2010. Long-term effects of early-lifedevelopment: evidence from the 1959–1961 China famine. NBER-EASE 19, 321–345.

Almond, D., Mazumder, B., 2011. Health capital and the prenatal envi-ronment: the effect of Ramadan observance during pregnancy. Amer-ican Economic Journal: Applied Economics 3, 56–85.

Ashton, B., Hill, K., Piazza, A., Zeitz, R., 1984. Famine in China, 1958–61.Population and Development Review 10, 613–645.

Barker, D.J.P., 1992. The fetal origins of diseases of old age. EuropeanJournal of Clinical Nutrition 46, S3–9.

Barker, D.J.P., 1995. Fetal origins of coronary heart disease. British MedicalJournal 311, 171.

Barker, D.J.P., Osmond, C., 1986. Infant mortality, childhood nutrition, andischaemic heart disease in England and Wales. The Lancet 327,1077–1081.

Bateson, P., Barker, D., Clutton-Brock, T., Deb, D., D’Udine, B., Foley, R.,Gluckman, P., Godfrey, K., Kirkwood, T., Lahr, M., et al., 2004.Developmental plasticity and human health. Nature 430, 419–421.

Becker, J., 1996. Hungry Ghosts: China’s Secret Famine. The Free Press,New York.

Blackwell, M., Honaker, J., King, G., 2011. Multiple overimputation: aunified approach to measurement error and missing data. Workingpaper.

Bongaarts, J., Cain, M., 1981. Demographic response to famine. In: Cahill,K. (Ed.), Famine. Orbis, New York, pp. 44–59.

Brown, J., 2011. Great leap city: surviving the famine in Tianjin. In:Manning, K., Wemheuer, F. (Eds.), Eating Bitterness: New Perspectiveson China’s Great Leap Forward and Famine. UBC Press, Vancouver,Canada, (Chapter 9), pp. 226–250.

Chai, J., 1992. Consumption and living standards in China. The ChinaQuarterly 131, 721–749.

Chen, J., Xie, Z., Liu, H., 2007. Son preference, use of maternal health care,and infant mortality in rural China, 1989–2000. Population Studies61, 161–183.

Chen, Y., Zhou, L., 2007. The long-term health and economic consequencesof the 1959–1961 famine in China. Journal of Health Economics 26,659–681.

Cochrane, J., Chen, H., Conigrave, K., Hao, W., 2003. Alcohol use in China.Alcohol and Alcoholism 38, 537–542.

Cole, S.R., Chu, H., Greenland, S., 2006. Multiple-imputation for measure-ment-error correction. International Journal of Epidemiology 35,1074–1081.

Davies, R.W., Wheatcroft, S.G., 2004. The Years of Hunger: Soviet Agricul-ture, 1931–1933, vol. 5. Palgrave Macmillan, London.

Desai, S., Alva, S., 1998. Maternal education and child health: is there astrong causal relationship? Demography 35, 71–81.

Development Data Group, The World Bank, 2011. 2011 World Develop-ment Indicators. The World Bank.

Dikotter, F., 2010. Mao’s Great Famine: The History of China’s MostDevastating Catastrophe, 1958–1962. Walker & Co.

Doyle, O., Harmon, C.P., Heckman, J.J., Tremblay, R.E., 2009. Investing inearly human development: timing and economic efficiency. Econom-ics & Human Biology 7, 1–6.

Frisbie, W., Song, S., Powers, D., Street, J., 2004. The increasing racialdisparity in infant mortality: respiratory distress syndrome and othercauses. Demography 41, 773–800.

Frisbie, W.P., 2005. Infant mortality. In: Poston, D.L., Micklin, M. (Eds.),Handbook of Population. Kluwer Academic/Plenum, New York, pp.251–282.

Fung, W., Ha, W., 2010. Intergenerational effects of the 1959–61 Chinafamine. In: Fuentes-Nieva, S. (Ed.), Risk, Shocks and Human Develop-ment: On the Brink. Palgrave-Macmillan, London, UK, pp. 222–254.

Gluckman, P.D., Hanson, M.A., 2006. The conceptual basis for the devel-opmental origins of health and disease. In: Gluckman, P., Hanson, M.(Eds.), Developmental Origins of Health and Disease. CambridgeUniversity Press, Cambridge, pp. 33–50.

G

G

H

H

H

H

H

Jo

K

K

K

L

L

L

L

L

L

L

L

M

M

M

O

S. Song / Economics and Human Biology 11 (2013) 474–487 487

luckman, P.D., Hanson, M.A., Spencer, H.G., Bateson, P., 2005. Environ-mental influences during development and their later consequencesfor health and disease: implications for the interpretation of empiricalstudies. Proceedings of the Royal Society B: Biological Sciences 272,671.

ortmaker, S., Wise, P., 1997. The first injustice: socioeconomic dispar-ities, health services technology, and infant mortality. Annual Reviewof Sociology 23, 147–170.

onaker, J., King, G., 2010. What to do about missing values in time-seriescross-section data. American Journal of Political Science 54,561–581.

onaker, J., King, G., Blackwell, M., 2011. Amelia II: a program for missingdata. Journal of Statistical Software 45, 1–47.

uang, C., Li, Z., Venkat Narayan, K., Williamson, D., Martorell, R., 2010.Bigger babies born to women survivors of the 1959–1961 Chinesefamine: a puzzle due to survival selection? Journal of DevelopmentalOrigins of Health and Disease 1, 412–418.

uang, C., Li, Z., Wang, M., Martorell, R., 2010. Early life exposure to the1959–1961 Chinese famine has long-term health consequences.Journal of Nutrition 140, 1874.

uxley, R., Neil, A., Collins, R., 2002. Unravelling the fetal origins hypoth-esis: is there really an inverse association between birthweight andsubsequent blood pressure? The Lancet 360, 659–665.

seph, K., Kramer, M., et al., 1996. Review of the evidence on fetal andearly childhood antecedents of adult chronic disease. EpidemiologicReviews 18, 158–174.

ane, P., 1988. Famine in China 1959–61. Demographic and Social Impli-cations. Macmillan Press, London, UK.

im, S., Deng, Q., Fleisher, B., Li, S., 2012. The China great leap forwardfamine: the lasting impact of mothers’ fetal malnutrition on theiroffspring. Working paper.

ing, G., Tomz, M., Wittenberg, J., 2000. Making the most of statisticalanalyses: improving interpretation and presentation. American Jour-nal of Political Science 44, 347–361.

ardy, N., 1984. Consumption and living standards in China, 1978–83. TheChina Quarterly 100, 849–865.

avely, W., Lee, J., Feng, W., 1990. Chinese demography: the state of thefield. The Journal of Asian Studies 49, 807–834.

i, Y., Jaddoe, V., Qi, L., He, Y., Lai, J., Wang, J., Zhang, J., Hu, Y., Ding, E., Yang,X., et al., 2011. Exposure to the Chinese famine in early life and therisk of hypertension in adulthood. Journal of Hypertension 29 .

i, Y., Jaddoe, V., Qi, L., He, Y., Wang, D., Lai, J., Zhang, J., Fu, P., Yang, X., Hu,F., 2011. Exposure to the Chinese famine in early life and the risk ofmetabolic syndrome in adulthood. Diabetes Care 34, 1014.

in, J.Y., Yang, D.T., 2000. Food availability, entitlements and the Chinesefamine of 1959–61. The Economic Journal 110, 136–158.

ucas, A., Fewtrell, M., Cole, T., 1999. Fetal origins of adult disease – thehypothesis revisited. BMJ 319, 245.

umey, L.H., Stein, A.D., 1997. In utero exposure to famine and subsequentfertility: the Dutch famine birth cohort study. American Journal ofPublic Health 87, 1962–1966.

uo, Z., Mu, R., Zhang, X., 2006. Famine and overweight in China. AppliedEconomic Perspectives and Policy 28, 296.

eng, X., Qian, N., 2009. The long term consequences of famine onsurvivors: evidence from a unique natural experiment using China’sgreat famine. NBER working paper.

innesota Population Center, 2010. Integrated Public Use MicrodataSeries, International: Version 6.0 [Machine-readable database]. Uni-versity of Minnesota, Minneapolis.

u, R., Zhang, X., 2011. Why does the Great Chinese Famine affect themale and female survivors differently? Mortality selection versus sonpreference. Economics & Human Biology 9, 92–105.

smani, S., Sen, A., 2003. The hidden penalties of gender inequality: fetalorigins of ill-health. Economics & Human Biology 1, 105–121.

Painter, R., Roseboom, T., Bleker, O., 2005. Prenatal exposure to the Dutchfamine and disease in later life: an overview. Reproductive Toxicology20, 345–352.

Painter, R.C., Westendorp, R.G.J., de Rooij, S.R., Osmond, C., Barker, D.J.P.,Roseboom, T.J., 2008. Increased reproductive success of womenafter prenatal undernutrition. Human Reproduction 23,2591–2595.

Pan, W., 2001. Akaike’s information criterion in generalized estimatingequations. Biometrics 57, 120–125.

Paneth, N., Susser, M., 1995. Early origin of coronary heart disease (the‘‘barker hypothesis’’). BMJ 310, 411–412.

Peng, X., 1987. Demographic consequences of the great leap forward inChina’s provinces. Population and Development Review 13, 639–670.

Rasmussen, K.M., 2001. The ‘‘fetal origins’’ hypothesis: challenges andopportunities for maternal and child nutrition. Annual Review ofNutrition 21, 73–95.

Ravelli, A., van der Meulen, J., Michels, R., Osmond, C., Barker, D., Hales, C.,Bleker, O., 1998. Glucose tolerance in adults after prenatal exposure tofamine. The Lancet 351, 173–177.

Razzaque, A., Alam, N., Wai, L., Foster, A., 1990. Sustained effects of the1974–5 famine on infant and child mortality in a rural area ofbangladesh. Population Studies 44, 145–154s.

Roseboom, T., Van der Meulen, J., Ravelli, A., Osmond, C., Barker, D., Bleker,O., 2001. Effects of prenatal exposure to the Dutch famine on adultdisease in later life: an overview. Molecular and Cellular Endocrinol-ogy 185, 93–98.

Rubin, D., 1987. Multiple Imputation for Nonresponse in Surveys. JohnWiley & Sons, New York.

Song, S., 2010. Mortality consequences of the 1959–1961 great leapforward famine in China: debilitation, selection, and mortality cross-overs. Social Science & Medicine 71, 551–558.

Song, S., 2013. Assessing the impact of in utero exposure to famine onfecundity: evidence from the 1959-61 famine in China, Population Studies67.

Song, S., Burgard, S., 2011. Dynamics of inequality: mother’s educationand infant mortality in China, 1970–2001. Journal of Health and SocialBehavior 52, 349–364.

Song, S., Wang, W., Hu, P., 2009. Famine, death, and madness: Schizo-phrenia in early adulthood after prenatal exposure to the Chinesegreat leap forward famine. Social Science & Medicine 68, 1315–1321.

Sotomayor, O., 2013. Fetal and infant origins of diabetes and ill health:evidence from Puerto Rico’s 1928 and 1932 hurricanes. Economics &Human Biology 11, 281–293.

Stearns, S.C., 1992. The evolution of life histories. Oxford University Press,Oxford.

Tu, Y., West, R., Ellison, G., Gilthorpe, M., 2005. Why evidence for the fetalorigins of adult disease might be a statistical artifact: the ‘‘reversalparadox’’ for the relation between birth weight and blood pressure inlater life. American journal of Epidemiology 161, 27.

Wise, P., 2003. The anatomy of a disparity in infant mortality. AnnualReview of Public Health 24, 341–362.

Wu, X., Treiman, D.J., 2004. The household registration system and socialstratification in China: 1955–1996. Demography 41, 363–384.

Yang, G., Fan, L., Tan, J., Qi, G., Zhang, Y., Samet, J., Taylor, C., Becker, K., Xu,J., 1999. Smoking in China. JAMA: The Journal of the American MedicalAssociation 282, 1247–1253.

Yang, J., 2012. Tombstone: the great Chinese famine, 1958–1962. Farrar,Straus, and Giroux, New York.

Yao, S., 1999. A note on the causal factors of China’s famine in 1959–1961.Journal of Political Economy 107, 1365–1369.

Zeger, S.L., Liang, K.Y., 1986. Longitudinal data analysis for discrete andcontinuous outcomes. Biometrics 42, 121–130.

Zhang, G., Zhao, Z., 2006. Reexamining China’s fertility puzzle: datacollection and quality over the last two decades. Population andDevelopment Review 32, 293–321.