The Origins of Savings Behavior
Henrik Cronqvist China Europe International Business School
Stephan Siegel University of Washington
Analyzing the savings behavior of a large sample of identical and fraternal twins, we find that genetic differences explain about 33 percent of the variation in savings propensities across individuals. Individuals are born with a persistent genetic predisposition to a specific savings behavior. Parenting contributes to the variation in savings rates among younger individuals, but its effect decays over time. The environment when growing up ðe.g., parents’ wealthÞ moderates genetic effects. Finally, savings behavior is genetically correlated with income growth, smoking, and obesity, suggesting that the genetic component of savings behavior reflects genetic variation in time preferences or self-control.
Economics is a branch of biology broadly interpreted. ðAlfred Marshall, 1920Þ I. Introduction There is enormous variation across individuals with respect to wealth accumulated at retirement age, even among those with very similar lifeWe are thankful for comments from three anonymous referees, the editor ðPhilip RenyÞ, seminar participants at the American Finance Association, Arizona State University, California State University, Fullerton, China Europe International Business School, Claremont Graduate University, 14th Congress of the International Society of Twin Studies, Duke University, Financial Management Association, Hong Kong University of Science and Technology, Johns Hopkins University, National Bureau of Economic Research Behavioral Econom[ Journal of Political Economy, 2015, vol. 123, no. 1] © 2015 by The University of Chicago. All rights reserved. 0022-3808/2015/12301-0005$10.00
123
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
124
journal of political economy
time incomes. Economists have found that this dispersion cannot be easily explained by asset allocation choices or socioeconomic characteristics ðe.g., Venti and Wise 1998Þ. Instead, an individual’s savings propensity, that is, the choice by an individual to consume or save earlier in life, appears to be a more important determinant of cross-sectional variation in wealth accumulation.1 This evidence raises several fundamental questions: Why do individuals differ with respect to their savings behavior? Are we genetically predisposed to a particular savings behavior? Is it formed by parents instilling their preferences into their children? Or is it the result of individual-specific life experiences? In this paper, we address these questions. Existing research has employed a variety of empirical approaches to analyzing savings behavior.2 The novel approach of this paper is to empirically decompose the variation in savings propensity across individuals into genetic and environmental components and to also examine geneenvironment interactions, that is, whether specific factors moderate genetic predispositions to a particular behavior. Thus, the approach to analyzing savings behavior employed in this paper blends economics and biology, an intersection of research fields considered to be important by several economists ðe.g., Becker 1976; Hirshleifer 1977; Knudsen et al. 2006Þ but one that has been relatively underresearched to date, particularly on the empirical side. In standard life cycle or precautionary savings models, variation in savings behavior across individuals is explained by heterogeneity in time and ics Working Group, Singapore International Conference on Finance, University of Miami Behavioral Finance Conference, University of Southern California, University of Western Ontario, and valuable discussions with Peter Bossaerts, Alon Brav, John Campbell, Christopher Carroll, Darrell Duffie, Daniel Hamermesh, Jarrad Harford, Larry Harris, Cam Harvey, Greg Hess, David Hirshleifer, Chris Hrdlicka, Andrew Karolyi, Camelia Kuhnen, Owen Lamont, Andy Lo, Robert Plomin, Ed Rice, Jay Ritter, Tano Santos, Mark Seasholes, Hersh Shefrin, Andy Siegel, Robert Shiller, Richard Thaler, James Weston, Frank Yu, and Paul Zak. We thank Harvey Cheong, Florian Mu¨nkel, and Lew Thorson for excellent research assistance; Jack Goldberg ðUniversity of Washington Twin RegistryÞ and Nancy Segal ðTwin Studies Center at California State University, FullertonÞ for advice related to twins studies; and Kara Ralston for editorial assistance. We acknowledge research funding from the Financial Economics Institute and the Lowe Institute of Political Economy at Claremont McKenna College as well as the Global Business Center and the CFO Forum at the University of Washington. This project was pursued in part when Cronqvist was Olof Stenhammar Visiting Professor at the Swedish Institute for Financial Research, which he thanks for its support, and while Siegel was visiting W. P. Carey School of Business at Arizona State University, which he thanks for its hospitality. Statistics Sweden and the Swedish Twin Registry ðSTRÞ provided the data for this study. STR is supported by grants from the Swedish Research Council, the Ministry of Higher Education, AstraZeneca, and the National Institute of Health ðgrants AG08724, DK066134, and CA085739Þ. Any errors or omissions are our own. 1 Already Friedman ð1953Þ concluded that “a large part of the existing inequality of wealth can be regarded as produced by men to satisfy their tastes” ð290Þ. 2 We refer to Browning and Lusardi ð1996Þ for an extensive review of micro-level research on savings behavior.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
origins of savings behavior
125
risk preferences ði.e., those with low personal discount rates and high risk aversion save moreÞ as well as differences in economic conditions ðe.g., income volatilityÞ. Theoretical work suggests that individual preferences are the outcome of natural selection ðe.g., Rogers 1994; Robson 2001; Brennan and Lo 2011Þ, suggesting that preferences, and therefore savings propensities, are at least partially genetic. On the empirical side, there is emerging evidence that preferences are partly genetic ðe.g., Eisenberg et al. ½2007 and Carpenter, Garcia, and Lum ½2011 for time preferences and Kuhnen and Chiao ½2009, Barnea, Cronqvist, and Siegel ½2010, and Cesarini et al. ½2010 for risk preferencesÞ. Other economists have questioned some of the assumptions embedded in standard life cycle models, for example, the self-control and willpower to execute a savings plan ðe.g., Thaler 1994; Benartzi and Thaler 2007Þ. Several studies have indeed found that nonstandard models and “behavioral factors” also explain variation in savings behavior ðe.g., Bernheim, Skinner, and Weinberg 2001; Lusardi 2001; Madrian and Shea 2001Þ. As a result, genetic variation in savings behavior may also reflect self-control or other factors being partly genetic. Considering social, rather than genetic, transmission of preferences and behavior, other economists have emphasized parents’ instilling behaviors into their children ðe.g., Cavalli-Sforza and Feldman 1981; Bisin and Verdier 2008Þ. Anecdotal evidence certainly suggests that some parents exert costly effort to teach their children a particular savings behavior, for example, by providing a weekly allowance, opening a savings account, and so on. While there exists work on state-sponsored financial education programs ðe.g., Bernheim, Garrett, and Maki 2001Þ, there exists little evidence on parenting and savings behavior. Research by Charles and Hurst ð2003Þ suggests that there are significant parent-child similarities in savings behavior. While similarity in income explains about half of the age-adjusted elasticity of child wealth with respect to parental wealth, Charles and Hurst hypothesize that similar savings propensities among parents and their children are another explanation. Their study is important as it suggests that parents pass on their savings propensities to their children, but it does not uncover the extent to which this similarity is genetic versus the result of social transmission of behavior from parents to their children.3 This paper attempts to fill this gap. 3 A large literature in economics has studied parent-child similarities in domains other than savings behavior. Borjas ð1992Þ and Solon ð1992Þ find a positive and significant intergenerational correlation in income. Chiteji and Stafford ð1999Þ examine asset allocations and find that children, as young adults, are more likely to own stocks when their parents owned such assets. Bowles and Gintis ð2002Þ report that socioeconomic status is persistent across generations. Mulligan ð1999Þ, Black, Devereux, and Salvanes ð2005Þ, and Gu¨ell, Rodrı´guez Mora, and Telmer ð2013Þ provide evidence of significant intergenerational transmission with respect to education.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
126
journal of political economy
There are several reasons why it is important for economists to examine the origins of savings behavior. First, many countries are transitioning from pensions and defined-benefit plans to defined-contribution plans such as 401ðkÞ and private social security accounts, causing individual workers to become responsible for their own savings ðe.g., Cronqvist and Thaler 2004; Poterba, Venti, and Wise 2007Þ. With increased autonomy with respect to savings, it becomes increasingly important to understand why some individuals choose to save while others do not. Current empirical models provide only a partial understanding of what determines individuals’ savings propensities. Second, our work has implications for the sensitivity of savings behavior to public policy intervention and therefore for policy design. As Bernheim ð2009, 12Þ recently concluded, “The discovery and analysis of a patience marker could shed light on the extent to which correlations between the wealth of parents and children reflect genetic predispositions rather than environmental factors that are presumably more amenable to policy interventions.” Analyzing data on identical and fraternal twins and their savings behavior, we are able to empirically decompose the variation in individuals’ propensity to save into genetic and environmental components. Our data on twins are from the Swedish Twin Registry, the world’s largest research database of twins, and matched with detailed data from the Swedish Tax Agency and other register-based data sets. Following Venti and Wise ð1998Þ, we control for effects of socioeconomic conditions and characteristics as well as asset allocation choices and estimate an adjusted savings rate ði.e., the residual from a regression with a large set of controlsÞ for each individual. The decomposition of this savings measure into genetic and environmental factors rests on an intuitive insight: Identical twins share 100 percent of their genes, while the average proportion of shared genes is only 50 percent for fraternal twins; so if identical twins have more similar savings behavior than fraternal twins, then there is evidence that the propensity to save, at least partly, originates from an individual’s genetic composition.4 Savings rates are indeed much more correlated among identical than fraternal twins in our data, with correlations of .33 versus .16. Our conclusion is confirmed by estimating a random-effects model with three effects ðone genetic, one parental or common, and one individual-specificÞ and a covariance structure imposed by genetic theory, an approach that is standard in contemporaneous quantitative behavioral genetics research ðe.g., Neale and Maes 2004Þ. Our empirical analysis produces several results. First, we find that genetic differences explain about 33 percent of the variation in savings behavior across individuals. That is, each individual is born with a ge-
4 Several studies in economics have previously examined data on twins in other contexts, e.g., Taubman ð1976Þ, Behrman and Taubman ð1989Þ, and Ashenfelter and Krueger ð1994Þ.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
origins of savings behavior
127
netic predisposition to a specific savings behavior, an effect that is found not to disappear later in life. Our finding is robust to a number of alternative savings rate measures and to robustness checks of the most important assumptions underlying the estimated model. In an extension, we infer individual consumption growth rates and estimate Euler equations at the individual level. While the individual-specific time preference parameter estimates are noisy, we still find evidence of a significant genetic component. Second, we find that parenting contributes to the variation in savings rates among younger individuals in our sample, but its effect has decayed significantly for middle-aged and older individuals. We also find that parenting explains more of the variation in savings behavior when parenting resources are likely to be less scarce. Third, we examine gene-environment interaction and find that the family environment when growing up ðas proxied by the wealth of the parentsÞ moderates genetic predispositions to a particular savings behavior, evidence that is consistent with theories that genetic effects are predicted to be stronger in more supportive environments. Finally, we examine why savings behavior is genetic and find that savings is genetically correlated with income growth, smoking, and obesity, suggesting that the genetic component of savings behavior reflects genetic differences across individuals with respect to time preferences or self-control. The paper is organized as follows. Section II provides an overview of related research. Section III describes our data sources, defines the savings measures and other variables, and reports summary statistics. Section IV presents the empirical approach and explains the identification assumptions. Section V reports our results and several robustness checks. Section VI reports further evidence and extensions. Section VII discusses important implications of the reported empirical evidence. Section VIII presents conclusions and raises some questions for future research. II. Overview of Related Research On the basis of theory as well as existing empirical evidence, there are reasons to hypothesize that an individual’s savings propensity may be explained by both genetic factors and parent-child socialization. In this section, we provide an overview of related research that motivates our analysis. A.
Genes and Savings Behavior
1.
Preferences
Since the seminal work by Modigliani and Brumberg ð1954Þ, the standard life cycle savings model has been a framework through which economists have studied individuals’ savings and consumption choices. This
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
128
journal of political economy
model has been extended to incorporate precautionary savings and bequest motives.5 Lusardi ð1998Þ provides empirical evidence that savings behavior is in part explained by individual heterogeneity in time, risk, and bequest preferences. This raises the question of whether such preferences are genetic. Rogers ð1994Þ, Robson ð1996Þ, Netzer ð2009Þ, Robson and Samuelson ð2009Þ, and Brennan and Lo ð2011Þ theorize that individual preferences are the outcome of natural selection, implying that preferences are partially genetic.6 A large number of empirical studies have found significant heterogeneity in time preferences across individuals ðe.g., Lawrance 1991; Barsky et al. 1997; Warner and Pleeter 2001; Harrison, Lau, and Williams 2002Þ, but little is known about the sources of such variation. In a gene candidate study, Eisenberg et al. ð2007Þ identify specific genes related to variation in personal discount rates across individuals. While this evidence suggests that time preferences have a genetic component, gene candidate studies cannot quantify the overall importance of genetic variation in explaining the observed heterogeneity. Several recent studies also report that heterogeneity in individual risk aversion is genetic ðe.g., Cesarini et al. 2009, 2010Þ. For example, Barnea et al. ð2010Þ examine individuals’ financial portfolios and find that about one-third of the crosssectional variation in the share in equities is explained by a genetic effect. Moreover, gene candidate studies have found that those with certain genes find risky behaviors more rewarding ðe.g., Dreber et al. 2009; Kuhnen and Chiao 2009; Zhong et al. 2009Þ. The aforementioned studies suggest a relation between, on the one hand, genes and, on the other hand, time and risk preferences. This preexisting research motivates our hypothesis that genetic factors partly explain variation in savings propensities across individuals. 2.
Behavioral Factors
Some economists, such as Thaler ð1994Þ and Benartzi and Thaler ð2007Þ, have questioned some of the assumptions embedded in standard models, for example, the self-control and willpower to execute a savings plan. Several studies have indeed found that nonstandard models and “behavioral factors” also explain variation in savings behavior ðe.g., Thaler and Shefrin 1981; O’Donoghue and Rabin 1999; Lusardi 2001; Ameriks, 5 For precautionary savings models, see, e.g., Skinner ð1988Þ, Kimball ð1990Þ, Carroll ð1992Þ, and Hubbard, Skinner, and Zeldes ð1995Þ. For models with bequest motives, see, e.g., Becker and Tomes ð1976Þ. 6 For a more extensive overview of research at the intersection of neurobiology, genetics, and economics, we refer to Camerer, Loewenstein, and Prelec ð2005Þ and Benjamin et al. ð2008Þ.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
origins of savings behavior
129
Caplin, and Leahy 2003; Duflo et al. 2006Þ. A large literature in broader social science research reports evidence of variation across individuals with respect to self-control. In the seminal “marshmallow experiment,” psychology researcher Walter Mischel found significant heterogeneity across preschool children in the propensity to forgo an immediate reward ði.e., consuming a marshmallowÞ for a larger, but delayed, reward ði.e., consuming two marshmallows when the experimenter returned to the roomÞ. Most importantly, studies of these children in adolescence showed that the self-control behavior estimated using the marshmallow experiment explained a series of outcomes later in life, including SAT scores, educational attainment, and social competence ðe.g., Mischel, Shoda, and Rodriguez 1989Þ. Several behaviors related to self-control, for example, smoking, obesity, and attention-deficit hyperactivity disorder, have a significant genetic component ðe.g., Barkley 1997Þ. These studies suggest a relation between genes, behavioral factors such as self-control, and individual savings propensities. B.
Parent-Child Socialization and Savings Behavior
Social transmission, as opposed to genetic transmission, is another potential explanation for parent-child similarities ðBisin and Verdier 2008Þ.7 For example, Becker and Mulligan ð1997Þ argue that “parents often spend resources on teaching their children to better plan for the future, resources that affect the children’s discount rates” ð736Þ. Bisin and Verdier ð2001Þ provide a theoretical model of parent-child socialization. A common assumption in such models is that children are born without defined preferences and that parents transmit their own preferences to their children through parenting. Altruism makes parents exert costly effort to socialize their children, but this altruism is paternalistic in that parents prefer to socialize their children to their own preferences. Models with these assumptions have been used to explain parent-child social transmission of, for example, religion and labor supply preferences ðe.g., Bisin, Topa, and Verdier 2004Þ. On the basis of this work, we hypothesize that models of social transmission from parents to their children may extend also to savings behavior.
7 To be precise, the socialization we study in this paper is “direct vertical socialization,” i.e., transmission from parents to their children ðe.g., Cavalli-Sforza and Feldman 1981; Boyd and Richerson 1985; Richerson and Boyd 2005Þ. We do not study “oblique socialization,” i.e., socialization outside the family, which takes place in society at large through learning from others in the population.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
130
journal of political economy
III. Data A.
Data Sources
The data set we use in this paper was constructed by matching a large number of identical and fraternal twins from the Swedish Twin Registry ðSTRÞ, the world’s largest twin registry, with annual data from these twins’ tax filings. Twins are registered at birth in Sweden, but the STR also collects additional data through in-depth interviews for use by researchers.8 Importantly, we have data on the zygosity of each twin pair: “monozygotic,” or identical, twins are genetically identical, while “dizygotic,” or fraternal, twins are genetically different.9 We also have data on, for example, smoking behavior and obesity, which we use in some of our analysis. Data from the twins’ tax filings are available because, until 2006, Swedish taxpayers were subject to a wealth tax, and banks, brokerage firms, and other financial institutions were required to report to the tax authorities annual data about assets and liabilities for each individual taxpayer. All data are managed by Statistics Sweden and may be used by researchers. Since Swedish tax law requires taxes to be filed individually and does not allow for joint filings, we observe all financial data at the individual level.10 For twins who are married, we also observe the corresponding data for the spouse. While we focus on savings decisions at the individual level, we also discuss household-level results. The combination of our data on income, assets, and debt provides an unusually detailed and complete characterization of the financial decisions of Swedish households. For a detailed discussion of the data from tax filings, see Calvet, Campbell, and Sodini ð2007Þ. B. Sample Selection After merging cross-sectional data on twins from the STR with annual tax data between 2003 and 2007, we compile our sample in several steps. First, the methodology we use requires us to drop twin pairs with incom8 The STR data used in this study were obtained through the Screening Across Lifespan Twin study for twins born between 1886 and 1958 and through the Swedish Twin Studies of Adults: Genes and Environment study for those born between 1959 and 1985. Participation rates were 60–70 percent. 9 Zygosity is based on questions about intrapair similarities in childhood. One of the survey questions is, Were you and your twin partner during childhood “as alike as two peas in a pod” or were you “no more alike than siblings in general” with regard to appearance? This method has been validated with DNA as having 98 percent or higher accuracy. For twin pairs for which DNA sampling has been conducted, zygosity status based on DNA analysis is used. 10 If accounts are owned jointly, financial institutions report the proportion owned by each individual. Similarly, values of real estate are reported proportional to ownership.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
origins of savings behavior
131
plete data. To focus our study on preretirement savings behavior, we select only individuals between 20 and 65 years of age ðat the end of our sample period in 2007Þ. To reduce the impact of significant changes in net worth unrelated to savings behavior, we exclude twins whose marital status changed during the sample period. We also require that an individual’s average annual net worth is positive, that the individual owns some financial or real assets other than the primary residence, and that the average annual disposable income is more than SEK 10,000 ðapproximately $1,400 at an average exchange rate of about SEK 7.00 per US dollarÞ. Finally, we drop individuals in the 1 percent tails of the savings rate distribution to reduce the potential impact of extreme individual savings rate values.11 Our final data set consists of 14,930 twins. Panel A of table 1 reports the number of twins by zygosity and gender; 30 percent are identical, while the rest are fraternal. Opposite-sex twins are most common ð38 percentÞ, while identical male twins are the least common ð13 percentÞ. The frequency of different types of twins in the table is consistent with what would be expected from populations of twins. Panel B of table 1 reports summary statistics for socioeconomic characteristics as well as asset allocation choices for the individuals in our sample. Unless otherwise indicated, the variables are measured at the individual, as opposed to the household, level. Detailed definitions for all the variables are available in Appendix table A1. While identical and fraternal twins appear similar with respect to characteristics such as age, marital status, health, income, net worth, and asset allocation, we observe substantial cross-sectional variation. Differences in the socioeconomic characteristics in the table, as well as asset allocation choices, may affect individual savings behavior. We return to this question after introducing our savings measure. C. Measuring Savings Behavior Our data set does not contain a predefined measure of savings behavior. As Dynan, Skinner, and Zeldes ð2004Þ point out, there are several empirical approaches to measuring individuals’ savings rates. Conceptually, we want to characterize an individual’s savings behavior by the relative amount of income that is not consumed but saved for future consumption or bequest. In practice, we measure savings as the change in an individual’s net worth between the end of 2003 and the end of 2007, 11 We have checked that our results are robust to alternative approaches for dealing with extreme valuesof thesavingsrate, s, including ðiÞ winsorizing pffiffiffiffiffi thedata atthe 1 percent tail values, ðiiÞ using a square rootptransformation, signðsÞ jsj, and ðiiiÞ using the hyperbolic sine ffiffiffiffiffiffiffiffiffiffiffiffiffi transformation, lnðs 1 s 2 1 1Þ.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
132
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
Age Married Divorced Widowed Number of siblings Children in household indicator Number of children in household ð4 - year averageÞ Poor health ð4 - year averageÞ Unemployed ð4 - year averageÞ
Variable
Number of twins ðN Þ Fraction ð%Þ 2,494 17
Female 4,482 30
Total
2,142 14
Same Sex: Male
2,622 18
Same Sex: Female
48.84 .51 .10 .02 1.38 .42 .86 .19 .12
14,930 14,930 14,930
Mean
.00 .00 .00
52.00 1.00 .00 .00 1.00 .00
Median
1.17 .32 .27
12.52 .50 .30 .12 1.33 .49
Standard Deviation
Identical Twins
.71 .21 .10
53.20 .58 .11 .02 1.51 .38
Mean
.00 .00 .00
56.00 1.00 .00 .00 1.00 .00
Median
1.07 .35 .26
9.98 .49 .32 .15 1.45 .48
Standard Deviation
Fraternal Twins
5,684 38
Opposite Sex
Fraternal Twins
B. Socioeconomic Characteristics and Asset Allocation
1,988 13
Male
14,930 14,930 14,930 14,930 14,930 14,930
All Twins (N )
14,930 100
All Twins
Identical Twins
A. Number of Twins by Zygosity and Sex
TABLE 1 Summary Statistics
Total 10,448 70
133
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
74,689 42,290 .23 .00
.18
14,930
14,930 14,930 14,930 14,930
.08
.22 135,522 .05 .37 .02 .34
14,930 14,930 14,930 14,930 14,930 14,930 14,930
12.18 .61 25,815 35,539
14,742 14,930 1,188 14,930
48,771 24,806 .12 2.05
.00
.00
.19 85,284 .00 .33 .00 .26
11.64 1.00 5,264 31,354
.22
.08
.20 142,876 .07 .35 .02 .33
11.82 .52 17,130 36,924
104,936 65,048 .53 .50
77,027 41,820 .25 .00
C. Savings Behavior
.32
.18
.14 166,222 .22 .31 .07 .30
2.69 .49 43,844 19,923
51,284 25,900 .13 2.05
.00
.00
.17 92,782 .00 .30 .00 .25
11.64 1.00 6,162 32,031
A. Number of Twins by Zygosity and Sex
110,743 65,938 .54 .51
.34
.18
.14 172,674 .25 .31 .07 .29
2.84 .50 28,687 23,144
Note.—Panel A provides information on the number of identical and nonidentical twins used in this study. Panel B provides summary statistics for several sociodemographic characteristics and asset allocation choices, separately for identical and nonidentical twins. Panel C reports summary statistics for the main measure of savings behavior; savings rate, which is defined as the 4 - year change in net worth adjusted for the change in home value ðzero for nonhomeownersÞ divided by the 4 - year disposable income; as well as the adjusted savings rate, which is the residual from regressions reported in App. table A2. All variables are defined in App. table A1.
Change in net worth ðUS$Þ Change in home value ðUS$Þ Savings rate Adjusted savings rate
Years of education ParentðsÞ still alive ð2003Þ Inheritance ðUS$, 2004 –7Þ Disposable income ðUS$, 4 - year averageÞ Standard deviation of log growth rate of disposable household income Net worth ðUS$, 4 - year averageÞ Business owner Equity/assets ðexcluding housingÞ ð4 - year averageÞ Bonds/assets ðexcluding housingÞ ð4 - year averageÞ Cash/assets ðexcluding housingÞ ð4 - year averageÞ Other financial assets/assets ðexcluding housingÞ ð4-year averageÞ Real assets ðexcluding housingÞ/assets ðexcluding housingÞ ð4 - year averageÞ
134
journal of political economy
excluding capital gains or losses related to the individual’s primary home.12 We scale the amount of savings by an individual’s disposable income over the same period. The savings rate, sij, for twinj ð1 or 2Þ in pair i is therefore defined as 2007 Net Worthij;2007 2 Net Worthij;2003 2 ot52004 DHome Valueij;t ; ð1Þ sij 5 2007 ot52004Disposable Incomeij;t where Net Worthij,t is the sum of the value of financial assets, including bank accounts, real estate, and other assets, less debt, at the end of year t;13 Disposable Incomeij,t is the sum of labor income, investment income, income from businesses, early retirement income, and net transfer payments ðe.g., child support and unemployment benefitsÞ, less taxes, at the end of year t; and DHome Valueij;t is a measure of capital gains or losses related to the individual’s primary home and captures the annual difference in reported home values, excluding changes that are associated with changes in home ownership status. Panel C of table 1 reports summary statistics for the savings rate of the individuals in our sample. For identical twins the average savings rate as defined in equation ð1Þ is 23 percent, while it is 25 percent for fraternal twins. The medians are about 12 percent for both groups. The variations across identical twins ð53 percentÞ and across fraternal twins ð54 percentÞ are very similar. We can compare the average savings rate of our sample to the numbers reported by, for example, Dynan et al. ð2004Þ, who report 22 percent and 21 percent for scaled changes in net worth based on Survey of Consumer Finances and Panel Study of Income Dynamics data ðacross all agesÞ. There are two points of caution regarding this comparison. First, our savings measure does not include home price changes. Second, the numbers were calculated over different time periods. Our data therefore suggest that Swedish individuals in our sample save a larger proportion of their disposable income than Americans did in the 1980s. Following Venti and Wise ð1998Þ, we use a linear regression model to control for individual socioeconomic conditions and characteristics that may enhance or reduce an individual’s savings ði.e., “circumstances,” “resource shocks,” or “chance” eventsÞ. Age is included to account for life cycle effects in savings behavior. Although marital status and the number of children are not necessarily chance events, we include them since, if 12 Consequently, we regard the individual’s primary home as consumption. The change in value of other real estate assets is regarded as investment income and is part of the savings measure. 13 Bank account balances of less than SEK 10,000 ðor for which the interest was less than SEK 100 during the yearÞ are not included. However, Statistics Sweden’s estimations suggest that 98 percent of all bank account balances are included in the data.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
origins of savings behavior
135
only because of economies of scale, marriage may increase the resources available for savings. The number of children may be a measure of a bequest motive for savings, and expenses related to raising children reduce the resources for savings. We also include proxies for health and unemployment status as poor health may reduce the amounts out of which savings could be taken, and so may unemployment. By Swedish family law, in case of death of a spouse, the surviving spouse inherits 100 percent of the assets of the deceased spouse, and the children inherit any assets only after the death of the last parent. Inheritances could mechanically lead to commonality in savings rates among twins, without necessarily reflecting commonality in preferences. We therefore include a proxy for received inheritances. We also control for the number of ðnontwinÞ siblings as more siblings may act as insurance, reducing savings. We also include 20 regional fixed effects. Similarly to Venti and Wise ð1998Þ, we do not include education and other characteristics that may correlate with time preferences. Our analysis relies on the assumption that savings decisions measured at the individual level reflect an individual’s preferences. We recognize that the decisions of a married individual are potentially affected by the presence and characteristics of a spouse ðe.g., Browning et al. 1994; Browning and Chiappori 1998Þ. In our regression analysis we therefore include socioeconomic characteristics of the spouse, such as age and health and employment status. We measure other variables, for example, income, income volatility, and business ownership at the household level, again to reflect that these household aggregates should affect a married individual’s savings decisions. In order to focus on savings behavior related to time preferences or self-control, we attempt to control for risk preferences, which as we discussed above have been shown to have a genetic component, by using an individual’s asset allocation as a proxy for risk aversion. This approach is motivated by standard models, such as Merton ð1969Þ and Samuelson ð1969Þ, which show that differences in risk preferences lead to crosssectional variation in the share of risky assets. The richness of our data allows us to include individuals’ asset allocation choices in detail, consisting of the fraction of assets held in equity securities, bonds, other financial securities, real estate assets, and cash. We are therefore able to remove ðsome ofÞ the variation in savings behavior that reflects crosssectional differences in risk preferences.14 As a result, genetic variation in this residual savings rate variable should more likely reflect variation in time than in risk preferences.
14 See chap. 4.3.2 in Wooldridge ð2001Þ for a discussion of using a proxy variable that is correlated with an unobservable and therefore omitted variable.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
136
journal of political economy
Appendix table A2 reports results from regressing the savings rate, sij, as defined in equation ð1Þ, on all the aforementioned individual- and household-level socioeconomic characteristics as well as asset allocation choices. Because these variables may have varying effects over the life cycle, we estimate separate regressions for those younger than 35, those older than 50, and those in the intermediate age category. Several of the coefficients are statistically significant and have the expected signs. In aggregate, these variables explain 9–14 percent of the cross-sectional variation of our savings rate, depending on age group. Having removed the variation in socioeconomic conditions and characteristics as well as asset allocation across individuals, we use the regression residuals as our adjusted savings rate, s^ij , in the following analysis of individual savings behavior. Panel C of table 1 provides summary statistics for the adjusted savings rate. By construction, the cross-sectional variation has been reduced. The distribution of the adjusted savings rate has also become more centered, with the median ð20.05Þ for both identical and fraternal twins now much closer to the mean ð0.00Þ. IV.
Empirical Approach and Identification
In order to decompose the cross-sectional variation in individuals’ propensity to save into genetic and environmental components, we model the adjusted savings rate, s^ij , for twin j ð1 or 2Þ of pair i as a function of three unobservable effects. Specifically, we assume that s^ij is a linear function of an additive genetic effect, aij, an effect of the environment common to both twins ðe.g., parentingÞ, ci, and an individual-specific effect, eij, which also captures individual-specific measurement error:15 s^ij 5 b0 1 aij 1 ci 1 eij ;
ð2Þ
where b0 is an intercept term. The additive genetic component aij represents the sum of the genotypic values of all genes that influence savings behavior ðsee, e.g., Falconer and Mackay ½1996 for a detailed discussionÞ. Each individual has two, potentially different, versions ðallelesÞ of each gene ðone from each parentÞ, and each version is assumed to have a specific, additive effect on savings behavior. The genotypic value of a gene is the sum of the effects of both alleles present in a given individual. Consider, for example, two different alleles A1 and A2 for a given gene and assume that the effect of the A1 allele on savings is of magnitude a1 , while the effect of the A2 allele is a2 . An individual with genotype A1A1 would experience the genetic effect 15 In quantitative behavioral genetics research, this model is referred to as an “ACE model”: A stands for additive genetic effects, C for common environment, and E for idiosyncratic environment.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
origins of savings behavior
137
of magnitude 2a1 , while genotype A1A2 would have a genetic effect of a1 1 a2 . The extent to which the effect of two different alleles deviates from the sum of their individual effects is called dominance deviation. While our baseline model above assumes no dominance deviation, we consider dominance deviations as a robustness check. The baseline model makes the additional assumptions that aij, ci, and eij are uncorrelated with one another and across twin pairs and normally distributed with zero means and variances j2a , j2c , and j2e , respectively, so that the total residual variance j2 is the sum of the three variance components ðj2 5 ja2 1 jc2 1 je2 Þ. We consider gene-environment interactions in Section VI. Identifying variation due to aij, ci, and eij separately is feasible because of constraints on the covariances. These constraints are motivated by the genetic similarity of twins as well as assumptions about parenting and other aspects of the common environment. Consider two twin pairs i 5 1, 2 with twins j 5 1, 2 in each pair, where the first is a pair of identical twins and the second is a pair of fraternal twins. The additive genetic 0 effects are a 5 ða11 ; a12 ; a 21 ; a 22 Þ . Identical and fraternal twin pairs differ in their genetic similarity; that is, the off-diagonal elements related to identical twins in the correlation matrix in ð3Þ are 1 as the proportion of shared additive genetic variation is 100 percent between identical twins. In contrast, for fraternal twins, the proportion of the shared additive genetic variation is on average only 50 percent; that is, the off-diagonal elements related to fraternal twins in the correlation matrix in ð3Þ are 1=2.16 As a result, for these two twin pairs, the variance-covariance matrix with respect to aij is 2 3 1 1 0 0 6 1 1 0 0 7 7 ð3Þ CovðaÞ 5 j2a 6 4 0 0 1 1=2 5: 0 0 1=2 1 0
The common environmental effects are c 5 ðc11 ; c 12 ; c 21 ; c 22 Þ . The model assumes that identical and fraternal twins experience the same degree of similarity in their common environments ðthe “equal environments assumption”Þ. That is, the off-diagonal elements related to either identical or fraternal twins in the matrix in ð4Þ are 1. Assuming that 16 For an intuitive explanation of the proportion of the shared additive genetic variation for fraternal twins as well as nontwin siblings, consider a single gene, of which one parent has alleles A1 and A2 while the other parent has alleles A3 and A4. Any of their offspring will have one of the following combinations as they get one allele from each parent: A1A3, A1A4, A2A3, or A2A4. Suppose that one fraternal twin is of A1A3 type. The overlap with the fraternal twin sibling will be 1 if the sibling is of A1A3 type, 1=2 if type A1A4, 1=2 if type A2A3, and 0 if type A2A4. This implies an average overlap of 1=2. For a formal derivation, see, e.g., Falconer and Mackay ð1996Þ.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
138
journal of political economy
identical and fraternal twins experience the same degree of similarity in their common environment, any excess similarity between identical twins is due to the greater proportion of genes shared by identical twins than by fraternal twins. As a result, for the two twin pairs, the variance-covariance matrix with respect to ci is 2 3 1 1 0 0 61 1 0 07 7 CovðcÞ 5 jc2 6 ð4Þ 4 0 0 1 1 5: 0 0 1 1 The individual-specific environmental effects are e 5 ðe11 ; e12 ; e 21 ; e 22 Þ0 . These error terms represent, for example, life experiences but also idiosyncratic measurement error. That is, the off-diagonal elements related to either identical or fraternal twins in the correlation matrix in ð5Þ are 0. As a result, for the two twin pairs, the variance-covariance matrix with respect to eij is 2 3 1 0 0 0 60 1 0 07 7 CovðeÞ 5 je2 6 ð5Þ 4 0 0 1 0 5: 0 0 0 1 We use maximum likelihood to estimate the model in equation ð2Þ. Finally, we calculate the variance components A, C, and E. The component A is the proportion of the total residual variance in the propensity to save that is due to an additive genetic factor: A5
ja2 ja2 5 2 : 2 j ja 1 jc2 1 je2
ð6Þ
The proportions attributable to the common environment ðC Þ and individual-specific environmental effects ðE Þ are computed analogously. Standard errors reported in the tables are bootstrapped with 1,000 repetitions. V.
Results
A.
Correlations for Identical versus Fraternal Twins
Before we report estimates from the above-specified model, we report evidence based on correlations of savings behavior for identical and fraternal twins. If identical twins are significantly more similar with respect to savings behavior than fraternal twins, then there is evidence that the propensity to save, at least partly, comes from an individual’s genes. In figure 1 we compute Pearson’s correlation coefficients for the savings rate
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
origins of savings behavior
139
F IG . 1.—Twin pair correlations of savings behavior by genetic similarity and gender
measure in equation ð1Þ. We find that the correlation for identical twins is .33, compared to only .16 for fraternal twins. This difference is statistically significant at the 1 percent level. While a first and intuitive indication that variation in savings behavior has a genetic component, it is important to emphasize that similarities in socioeconomic characteristics, also related to savings behavior, can explain these correlation results. In the rest of the paper we therefore calculate the adjusted savings rate using linear regression as described above, and then we perform a formal empirical decomposition of the cross-sectional variation in this measure using the model in equation ð2Þ. B.
Estimates from a Random-Effects Model
Table 2 reports estimates of variance components A, C, and E as defined in equation ð6Þ. As a benchmark for model fit, the first row of table 2, panel A, reports an E model in which both A and C are constrained to zero, and the second row reports a CE model in which only A is constrained to zero. The final row reports a full ACE model. We also report the Akaike information criterion ðAICÞ to compare fit across models, and we have performed likelihood ratio ðLRÞ tests to compare the E model against the CE model and the CE model against the ACE model. Our analysis produces several results. First, on the basis of AIC, the full ACE model is preferred; on the basis of the LR tests, the full ACE model is preferred over a CE model, which is preferred over an E model. That
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
140
journal of political economy TABLE 2 Decomposition of Savings Behavior Variance Components AIC
Model
A
C
E
A. All Twins ðN 5 14,930Þ E
21,850
CE
21,526
ACE
21,470
.327 ð.031Þ
.207 ð.016Þ .000 ð.011Þ
1.000 ð.000Þ .793 ð.016Þ .673 ð.026Þ
.000 ð.000Þ .078 ð.072Þ
.646 ð.038Þ .692 ð.044Þ
B. By Sex Male twins ðN 5 4,130Þ Female twins ðN 5 5,116Þ
.354 ð.038Þ .231 ð.098Þ
Note.—The table reports results from maximum likelihood estimation of linear randomeffects models. The adjusted savings rate is modeled as a linear function of up to three random effects representing additive genetic effects ðAÞ, shared environmental effects ðC Þ, as well as an individual-specific error ðE Þ. In panel A, we report results for a model that allows for only an individual-specific random effect ðE modelÞ, a model that also allows for a shared environmental effect ðCE modelÞ, and a model that also allows for an additive genetic effect ðACE modelÞ. The model is estimated using all 14,930 twins in our data set. In each case, we report Akaike’s information criterion ðAICÞ, the variance fraction of the combined error term explained by each random effect ðA for the additive genetic effects, C for shared environmental effects, and E for the individual-specific random effectÞ as well as the corresponding standard errors. All standard errors have been bootstrapped with 1,000 resamples. We perform likelihood ratio tests and at the 1 percent level reject the E model in favor of the CE model and the CE model in favor of the ACE model. Panel B reports results for ACE models estimated separately for male and female twins. N provides the number of observations used in each estimation.
is, modeling a genetic factor significantly improves the fit of a model that explains cross-sectional variation in individual savings behavior. Second, we quantify the proportion of the variation in savings behavior attributable to a genetic effect A and find that it is about 33 percent ðstatistically significant at the 1 percent levelÞ. That is, variation in savings behavior originates to a large extent from genetic variation across individuals.17 This evidence supports some economists’ recent references to a “patience gene” ðBernheim 2009Þ such that those equipped with such genes would then be predicted to save more. In contrast, C is estimated to be zero, suggesting that, on average, in our sample, savings behavior is not explained by differences in the common parental environment in which children grow up. Finally, we find that idiosyncratic environmental effects
17 We have reestimated the model for same-sex twins only and find very similar results ðuntabulatedÞ.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
origins of savings behavior
141
E contribute substantially to the variation in savings behavior: E is 67 percent and is statistically significant at all levels. This is the largest component in the model but captures an entire set of possible individual-specific life experiences and other nongenetic circumstances, in addition to measurement error. If the head of the household is male and influences the savings decisions of the spouse ðpossibly because of higher income or more experience with financial decisionsÞ, we expect different results for male and female twins. Panel B of table 2 reports results from separate analyses by gender. We find that for men, the A component is 35 percent and larger than for women, for which it is 23 percent. That is, the relative genetic variation of men’s propensity to save is about 50 percent larger than that of women. The C component is zero for men but 7.8 percent for women, albeit not statistically significant. We conclude that for men the variation in savings behavior is more attributable to genetic variation than for women, while for women variation in environmental factors, for example through their spouses, plays a larger role than for men. C. Alternative Measures of Savings Behavior We have also analyzed several alternative measures of savings behavior. We report the corresponding results in table 3, panel A. First, we recognize that our ability to empirically distinguish between risk and time preferences as well as other determinants of savings decisions is imperfect. Instead of relying on the assumptions discussed above to isolate time preferences by means of our regression model, we decompose the variation in the “unadjusted” ðas opposed to adjustedÞ savings measure, referred to as the unadjusted savings rate. We find that about 37 percent of the variation in savings is due to genetic variation, as opposed to 33 percent in the case of the main adjusted savings measure that we use throughout the paper.18 If we wanted to measure the total genetic component of savings behavior, regardless of the specific source of the genetic component, the 37 percent estimate would be most relevant. If we want to measure the genetic component of savings behavior that is not related to risk preferences but plausibly explained by time preferences, then we must control for the cross-sectional variation in attitudes to risk ðby using proxies such as the percentage allocation to risky assetsÞ. The 33 percent estimate above is then more relevant. Second, we calculate a total savings rate, which, in contrast with our previous measure, includes capital gains or losses related
18 We find that 35 percent of the variation in savings behavior is due to genetic variation if we do not include proxies for risk preferences ðuntabulatedÞ. This finding is consistent with the existing evidence of a substantial genetic effect with respect to risk aversion.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
142
journal of political economy TABLE 3 Measuring Savings Behavior Variance Components N
Model
A
C
E
A. Alternative Measures of Savings Behavior Unadjusted savings rate
14,930
Total savings rate
14,350
Active savings rate
14,370
Wealth-to-income ratio
14,362
.374 ð.037Þ .322 ð.041Þ .210 ð.045Þ .397 ð.049Þ
.000 ð.020Þ .008 ð.025Þ .010 ð.029Þ .032 ð.032Þ
.626 ð.026Þ .670 ð.023Þ .781 ð.024Þ .572 ð.024Þ
B. Controlling for Private Employer Pensions Not controlling for occupation
7,886
Controlling for occupation
7,886
.277 ð.035Þ .249 ð.043Þ
.000 ð.000Þ .000 ð.014Þ
.723 ð.035Þ .751 ð.036Þ
C. Measuring Savings at the Household Level Savings rate—household level
14,634
.253 ð.060Þ
.039 ð.039Þ
.708 ð.029Þ
Note.—The table reports results from maximum likelihood estimation of linear randomeffects models. Savings behavior is modeled as a linear function of three random effects representing additive genetic effects ðAÞ, shared environmental effects ðC Þ, as well as an individual-specific error ðE Þ. Panel A reports results for alternative proxies for savings behavior. See App. table A1 for a definition of unadjusted savings rate, total savings rate, active savings rate, and the wealth-to-income ratio. In panel B, we report results for the subset of 7,886 twins for whom we have data on their occupation. We first report estimates when not controlling for occupation. We then report estimates after also including 93 occupational fixed effects in the regressions otherwise specified in App. table A2. Panel C reports results when measuring savings at the household level but otherwise using the same definition as for our main adjusted savings rate. In each case, we report the variance fraction of the combined error term explained by each random effect ðA for the additive genetic effects, C for shared environmental effects, and E for the individual-specific random effectÞ as well as the corresponding standard errors. All standard errors have been bootstrapped with 1,000 resamples. N provides the number of observations used in each estimation.
to an individual’s primary home. We again remove variation in socioeconomic characteristics and asset allocation choices by regression, and we decompose the adjusted savings rate.19 We find that the genetic component ðAÞ explains 32 percent of the variation. Next, we construct an active savings rate, which excludes wealth changes due to the appreciation 19 In contrast to the regression specification shown in App. table A2, we also include the individual’s investment in real estate other than the primary home, scaled by total assets ðincluding home assetsÞ. The R 2’s of the three regressions are between 15 and 24 percent ðuntabulatedÞ.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
origins of savings behavior
143
of real estate or financial assets and thereby focuses on the “active” decision to save from disposable income. The genetic share is slightly lower at 21 percent. Finally, we define savings as a stock variable by calculating the wealth-to-income ratio, that is, dividing an individual’s net worth at the end of 2007 by the cumulative disposable income. This measure is similar to the definition used by Bernheim, Skinner, and Weinberg ð2001Þ and reveals a genetic component of 40 percent. Overall, we conclude that genetic variation ðAÞ accounts for about one-third of total variation, while the effect of the common environment ðCÞ, including parenting, is small and statistically not significantly different from zero. Another concern is that our savings rate measure does not capture expected benefits from pension and defined-benefit plans. This is problematic to completely address because we do not have data on expected pensions upon retirement. However, public pensions in Sweden are based on income, which we control for, and private pensions by employers vary largely by occupation. To the extent that occupational choices are more similar for identical than for fraternal twins, our results may be confounded. We therefore examine data on occupation, based on the International Standard Classification of Occupations by the International Labor Organization, which are available for a subset of our data. In panel B of table 3, we report results for this subset of twins, first without controlling for occupation and then when controlling for 94 occupation fixed effects. Without controlling for occupation, we find that the genetic component, A, is 28 percent in this subset of the data. When controlling for occupational choices, the genetic component drops by about 3 percentage points. We conclude that our results do not seem to simply represent genetic variation in occupational choices and related pensions.20 So far we have measured savings behavior at the individual level and we control for spousal influences in our regressions ðsee App. table A2Þ. In the absence of perfect assortative mating, the saving decisions measured at the household level likely reflect preferences of both spouses in a household. In panel C of table 3, we report our results when using household-level data as opposed to individual-level data. While we still find a large and statistically significant genetic effect, A drops to 25 percent when we use household-level data, consistent with the householdlevel measure being a less precise measure of an individual’s preferences.
20 As a complement to defined-benefit plans, Sweden also has had a privatized Social Security system since year 2000 ðe.g., Cronqvist and Thaler 2004Þ, where required contributions are a small portion ðabout 2.5 percentÞ of labor income, which we control for in our regressions. The amounts in these accounts are excluded from our net worth measure but represent a very small portion of wealth. Also, because of their age, most individuals in our sample are not affected by the privatized Social Security system.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
144 D.
journal of political economy Robustness
In this subsection, we discuss the robustness of our results with respect to the most important assumptions underlying the model in equation ð2Þ. 1.
Measurement Error
A well-recognized problem for any empirical study of savings behavior is measurement error ðsee, e.g., Dynan et al. 2004Þ. The data used to construct the savings rate in this study are not survey based but are provided by the Swedish Tax Agency. Because misreporting by financial institutions or individuals is subject to prosecution, we consider our data to be of good quality. A potentially more severe problem, which is relevant even with perfectly reported data, is that some of the data may reflect only transitory values because of specific circumstances in a particular year, for example, temporary unemployment. We address this concern in several ways. We measure savings over a 4-year period, so some transitory effects are expected to average out. In addition, we control for several temporary circumstances, for example, unemployment status, which may explain savings behavior but is not related to savings preferences. In table 3, panel A above, we have also considered four different savings measures and found that a significant genetic component explains each of them. Finally, it should be emphasized that idiosyncratic measurement error is captured by eij in the model in equation ð2Þ, which has two consequences. First, the genetic effect, aij, may be underestimated. Second, we have to be careful not to interpret the E component as capturing only effects of individual-specific life experiences on savings propensities because it also captures individual-specific measurement error. 2.
Model Misspecification
Several of the reported C components in the reported tables are zero, reflecting a corner solution as we constrain the variance components to be nonnegative. As a robustness check, we have reestimated the model in equation ð2Þ, but without the nonnegativity constraint on the individual variance components ðuntabulatedÞ. We find that the resulting negative C components are small in magnitude and generally not statistically significant from zero, reducing concerns about misspecification bias.21 Another concern arises when the identical twin correlation is more than twice the fraternal twin correlation. A genetic dominance deviation
21 Only in one out of 10 cases do we find a marginally significantly negative C component.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
origins of savings behavior
145
may then affect behavior. Such effects cannot be captured with an additive genetic effect of our baseline model in equation ð2Þ but require an additional factor for dominance genetic variation. While a dominance deviation effect, dij, can be added, a model with both C and D components is not identified. The model we can estimate is22 ^sij 5 b0 1 aij 1 dij 1 eij ;
ð7Þ
where the definitions are as before. The dominant genetic effects are d 5 ðd11 ; d 12 ; d 21 ; d 22 Þ0 . The off-diagonal elements related to identical twins in the matrix in ð8Þ are 1 ðas previouslyÞ. In contrast, for fraternal twins, the off-diagonal elements related to fraternal twins in the matrix in ð3Þ are 0.25.23 As a result, for the two twin pairs, the variance-covariance matrix with respect to dij is 2 3 1 0 0 0 6 0 1 0 0 7 7 ð8Þ CovðdÞ 5 j2d 6 4 0 0 1 1=4 5: 0 0 1=4 1 As a robustness check, we have estimated the model in equation ð7Þ. We find that the resulting D 5 ja2 =ðja2 1 jd2 1 je2 Þ component is small in magnitude ði.e., 5.3 percentÞ and not statistically significant from zero ðuntabulatedÞ, while the A component accounts for about 28.3 percent of the variation. This reduces concerns that our reported models are misspecified. Finally, we note that positive assortative mating between our twins’ parents would make fraternal twins more similar relative to identical twins and in our model would bias the estimate of the genetic ðAÞ component downward ðe.g., Neale and Maes 2004Þ. 3.
Equal Environments Assumption
If parents and others in an individual’s environment treat identical twins more similarly compared to fraternal twins ðalong the dimensions that are relevant for savings behaviorÞ, then estimates of A may be biased upward. This is a well-recognized problem in twin research, and substantial resources have been devoted to address the equal environments as-
22 This model is often referred to as an “ADE model” and is commonly used in quantitative genetics research ðe.g., Neale and Maes 2004Þ. 23 The 25 percent shared variation for fraternal twins is due to the fact that dominance deviation depends on both alleles. Since fraternal twins have only a 25 percent probability of sharing both alleles in a given locus, they will share only 25 percent of the dominance genetic variation ðfor further details, we refer to Plomin et al. ½2008Þ.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
146
journal of political economy
sumption ðEEAÞ.24 From research on IQ and personality traits, where the EEA has been addressed most rigorously, the evidence suggests that any bias from violations of the EEA is not of first-order importance ðe.g., Bouchard 1998Þ. Specifically, researchers have studied twins reared apart, that is, twins separated at birth or early in life, for whom there is no common parental environment. Such studies often produce heritability estimates similar to those using twins reared together ðe.g., Bouchard et al. 1990Þ.25 Furthermore, in the relatively rare cases in which parents miscategorize their twins as identical instead of fraternal ðor vice versaÞ, little evidence was found that parents’ perceived zygosity influences twin resemblance with respect to IQ, personality, or mental disorders ðe.g., Kendler et al. 1994Þ. Another concern has been that the matched physical appearance of identical twins results in more similar treatment by those who are a part of these individuals’ environments, in the end causing more similar outcomes. Using a novel research design, Segal ð2013Þ studies unrelated look-alike individuals and finds that their correlations for personality measures are much lower than for identical twins, suggesting that identical twins’ similarity mostly reflects similarity in their genes and not similar treatments by others. Finally, recent progress has enabled researchers to construct DNA-based measures of pairwise genetic relatedness, which were then compared to pairwise similarities with respect to IQ ðDavies et al. 2011Þ and height ð Jian et al. 2010Þ. In contrast to twin studies, these studies use unrelated subjects and do not rely on assumptions such as the EEA. Importantly, theses studies have estimated the relative amount of genetic variation to be consistent with existing evidence from traditional twin studies.
24 See, e.g., Goldberger ð1979Þ for a discussion of common concerns related to twin studies. 25 Phillips ð1993Þ points out that identical and fraternal twins may also differ with respect to their prenatal environments. Specifically, about two-thirds of identical twins share the same chorion ðand share blood supply and are exposed to the same hormonesÞ, while onethird of identical and all fraternal twins have their own chorion. If the prenatal environment is important for postnatal behavior, then chorionic variation could inflate estimates of genetic effects. Most twin data sets, including ours, do not provide information on whether identical twins are mono- or dichorionic. Sokol et al. ð1995Þ study 44 twins and find that identical twins with a shared chorion are more similar than dichorionic twins with respect to some personality traits, including impulsivity, but generally not with respect to cognitive ability. In a more recent study of social behavior, Hur ð2007Þ finds no impact of chorionic variation. While we cannot empirically assess the importance of chorionic variation for savings behavior, we follow Prescott, Johnson, and McArdle ð1999Þ and estimate that the possible upward bias of our estimate of the genetic effect is about 4 percentage points. For this calculation, we assume that the relative variance of the chorionic effect equals 0.20 ðSokol et al. 1995Þ and that two-thirds of identical twins are monochorionic. We also use the fact that about 30 percent of the twins in our data are identical. Thus, chorionic variation seems to have at most a modest effect on our point estimates.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
origins of savings behavior VI. A.
147
Further Evidence and Extensions Effects of Parenting on Savings Behavior
Social transmission, as opposed to genetic transmission, is a potential explanation for parent-child similarity in savings behavior ðBisin and Verdier 2008Þ. One intriguing result so far is that parenting seems to affect savings behavior very little; that is, the C component is close to zero and statistically insignificant in most models we have estimated. This is surprising in the sense that social transmission of behavior from parents to their children seems to be important for, for example, religion and labor supply preferences ðe.g., Bisin et al. 2004; Fernandez, Fogli, and Olivetti 2004; Knudsen et al. 2006Þ. In this section, we extend the previous analysis by examining whether the strength of the parenting effect varies with an individual’s age and scarce resources available for parenting. 1.
Age
In the analysis in this subsection, we use a standard interaction model ðe.g., Harden, Turkheimer, and Loehlin 2007Þ in which age and age squared act as variance moderators. Appendix table A3 reports all model parameters. We find that the effect of parenting on the savings propensity in fact varies with age. Figure 2 shows the decay in the estimated parenting
F IG . 2.—Fraction of cross-sectional variation explained by parenting ðcaptured by shared environmental effects ½C Þ by age.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
148
journal of political economy
effect. In particular, we find that parenting explains up to half of the variation in savings rates for the youngest individuals in our sample ð20– 25-year-oldsÞ, but parenting does not have a lifelong impact on an individual’s savings behavior as the effect decays significantly over time. That is, while parents seem to strongly influence their children’s savings behavior earlier in life, the effect disappears over time as their children gain their own individual life experiences. At the same time, we find that genes still explain 28 percent of the cross-sectional variation in saving propensities among those older than 50, so in contrast to parenting effects, the genetic effect does not disappear ðuntabulatedÞ. 2.
Scarce Resources for Parenting
Becker and Mulligan ð1997Þ argue that parents often spend resources on teaching their children to better plan for the future, resources that might affect the children’s personal discount rates. Parenting involves scarce resources such as time and is often difficult to delegate. We therefore predict that the common family environment results in a smaller effect on a child’s savings behavior when parenting is relatively more scarce, all else equal. In particular, the presence of more children in the family may reduce the time parents are able to commit to each individual child. That is, we predict that the presence of siblings reduces the average effect of parenting.26 In panel A of table 4, we estimate and report separate models for twins who grew up in families with and without other siblings. We find that parenting explains about 12 percent of the cross-sectional variation in the propensity to save among individuals with no siblings.27 In contrast, we find no effect of parenting when there are siblings in the family. We do not find that this effect decays with the number of children, either because the main difference is between twins and one extra child or because the decay is difficult to estimate precisely as there are only a small number of families with four or more children in our data set. That is, the propensity to save in adult life of those who grew up with no siblings will be more affected by their parents’ savings behavior. We conclude that the effects of parenting on savings behavior seem to be smaller when resources for parenting are likely more scarce. B. Gene-Environment Interactions Research in other fields within the social sciences has shown that it is rare that genetics or environments are solely responsible for produc26 We recognize a potential sample selection problem if only those with particularly effective parenting skills endogenously choose to have a larger number of children. 27 We recognize that with a p -value of 14 percent, the effect is not statistically significant at conventional levels.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
origins of savings behavior
149
TABLE 4 Further Evidence and Extensions Variance Components N
Model
A
C
E
A. Scarce Parenting No siblings
3,682
Siblings
11,242
.207 ð.124Þ .312 ð.027Þ
.125 ð.084Þ .000 ð.000Þ
.668 ð.056Þ .688 ð.027Þ
B. Parents’ Net Worth Below median net worth
4,036
Above median net worth
4,054
.178 ð.075Þ .254 ð.105Þ
.000 ð.033Þ .142 ð.0720
.822 ð.057Þ .604 ð.046Þ
Note.—The table reports results from maximum likelihood estimation of linear randomeffects models. The adjusted savings rate is modeled as a linear function of three random effects representing additive genetic effects ðAÞ, shared environmental effects ðC Þ, as well as an individual-specific error ðE Þ. In panel A, we report results separately for those twins who have no other siblings and those who have at least on nontwin sibling. Panel B reports results separately for twins whose parents at the end of 2002 have net worth below and above the sample median. In each case, we report the variance fraction of the combined error term explained by each random effect ðA for the additive genetic effects, C for shared environmental effects, and E for the individual-specific random effectÞ as well as the corresponding standard errors. All standard errors have been bootstrapped with 1,000 resamples. N provides the number of observations used in each estimation.
ing individual variation and that most studied behaviors involve geneenvironment interactions. For a more extensive overview of research on the interactions of genes and environments, we refer to Rutter ð2006Þ. Some economists, such as Cunha and Heckman ð2010Þ, have gone as far as concluding that “the nature versus nurture distinction is obsolete” ð3Þ, and they argue that the notion that genes are moderated by environments should receive more attention in economic research. In principle, environments can moderate genetic predispositions by either enhancing or suppressing genetic effects. The bioecological theory of gene-environment interactions suggests that the expression of a specific genetic predisposition can become stronger in more supportive environments ðBronfenbrenner and Ceci 1994Þ. This model can, for example, explain the evidence in Taylor et al. ð2010Þ, who show that the genetic effect on reading fluency among first and second graders increases as the quality of the children’s teacher increases ðmeasured by reading gain among nontwin classmatesÞ. In this subsection, we extend the previous analysis by examining whether more supporting environments, in a broad sense, enhance genetic predispositions also in the context of savings propensities. We examine the extent to which the individual’s family environment when growing up moderates genetic propensities. One characteristic of
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
150
journal of political economy
the family environment is the socioeconomic status of the individual’s parents. We use net worth as a proxy, and because we are not able to measure parents’ net worth during the upbringing, we use net worth at the end of 2003. Panel B of table 4 reports that an individual’s savings behavior in adult life is more affected by genes if growing up in a wealthier family. In particular, genes explain 25 percent of the crosssectional variation in savings propensity among those who grew up in wealthier families, compared to 18 percent among those who grew up in a family that was relatively less well off. That is, a wealthier upbringing results in a stronger expression of genetic predispositions to a specific savings propensity. We also find a significant parent or common environmental component, C, for those twins with wealthier parents. This finding is consistent with the above evidence related to scarce resources for parenting.28 We conclude that a more supportive socioeconomic environment ðproxied by net worthÞ when growing up seems to moderate a genetic effect on savings behavior. C. Why Is Savings Behavior Genetic? To understand what explains the genetic component of savings behavior, we examine whether our measure of savings behavior is correlated with other behaviors and outcomes that may reflect time preferences, impatience, or self-control, such as income growth, education, smoking, and obesity.29 In panel A of table 5, we report the correlation between the propensity to save and these other measures. We compute the average log income growth rate for each individual between the end of 2003 and the end of 2007. We find a significantly positive correlation between an individual’s savings rate and income growth, consistent with the common prediction that patient individuals commonly experience higher income growth rates. Importantly, for a subset of individuals, data on education ðnumber of yearsÞ, smoking ðnumber of cigarettes smoked
28 We also examine whether an individual’s current socioeconomic status, proxied by own net worth, moderates genetic effects on savings behavior and find that genes explain about 37 percent of the cross-sectional variation in the propensity to save among wealthier individuals, compared to only 21 percent among those who are relatively less well off. One interpretation of this evidence is that while the wealthier are able to choose a consumptionsavings behavior that reflects their genetic predisposition and preferences ði.e., either save a lot or save only a littleÞ, the savings behavior of the relatively poor is much more constrained and is governed largely by the individual-specific life experiences that, by chance, they were exposed to. Concerns when interpreting this evidence include that net worth may proxy for many things other than socioeconomic status and that the main savings measure is constructed as a difference in net worth. 29 In each case, we use the residuals from a regression of each measure on the same control variables included in our savings regression.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
origins of savings behavior
151
TABLE 5 Genetic and Environmental Correlations Log Income Growth
Years of Education
Number of Cigarettes
BMI
A. Correlation with the Adjusted Savings Rate Total correlation Observations
.047 ð.008Þ 14,726
.000 ð.009Þ 14,590
2.065 ð.013Þ 6,216
2.059 ð.012Þ 7,746
B. Genetic and Environmental Correlations Correlation of genetic effects Correlation of individual-specific environments
.139 ð.049Þ
.018 ð.027Þ
2.087 ð.058Þ
2.115 ð.042Þ
.017 ð.016Þ
2.018 ð.019Þ
2.056 ð.025Þ
2.020 ð.028Þ
Note.—Panel A reports maximum likelihood estimates of the correlation between the adjusted savings rate and the average ðadjustedÞ log income growth rate between 2002 and 2006, the ðadjustedÞ number of years of education, the ðadjustedÞ number of cigarettes smoked per day, and the ðadjustedÞ body mass index ðBMIÞ. Panel B reports the correlation between the corresponding genetic and individual-specific environmental effects. The correlation due to common factors is not defined as the C component of savings behavior is zero. All standard errors have been bootstrapped with 1,000 resamples.
per dayÞ, and body mass index ðBMI; weight relative to squared heightÞ are available from the STR’s interviews.30 We find no statistically significant correlation between education and savings behavior but significantly negative correlations between an individual’s savings rate and smoking and BMI. That is, those who save less are found to smoke more and are more likely to be obese on the basis of their BMI. Overall, this evidence, while based on relatively crude proxies of individual preferences, suggests a certain consistency in behavior ðBarsky et al. 1997Þ because of an individual’s time preferences and level of self-control. We next decompose the covariance matrices for these outcomes into components corresponding to genetic as well as environmental effects.31 On the basis of the decomposed covariance matrices, we compute the genetic and environmental correlations between savings behavior and the other behaviors and outcomes. The correlation due to common factors is not defined as the C component of savings behavior, but not necessarily of the other measures, is zero. Table 5, panel B, shows that the correlations between savings behavior and each of the other behaviors
30 These data were collected between 1998 and 2002 and are available only for twins born before 1958. 31 For this analysis, we use a bivariate Cholesky decomposition; see Neale and Maes ð2004Þ for details.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
152
journal of political economy
and outcomes are largely genetic. For example, lack of savings and obesity are correlated because of genes, not environmental conditions. Only our measure of smoking has a significant environmental correlation with savings behavior, while the genetic correlation is 25.6 percent with a pvalue of 13 percent. One explanation for why savings behavior is genetic appears to be that an individual’s time preferences are partly genetic. Our evidence of a significant positive genetic correlation between an individual’s savings and income growth rates supports such an explanation. Some individuals are born to be more patient, and this affects these individuals’ savings behavior, as well as other outcomes, for example, the choice of income process. Moreover, the negative and significant genetic correlation between savings rate and both smoking and body weight suggests that behavioral factors such as lack of self-control may also affect savings behavior. For example, to the extent that a high BMI and obesity may be interpreted as an expression of lack of self-control, we conclude that lack of savings correlates with lack of self-control, and this correlation is mainly found to be genetic. D.
Evidence from Consumption Data
So far, we have identified variation in time preferences through residual variation in savings behavior that is not explained by financial resources, circumstances, or investment choices. Our data allow us to impute consumption growth rates at the individual level, which, in principle, can be used to estimate individual-specific Euler equations to obtain subjective discount rates for the twins in our data set. While we expect that substantial measurement error will lead to a reduced A component and an increased E component, which by design accounts for individual-specific measurement error, we explore this alternative route. Specifically, using an extended sample with data between 1999 and 2007, we impute consumption by subtracting the amount of active saving from disposable income. For details of the construction of the consumption data, see Appendix table A1. We obtain up to seven consumption growth rates for each individual. We adjust the annual log consumption growth rate for age, change in family size, and change in marital status and drop the top and bottom 1 percent of the distribution of the consumption growth rate. The average log consumption growth rate of those with at least 3 years of nonmissing consumption growth rates is 0.054 with a crosssectional standard deviation of 0.203. In table 6, we report results from two separate analyses of these data. First, we decompose the cross-sectional variation of average log consumption growth into genetic ðAÞ and environmental ðC and EÞ components,
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
origins of savings behavior
153
TABLE 6 Consumption Growth and Time Preferences Variance Components N
Model
A
C
E
A. Average Consumption Growth Rates ðlnÞ At least 3 years of data
12,922
At least 4 years of data
11,074
At least 5 years of data
9,160
.126 ð.041Þ .136 ð.048Þ .178 ð.047Þ
.003 ð.026Þ .019 ð.033Þ .003 ð.032Þ
.871 ð.022Þ .846 ð.023Þ .820 ð.025Þ
B. Time Preferences Standard Euler equation Double-difference Euler equation
9,700 10,042
.136 ð.025Þ .099 ð.026Þ
.000 ð.006Þ .000 ð.011Þ
.864 ð.024Þ .901 ð.022Þ
Note.—The table reports results from maximum likelihood estimation of linear randomeffects models. The average log consumption growth rate ðpanel AÞ and the subjective discount rate ðpanel BÞ are modeled as a linear function of age, gender, and three random effects representing additive ðAÞ genetic effects, shared environmental effects ðC Þ, as well as an individual-specific error ðE Þ. Panel A reports different results for averages formed over at least 3, 4, or 5 years of data. All models in panel A account for an individual’s equity allocation. In panel B, the subjective discount rate is estimated from either a standard Euler equation or a double-difference Euler equation ðsee Alan et al. 2009, eq. ½10Þ. All standard error have been bootstrapped with 1,000 resamples. See App. table A1 for a definition of all variables. N provides the number of observations used in each estimation.
accounting for age, gender, and equity allocation, which we use as a simple proxy for risk aversion. We report these results in table 6, panel A. Assuming that the remaining variation is mainly driven by differences in subjective discount rates as well as measurement error, we find that 13–18 percent of the remaining variation is due to genetic factors, with the remainder being due to environmental influences, including measurement error. As we would expect, measurement error and therefore the E component decrease, as we increase the required minimum number of years of consumption growth from 3 to 5 years. In our second approach, we use nonlinear Euler equations to identify individual-specific time preference parameters, dij , where as above j denotes twin 1 or 2 in pair i. As far as we know, this is one of the first attempts to use individual-specific Euler equations to estimate time preference parameters for a large cross section of individuals. In particular, we assume standard constant relative risk aversion preferences, with gij denoting the coefficient of relative risk aversion, applied to individual annual consumption, Cij,t, and the risk-free rate, rt, as the only asset. We use the standard nonlinear Euler equation:
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
154 E
Cij;t11 Cij;t
2gij
journal of political economy ð9Þ ð1 1 rt11 Þdij 2 1 5 0
as well as an alternative expression that Alan, Attanasio, and Browning ð2009Þ derive to address measurement error ðsee their eq. ½10Þ: E 2
Cij;t12 Cij;t
Cij;t11 Cij;t
2gij
2gij
ð1 1 rt11 Þdij ð10Þ
ð1 1 rt11 Þð1 1 rt12 Þd
2 ij
5 0:
In both cases, we use the lagged risk-free rate as an instrument and drop individuals for whom d estimates take on boundary values.32 Using equation ð9Þ above, we find an average estimate of d of 0.87 with a standard deviation of 0.21. Using equation ð10Þ yields an average d estimate of 1.06 with a standard deviation of 0.23. We then decompose the cross-sectional variation of the two types of d estimates into genetic and environmental components, controlling for age and gender. Panel B of table 6 reports the results. The genetic component ðAÞ accounts for about 10–14 percent of the total variation and is significantly different from zero. The individual-specific E component accounts for all the remaining variation. Overall, we conclude that while imputing individual consumption growth rates and estimating individual-specific time preference parameters is challenging because of substantial measurement error and a short time series of data, we obtain results that confirm our main finding of a substantial genetic component of the variation of time preferences. The lower point estimates of 10–14 percent relative to the 30 percent, which we find in our main analysis, are likely due to substantial measurement error. VII.
Discussion
In this section, we discuss important implications of the reported empirical evidence.
32 To obtain parameter estimates, we perform a grid search for d and g in which d can take on values between 0.50 and 1.50 ðwith a spacing of size 0.001Þ and g values between 1.0 and 10.0 ðwith a spacing of size 0.01Þ. We have verified that our decomposition results are robust to coarser and finer spacings as well as to using lagged consumption growth or lagged income as instruments.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
origins of savings behavior A.
155
Wealth at Retirement
Accumulated wealth at the time of retirement is a function of income, savings propensity, and asset allocation choices over many years. If ðiÞ abilities that affect an individual’s income are partly genetic ðe.g., Taubman 1976; Behrman and Taubman 1989Þ, ðiiÞ time preferences that affect savings propensities are genetic as reported in this paper, and ðiiiÞ risk aversion that affects asset allocation choices is also partly genetic ðe.g., Barnea et al. 2010; Cesarini et al. 2010Þ, then the cross-sectional variation in wealth at retirement age is predicted to have a significant genetic component. In table 7, we estimate a random-effect model as in equation ð2Þ for net worth around retirement age ð60–69 years oldÞ in 2007.33 We find that about 39 percent of the cross-sectional variation in wealth accumulated up to retirement is explained by genetic variation. As discussed above, under positive assortative mating, this estimate would be a lower bound of the variation in net worth due to genetic variation. The effect of the common family environment and upbringing, which by model construction also reflects wealth inherited from parents, explains only 7 percent of the cross-sectional variation ðp -value of 12 percentÞ. This evidence suggests that parent-child similarities with respect to economic status, as proxied for by net worth, reflect mainly genetic rather than financial endowments that children receive from their parents. We note, though, that about half of the variation in net worth reflects individual-specific events and circumstances. Overall, these findings have implications for our understanding of the causes behind persistent inequalities in wealth in society ðClark 2014Þ. B.
Variation across Countries and Cultures
Different countries have different degrees of genetic and environmental variation. Recent studies have examined genetic diversity and conclude that the country examined in this paper, Sweden, is at least as genetically diverse as many other countries ðe.g., Nelis et al. 2009; Humphreys et al. 2011Þ. We also note that the genetic component of 33 percent reported in this paper is not a universal biological constant, but an estimate relative to the amount of environmental variation in our sample. For example, if there is no variation in environmental factors, then the genetic component by definition will be large. Sweden is sometimes perceived to have high cultural homogeneity. Since we observe individuals from only
33 We include all individuals who in 2007 were between 60 and 69 years old and had nonmissing wealth data between 2004 and 2007. To avoid that our results are affected by extreme values, we drop individuals in the bottom and top 1 percent of the distribution.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
156
journal of political economy TABLE 7 Decomposition of Wealth Distribution Variance Components Model Net worth ð2007Þ
N
A
C
E
11,992
.3930 ð.0670Þ
.0690 ð.0440Þ
.5380 ð.0300Þ
Note.—The table reports results from maximum likelihood estimation of linear random-effect models. Net worth ðat the end of 2007Þ of twins between ages 60 and 69 is modeled as a linear function of three random effects representing additive genetic effects ðAÞ, shared environmental effects ðC Þ, as well as an individual-specific error ðE Þ. We report the variance fraction of the combined error term explained by each random effect ðA for the additive genetic effects, C for shared environmental effects, and E for the individual-specific random effectÞ as well as the corresponding standard errors. All standard errors have been bootstrapped with 1,000 resamples. N provides the number of observations.
one country at one point in time, our empirical analysis cannot capture the relative importance of country characteristics such as ðcountrywideÞ culture. Also, the analysis of only one country reduces the amount of total variation in environmental factors. To get some sense for the importance of environmental variation, we have reestimated our model for ðiÞ urban areas with average population density of 1,030 residents per square kilometer and ðiiÞ rural areas with average population density of six residents per square kilometer ðuntabulatedÞ. The results show that the individual-specific environment ðE Þ explains a larger proportion of the variation in urban areas ð65 percentÞ than in rural areas ð52 percentÞ. One interpretation of this evidence is that urban life results in exposure to more individual-specific environmental factors, which determine an individual’s savings behavior. In contrast, rural life may involve exposure to relatively fewer individual-specific environmental factors, and as a result, savings propensities are more affected by genetic predispositions in rural areas. That is, our evidence may have implications for the determinants of savings behavior as urbanization increases in many countries. C. Social Interaction Our evidence should not be interpreted as meaning that social interaction effects are unimportant for savings behavior.34 Examining the data on twin-twin communication, we find that social interactions are important. Twins who interact more than once per week have more similar 34 Several studies in economics and finance suggest that social interaction and peer effects are important in the financial domain ðe.g., Hong, Kubik, and Stein 2004; Brown et al. 2008Þ.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
origins of savings behavior
157
savings behavior as captured by a significant common environmental component C that explains about 12 percent of the variation in savings behavior ðuntabulatedÞ. This result is consistent with social interaction affecting behavior in general ðe.g., Bikhchandani, Hirshleifer, and Welch 1992; Shiller 1995Þ and savings propensities in particular. Recent work has shown that the kindergarten environment explains savings behavior in adulthood ðChetty et al. 2011Þ. Future research should examine which specific characteristics of social networks are important for savings behavior and also whether this is an environmental factor that moderates genetic predispositions, for example, whether social interaction with similar peers enhances genetic effects in the context of savings behavior. D.
Effectiveness of Public Policy
Economists and policy makers have recently devoted significant effort to examining financial literacy and methods to change individual savings behavior ðe.g., Bernheim, Garrett, and Maki 2001; Thaler and Benartzi 2004Þ. One important question is therefore whether our evidence of a genetic component of savings behavior means that policy in the domain of savings behavior is meaningless. While environmental factors are possibly more amenable to policy than innate predispositions ðsee, e.g., Bernheim 2009Þ, a genetic component of savings behavior would likely have implications for the specific design of public policy. Furthermore, additional research on gene-environment interactions may identify circumstances in which genetic predisposition matters more or less, thereby contributing to developing more effective policies. Moreover, about two-thirds of the variation in savings behavior across individuals is due to the individual-specific environmental component. Finally, we note that public policies that affect the average amount of savings without affecting the variation across individuals, for example, through changes of a country’s “savings culture,” are not captured in our variance decomposition model. VIII. Conclusion Our evidence suggests that there is not a simple answer to the question of what explains variation in savings behavior across individuals. An individual’s savings propensity is governed by genetic predispositions, social transmission of behavior from parents to their children, and geneenvironment interactions in which environmental conditions moderate genetic effects. We find that genetic differences explain about 33 percent of the variation in savings behavior across individuals. Each individual is born with a genetic predisposition to a specific savings behavior, an effect that
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
158
journal of political economy
is found not to disappear later in life. Parenting contributes to the variation in savings rates among younger individuals in our sample, but its effect has decayed significantly for middle-aged and older individuals; that is, parenting does not have a lifelong impact on their children’s savings propensities. We also find that the family environment when growing up ðe.g., parents’ wealthÞ moderates genetic predispositions to a particular savings behavior, evidence that is consistent with theories that genetic effects are stronger in more supportive environments. Finally, we examine why savings behavior is genetic and find that savings behavior is genetically correlated with income growth, smoking, and obesity, suggesting that the genetic component of savings behavior reflects genetic differences across individuals with respect to time preferences or self-control. Our work raises a number of currently unanswered questions for future research. What are the important events in life that systematically affect an individual’s savings behavior? We find that identical twins start off early in life exhibiting very similar savings behaviors, but over their lives their savings behavior often diverges. What are the differential life experiences that result in different savings behavior even among genetically identical individuals with a common family environment when growing up? Another question relates to gene-environment interactions, in particular, to better understanding what experiences or interventions moderate genetic propensities to a specific savings behavior.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
Active savings rate
Total savings rate
Measures of savings behavior: Savings rate ðunadjusted, adjusted, household-levelÞ
Nonidentical twins
Types of twins: Identical twins
Variable
Appendix
Description
Savings rate ðunadjustedÞ is calculated as the change in net worth between the end of 2003 and the end of 2007, less capital gains or losses related to an individual’s primary residence, divided by the total disposable income for 2004 –7. The top and bottom 1 percent of the distribution have been dropped. We use the residual from a linear regression model ðsee App. table A2Þ of the savings rate onto individual circumstances and asset allocation choices as the adjusted savings rate. The adjusted savings rate is the main empirical proxy for an individual’s savings behavior. Unless otherwise indicated, the savings rate is calculated at the individual level, not at the household level. Total savings rate is calculated as the change in net worth between the end of 2003 and the end of 2007 divided by the total disposable income for 2004 –7. The top and bottom 1 percent of the distribution have been dropped. We use the residual from a linear regression model of the savings rate onto individual circumstances and asset allocation choices as the adjusted total savings rate. The regression specification differs from the one in App. table A2, as we also include the relative allocation to other real assets. All allocations are scaled by total assets that include home assets. Active savings rate is calculated as the change in net worth between the end of 2003 and the end of 2007, less capital gains or losses related to an individual’s real estate or risky financial assets, divided by the total disposable income for 2004 – 7. For real estate assets, capital gains or losses are calculated as the change in strictly positive prices over the course of 1 year. For risky financial assets, we calculate capital/ gain and losses by applying the change of the European price index ðin Swedish kronor ½SEKÞ to the value of risky financial assets at the beginning of a given year. The top and bottom 1 percent of the distribution have been dropped. We use the residual from a linear regression model of the savings rate onto individual circumstances and asset allocation choices as the adjusted active savings rate. The regression specification is the same as in App. table A2.
Twins that are genetically identical, also called monozygotic twins. Zygosity is determined by the STR on the basis of questions about intrapair similarities in childhood. Twins that share, on average, 50 percent of their genes, also called dizygotic or fraternal twins. Nonidentical twins can be of the same sex or the opposite sex. Zygosity is determined by the STR on the basis of questions about intrapair similarities in childhood.
TABLE A1 Definition of All Variables
160
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
Widowed
Divorced
Married
Age Age ðspouseÞ
Socioeconomic characteristics: Male
Time preference parameter
Consumption growth rate ðlnÞ
Wealth-to-income ratio
Variable
Description
An indicator variable that equals one if an individual is male and zero otherwise. Gender is obtained from Statistics Sweden. An individual’s age on December 31, 2007, as reported by Statistics Sweden. The age of the spouse on December 31, 2007, as reported by Statistics Sweden. For unmarried individuals, the variable is zero. An indicator variable that equals one if an individual is married in all years between 2004 and 2007 and zero otherwise. It is obtained from Statistics Sweden. An indicator variable that equals one if an individual is divorced in all years between 2004 and 2007 and zero otherwise. It is obtained from Statistics Sweden. An indicator variable that equals one if an individual is widowed in all years between 2004 and 2007 and zero otherwise. It is obtained from Statistics Sweden.
Wealth-to-income ratio is calculated by dividing an individual’s net worth at the end of 2007 by the cumulative disposable income for 1999–2007. We use the residual from a linear regression model of the wealth-toincome ratio onto individual circumstances and asset allocation choices as the adjusted wealth-to-income ratio. The regression specification differs from the one in App. table A2, as we also include the relative allocation to other real assets. All allocations are scaled by total assets that include home assets. The adjusted wealth-to-income ratio is an alternative empirical proxy for an individual’s savings behavior. Consumption growth rate is calculated at the individual level. For each twin and year between 2000 and 2007, we subtract the amount of annual active saving ðsee active savings rate for detailsÞ from disposable income. If imputed consumption is not strictly positive, we set it to missing. We adjust the annual log consumption growth rate for age, change in family size, and change in marital status and drop the top and bottom 1 percent of the distribution of the consumption growth rate. We then form averages, requiring at least 3, 4, or 5 years of data. We estimate time preference parameters separately for each twin. The time preference parameter ðsubjective discount factorÞ is found for each twin by performing a grid search for the objective function g 0 g , where g contains the Euler equation expectational error interacted with a constant and an instrument. The grid search evaluates the objective function for values of the time preference parameter ðbÞ between 0.5 and 1.5 and the risk aversion parameter ðgÞ between 1.0 and 10.0, where the step size is .001 for b and .01 for g and selects the parameter values associated with the smallest value of the objective function. The Euler expectational error is modeled after eqq. ð2Þ ðstandard Euler equationÞ and ð10Þ ðdouble-difference Euler equationÞ in Alan et al. ð2009Þ. For an individual to be included, we require at least 3 years of consumption growth. We drop cases in which the time preference parameter takes on a boundary value.
TABLE A1 (Continued )
161
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
Disposable household income
Disposable income
Inheritance
ParentðsÞ still alive ð2003Þ
Years of education
Unemployed ðspouseÞ
Unemployed
Poor health ðspouseÞ
Poor health
Number of children in household
Number of siblings Children in household indicator
The number of nontwin siblings an individual has. The number is obtained from Statistics Sweden. An indicator that equals one if an individual has any children living in the household and zero otherwise. The information is obtained from Statistics Sweden. The number of children living in the household. The number is obtained annually from Statistics Sweden and averaged for 2004 –7. An indicator variable that equals one if an individual receives payments due to illness, injury, or disability and zero otherwise. The data are obtained annually from Statistics Sweden and averaged for 2004 –7. An indicator variable that equals one if an individual’s spouse receives payments due to illness, injury, or disability and zero otherwise. The data are obtained annually from Statistics Sweden and averaged for 2004 –7. For unmarried individuals, the variable equals zero. An indicator variable that equals one if an individual receives payments due to unemployment and zero otherwise. The data are obtained annually from Statistics Sweden and averaged for 2004 –7. An indicator variable that equals one if an individual’s spouse receives payments due to unemployment and zero otherwise. The data are obtained annually from Statistics Sweden and averaged for 2004 –7. For unmarried individuals, the variable equals zero. Years of education is based on the highest completed degree. For a subset of the sample, the variable is obtained from the STR. We use a linear regression model to extend the variable to the rest of our sample. Specifically, we regress the years of education onto indicator variables high school and college or more ðavailable for most individuals in our data set from Statistics SwedenÞ and then predict years of education for those for whom years of education is missing. An indicator that equals one if at least one parent is still alive in 2003, otherwise zero. The data are obtained from Statistics Sweden. Inheritance is calculated by identifying cases in which the last parent dies between the end of 2003 and 2007 and dividing the parent’s net worth at the end of 2003 ðif positiveÞ by the number of children ði.e., two twins plus number of siblingsÞ. The variable equals zero in all other cases. When reported in US$, SEK amounts have been converted at SEK/US$ 6.9753, the average end-of-year exchange rate for 2004 –7. All data are from Statistics Sweden. The average disposable income of the individual for the period 2004 –7, as defined by Statistics Sweden, which is the sum of income from labor, business, and investment, plus received transfers, less taxes and alimony payments. When reported in US$, SEK amounts have been converted at SEK/US$ 6.9753, the average end-of-year exchange rate for 2004 –7. The data are obtained from Statistics Sweden. The average disposable income of the individual and, if married, her spouse for the period 2004 –7, as defined by Statistics Sweden, i.e., the sum of income from labor, business, and investment, plus received transfers, less taxes and alimony payments. The data are obtained from Statistics Sweden.
Variable
162
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
Other financial assets/assets ðexcluding housingÞ Real assets ðexcluding housingÞ/assets ðexcluding housingÞ
Cash/assets ðexcluding housingÞ
Bonds/assets ðexcluding housingÞ
Equity/assets ðexcluding housingÞ
Asset allocation: Business owner
Change in home value
Change in net worth
Net worth
Standard deviation of log growth rate of disposable household income
Variable
Description
An indicator that equals one if in at least 1 year between 2004 and 2007 an individual or her spouse has income from active business activity that exceeds 50 percent of the labor income. The indicator equals zero otherwise. Income data are obtained from Statistics Sweden. The amount of direct and indirect equity investments scaled by all assets ðexcluding home assetsÞ. The data are obtained annually from Statistics Sweden and averaged for 2004 –7. The amount of fixed income investments scaled by all assets ðexcluding home assetsÞ. The data are obtained annually from Statistics Sweden and averaged for 2004 –7. The amount held in bank accounts and money market funds scaled by all assets ðexcluding home assetsÞ. The data are obtained annually from Statistics Sweden and averaged for 2004 –7. The amount invested in other financial assets, such as derivatives and insurance products, scaled by all assets ðexcluding home assetsÞ. The data are obtained annually from Statistics Sweden and averaged for 2004 –7. The amount invested in real assets, such as vacation and rental properties, but excluding home assets scaled by all assets ðexcluding home assetsÞ. The data are obtained annually from Statistics Sweden and averaged for 2004 –7.
The time-series standard deviation of the log growth rate of disposable household income between 2000 and 2007. The variable is missing if four or more of the log growth rates are missing. The top and bottom one percent of the log growth rate distribution are set to missing. The average difference between the market value of an individual’s assets and her liabilities, calculated by Statistics Sweden at the end of each year between 2004 and 2007. When reported in US$, SEK amounts have been converted at SEK/US$ 6.9753, the average end-of-year exchange rate for 2004 –7. The data are obtained from Statistics Sweden. The change in an individual’s net worth between the end of 2003 and the end of 2007. When reported in US$, SEK amounts have been converted at SEK/US$ 6.9753, the average end-of-year exchange rate for 2004 –7. The data are obtained from Statistics Sweden. Capital gains and losses associated with the owner-occupied house between the end of 2003 and the end of 2007 When reported in US$, SEK amounts have been converted at SEK/US$ 6.9753, the average end-ofyear exchange rate for 2004 –7. The data are obtained from Statistics Sweden.
TABLE A1 (Continued )
TABLE A2 Controlling for Socioeconomic Characteristics and Asset Allocation
Intercept Male Age Age2 Age ðspouseÞ Married Divorced
Young ð18 – 35Þ
Middle Age ð36 – 50Þ
Older ð51– 65Þ
2.6719 ð1.0379Þ .0453 ð.0327Þ .1535** ð.0715Þ 2.0026** ð.0012Þ .0000 ð.0038Þ .1379 ð.1166Þ 2.1412 ð.3600Þ 2.1450 ð.0905Þ .0180 ð.0338Þ 2.0467 ð.0867Þ 2.2343 ð.2854Þ .0045 ð.0553Þ 2.0033 ð.2128Þ 2.0526*** ð.0195Þ .0766 ð.0506Þ 2.0177 ð.0252Þ
2.6252 ð.9645Þ 2.0037 ð.0171Þ .0093 ð.0425Þ 2.0001 ð.0005Þ 2.0003 ð.0009Þ 2.1034*** ð.0319Þ 2.1042*** ð.0348Þ 2.0275 ð.1654Þ 2.0328 ð.0334Þ 2.0177 ð.0112Þ 2.0900*** ð.0323Þ 2.1027** ð.0487Þ 2.0274 ð.0343Þ 2.0747 ð.0538Þ 2.0180** ð.0071Þ 2.0087 ð.0245Þ .0144*** ð.0042Þ
22.8047** ð1.0936Þ 2.0224** ð.0105Þ .1340*** ð.0366Þ 2.0011*** ð.0003Þ .0002 ð.0004Þ 2.1042*** ð.0189Þ 2.1571*** ð.0179Þ 2.1847*** ð.0323Þ 2.0304 ð.0235Þ 2.0079 ð.0138Þ 2.1066*** ð.0140Þ 2.0269 ð.0207Þ 2.044** ð.0193Þ 2.0395 ð.0288Þ 2.0132*** ð.0037Þ .0198* ð.0120Þ .0078*** ð.0017Þ
2.0931** ð.0445Þ
.0758*** ð.0271Þ
2.0304** ð.0135Þ
.2514** ð.1240Þ .1997** ð.1013Þ 2.5648*** ð.0837Þ 2.7458*** ð.2510Þ 2.5802*** ð.0854Þ
.1910*** ð.0689Þ .2289*** ð.0358Þ 2.3943*** ð.0280Þ 2.3615*** ð.1247Þ 2.4961*** ð.0306Þ
.1594*** ð.0427Þ .2295*** ð.0220Þ 2.4069*** ð.0182Þ 2.3657*** ð.0702Þ 2.5328*** ð.0189Þ
2.7108*** ð.1145Þ Yes
2.4936*** ð.0449Þ Yes
2.4478*** ð.0299Þ Yes
Widowed Children indicator Number of children ð4 - year averageÞ Poor health ð4 - year averageÞ Poor health ðspouseÞ ð4 - year averageÞ Unemployed ð4 - year averageÞ Unemployed ðspouseÞ ð4 - year averageÞ Number of siblings ParentðsÞ still alive ð2003Þ Inheritance ðlnÞ Disposable household income ðln, 4 - year averageÞ Standard deviation of log growth rate of disposable household income Business owner Equity/assets ðexcluding housingÞ Bonds/assets ðexcluding housingÞ Cash/assets ðexcluding housingÞ Other financial assets/assets ðexcluding housingÞ Regional fixed effects ð21 regionsÞ
163
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
164
journal of political economy TABLE A2 (Continued )
Observations R2
Young ð18 – 35Þ
Middle Age ð36 – 50Þ
Older ð51– 65Þ
1,542 .09
3,678 .14
9,710 .14
Note.—The table reports ordinary least squares estimates and standard errors for three linear regressions ðone per age groupÞ of the savings rate onto an individual’s socioeconomic characteristics and asset allocation choices. In addition to the variables listed in the table, we include 20 regional fixed effects in all regressions. R 2 is the coefficient of determination. * Significant at the 10 percent level. ** Significant at the 5 percent level. *** Significant at the 1 percent level. TABLE A3 Age Effects
Model Parameters Mean: Intercept Age Age2 Variance: Genetic component: a0 a1 a2 Common environment: c0 c1 c2 Individual-specific environment: e0 e1 e2
Estimate
Standard Error
2.002 .010 2.013
.086 .367 .379
21.123*** 5.824*** 25.893***
.302 1.098 1.002
1.233*** 24.196*** 3.495***
.234 1.193 1.281
.678*** 21.045*** .997***
.068 .310 .336
Note.—The table reports parameter estimates and standard errors from maximum likelihood estimation of a random-effects model in which variance components vary with age and age2. We model the variance of the genetic effect, VarðaÞ, as ða 0 1 a 1age 1 a 2age2Þ2, the variance of the common ðparentingÞ environment, VarðcÞ, as ðc 0 1 c1age 1 c 2age2Þ2, and the variance of the individual-specific environments, Var(e), as (e0 + e1age + e2age2)2. Age is measured in units of 100 years. N 5 14,930. * Significant at the 10 percent level. ** Significant at the 5 percent level. *** Significant at the 1 percent level.
References Alan, S., O. Attanasio, and M. Browning. 2009. “Estimating Euler Equations with Noisy Data: Two Exact GMM Estimators.” J. Appl. Econometrics 24:309–24. Ameriks, J., A. Caplin, and J. Leahy. 2003. “Wealth Accumulation and the Propensity to Plan.” Q.J.E. 118 ð3Þ: 1007–47.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
origins of savings behavior
165
Ashenfelter, O., and A. B. Krueger. 1994. “Estimates of the Economic Returns to Schooling from a New Sample of Twins.” A.E.R. 84 ð5Þ: 1157–73. Barkley, R. A. 1997. ADHD and the Nature of Self-Control. New York: Guilford. Barnea, A., H. Cronqvist, and S. Siegel. 2010. “Nature or Nurture: What Determines Investor Behavior?” J. Financial Econ. 98:583–604. Barsky, R. B., F. T. Juster, M. S. Kimball, and M. D. Shapiro. 1997. “Preference Parameters and Behavioral Heterogeneity: An Experimental Approach in the Health and Retirement Study.” Q.J.E. 112:537–79. Becker, G. 1976. “Altruism, Egoism, and Genetic Fitness: Economics and Sociobiology.” J. Econ. Literature 14:817–26. Becker, G., and C. Mulligan. 1997. “The Endogenous Determination of Time Preference.” Q.J.E. 112 ð3Þ: 729–58. Becker, G., and N. Tomes. 1976. “Child Endowments and the Quantity and Quality of Children.” J.P.E. 84 ð4Þ: 143–62. Behrman, J., and P. Taubman. 1989. “Is Schooling ‘Mostly in the Genes’? NatureNurture Decomposition Using Data on Relatives.” J.P.E. 97:1425–46. Benartzi, S., and R. H. Thaler. 2007. “Heuristics and Biases in Retirement Savings Behavior.” J. Econ. Perspectives 21:81–104. Benjamin, D., et al. 2008. “Genoeconomics.” In Biosocial Surveys, edited by Maxine Weinstein, James W. Vaupel, and Kenneth W. Wachter, 304–35. Washington, DC: Nat. Acad. Press. Bernheim, B. D. 2009. “On the Potential of Neuroeconomics: A Critical ðbut HopefulÞ Appraisal.” American Econ. J.: Microeconomics 1 ð2Þ: 1–41. Bernheim, B. D., D. M. Garrett, and D. M. Maki. 2001. “Education and Saving: The Long-Term Effects of High School Financial Curriculum Mandates.” J. Public Econ. 80:435–65. Bernheim, B. D., J. Skinner, and S. Weinberg. 2001. “What Accounts for the Variation in Retirement Wealth among U.S. Households?” A.E.R. 91 ð4Þ: 832– 57. Bikhchandani, S., D. Hirshleifer, and I. Welch. 1992. “A Theory of Fads, Fashion, Custom, and Cultural Change as Informational Cascades.” J.P.E. 100 ð5Þ: 992– 1026. Bisin, A., G. Topa, and T. Verdier. 2004. “Religious Intermarriage and Socialization in the United States.” J.P.E. 112 ð3Þ: 615–64. Bisin, A., and T. Verdier. 2001. “The Economics of Cultural Transmission and the Dynamics of Preferences.” J. Econ. Theory 97 ð2Þ: 298–319. ———. 2008. “Cultural Transmission.” In The New Palgrave Dictionary of Economics, edited by S. N. Durlauf and L. E. Blume. Basingstoke, UK: Palgrave Macmillan. Black, S. E., P. J. Devereux, and K. G. Salvanes. 2005. “Why the Apple Doesn’t Fall Far: Understanding Intergenerational Transmission of Human Capital.” A.E.R. 95 ð1Þ: 437–49. Borjas, G. J. 1992. “Ethnic Capital and Intergenerational Mobility.” Q.J.E. 107 ð1Þ: 123–50. Bouchard, T. J. 1998. “Genetic and Environmental Influences on Adult Intelligence and Special Mental Abilities.” Human Biology 70:257–79. Bouchard, T. J., D. T. Lykken, M. McGue, N. L. Segal, and A. Tellegen. 1990. “Sources of Human Psychological Differences: The Minnesota Study of Twins Reared Apart.” Science 250:223–28. Bowles, S., and H. Gintis. 2002. “The Inheritance of Inequality.” J. Econ. Perspectives 16 ð3Þ: 3–30. Boyd, R., and P. Richerson. 1985. Culture and the Evolutionary Process. Chicago: Univ. Chicago Press.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
166
journal of political economy
Brennan, T. J., and A. W. Lo. 2011. “The Origin of Behavior.” Q. J. Finance 1:55– 108. Bronfenbrenner, U., and S. J. Ceci. 1994. “Nature-Nurture in Developmental Perspective: A Bioecological Theory.” Psychological Rev. 101:568–86. Brown, J. R., Z. Ivkovic´, P. A. Smith, and S. Weisbenner. 2008. “Neighbors Matter: Causal Community Effects and Stock Market Participation.” J. Finance 63:1509–32. Browning, M., F. Bourguignon, P.-A. Chiappori, and V. Lechene. 1994. “Income and Outcomes: A Structural Model of Intrahousehold Allocation.” J.P.E. 102:1067–96. Browning, M., and P.-A. Chiappori. 1998. “Efficient Intra-household Allocations: A General Characterization and Empirical Tests.” Econometrica 66:1241– 78. Browning, M., and A. Lusardi. 1996. “Household Saving: Micro Theories and Micro Facts.” J. Econ. Literature 34:1797–1855. Calvet, L. E., J. Y. Campbell, and P. Sodini. 2007. “Down or Out: Assessing the Welfare Costs of Household Investment Mistakes.” J.P.E. 115:707–47. Camerer, C., G. Loewenstein, and D. Prelec. 2005. “Neuroeconomics: How Neuroscience Can Inform Economics.” J. Econ. Literature 43:9 –64. Carpenter, J. P., J. R. Garcia, and J. K. Lum. 2011. “Dopamine Receptor Genes Predict Risk Preferences, Time Preferences, and Related Economic Choices.” J. Risk and Uncertainty 42 ð3Þ: 233–61. Carroll, C. D. 1992. “The Buffer-Stock Theory of Saving: Some Macroeconomic Evidence.” Brookings Papers Econ. Activity 23 ð2Þ: 61–156. Cavalli-Sforza, L., and M. Feldman. 1981. Cultural Transmission and Evolution: A Quantitative Approach. Princeton, NJ: Princeton Univ. Press. Cesarini, D., C. T. Dawes, M. Johannesson, P. Lichtenstein, and B. Wallace. 2009. “Genetic Variation in Preferences for Giving and Risk Taking.” Q.J.E. 124:809– 42. Cesarini, D., M. Johannesson, P. Lichtenstein, O. Sandewall, and B. Wallace. 2010. “Genetic Variation in Financial Decision Making.” J. Finance 65:1725–54. Charles, K. K., and E. Hurst. 2003. “The Correlation of Wealth across Generations.” J.P.E. 111 ð6Þ: 1155–82. Chetty, R., J. N. Friedman, N. Hilger, E. Saez, D. W. Schanzenbach, and D. Yagan. 2011. “How Does Your Kindergarten Classroom Affect Your Earnings? Evidence from Project STAR.” Q.J.E. 126 ð4Þ: 1593–1660. Chiteji, N. S., and F. P. Stafford. 1999. “Portfolio Choices of Parents and Their Children as Young Adults: Asset Accumulation by African-Americans.” A.E.R. 89:377–80. Clark, G. 2014. The Son Also Rises: Surnames and the History of Social Mobility. Princeton, NJ: Princeton Univ. Press. Cronqvist, H., and R. H. Thaler. 2004. “Design Choices in Privatized SocialSecurity Systems: Learning from the Swedish Experience.” A.E.R. 94 ð2Þ: 424– 28. Cunha, F., and J. Heckman. 2010. “Investing in Our Young People.” In Cost Effective Early Childhood Programs in the First Decade: A Human Capital Integration, edited by A. Reynolds. Cambridge: Cambridge Univ. Press. Davies, G., et al. 2011. “Genome-Wide Association Studies Establish That Human Intelligence Is Highly Heritable and Polygenic.” Molecular Psychiatry 16 ð10Þ: 996–1005. Dreber, A., C. L. Apicella, D. T. A. Eisenberg, J. R. Garcia, and R. S. Zamore. 2009. “The 7R Polymorphism in the Dopamine Receptor D4 Gene ðDRD4Þ Is Asso-
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
origins of savings behavior
167
ciated with Financial Risk-Taking in Men.” Evolution and Human Behavior 30: 85–92. Duflo, E., W. Gale, J. Liebman, P. Orszag, and E. Saez. 2006. “Saving Incentives for Low- and Middle-Income Families: Evidence from a Field Experiment with H&R Block.” Q.J.E. 121 ð4Þ: 1311–46. Dynan, K. E., J. Skinner, and S. P. Zeldes. 2004. “Do the Rich Save More?” J.P.E. 112 ð2Þ: 397–444. Eisenberg, D., et al. 2007. “Examining Impulsivity as an Endophenotype Using a Behavioral Approach: A DRD2 TaqI A and DRD4 48-bp VNTR Association Study.” Behavioral and Brain Functions 3:2. Falconer, D. S., and T. F. C. Mackay. 1996. Introduction to Quantitative Genetics. 4th ed. Essex, UK: Longman. Fernandez, R., A. Fogli, and C. Olivetti. 2004. “Mothers and Sons: Preference Formation and Female Labor Force Dynamics.” Q.J.E. 119 ð4Þ: 1249–99. Friedman, M. 1953. “Choice, Chance, and the Personal Distribution of Income.” J.P.E. 61 ð4Þ: 277–90. Goldberger, A. 1979. “Heritability.” Economica 46 ð184Þ: 327–47. Gu¨ell, M., J. V. Rodrı´guez Mora, and C. Telmer. 2013. “Intergenerational Mobility and the Informative Content of Surnames.” Working paper, Tepper School Bus., Carnegie Mellon Univ. Harden, K., E. Turkheimer, and J. Loehlin. 2007. “Genotype by Environment Interaction in Adolescents’ Cognitive Aptitude.” Behavior Genetics 37:273–83. Harrison, G. W., M. I. Lau, and M. B. Williams. 2002. “Estimating Individual Discount Rates in Denmark: A Field Experiment.” A.E.R. 92 ð5Þ: 1606–17. Hirshleifer, J. 1977. “Economics from a Biological Viewpoint.” J. Law and Econ. 20:1–52. Hong, H., J. D. Kubik, and J. C. Stein. 2004. “Social Interaction and Stock-Market Participation.” J. Finance 59 ð1Þ: 137–63. Hubbard, R., J. Skinner, and S. Zeldes. 1995. “Precautionary Saving and Social Insurance.” J.P.E. 103:360–99. Humphreys, K., et al. 2011. “The Genetic Structure of the Swedish Population.” PLoS ONE 6 ð8Þ. Hur, Y.-M. 2007. “Effects of the Chorion Type on Prosocial Behavior in Young South Korean Twins.” Twin Res. and Human Genetics 10 ð5Þ: 773–77. Jian, Y., et al. 2010. “Common SNPs Explain a Large Proportion of the Heritability for Human Height.” Nature Genetics 42 ð7Þ: 565–69. Kendler, K. S., M. C. Neale, R. C. Kessler, A. C. Heath, and L. J. Eaves. 1994. “Parental Treatment and the Equal Environment Assumption in Twin Studies of Psychiatric Illness.” Psychological Medicine 24:579–90. Kimball, M. 1990. “Precautionary Saving in the Small and in the Large.” Econometrica 58 ð1Þ: 53–73. Knudsen, E. I., J. J. Heckman, J. L. Cameron, and J. P. Shonkoff. 2006. “Economic, Neurobiological, and Behavioral Perspectives on Building America’s Future Workforce.” Proc. Nat. Acad. Sci. 103 ð27Þ: 10155–62. Kuhnen, C. M., and J. Chiao. 2009. “Genetic Determinants of Financial Risk Taking.” PLoS ONE 4 ð2Þ. Lawrance, E. C. 1991. “Poverty and the Rate of Time Preference: Evidence from Panel Data.” J.P.E. 99 ð1Þ: 54–77. Lusardi, A. 1998. “On the Importance of the Precautionary Saving Motive.” A.E.R. 88 ð2Þ: 449–53. ———. 2001. “Explaining Why So Many People Do Not Save.” Working paper, Dept. Econ., Dartmouth Coll.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
168
journal of political economy
Madrian, B. C., and D. F. Shea. 2001. “The Power of Suggestion: Inertia in 401ðkÞ Participation and Savings Behavior.” Q.J.E. 116:1149–1525. Merton, R. C. 1969. “Lifetime Portfolio Selection under Uncertainty: The Continuous-Time Case.” Rev. Econ. and Statis. 51:247–57. Mischel, W., Y. Shoda, and M. Rodriguez. 1989. “Delay of Gratification in Children.” Science 244 ð4907Þ: 933–38. Modigliani, F., and R. Brumberg. 1954. “Utility Analysis and the Consumption Function: An Interpretation of Cross-Section Data.” In Post-Keynesian Economics, edited by K. K. Kurihara. New Brunswick, NJ: Rutgers Univ. Press. Mulligan, C. 1999. “Galton versus the Human Capital Approach to Inheritance.” J.P.E. 107:184–224. Neale, M. C., and H. H. M. Maes. 2004. Methodology for Genetic Studies of Twins and Families. NATO ASI Series D: Behavioural and Social Science. Dordrecht, Neth.: Kluwer. Nelis, M., et al. 2009. “Genetic Structure of Europeans: A View from the NorthEast.” PLoS ONE 4 ð5Þ. Netzer, N. 2009. “Evolution of Time Preferences and Attitudes toward Risk.” A.E.R. 99:937–55. O’Donoghue, T., and M. Rabin. 1999. “Doing It Now or Later.” A.E.R. 89:103–24. Phillips, D. I. W. 1993. “Twin Studies in Medical Research: Can They Tell Us Whether Diseases Are Genetically Determined?” Lancet 341 ð8851Þ: 1008–9. Plomin, R., J. C. Defries, G. E. McClearn, and P. McGuffin. 2008. Behavioral Genetics. New York: Worth. Poterba, J., S. Venti, and D. Wise. 2007. “The Decline of Defined Benefit Retirement Plans and Asset Flows.” Working paper, NBER, Cambridge, MA. Prescott, C. A., R. C. Johnson, and J. J. McArdle. 1999. “Chorion Type as a Possible Influence on the Results and Interpretation of Twin Study Data.” Twin Res. 2:244–49. Richerson, P., and R. Boyd. 2005. Not by Genes Alone. Chicago: Univ. Chicago Press. Robson, A. J. 1996. “The Evolution of Attitudes to Risk: Lottery Tickets and Relative Wealth.” Games and Econ. Behavior 14:190–207. ———. 2001. “Why Would Nature Give Individuals Utility Functions?” J.P.E. 109:900–914. Robson, A. J., and L. Samuelson. 2009. “The Evolution of Time Preference with Aggregate Uncertainty.” A.E.R. 99 ð5Þ: 1925–53. Rogers, A. R. 1994. “Evolution of Time Preference by Natural Selection.” A.E.R. 84 ð3Þ: 460–81. Rutter, M. 2006. Genes and Behavior: Nature/Nurture Interplay Explained. Oxford: Blackwell. Samuelson, P. A. 1969. “Lifetime Portfolio Selection by Dynamic Stochastic Programming.” Rev. Econ. and Statis. 51:239–46. Segal, N. L. 2013. “Personality Similarity in Unrelated Look-Alike Pairs: Addressing a Twin Study Challenge.” Personality and Individual Differences 54:23– 28. Shiller, R. J. 1995. “Conversation, Information, and Herd Behavior.” A.E.R. 85 ð2Þ: 181–85. Skinner, J. 1988. “Risky Income, Life Cycle Consumption, and Precautionary Savings.” J. Monetary Econ. 22 ð2Þ: 237–55. Sokol, D., C. Moore, R. Rose, C. Williams, T. Reed, and J. Christian. 1995. “Intrapair Differences in Personality and Cognitive Ability among Young Monozygotic Twins Distinguished by Chorion Type.” Behavior Genetics 25:457–66.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
origins of savings behavior
169
Solon, G. 1992. “Intergenerational Income Mobility in the United States.” A.E.R. 82 ð3Þ: 393–408. Taubman, P. 1976. “The Determinants of Earnings: Genetics, Family, and Other Environments; a Study of White Male Twins.” A.E.R. 66 ð5Þ: 858–70. Taylor, J., A. D. Roehrig, B. S. Hensler, C. M. Connor, and C. Schatschneider. 2010. “Teacher Quality Moderates the Genetic Effects on Early Reading.” Science 328 ð5977Þ: 512–14. Thaler, R. H. 1994. “Psychology and Savings Policies.” A.E.R. 84 ð2Þ: 186–92. Thaler, R. H., and S. Benartzi. 2004. “Save More Tomorrow: Using Behavioral Economics to Increase Employee Savings.” J.P.E. 112:164–87. Thaler, R. H., and H. M. Shefrin. 1981. “An Economic Theory of Self-Control.” J.P.E. 89 ð2Þ: 392–406. Venti, S. F., and D. A. Wise. 1998. “The Cause of Wealth Dispersion at Retirement: Choice or Chance?” A.E.R. 88 ð2Þ: 185–91. Warner, J. T., and S. Pleeter. 2001. “The Personal Discount Rate: Evidence from Military Downsizing Programs.” A.E.R. 91 ð1Þ: 33–53. Wooldridge, J. M. 2001. Econometric Analysis of Cross Section and Panel Data. Cambridge, MA: MIT Press. Zhong, S., S. Israel, H. Xue, R. P. Ebstein, and S. H. Chew. 2009. “Monoamine Oxidase A Gene ðMAOAÞ Associated with Attitude Towards Longshot Risks.” PLoS ONE 4 ð12Þ.
This content downloaded from 116.246.20.194 on Tue, 10 Feb 2015 03:50:03 AM All use subject to JSTOR Terms and Conditions
