Download Smart Money: The Effect of Education on Financial Behavior

Smart Money: The E¤ect of Education on Financial Behavior
Shawn Cole, Anna Paulson, and Gauri Kartini Shastry
March 2012
Abstract
Household …nancial decisions are important for both households and the greater
economy. Yet, our understanding of the process of …nancial decision-making is limited. Applying standard and two-sample instrumental variables strategies to census
and credit bureau data, we provide the …rst precise, causal estimates of the e¤ects
of education on …nancial behavior. Education has large e¤ects on …nancial market participation and smaller, but statistically and economically signi…cant e¤ects
on …nancial management. We …nd that education improves credit scores, and dramatically reduces the probability of declaring bankruptcy or su¤ering foreclosure
during the …nancial crisis. Examining mechanisms, we show that cognitive ability increases …nancial participation, and discuss how education may a¤ect decisionmaking through: attitudes, borrowing behavior, discount rates, risk-aversion, and
the in‡uence of coworkers and neighbors.
Harvard Business School ([email protected]), Federal Reserve Bank of Chicago ([email protected]), and
Wellesley College ([email protected]), respectively. We thank the editor, associate editor, two referees, and
Josh Angrist, Malcolm Baker, Daniel Bergstresser, Carol Bertaut, David Cutler, Robin Greenwood, Campbell
Harvey, Caroline Hoxby, Michael Kremer, Annamaria Lusardi, Erik Sta¤ord, Jeremy Tobacman, Petia Topalova,
Peter Tufano, and workshop participants at Harvard, the Federal Reserve Board of Governors, the University of
Virginia and the Boston Federal Reserve for comments and suggestions. Paymon Khorrami and Wentao Xiong
provided excellent research assistance. The views presented in this paper are those of the authors and do not
necessarily re‡ect those of the Federal Reserve Bank of Chicago.
1
Introduction
Individuals face an increasingly complex menu of …nancial products. On the asset side of the
balance sheet, the shift from de…ned bene…t to de…ned contribution pension plans, and the
growing importance of private retirement accounts, require individuals to choose the amount
they save, as well as the mix of assets in which they invest. Yet, participation in …nancial
markets is far from universal in the United States, and we have only a limited understanding
of what factors in‡uence participation. On the liabilities side, a dramatic increase in the range
and complexity of credit products to households has been accompanied by increased default,
bankruptcy, and foreclosures. These trends have sparked a vigorous debate about whether
individuals are well-equipped to make informed …nancial decisions.
Using data and estimation techniques new to the literature, this paper provides precise,
causal estimates of the e¤ect of education on …nancial market participation and …nancial management. We exploit exogenous variation in education caused by changes in compulsory schooling laws. We …nd an additional year of education increases the probability of …nancial market
participation by 7-8 percentage points, holding constant other factors, including income. The
size of this e¤ect is economically important both on its own and in the context of previously
identi…ed correlates of …nancial participation, such as trust (Guiso, Sapienza, and Zingales,
2008), peer e¤ects (Hong, Kubik, and Stein, 2004), prior life experience with the stock market
(Malmendier and Nagel, 2011), or institutional quality (Osili and Paulson, 2008).
To study the e¤ect of education on other aspects of …nancial behavior, we employ a twosample instrumental variables strategy using the same compulsory schooling laws, together with
a new data set on consumer credit behavior. We …nd that exogenous increases in education lead to
higher credit scores, more responsible …nancial behavior (fewer delinquent credit-card payments),
and importantly, substantial reductions in the probability of bankruptcy and foreclosure. This
e¤ect is particularly pronounced during the recent …nancial crisis.
We then explore why education in‡uences …nancial behavior. Agarwal et al (2007) and
1
Agarwal and Mazumder (2010) demonstrate the importance of cognitive ability for sound …nancial decision making. By exploiting within-sibling group variation in cognitive ability, we
show that indeed higher levels of cognitive ability lead to greater …nancial market participation. Importantly, these estimates are not confounded by unobserved background and family
characteristics.
Finally, we describe ways in which education might a¤ect …nancial behavior. We …nd education e¤ects a measure of con…dence; borrowing decisions (such as whether to take a second
mortgage); the probability one has a pension, through occupational choices; and the type of
peers one has, through residential choices. Education does not a¤ect the probability of moving
to another city, which may correlate with willingness to take risks.
Financial management is important for many reasons. For the household, it facilitates asset
accumulation and consumption smoothing, with potentially signi…cant e¤ects on welfare. For
the …nancial system as a whole, the depth and breadth of …nancial market participation are
important determinants of the equity premium, the volatility of markets, and household expenditure (Mankiw and Zeldes, 1991; Heaton and Lucas, 1999; Vissing-Jorgensen, 2002; and Brav,
Constantinides, and Gezcy, 2002). Financial behavior may also a¤ect the political economy
of …nancial regulation, as those holding …nancial assets may have di¤erent attitudes towards
corporate and investment income tax policy, as well as di¤erent attitudes towards risk-sharing
and redistribution.
Several aspects of …nancial behavior, such as limited equity market participation, low savings
rates, and, more recently, a high incidence of bankruptcy and foreclosures have drawn attention
from economists as potentially inconsistent with standard models of optimizing behavior. While
survey evidence has proven useful in demonstrating what factors are correlated with such behaviors1 , there is much less understanding of what the causal drivers are. This paper contributes
1
Previous work has demonstrated that …nancial behavior is, not surprisingly, correlated with income, as well as
education (Bertaut and Starr-McCluer, 2001, among others), measured …nancial literacy (Lusardi and Mitchell,
2007), social connections (Hong, Kubik, and Stein, 2004), trust (Guiso, Sapienza, and Zingales, 2008), experience
with the stock market (Malmendier and Nagel, 2011), and cognitive ability (Grinblatt, Keloharju, and Linnainmaa,
2011).
2
to the literature by demonstrating an important causal determinant of …nancial behaviors that
have been poorly understood.
In 2004 only 48.6% of households held stocks, either directly or indirectly (Bucks, Kennickell, and Moore, 2006). Some view this limited participation in the stock market as a puzzle:
Haliassos and Bertaut (1995) consider and reject risk aversion, belief heterogeneity, and other
potential explanations, instead favoring “departures from expected-utility maximization.” Our
paper shows that low levels of education may help explain the limited participation puzzle.
At the lower end of the income spectrum, economists have focused on individuals’low savings
rates and propensity to declare bankruptcy, and take on mortgages they cannot repay. Gross
and Souleles (2002) note individuals borrow from credit cards when holding large bank account
balances. Stango and Zinman (2009) show households systematically underestimate the returns
to saving. Lusardi et al. (2011) …nd that a quarter of Americans would be unable to come up
with $2,000 if needed within 30 days. Our second set of results sheds light on this …nancial
management puzzle. We show exogenous increases in education improve individuals’ credit
scores. More educated individuals pay o¤ a greater share of their outstanding credit balance,
are less likely to be delinquent on their credit card bills, and are less likely to declare bankruptcy
or experience a foreclosure.
2
Data
This paper uses three complementary data sets: the U.S. Census, the Federal Reserve Bank of
New York Consumer Credit Panel/Equifax dataset, and the National Longitudinal Survey of
Youth (NLSY). Summary statistics for the …rst two are presented in Table I, while summary
statistics for the NLSY are presented in Online Appendix Table A1.
We use a 5 percent sample from the 1980, 1990, and 2000 Public Use Census Data2 , representing a random draw of the US population. The key advantage of this data set is its size: with
2
The 2010 census did not include a “long form,” and hence does not have information on investment income.
3
over 14 million observations, we can use non-parametric controls, obtain precise estimates, and
most importantly, use an instrumental variable strategies that would simply not be possible in
most other, smaller, data sets.
The Census does not collect any information on …nancial wealth, and is very rarely used to
study …nancial decision-making (an exception is Carroll, Rhee and Rhee, 1999). However, it does
collect detailed income data. Thus, the main measure of …nancial market participation we will
use is “income from interest, dividends, net rental income, royalty income, or income from estates
and trusts,”received during the previous year, which we term “investment income.”Households
are instructed to “report even small amounts credited to an account.” (Ruggles et al., 2004). A
second type of income we use is “retirement, survivor, or disability pensions,” received during
the previous year, which we term “retirement income.”This is distinct from Social Security and
Supplemental Security Income, both of which are reported on separate lines.3
A limitation of using the amount of investment income, rather than the amount invested,
is that it is only partially informative about the amount and type of investments held by the
respondent. This would make it di¢ cult to use the data for structural estimates of investment
levels (such as calibrating models of participation costs). In this work, however, we focus primarily on the decision to participate in …nancial markets, for which we de…ne a dummy variable
equal to one if the household reports any non-zero investment income. Approximately 22% of
respondents do so, which is close to the 21.3% of families that report holding equity in the 2001
Survey of Consumer Finances (Bucks, Kennickell, and Moore, 2006), but lower than the 33% of
households reporting any investment income in the 2001 SCF.
In online appendix tables A2 and A3, we compare our census data to data from the SCF.
We …nd the census data yield very similar estimates of means, medians and percentiles for
our measures of participation and investment and retirement income. Regressions of investment
3
One may be concerned that small amounts of investment income simply represent interest from savings
accounts. As a robustness check, we rerun our analysis considering only those who receive income greater than
$500 (or, alternatively, $1000) in investment income as "participating." The results are very similar (available on
request).
4
income and …nancial market participation on individual characteristics in the two data sets yield
strikingly similar coe¢ cients. Finally, we explore the relationship between reported investment
income and more traditional measures of …nancial market participation. In particular, we …nd a
large jump in the use of transactions accounts as individuals move from 0 to any positive amount
of investment income. For example, 78.4% of households reporting no investment income possess
a checking account, while 91.9% of those reporting income between 1 and 100 dollars have
checking accounts. There is a similar, strongly positive and nearly linear relationship between
reported investment income and participation in equity markets.
The FRBNY Consumer Credit Panel/Equifax dataset is a quarterly longitudinal panel of
individual credit bureau data, similar to information that would be contained in an individual’s
credit report. It is described in detail in Lee and van der Klaauw (2010). The panel begins in
the …rst quarter of 1999 and continues to the third quarter of 2011. The primary sample is made
up of a random 5% sample of all U.S. residents aged 18 and up who have a credit report. The
sample selection procedures ensure that, in any given quarter, there is a nationally representative
cross-section, conditional on having a credit report. We restrict attention to individuals who are
aged 36 to 75 in the third quarter of 2000, matching our census sample. Ultimately, we have
approximately 5 million observations per quarter.
We focus on …ve key outcome variables from this data set: credit score, the proportion of
an individual’s credit card debt that is not delinquent, the proportion of quarters in which an
individual has any delinquent credit card balance, a bankruptcy indicator, and a foreclosure
indicator. The credit score, similar to a FICO score, predicts the likelihood of being 90 or
more days delinquent over the next 24 months. Credit scores range from 280 to 850, and
higher scores imply a lower probability of being seriously delinquent in the future. Both credit
score and the proportion of an individual’s credit card debt that is not delinquent are averaged
across all quarters. The bankruptcy and foreclosure variables indicate whether an individual has
undergone bankruptcy or foreclosure at least once, respectively, between 1992 and 2011. These
5
indicators are able to track bankruptcies and foreclosures back through 1992 because credit
bureaus maintain records on bankruptcy and foreclosure proceedings for 7 years.
The NLSY79 is a survey of 12,686 Americans aged 14 to 22 in 1979, with annual follow-ups
until 1994, and biennial follow-ups afterwards. Each survey contained detailed questions on the
individual’s savings decisions and accumulated assets. Importantly, in 1980 survey respondents
completed the Armed Services Vocational Aptitude Battery (ASVAB), a set of 10 exams that
measure cognitive ability and knowledge. More details on these data are included in an online
appendix.
3
The E¤ect of Education on Financial Market Participation
3.1
Empirical Strategy
Results from previous literature and simple regressions on our data strongly suggest that households with higher levels of education are more likely to participate in …nancial markets. Campbell
(2006), for example, notes that educated households in Sweden diversify their portfolios more
e¢ ciently.4 However, the simple relationship between …nancial decisions and education levels
omits many other important factors, such as ability or family background, that are also likely
to in‡uence …nancial decisions. In this section we exploit an instrumental variables strategy developed by Acemoglu and Angrist (2000) - changes in state compulsory education laws - which
provides exogenous variation in education. Revisions in state laws a¤ected individuals’education attainment, but are not correlated with individual ability, parental characteristics, or other
potentially confounding factors.
In particular, we follow the strategy laid out by Lochner and Moretti (2004, hereafter LM),
who use changes in schooling requirements between 1914 and 1978 to measure the e¤ect of
education on incarceration rates. The principal advantage of following LM closely is that they
4
At the same time, Tortorice (2012) …nds that education only slightly reduces the likelihood that individuals
make expectational errors regarding macroeconomic variables, and that these errors a¤ect buying attitudes and
…nancial decisions.
6
have conducted a battery of speci…cation checks, demonstrating the validity of using compulsory
schooling laws as a natural experiment. For example, LM show that there is no clear trend in
educational attainment in the years prior to changes in schooling laws and that compulsory
schooling laws do not a¤ect college attendance.
The structural equation of interest is the following,
yi =
+ si + Xi + "i
(1)
where si is years of education for individual i, and Xi is a set of controls, including age, gender,
race, state of birth, state of residence, census year, cohort of birth …xed e¤ects and a cubic
polynomial in earned income. Age e¤ects are de…ned as dummies for each 3-year age group
from 20 to 75, while year e¤ects are dummies for each census year. Following LM, we exclude
people born in Alaska and Hawaii but include those born in the District of Columbia; thus we
have 49 state of birth dummies, but 51 state of residence dummies. Again following LM, we
include state of birth dummies interacted with a dummy variable for cohorts born in the South
who turn 14 in or after 1958 to allow for the impact of the Brown vs. Board of Education
decision. Cohort of birth is de…ned as 10-year birth intervals. Standard errors are corrected for
intracluster correlation within state of birth
year of birth. The outcome variable is either
an indicator for having any investment or retirement income or the actual level of investment
or retirement income. When studying the amount of income, we drop observations that were
top-coded by the survey; in 1980 (1990; 2000) these individuals reported amounts greater than
$75,000 ($40,000; $50,000) for investment income and in 1990 (2000)5 these individuals reported
amounts greater than $52,000 ($30,000) for retirement income.6
5
The 1980 Census did not separate retirement income from other sources of income.
To preclude the possibility of revealing personal information, the Census “top-codes” values for very rich
individuals. Speci…cally, they replace the income variable for individuals with investment income or retirement
income above a year-speci…c limit with the median income of all individuals in that state earning above that
limit. The percentage of topcoded observations is very low: 0.47% for investment income and 0.22% for retirement
income. Of course, using as a dependent variable “any investment income”avoids the top-coding problem entirely.
Nevertheless, as an alternative approach, we run Tobit regressions to account for top-coding, and …nd very similar
results (available upon request).
6
7
Following Acemoglu and Angrist (2000) and LM, we create dummy variables for whether
the years of required schooling are 8 or less, 9, 10, and 11 or more.7 These dummies are based
on the law in place in an individual’s state of birth when an individual turns 14 years of age.
As LM note, migration between birth and age 14 will add noise to this estimation, but the IV
strategy is still valid. The …rst stage for the IV strategy can then be written as
si =
+
9 Comp9
+
10 Comp10
+
11 Comp11
+ Xi + "i ;
(2)
where si is years of schooling, Comp9 - Comp11 are dummy variables that indicate the required
number of years of schooling that individual i was exposed to, and Xi is the same set of controls
de…ned above.
Compulsory schooling laws were changed numerous times from 1914 to 1978, even within a
state and not always in the same direction. We use data from the 1980, 1990, and 2000 censuses,
and focus on individuals who are between 18 and 75 years old, and born on or before 1964.8 The
census does not code a continuous measure of years of schooling, but rather identi…es categories
of educational attainment: preschool, grades 1-4, grades 5-8, grade 9, grade 10, grade 11, grade
12, 1-3 years of college, and college degree or more. We translate these categories into years of
schooling by assigning each range of grades the highest number of years of schooling for that
category. This should not a¤ect our estimates since individuals who fall within the ranges of
grades 1-8 and 1-3 years of college will not be in‡uenced by the compulsory schooling laws.
Finally, it is worth noting that the estimates produced here are Local Average Treatment
E¤ects, which measure the e¤ect of education on participation for those who were a¤ected by the
compulsory education laws.9 We note that those who are in fact a¤ected by the laws are likely
to have low levels of participation, and thus constitute a relevant study population. Moreover,
7
Speci…cally, we de…ne the years of mandated schooling as the di¤erence between the latest age an individual
is required to stay in school and earliest age she is required to enroll when states do not set the minimum required
years of schooling. When these two measures disagree, we take the larger value.
8
LM use the 1960, 1970, and 1980 censuses, which contain information on correctional facility residence, and
focus on a narrower age group, ages 20-60.
9
Imbens and Angrist (1994) provides a discussion of Local Average Treatment E¤ects.
8
we draw some comfort from Oreopoulos (2006), who studies a compulsory schooling reform that
a¤ected a very large fraction of the population in the United Kingdom. Studying the e¤ect of
education on earnings, Oreopoulos …nds the LATE e¤ect estimated in the United States from a
small fraction of the population is quite similar to the estimated e¤ect in the UK from a very
large fraction of the population.
3.2
Education and Financial Market Participation
OLS estimates of equation (1) are presented in Table II. Panel A presents the results for the
linear probability model, using “any income”as the dependent variable, and panel B studies the
level of total income, and includes the controls described in the previous section. The OLS results
have the expected signs: an additional year of education is associated with a 3.54 percentage
point higher probability of …nancial market participation, $271 more in investment income, and
$548 more in retirement income. We caution that these estimates are likely plagued by omitted
variables bias - educational attainment is correlated with unobserved individual characteristics
that may also a¤ect savings.
In Table III, we present the …rst stage, demonstrating that compulsory schooling laws did
increase human capital accumulation. Clearly, the state laws do in‡uence some individuals when states mandate a greater number of years of schooling, some individuals are required to
obtain more education than they otherwise would have acquired. A 9th year or 10th year of
mandated schooling increases average years of completed education by 0.2 years, while requiring
11 years of education increases education by 0.26 years. Requiring students to remain in school
for even one more year (9 years of required schooling) increases the probability of graduating
high school by 3.9%.10
Table IV presents 2SLS estimates of equation (1). Panel A reveals that an additional year
of schooling increases the probability of having any investment income by 7.5 percentage points.
10
‘Weak instruments’bias is not a problem in this context. We report the F-statistics of the excluded instruments
in Table IV. The F-statistics range from 44.5 to 49.9, well above the critical values proposed by Stock and Yogo
(2005).
9
For retirement investments, an additional year of schooling increases the probability of non-zero
income by about 5.9 percentage points. These estimates are somewhat larger than the OLS
estimates in Table II, suggesting a downward bias in the OLS.
In Panel B, we study the amount of income from these assets and …nd a large and signi…cant
e¤ect on both types of investment income. The magnitudes are substantially larger than the OLS
estimates: an additional year of schooling increases investment income and retirement income
by $1760 and $966 respectively.11 We …nd similar e¤ects if we use high school completion as the
measure of schooling. Including a cubic in earned income (which includes wages and income from
one’s own business or farm) as a control does not a¤ect the results appreciably. The striking
fact is that no matter what type of income control we include, we …nd a persistent and large
impact of education on participation.
Our analysis of the Survey of Consumer Finances suggests a magnitude in terms of equity
market participation. In the SCF data, an increase in investment income of $1,760 is associated
with a 5 percentage point increase in the probability of owning stocks.12
Another way to get at the economic importance of these estimates is to conduct the following
back-of-the-envelope calibration exercise. This calibration also helps us to understand the source
of the increase: does education raise investment earnings simply because households earn more
money, while keeping the fraction of income saved constant, or does it a¤ect the savings rate as
well?
The average individual in our sample is 49 years old. To simplify the algebra, we assume he
earned a constant $20,000 (the average income for high school graduates in our sample) since he
was 20 years old,13 saved a constant 10% of his income at the end of each year and earned a 5%
return on his assets. Assuming one additional year of schooling boosts wage income by 7% (an
estimate from Acemoglu and Angrist 2000), if the individual’s savings rate did not vary with
11
Using IV Tobit for investment income yields very similar results; results are available on request.
The coe¤ecient of a regression of equity ownership on investment income (regression not reported) is .0000287
(t-stat 3.1), and $1,760*.0000287=.055.
13
Using the average income at each age gives very similar estimates.
12
10
schooling, an additional year would increase his savings by $140 per year, though he would earn
income for one fewer year. At the age of 49, his accumulated savings would be $493 higher, and
his income from these assets approximately $25 greater.14
In contrast, if we assume that the year of education also increased our hypothetical individual’s income by 7 percent and his savings rate by 2 percentage points, an additional year
of schooling would increase his annual savings contribution by $568, yielding an approximately
$27,167 greater asset base by age 49, and a corresponding increase in investment income of
$1,358.15
The point estimates on investment income, $1,759 per year, are much closer to this latter
…gure, suggesting education increased the savings rate. Finally, it is also possible that education
a¤ects the choice of asset allocation: better educated individuals may choose portfolios that
yield higher returns, perhaps with lower fees and less tax impact.
4
Education and Financial Management
Analysis of the e¤ects of education on personal …nancial management is complicated by the
fact that data on credit and education are derived from two distinct data sets; as the census
data contain no personal identi…ers (and the FRBNY Consumer Credit/Equifax panel has only
anonymous identi…ers), it is not possible to match individuals across the data sets. We therefore
follow Angrist (1990) and adopt a two-sample instrumental variables approach.16 We use the
census data to estimate the relationship between compulsory schooling laws and education, and
the FRBNY Consumer Credit Panel/Equifax to calculate the reduced form relationship between
compulsory schooling laws and …nancial management.17
The reduced form is estimated using one dataset, the FRBNY Consumer Credit Panel/Equifax,
14
2140
15
2568
1:0529 1
:05
(1+:05)29
:05
2000
1
2000
1:0528 1
= $492:91
:05
(1+:05)30 1
= $27; 167:
:05
16
Two-sample IV is relatively rare in …nance, but is used in Bitler, Moskowitz, and Vissing-Jorgensen (2005).
For a detailed discussion of the two-sample instrumental variables technique, please see section 4.4 of Angrist
and Pischke (2008).
17
11
which has data on individuals’state of residence and year of birth.18 Table V provides reducedform estimates of the e¤ect of compulsory education on credit score and credit management, and
the probability of …ling for bankruptcy or experiencing a foreclosure. The speci…cation parallels
that of equation 1 except that the dependent variable is now a measure of credit management
and the set of control variables does not include race, gender, census year or the cubic polynomial in earned income, since this information is not available in the credit bureau data. Column
(1) through (3) present strong evidence that compulsory schooling laws improve an individual’s
credit score. The reduced form e¤ect indicates that cohorts who are required to attend school
through the 11th grade have on average credit scores that are 1.7 points higher compared to cohorts not required to attend school beyond the 8th grade. The compulsory attendance dummies
are jointly signi…cant at the one-percent level in every speci…cation. Using years of schooling
required rather than dummy variables yields an estimate that each year of required schooling
increases credit scores by 0.253 points, signi…cant at the …ve percent level. Column (3) adds
zipcode level …xed e¤ects, which control for geographic heterogeneity at a very …ne level (there
are approximately 43,000 zip codes in the U.S.) The point estimate is smaller, but years of
compulsory attendance is signi…cant at the 1 percent level.
Columns (4)-(9) examine the reduced form relationship between compulsory education laws
and credit behavior, studying both the fraction of borrower balance that is non-delinquent
(averaged over the period for which we have credit bureau data, 1999-2011), and the fraction
of quarters a borrower has any delinquent credit. We …nd statistically signi…cant e¤ects on
both. Finally, columns (10)-(15) study the e¤ect of compulsory schooling on the probability a
household declares bankruptcy or experiences a foreclosure. Relative to those who were able to
drop out before 9th grade, cohorts in states that required attendance through the 11th grade
were one percentage point less likely to have declared bankruptcy, and one percentage point less
likely to experience a foreclosure.
18
The FRBNY CCP/Equifax data do not include state of birth information, so we construct a state-of-birth
proxy from an individual’s state-of-residence in the …rst quarter of their inclusion in the panel. This creates
attenuation bias, making it more di¢ cult to …nd an e¤ect of education on …nancial outcomes.
12
The reduced form results provide the average e¤ect on the exposed cohort. To understand
the structural e¤ect of education on …nancial management, we turn to instrumental variables
estimation, which provides the e¤ect of education on the individuals who are a¤ected by compulsory schooling. As stated above, because the credit score and bankruptcy data come from a
di¤erent source than the education data, we cannot use a conventional instrumental variables
strategy. Instead, we use a two-sample IV approach.19 The …rst-stage regression speci…cation,
equation 2, is the same as used in Table III, from the census data set, except that we only use
data from the 2000 Census to estimate it in order to match the credit bureau data. The split
sample IV estimates are constructed by combining moments from the …rst-stage with moments
from the credit bureau dataset. We estimate standard errors in two ways. First, we provide
robust standard errors, as described by Murphy and Topel (1985). Second, we use a block
bootstrap technique to account for possible correlation within birthyear-state groups.
Results are presented in Table VI. The point estimates using either estimation technique are
quite similar, and suggest that education has important causal e¤ects on …nancial management.
The point estimate on the coe¢ cient for years of schooling, 7.7, is signi…cant at the one percent
level using Murphy and Topel standard errors, suggesting that a one standard deviation increase
in education (2.7 years) would raise an individual’s credit score by 20 points. This result is
signi…cant at the ten percent level using a block bootstrap estimation technique.
A 20 point movement in the credit score is less than one standard deviation in credit score,
but there are certainly ranges where such perturbations can be very important. For example,
Chomsisengphet and Pennington-Cross (2006) document how a 20 point di¤erence in credit score
can impact both the cost and availability of certain home mortgage products. The e¤ect sizes
for credit management are similar: a one-standard deviation increase in education increases the
fraction of credit card balances kept current by 1.4 percentage points, relative to an unconditional
average of 95.6%, and reduces the number of quarters delinquent by 3.5 percentage points, from
a mean of 7.5 percentage points.
19
We thank the editor for this suggestion.
13
The point estimates of the e¤ect of education on bankruptcy and foreclosure are striking.
Over the sample period, 14.4% of individuals declare bankruptcy, and 5.8% experience at least
one foreclosure. An additional year of education reduces the probability of …ling for bankruptcy
by 3.3 percentage points, and foreclosure by 5.7 percentage points. These e¤ects are signi…cant
at the 1 percent level (using Murphy and Topel standard errors), but the con…dence intervals
do admit smaller e¤ect sizes, as small as 1.15 percentage points for bankruptcy and 2.18 percent
for foreclosure.
In Panel B of Table VI, we analyze whether education a¤ects bankruptcy and foreclosure
similarly throughout our sample, or whether there is a di¤erential e¤ect during the recent …nancial crisis. We …nd that during non-crisis periods (1999Q2-2007Q3), education does not reduce
bankruptcies or foreclosures. However, during the period since the …nancial crisis (2007Q32011Q4), we observe substantial e¤ects: an additional year of education reduces the probability
of bankruptcy by 2.3 percentage points, and the probability of foreclosure by 4.6 percentage
points.
Because these outcomes are “worst-case” scenarios, they may be particularly relevant for
the group of individuals a¤ected by our instrument. In this case, and in contrast to estimates
that examine the e¤ect of education on income, the LATE estimates may not characterize the
population parameter.
The economic implications of these results are important. Bankruptcy is costly to individuals,
as it results in lower credit scores, and reduced access to credit, and to society, through the
deadweight costs of debt collection (Cohen-Cole et al., 2009). Perhaps of even greater importance
are the costs of foreclosure. Campbell, Giglio, and Pathak (2011) estimate that a foreclosure
reduces the value of the foreclosed house by $44,000, but depresses the value of neighboring
houses by a total of $148,000-$477,000.
14
5
Cognitive Ability and Savings
Having established the impact of education on …nancial behavior, we turn to examine some possible mechanisms. Recent evidence suggests that the primary value of education is to increase
cognitive ability (Hanushek and Woessman, 2008). Financial decisions are often complicated and
cognitive ability may play an important role in helping households navigate these complications.
For example, household mortgage decisions are tremendously important, yet individuals regularly make costly mistakes when deciding whether to re…nance their mortgage (Schwartz, 2007).
Even decisions such as which credit card to use, which bank to use, or in which mutual fund to
invest, can involve complex trade-o¤s that require a nuanced understanding of probability and
compound interest.
Some evidence in favor of the hypothesis that cognitive ability matters for …nancial decision
making has already been documented. Chevalier and Ellison (1999) …nd that mutual fund
managers who graduated from institutions with high average SAT scores outperform those who
graduated from less selective institutions. Stango and Zinman (2009) show that households
that exhibit the cognitive bias of systematically miscalculating interest rates from information
on nominal repayment levels hold loans with higher interest rates, controlling for individual
characteristics. In a study closely related to this section, Grinblatt et al. (2011a) …nd that
Finnish individuals with higher IQs are more likely to participate in equity markets.
Only two other studies, to our knowledge, links actual measures of cognitive ability to investment decisions. Christelis, Jappelli, and Padula (2006) use a survey of households in Europe,
which directly measured household cognitive ability using math, verbal, and recall tests. They
…nd that cognitive abilities are strongly correlated with stock market participations. These results are correlations, and the degree to which causal interpretation may be assigned depends
on the determinants of cognitive ability. Grinblatt et al (2011b) …nd that high-IQ traders select
better stocks and exhibit fewer behavioral biases than low-IQ traders.
A limitation of this approach is that cognitive ability itself is correlated with other factors
15
that also a¤ect …nancial decision making. Bias could occur if, for example, measured cognitive
ability is correlated with wealth or the transfer of human capital from parent to child. This is
likely the case. Plomin and Petrill (1997), in a survey of the literature, …nd that both genetic
variation and shared environment play a signi…cant role in explaining variation in measured
cognitive ability.20 The importance of family background suggests that the coe¢ cient from a
regression of investment behavior on measured IQ which does not correctly control for parental
circumstances may be biased upwards.21
5.1
Empirical Strategy
One compelling strategy to remove the potential confound of family environment is to study
siblings, who grew up with similar backgrounds. Labor economists have used this technique
extensively to identify the e¤ect of education on earnings (see, e.g., Ashenfelter and Rouse
1998). Including a sibling group …xed-e¤ect provides a substantial advantage, as it controls
for a wide range of observed and unobserved characteristics. Most of the remaining variation in
cognitive ability is thus attributable to the random allocation of genes to each particular child.22
There are limitations to this approach as well. Children without siblings are of course excluded. The errors-in-variables bias is potentially exacerbated when di¤erencing between siblings
(Griliches 1979). Finally, as demonstrated in Bound and Solon (1999), if the endogenous variation is not eliminated when comparing between siblings, the resulting bias may constitute an
even larger proportion of the remaining variation than in traditional cross-sectional studies.
This concern may be less severe in the case of cognitive ability when measured at an early age,
because individuals do not choose cognitive ability in the way they choose how many years of
20
For example, the correlation between parental IQ and children reared apart is approximately 0.24, providing evidence that genes in‡uence IQ. Similarly, the correlation between two unrelated individuals (at least one
adopted) raised in the same household is approximately 0.25.
21
Mayer (2002) surveys evidence on the relationship between parental income and childhood outcomes, and
describes a strong consensus that higher parental income and education is associated with higher measured
cognitive ability among children.
22
Plomin and Petrill (1997) note that the correlation in IQ of monozygotic (identical) twins raised together is
much higher than dizygotic (fraternal) twins raised together.
16
schooling to obtain. While unobserved characteristics such as motivation and discount rates
may a¤ect educational attainment, they are unlikely to a¤ect measures of childhood cognitive
ability.
Benjamin and Shapiro (2007) employ this method to study how cognitive ability is correlated
with various behaviors, including …nancial market participation, using data from the National
Longitudinal Survey of Youth (NLSY). They regress a dummy for stock market participation
on a set of controls, a sibling group …xed-e¤ect, and a measure of cognitive ability. We expand
this analysis in several directions. We look at a range of …nancial assets, considering both
the extensive and intensive margins, and …nally unpack cognitive ability into two components,
knowledge and ability. The former is meant to capture factual aspects of cognitive ability
that are taught, such as general science (what is an eclipse?). The latter captures functional
abilities, which may or may not be taught: mathematical skills, or coding speed (how quickly
the respondent can look up a number in a table).
Following Benjamin and Shapiro (hereafter, BS), we use the National Longitudinal Survey of
Youth from 1979, as described above. In 1980, respondents took the Armed Services Vocational
Aptitude Battery (ASVAB), a set of 10 exams that measure ability and knowledge, which yields
an estimate of the respondent’s percentile score in the Armed Forces Qualifying Test (AFQT).
The AFQT comprises mostly questions that measure reasoning abilities, such as math skills,
paragraph comprehension and numerical operations. To calculate a measure of knowledge that
may have been acquired in school, we include ASVAB test scores such as general science, auto and
shop information and electronics information. These scores are then normalized by subtracting
the mean and dividing by the standard deviation. Further details are provided in the online
data appendix.
Using these test scores, we estimate the e¤ect of cognitive ability, knowledge and education
on …nancial decision making (yit ) with the following equation
yit =
1 knowledgei
+
2 abilityi
+ educationit + Xit + SGi + "it
17
(3)
where abilityi is a measure of innate ability, knowledgei is a measure of acquired knowledge,
educationit is the highest grade individual i has completed by year t; Xit includes age, race,
gender and survey year e¤ects, and SGi are sibling-group …xed e¤ects. Standard errors are
corrected for intracluster correlation within an individual over time. We proxy for permanent
income by controlling for the log of family income23 in every available survey year from 1979 to
2002, and including dummy variables for missing data.24
5.2
Results
Results are presented in Table VII. In the …rst column, the outcome variable is equal to one
if the respondent answers "something left over" to the following NLSY question:
“Suppose
you [and your spouse] were to sell all of your major possessions (including your home), turn
all of your investments and other assets into cash, and pay all of your debts. Would you have
something left over, break even, or be in debt?”The other answers are coded as zero. We …nd a
signi…cantly positive e¤ect of both knowledge and ability - an increase of one standard deviation
in knowledge (22 points out of 120 or 18%) increases the propensity to have accumulated assets
by about 2.6 percentage points, while an increase in one standard deviation in ability (41 points
out of 214 or 19%) increases the propensity by about 3.6 percentage points. Note that this
result includes controls for education. The point estimate on education alone is not statistically
signi…cant. Respondents were then asked to estimate how much money would be left over - we
…nd that neither ability nor knowledge has an e¤ect on this amount (Column (1) in panel B).
The second column in Table VII examines stock market participation. The NLSY question is:
“Not counting any individual retirement accounts (IRA or Keogh) 401K or pre-tax annuities...
23
We use log (family income + $1) to include individuals with zero income.
We also drop all observations which are top-coded; the cut-o¤ varies by year and outcome variable, but
typically does not exclude many individuals. We do not include individuals who are cousins, step-siblings, adopted
siblings, or only related by marriage or households with only one respondent.
To ensure that our results are not driven by large cognitive di¤erences between siblings due to mental handicaps,
we cut the data in two ways. Our results are robust to dropping all households where any individual is determined
to be mentally handicapped at any time between 1988 and 1992 when the question was asked. In addition, our
results are robust to dropping siblings with a cognitive ability di¤erence greater than 1 standard deviation of the
sample by race.
24
18
Do you [or your spouse] have any common stock, preferred stock, stock options, corporate or
government bonds, or mutual funds?” Knowledge and education have positive and signi…cant
e¤ects: a one standard deviation increase in knowledge or ability increases the participation
margin by 3.4 and 1.8 percentage points, respectively. An additional year of education increases
stock market participation by 1.5 percentage points. Knowledge and ability are not signi…cantly
associated with how much money an individual has in stocks, but education is.
We extend the analysis in BS by studying a number of other outcomes regarding whether
and how much individuals save in di¤erent …nancial instruments. In Column (3) we study how
respondents answer the question: “Do you [and your spouse] have any money in savings or
checking accounts, savings & loan companies, money market funds, credit unions, U.S. savings
bonds, individual retirement accounts (IRA or Keogh), or certi…cates of deposit, common stock,
stock options, bonds, mutual funds, rights to an estate or investment trust, or personal loans to
others or mortgages you hold (money owed to you by other people)?”25 Innate ability increases
an individual’s propensity to save: one standard deviation increases the propensity to save by 5
percentage points. An additional year of education increases the share with positive savings by
1.65 percentage points.
We …nd similar results when we focus on savings in 401Ks and pre-tax accounts. Ability
and knowledge are jointly signi…cant at the ten percent level. Education has a signi…cant e¤ect
on savings in IRAs and Keogh accounts (Column (4)). Ability increases participation in taxdeferred accounts such as 401Ks by 5 percentage points. One year of schooling increases both
participation in IRAs and Keogh accounts by 1.1 percentage points and participation in taxdeferred accounts by 1.3 percentage points. The e¤ects are substantially smaller for certi…cates
of deposit, loans and mortgage assets (Column (6)).
Our results might be confounded by strategic parents, who increase or decrease parental
25
In following years, respondents were asked a variant of this question - each few years, the list of types of
savings changes slightly. For example, in 1988 and 1989, respondents were no longer asked about savings and
loan companies while stocks, bonds and mutual funds were asked in a separate question. While our survey year
…xed e¤ects should take these changes into account, we also test the robustness of this speci…cation by recoding
a new variable with a consistent list of assets. The estimates are nearly identical.
19
transfers to children as a function of their cognitive ability. Column (7), which examines respondents’anticipated transfers, shows that this does not happen.
Finally, in Column (8) we look at an outcome variable, classi…ed as “other income” from
1979 to 2002, which includes income from investment and other sources of income,26 which
corresponds closely to our measure of investment income from the Census. Ability, knowledge
and education all have a positive and signi…cant e¤ect on income from these sources: one standard
deviation in knowledge increases the probability of having any such income by 5.3 percentage
points, one standard deviation in ability increases the probability by 4.1 percentage points, and
one year of schooling by 1.5 percentage points.
These results suggest that education, ability and knowledge acquired in school increase participation in …nancial markets.27 Acquired knowledge matters only for one investment class
(stocks, bonds, and mutual funds), while cognitive ability is associated with all assets and methods of investing measured in the data. The F-test reported at the bottom of Panel A indicate
that knowledge and ability are jointly signi…cant at either the …ve or ten percent level.
Our …nding that cognitive ability is more important than acquired knowledge is consistent
with a growing recognition of the key role of cognitive ability in determining economic outcomes
(Hanushek and Woessman, 2008). Our analysis suggests one channel through which schooling
may matter: it a¤ects cognitive ability, which in turn a¤ects savings and investment decisions.
The magnitudes of the e¤ects we identify are large, and may well account for a substantial
fraction of unexplained variation in …nancial behavior.
26
The question asks “(Aside from the things you have already told me about,) During [year], did you [or your
(husband/wife) receive any money, even if only a small amount, from any other sources such as the ones on this
card? For example: things like interest on savings, payments from social security, net rental income, or any other
regular or periodic sources of income.”
The list of assets changes slightly from year to year, but always includes interest on savings, net rental income,
any regular or periodic sources of income. In 1987, the question also lists worker’s compensation, veteran’s
bene…ts, estates or trusts and up until 1987, also includes payments from social security. From 1987 to 2002, the
interviewer also listed interest on bonds, dividends, pensions or annuities, royalties.
Due to the wording of the question (asking for “any other source”of income), we treat this question as constant.
The results are robust to focusing only on questions which ask about precisely the same set of assets.
27
Columns (1)-(3) of Appendix Table A4 demonstrate that the relationship between schooling and cognitive
ability holds in sibling pairs.
20
6
Other Mechanisms
How else might education a¤ect …nancial management? One possibility that has received some
attention is the fact that high school students in many states are required to attend …nancial
education courses. Bernheim, Garrett, and Maki (2001) study mandatory high school …nancial
education requirements, …nding large e¤ects. However, Cole and Shastry (2011) revisit this
question using the U.S. Census, and provide evidence that high school …nancial education did
not in fact have any e¤ect on individuals’…nancial behavior.
We begin our search for mechanisms by exploring whether the impact of education works
through changing the set of people an individual interacts with, either at home or work. Finally,
we examine whether education may a¤ect preferences and beliefs, such as attitudes towards risk
and feelings of control.
Education changes the set of job opportunities available to individuals. For example, a highschool degree may lead an employee to a salaried job at a large corporation, which facilitates
…nancial market participation. We test for this in the following manner. Using data from the
1970 census, we identify the share of individuals aged 65-70 in each occupation in each state
receiving a pension. We use the 1970 census because it includes the individuals’ occupation
from 5 years prior to the survey.28 We use this fraction as a measure of the pension probability
for each individual in our dataset from 1980, 1990, and 2000, and regress this probability on
education, using the state laws as instruments, as in equation (1). The result is presented in
Table VIII, Column (1). We …nd a positive relationship between education and the probability
of …nding a job in which a pension is o¤ered, statistically signi…cant at the 1 percent level. One
year of schooling increases the probability of receiving a pension by 1%. All of the estimates in
Table VIII mimic the education speci…cation, with controls for age, cohort, birth state, state of
28
The 1970 census does not include the retirement income variable we have been using this far. Instead, it
groups pension income into "income from other sources," such as unemployment compensation, child support
and alimony. We therefore de…ne an individual over 65 as having a pension if they received more than $1000 (in
1970 dollars) in other income during the previous year. The results are robust to using $2,500, $5,000 or $10,000
instead.
21
residence, gender, race, and income.
Hong et al. (2004) …nd that peer e¤ects are important determinants of …nancial market
participation. To test this channel, we use a similar approach: we calculate the percent of
individuals aged 65 and older in every "neighborhood" in the U.S. who received retirement
income, and use this as the dependent variable in equation (1). Neighborhoods are de…ned as
county groups, single counties or census-de…ned "places" with a population of approximately
100,000. Results are presented in Column (2) of Table VIII. We …nd a remarkably similar
e¤ect to that in Column (1): one year of schooling increases the share of retired neighbors with
retirement income other than Social Security income by 1 percentage point.29
A commonly advanced view is that education tempers impatience. Indeed, in a …eld study,
Harrison et al. (2002) …nd that discount rates are strongly negatively correlated with levels of
education. Of course, this correlation is hard to interpret: does education reduce discount rates,
or do more impatient individuals select to enter the labor market earlier? While we cannot measure discount rates, we do observe whether households take …rst or second mortgages. We …nd
that education does not have an e¤ect on whether a household takes out a …rst mortgage (Column (3)), but does signi…cantly reduce the likelihood a household takes out a second mortgage
(Column (4)).
As a …nal direct mechanism, we explore whether education a¤ects individuals’willingness to
take risks. Halek and Eisenhauer (2001) …nd a strong negative correlation between risk aversion
and education. We do not have a good measure of attitudes towards risk from the census.
However, one important risk an individual can take is to move in search of better opportunities.
Heitmueller (2005), for example, argues that risk aversion is an important determination of
within-EU migration. We …nd no evidence that more educated individuals are more likely to
move away from their city (Column (5)) or state (not reported) in the past …ve years.30 We
29
The F-statistic of the excluded instruments in this column is much lower than that in previous results because
we lose data from 1980 when more people were a¤ected by the laws. The 1980 census does not include the public
use microdata area identi…ers. This suggests this result may su¤er from weak instruments bias.
30
Information on whether an individual moved to a di¤erent city is available only in the 1980 census.
22
…nd evidence that more educated individuals are less likely to have moved into a di¤erent house
within the city in the previous …ve years.
It is also possible that education a¤ects …nancial behavior through beliefs and attitudes.
Graham et al. (2005) …nd that educated investors report higher levels of con…dence and invest
more abroad. Puri and Robinson (2007) show that optimistic individuals invest a greater share
of their portfolio in equities, as compared to other …nancial instruments. We do not have a view
on how education a¤ects optimism; it may well foster discipline and views on achieving speci…c
goals, by changing individuals’beliefs and self-control. While few datasets consider personality
and investment decisions in detail, the NLSY does ask respondents to indicate their agreement
with the statement “I have little control over the things that happen to me,” with 1 indicating
strong disagreement and 4 indicating strong agreement. Individuals who feel more in control
(or have greater self-control) may well be more likely to participate in …nancial markets.
Online Appendix Table A4, using the same within-family identi…cation strategy, provides evidence from the NLSY that feelings of lack of control are greater among less educated individuals,
and weakly greater among individuals with lower levels of cognitive ability.
To examine the relationship between control and …nancial decisions, we focus on investment
decisions made after 1993, the year the personality measure was taken, using the same identi…cation strategy as for cognitive ability. Results are presented in Table IX. Comparing two siblings
within the same family, we …nd that those who report feelings of lack of control are less likely to
have money left at the end of the month, less likely to report investment income, and less likely
to report having a positive savings balance (Panel A). The magnitudes are quite substantial:
moving from strong disagreement to strong agreement with the statement is associated with an
individual being 4.3 percentage points less likely to have investment income, and 7.5 percentage
points less likely to report having money left at the end of the month.
23
7
Conclusion
Household …nancial management is an important determinant of …nancial welfare (Campbell,
2006). Participation in …nancial markets is limited. While over 90% of households have transactions accounts, the fraction of families that own bonds (17.6%), stock (20.7%), and other assets
is relatively small. Low levels and low returns on savings may well be an important contributing
cause to consumer bankruptcy.
This paper contributes to a growing body of literature exploring the importance of nonneo-classical factors to household investment decisions. We explore how education a¤ects …nancial management, with a focus on discovering causal mechanisms. We …rst show education
signi…cantly increases investment income. Individuals with one more year of schooling are 7.5
percentage points more likely to report positive investment income. Similarly, those with more
years of schooling are signi…cantly more likely to report income from retirement savings.
Second, we study how education a¤ects consumers’borrowing and credit behavior. We …nd
that cohorts induced to receive higher levels of education have higher credit scores on average and
are signi…cantly less likely to be delinquent, declare bankruptcy or experience a foreclosure. Some
of these e¤ects are less dramatic than the e¤ect of education on …nancial market participation:
an additional year of schooling raises an individual’s credit score by 8 points (roughly 9% of
a standard deviation). Others are even more dramatic: one year of schooling reduces the
probability of bankruptcy by 3.3 percentage points, from a base of 14.4%.
Examining mechanisms, we …nd that cognitive ability itself is an important determinant of
…nancial behavior. Controlling for family background, those with higher test scores are more
likely to hold a wide variety of …nancial instruments, including stocks, bonds, mutual funds,
savings accounts, tax-deferred accounts, and CDs. When cognitive ability is decomposed into
innate abilities and acquired abilities or knowledge, the innate abilities matter for a greater
number of …nancial instruments, but both types of ability a¤ect key measures of …nancial market
participation such as having any accumulated assets and owning any stocks, bonds or mutual
24
funds.
The point estimates on education suggest that it is a very important determinant of behavior.
A convenient metric to compare the relative importance across di¤erent studies is the “e¤ect
size”, which is the e¤ect of a one standard deviation change in the dependent variable on
participation. The “e¤ect size” of education is 19.8 percentage points, which compares to an
e¤ect size of trust (Guiso, Sapienza, and Zingales) of 4 percentage points, peer e¤ects (Hong
and Stein) of 1.15 percentage points, and experience with stock market returns (Malmendier
and Nagel) of 4.2 percentage points.
Three studies serve as potential benchmarks for these e¤ects. Du‡o and Saez (2003) present
evidence from a randomized evaluation that minor incentives ($20 for university sta¤ attending
a bene…ts fair) can increase TDA participation rates by 1.25 percentage points. Du‡o et al.
(2006) o¤ered low-income tax …lers randomly assigned amounts of matching to contribute to
IRAs. They …nd that an o¤er of a 50 percent match increased participation by 14 percentage
points, which is comparable to two years of education in our analysis. However, no determinants
of participation have been found to be more e¤ective than simply changing the default enrollment
status for 401(k) plans. Beshears et al. (2006) …nd changing the default to “enroll” increases
participation by as much as 35 percentage points.
Concern about …nancial decision-making is not often cited as an important determinant of
educational policy. Yet, it is worth pointing out that because education a¤ects …nancial market
participation, studies that focus on wage earnings may in fact underestimate the returns to investment in human capital. Such estimates miss the reduction in the probability of bankruptcy
and foreclosure. This suggests adjusting earlier cost-bene…t analyses of educational programs.
Moreover, a growing body of evidence suggests that individuals do often make …nancial mistakes (Agarwal et al., 2007). Both micro evidence (Agarwal and Mazumder, 2010) and recent
experience suggest that some of these mistakes can be quite costly. Increasing educational attainment in the US could dramatically improve …nancial management, with important e¤ects
25
on bankruptcy and default, and may even facilitate a more stable …nancial system (Mian and
Su…, 2011).
8
Bibliography
Acemoglu, Daron, and Joshua Angrist, 2000, How Large are Human-Capital Externalities?
Evidence from Compulsory Schooling Laws, in Ben S. Bernanke and Kenneth Rogo¤ eds.:
NBER Macroeconomics Annual 2000. (MIT Press, Volume 15. Cambridge and London)
9-59.
Agarwal, Sumit, and Bhashkar Mazumder, 2010, Cognitive Abilities and Household Financial
Decision Making, Federal Reserve Bank of Chicago Working Paper 2010-16.
Agarwal, Sumit, John Driscoll, Xavier Gabaix, and David Laibson, 2007, The Age of Reason:
Financial Decisions Over the Lifecycle, NBER Working Paper Number 13191.
Angrist, Josh, 1990, Lifetime Earnings and the Vietnam Era Draft Lottery: Evidence from
Social Security Administrative Records, American Economic Review 80, 1535-1558.
Angrist, Josh, and Joern-Ste¤en Pischke, 2008, Mostly Harmless Econometrics: An Empiricist’s
Companion, Princeton University Press: Princeton, New Jersey.
Ashenfelter, Orley, and Cecilia Rouse, 1998, Income, Schooling, and Ability: Evidence from a
New Sample of Identical Twins, Quarterly Journal of Economics 113, 253-284.
Benjamin, Daniel, and Jesse Shapiro, 2007, Who is ’Behavioral?’ Cognitive Ability and Anomalous Preferences, Working Paper, University of Chicago Graduate School of Business.
Bernheim, B. D., Daniel M. Garrett, and Dean M. Maki, 2001, Education and Saving: The
Long-Term E¤ects of High School Financial Curriculum Mandates, Journal of Public Economics 80, 435-465.
Bertaut, Carol C., and Martha Starr-McCluer, 2001, Household Portfolios in the United States,
in Luigi Guiso, Michael Haliassos, and Tullio Jappelli eds.: Household Portfolios. (MIT
Press, Cambridge and London) 181-217.
Beshears, John, Jame Choi, David Laibson, and Brigitte Madrian, 2006, The Importance of
Default Options for Retirement Savings Outcomes: Evidence from the United States,
mimeo, Harvard University.
Brav, Alon, George M. Constantinides, and Christopher C. Geczy, 2002, Asset Participation
with Heterogenous Consumers and Limited Participation: Empirical Evidence, Journal of
Political Economy, 110, 793-824.
Bound, John, and Gary Solon, 1999, Double Trouble: On the Value of Twins-Based Estimation
of the Return to Schooling, Economics of Education Review 18, 169-182.
26
Bucks, Brian K., Arthur B. Kennickell, and Kevin B. Moore, 2006, Recent Changes in U.S.
Family Finances: Evidence from the 2001 and 2004 Survey of Consumer Finances, Federal
Reserve Bulletin 92, A1-38.
Calvet, Laurent E., John Y. Campbell, and Paolo Sodini, 2007, Down Or Out: Assessing
the Welfare Costs of Household Investment Mistakes, Journal of Political Economy 115,
707-747.
Campbell, John Y., 2006, Household Finance, Journal of Finance 61, 1553-1604.
Campbell, John Y., Stefano Giglio, and Parag Pathak. 2011, Forced Sales and House Prices,
American Economic Review, 101, 2108–31.
Carroll, Christoper D., Byung-Khun Rhee and Changyoung Rhee, 1999, Does Cultural Origin
A¤ect Savings Behavior? Evidence from Immigrants, Economic Development and Cultural
Change 48, 33-50.
Chevalier, Judith, and Glenn Ellison, 1999, Are some Mutual Fund Managers Better than
Others? Cross-Sectional Patterns in Behavior and Performance, Journal of Finance 54,
875-899.
Chomsisengphet, Souphala and Anthony Pennington-Cross, 2006, The Evolution of the Subprime Mortgage Market, Federal Reserve Bank of St. Louis Review 88, 31-56.
Christelis, Dimitris, Tullio Jappelli, and Mario Padula, 2006, Cognitive Abilities and Portfolio
Choice, Manuscript, University of Salerno.
Cohen-Cole, Ethan, Burcu Duygan-Bump, and Judit Montoriol-Garriga, 2009, Forgive and
Forget: Who Gets Credit after Bankruptcy and Why? manuscript, University of Maryland.
Cole, Shawn, and Kartini Shastry, 2011, Is High School the Right Time to Teach Savings Behavior? The E¤ect of Financial Education and Mathematics Courses on Savings, manuscript,
Harvard Business School.
Du‡o, Esther, and Emmanuel Saez, 2003, The Role of Information and Social Interactions in
Retirement Plan Decisions: Evidence from a Randomized Experiment, Quarterly Journal
of Economics 118, 815-842.
Du‡o, Esther, William Gale, Je¤rey Liebman, Peter Orszag, and Emmanuel Saez, 2006, Saving
Incentives for Low- and Middle-Income Families: Evidence from a Field Experiment with
H&R Block, Quarterly Journal of Economics 121, 1311-1346.
Graham, John R., Campbell R. Harvey, and Hai Huang, 2005, Investor Competence, Trading
Frequency, and Home Bias, Manuscript.
Griliches, Zvi, 1979, Sibling Models and Data in Economics: Beginnings of a Survey, Journal
of Political Economy 87, S37-64.
Grinblatt, Mark, Matti Keloharju, and Juhani Linnainmaa, 2011a, IQ and Stock Market Participation, Journal of Finance 66, 2119-2164.
27
Grinblatt, Mark, Matti Keloharju, and Juhani Linnainmaa, 2011b, IQ, trading behavior, and
performance, forthcoming, Journal of Financial Economics.
Guiso, Luigi, Paola Sapienza, and Luigi Zingales, 2008, Trusting the Stock Market, Journal of
Finance 63: 2557-2600.
Hanushek, Eric, and Ludger Woessmann, 2008, The Role of Cognitive Skills in Economic
Development, Journal of Economic Literature 46, 607-668.
Haliassos, Michael, and Carol C. Bertaut, 1995, Why do so Few Hold Stocks? Economic
Journal 105, 1110-1129.
Halek, M., and J.G. Eisenhauer, 2001, Demography of Risk Aversion, Journal of Risk and
Insurance 68:1-24.
Heaton, John, and Deborah Lucas, 1999, Stock prices and fundamentals, NBER Macroeconomics Annual 14, 213-242.
Harrison, Glenn W., Morten I. Lau, and Melonie B. Williams, 2002, Estimating Individual
Discount Rates in Denmark: A Field Experiment, American Economic Review 92, 16061617.
Heitmueller, Axel, 2005. Unemployment Bene…ts, Risk Aversion, and Migration Incentives,
Journal of Population Economics 18, 93-112.
Hong, Harrison, Je¤rey D. Kubik, and Jeremy C. Stein, 2004, Social Interaction and StockMarket Participation, Journal of Finance 59, 137-163.
Imbens, Guido W., and Joshua D. Angrist, 1994, Identi…cation and Estimation of Local Average
Treatment E¤ects, Econometrica 62, 467-475.
Lochner, Lance, and Enrico Moretti, 2004, The E¤ect of Education on Crime: Evidence from
Prison Inmates, Arrests, and Self-Reports, American Economic Review 94, 155-189.
Lee, Donghoon and Wilbert van der Klaauw, 2010, An Introduction to the FRBNY Consumer
Credit Panel, Federal Reserve Bank of New York Sta¤ Reports, no. 479.
Lusardi, Annamaria, and Olivia S. Mitchell, 2007, Baby Boomer Retirement Security: The
Roles of Planning, Financial Literacy, and Housing Wealth, Journal of Monetary Economics 54, 205-224.
Lusardi, Annamaria, Dan Schneider, and Peter Tufano, 2011, Financially Fragile Households:
Evidence and Implications, NBER Working paper 17072, Cambridge, MA.
Malmendier, Ulrike, and Stefan Nagel, 2011, Depression Babies: Do Macroeconomic Experiences A¤ect Risk-Taking? Quarterly Journal of Economics, 126, 373-416.
Mankiw, N. G., and Stephen P. Zeldes, 1991, The Consumption of Stockholders and Nonstockholders, Journal of Financial Economics, 29, 97-112.
28
Mayer, Susan, 2002, The In‡uence of Parental Income on Children’s Outcomes, Ministry of
Social Development (Wellington, New Zealand).
Mian, Atif, and Amir Su…, 2011. “House Prices, Home Equity-Based Borrowing, and the US
Household Leverage Crisis," American Economic Review, 101, 2132-56.
Murphy, Kevin, and Robert Topel, 1985, Estimation and Inference in Two-Step Econometric
Models, Journal of Business & Economic Statistics, 3, 370-379.
Oreopoulos, Philip, 2006, Estimating Average and Local Average Treatment E¤ects of Education when Compulsory Schooling Laws really Matter, American Economic Review 96,
152-175.
Osili, Una, and Anna Paulson, 2008, Institutions and Financial Development: Evidence from
International Migrants in the United States, Review of Economics and Statistics 90 (3),
498 - 517.
Plomin, Robert, and Stephen A. Petrill, 1997, Genetics and Intelligence: What’s New?, Intelligence, 24, 53-77.
Puri, Manju, and David T. Robinson, 2007, Optimism and Economic Choice, Journal of Financial Economics 86, 71-99.
Ruggles, Steven, Matthew Sobek, Trent Alexander, Catherine A. Fitch, Ronald Goeken, Patricia Kelly Hall, Miriam King, and Chad Ronnander, 2004, Integrated Public use Microdata
Series: Version 3.0, Machine-Readable Database (Minneapolis, MN: Minnesota Population Center).
Schwartz, Allie, 2007, Household Re…nancing Behavior in Fixed Rate Mortgages, Mimeo Harvard University.
Gross, David, and Nicholas Souleles, 2002, Do Liquidity Constraints And Interest Rates Matter For Consumer Behavior? Evidence From Credit Card Data, Quarterly Journal of
Economics 117(1), 149-185.
Stango, Victor, and Johathon Zinman, 2009, Exponential Growth Bias and Household Finance,
64, 2807-2849.
Stock, James H., and Motohiro Yogo, 2005, Testing for Weak Instruments in Linear IV Regression, in Donald W. K. Andrews and James H. Stock eds.: Identi…cation and Inference
for Econometric Models: Essays in Honor of Thomas Rothenberg (Cambridge University
Press) 80-108.
Tortorice, Daniel L., 2012, Unemployment Expectations and the Business Cycle, The B.E.
Journal of Macroeconomics 12,1–47.
Vissing-Jogensen, Annette, 2002, Limited Asset Market Participation and the Elasticity of
Intertemporal Substitution, Journal of Political Economy 110, 825-853.
29
Table I
Summary Statistics
This table reports summary statistics for data used in this paper. Panel A reports summary statistics from the 5% sample of the census, from 1980, 1990, and 2000. Panel B reports
summary statistics for data from the FRBNY Consumer Credit Panel/Equifax. The sample comprises a 5% panel of American borrowers, restricted to borrowers who have data in
every quarter of the panel from 1999 to 2011. In both panels, means are given by the years of compulsory schooling associated with each observation, as well as for the entire
sample. The sample is limited to individuals aged 36-75. Credit Score is averaged for each individual accross all quarters of data, and it can range from 280 to 850. The % of
Balance Current represents the non-delinquent balance on credit cards divided by the total credit card balance, averaged over the entire panel. The % of Quarters Delinquent
represents the proportion of quarters an individual has any delinquent balance on his/her credit card bills. Bankruptcy and Foreclosure are indicators for having undergone
bankruptcy or foreclosure at least once, respectively, between 1992 and 2011.
Panel A: Census Data
Demographic
Compulsory Attendence <= 8
Compulsory Attendence==9
Compulsory Attendence==10
Compulsory Attendence==11
Entire Sample
N
Years of
Schooling
12.39
(3.11)
12.86
(2.71)
13.06
(2.57)
13.17
(2.47)
12.91
(2.69)
14,913,356
Credit Score
Compulsory Attendence <= 8
Compulsory Attendence==9
Compulsory Attendence==10
Compulsory Attendence==11
Entire Sample
N
723.87
(85.30)
717.12
(89.63)
699.34
(94.88)
695.09
(96.04)
714.67
(90.57)
5,732,690
Age
51.30
(15.98)
47.13
(14.67)
42.34
(14.48)
42.20
(13.17)
45.60
(14.74)
14,913,356
% Balance
Current
0.959
(0.109)
0.957
(0.113)
0.947
(0.123)
0.952
(0.116)
0.956
(0.113)
5,329,619
Income from Investments
Male
0.470
(0.499)
0.482
(0.500)
0.484
(0.500)
0.489
(0.500)
0.483
(0.500)
14,913,356
Black
0.124
(0.330)
0.103
(0.304)
0.096
(0.295)
0.079
(0.270)
0.097
(0.296)
14,913,356
Any
Amount
30.049
2325.77
(45.847) (10297.52)
30.272
1970.70
(45.944) (9457.40)
25.823
1365.43
(43.766) (8057.58)
27.548
1522.60
(44.676) (8811.94)
28.858
1810.62
(45.310) (9250.44)
14,913,356 14,913,356
Panel B: FRBNY Consumer Credit Panel/Equifax
% Quarters
Bankruptcy
Delinquent
0.068
0.146
(0.144)
(0.352)
0.072
0.131
(0.150)
(0.338)
0.093
0.143
(0.169)
(0.350)
0.085
0.182
(0.156)
(0.386)
0.075
0.144
(0.152)
(0.351)
5,750,005
5,750,005
Income from Retirement
Savings
Any
25.994
(43.860)
22.400
(41.692)
23.843
(42.612)
20.199
(40.149)
22.431
(41.713)
4,150,828
Foreclosure
0.069
(0.253)
0.052
(0.222)
0.065
(0.247)
0.063
(0.243)
0.058
(0.234)
5,750,005
Amount
3554.62
(10209.82)
3249.02
(10409.12)
3436.47
(10702.68)
3253.61
(11156.37)
3315.67
(10635.99)
4,150,828
Table II
OLS Estimates of the Effect of Schooling on Income from Various Sources
This table reports results from a regression of investment income on education, gender, race, age (3year age groups), birth cohort (10 year cohorts), state of birth, state of residence, census year and a
cubic polynomial in earned income. Only the education coefficient is reported. The sample
comprises individuals reported in the 5% samples of the 1980, 1990, and 2000 census. We include
18-75 year olds (50-75 year olds when considering retirement income). The dependent variable of
interest is whether the household receives income from investments or retirement savings (Panel A)
and the amount (Panel B). Regressions also include state of residence fixed effects interacted with a
dummy variable for being born in the South and turning age 14 in 1958 or later to account for the
impact of Brown v. Board of Education for blacks. Top-coded individuals (see text) are dropped in
panel B. Standard errors, corrected for arbitrary correlation within state of birth-year of birth, are in
parentheses. (Numbers with *** indicate significance at the 1-percent level.)
Income from
Income from Retirement
Investments
Savings
(1)
(2)
Panel A: Any Investment Income
Years of schooling
3.54 ***
2.40 ***
(0.01)
(0.02)
Num of Observations
R-Squared
Panel B: Amount of Investment Income
Years of schooling
Num of Observations
R-Squared
14,913,356
0.184
271.10 ***
(5.02)
14,842,001
0.091
4,150,828
0.177
548.42 ***
(4.81)
4,117,987
0.147
Table III
Estimates of the Effect of Compulsory Schooling Laws on Education
This table reports the first-stage relationship between compulsory school laws and educational
attainment. The sample comprises individuals reported in the 5% samples of the 1980, 1990, and
2000 census. We include 18-75 year olds. The dependent variables of interest are the number of
years of schooling attained in column (1) and an indicator for whether the individual graduated
high school in column (2). The independent variables of interest indicate whether the state in
which the individual was born prohibited drop-out until a child had completed 9th grade, 10th
grade, or 11th grade and higher. Other controls include fixed effects for gender, race, 3-year age
groups, 10-year birth cohorts, state of birth, state of residence, census year and a cublic
polynomial in earned income. Regressions also include state of residence fixed effect interacted
with a dummy variable for being born in the South and turning age 14 in 1958 or later, to account
for the impact of Brown v. Board of Education for blacks. Standard errors, corrected for arbitrary
correlation within state of birth-year of birth, are in parentheses. (Numbers with *** indicate
significance at the 1-percent level.)
High school
Years of schooling
(1)
(2)
Compulsory Attendence = 9
Compulsory Attendence = 10
Compulsory Attendence = 11
Num of Observations
R-Squared
F-Stat of Instruments
0.214 ***
(0.018)
0.199 ***
(0.024)
0.266 ***
(0.028)
14,913,356
0.234
47.2
0.039 ***
(0.003)
0.041 ***
(0.004)
0.055 ***
(0.005)
14,913,356
0.178
52.4
Table IV
2SLS Estimates of the Effect of Schooling on Income from Various Sources
This table reports the second-stage relationship between financial market participation and educational attainment.
The sample comprises individuals reported in the 5% samples of the 1980, 1990, and 2000 census. We include 1875 year olds (50-75 year olds when considering retirement income). The dependent variable of interest is whether
the household receives income from investments or retirement savings (Panel A) and the amount (Panel B). Years
of schooling is instrumented with compulsory schooling laws. In addition, we include as controls, but do not report,
fixed effects for gender, race, 3-year age groups, 10-year birth cohorts, state of birth, state of residence, census year
and a cubic polynomial in earned income. Regressions also include state of residence fixed effects interacted with a
dummy variable for being born in the South and turning age 14 in 1958 or later to account for the impact of Brown
v. Board of Education for blacks. Top-coded individuals (see text) are dropped in panel B. Standard errors,
corrected for arbitrary correlation within state of birth-year of birth, are in parentheses. (Numbers with *** indicate
significance at the 1-percent level.)
Income from
Income from Retirement
Investments
Savings
(1)
(2)
Panel A: Any Investment Income
Years of schooling
7.50 ***
5.85 ***
(0.52)
(1.04)
Num of Observations
F-stat of excluded instruments
Panel B: Amount of Investment Income
Years of schooling
14,913,356
47.2
1759.41 ***
(128.29)
4,150,828
45.0
965.59 ***
(129.66)
Num of Observations
F-stat of excluded instruments
14,842,001
47.2
4,117,987
44.5
F-stat of excluded instruments
50.0
43.9
Table V
Reduced-Form Estimates of the Effect of Eduction on Credit Outcomes, FRBNY Consumer Credit Panel/Equifax
This table shows cross-sectional regressions of credit outcomes on education, measured by changes in compulsory attendance laws. The sample comprises a 5% panel of American borrowers, restricted to borrowers who have
data in every quarter of the panel from 1999 to 2011. Credit Score is averaged for each individual accross all quarters of data, and it can range from 280 to 850. The % of Balance Current represents the non-delinquent balance
on credit cards divided by the total credit card balance, averaged over the entire panel. The % of Quarters Delinquent represents the proportion of quarters an individual has any delinquent balance on his/her credit card bills.
Bankruptcy and Foreclosure are indicators for having undergone bankruptcy or foreclosure at least once, respectively, between 1992 and 2011. We include 35-75 year olds. The independent variables of interest indicate whether
the state in which the individual was born prohibited drop-out until a child had completed 9th, 10th, or 11th grade. Control variables included (coefficients not reported) in these regressions were dummies for 3-year age
cohorts, 10-year birth cohorts, state-of-residence, and a dummy for being born in a southern state and turning age 14 in 1958 or later. State-of-birth is proxied by an individual's state-of-residence in the first quarter of 1999.
Standard errors, corrected for arbitrary correlation within state of birth-year of birth, are in parentheses. (Numbers with *, **, or *** indicate significance at the 10-, 5-, or 1-percent level, respectively.)
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
% Quarters Delinquent
Credit Score
% Balance Current
Compulsory Attendance = 9
Compulsory Attendance = 10
Compulsory Attendance = 11
-1.096 **
(0.510)
1.461 **
(0.619)
1.669 ***
(0.480)
Years of Compulsory Schooling
Num of Observations
R-squared
p-value for F-stat of Compulsory Attendance
Additional fixed effects
0.253 **
(0.106)
5,732,690
0.141
0.000
none
(10)
Compulsory Attendance = 9
Compulsory Attendance = 10
Compulsory Attendance = 11
Additional fixed effects
5,732,690
0.141
none
(11)
Bankruptcy
0.056 ***
(0.106)
5,732,690
0.231
zipcode
(12)
-0.0024 *
(0.0014)
-0.0132 ***
(0.0027)
-0.0098 ***
(0.0017)
Years of Compulsory Schooling
Num of Observations
R-squared
p-value for F-stat of Compulsory Attendance
0.0003
(0.0004)
0.0006
(0.0004)
0.0012 ***
(0.0003)
none
0.0002 **
(0.0001)
5,329,619
0.020
0.003
none
(13)
5,329,619
0.020
none
(14)
Foreclosure
0.0001 ***
(0.0001)
5,329,619
0.051
zipcode
(15)
-0.0036 ***
(0.0012)
-0.0061 ***
(0.0015)
-0.0122 ***
(0.0021)
-0.0021 ***
(0.0003)
5,750,005
0.032
0.000
0.0007
(0.0006)
0.0002
(0.0007)
-0.0021 ***
(0.0005)
5,750,005
0.032
none
-0.0019 ***
(0.0003)
5,750,005
0.057
zipcode
-0.0022 ***
(0.0004)
5,750,005
0.020
0.000
none
5,750,005
0.020
none
-0.0020 ***
(0.0003)
5,750,005
0.038
zipcode
-0.0003 **
(0.0001)
5,750,005
0.045
0.000
none
5,750,005
0.045
none
-0.0001
(0.0001)
5,750,005
0.083
zipcode
Table VI
Two Sample IV Estimates of the Effect of Schooling on Credit Outcomes, FRBNY Consumer Credit Panel/Equifax
This table shows cross-sectional second stage two sample IV regressions of credit outcomes on education. The sample comprises a 5% panel of
American borrowers, restricted to borrowers who have data in every quarter of the panel from 1999 to 2011. Credit Score is averaged for each
individual accross all quarters of data, and it can range from 280 to 850. The % of Balance Current represents the non-delinquent balance on
credit cards divided by the total credit card balance, averaged over the entire panel. The % of Quarters Delinquent represents the proportion of
quarters an individual has any delinquent balance on his/her credit card bills. Bankruptcy and Foreclosure are indicators for having undergone
bankruptcy or foreclosure at least once, respectively, between 1992 and 2011. We include 35-75 year olds. The independent variable, years of
schooling, is instrumented using compulsory schooling laws. Control variables included (coefficients not reported) in these regressions were
dummies for 3-year age cohorts, 10-year birth cohorts, state-of-residence, and a dummy for being born in a southern state and turning age 14 in
1958 or later. State-of-birth is proxied by an individual's state-of-residence in the first quarter of 1999. Panel A reports second stage results for
variables using all available quarters in the panel dataset. Panel B reports results for bankruptcy and foreclosure indicators separately for the precrisis period (1999Q2-2007Q3) and the post-crisis period (2007Q3-2011Q4). The top half of each panel reports robust standard errors, following
Murphy and Topel (1985). The bottom half of each panel reports second stage results using 100 bootstraps. Standard errors, corrected for
arbitrary correlation within state of birth-year of birth, are in parentheses. (Numbers with *, **, or *** indicate significance at the 10-, 5-, or 1percent level, respectively.)
% Quarters
Dependent Variable:
Credit Score
% Balance Current
Bankruptcy
Foreclosure
Delinquent
(1)
(2)
(3)
(4)
(5)
Panel A: Entire Sample
Murphy Topel Standard Errors
Years of Schooling
7.705 ***
0.0052 **
-0.0133 ***
-0.033 ***
-0.057 ***
(2.781)
(0.0023)
(0.0035)
(0.011)
(0.018)
Num of Observations
Bootstrap Estimates
Years of Schooling
Panel B: Pre- and Post-Crisis
Murphy Topel Standard Errors
Years of Schooling
Num of Observations
Bootstrap Estimates
Years of Schooling
5,182,364
8.853 *
(4.762)
4,852,175
0.0056 *
(0.0031)
1999Q2-2007Q3
Bankruptcy
Foreclosure
(6)
0.009
(0.011)
(7)
0.006
(0.007)
5,198,529
5,198,529
0.007
(0.011)
0.005
(0.006)
5,198,529
-0.0137 ***
(0.0053)
5,198,529
-0.037
(0.023)
5,198,529
-0.058 ***
(0.010)
2007Q3-2011Q4
Bankruptcy
(8)
-0.023 ***
(0.009)
4,822,592
-0.024 ***
(0.005)
Foreclosure
(9)
-0.046 *
(0.023)
5,051,449
-0.046 ***
(0.008)
Table VII
Estimates of the Effect of Knowledge and Ability on Savings, NLSY
Data are from the National Longitudinal Survey of Youth. Panel A reports whether the individual has any money left at the end of the month (column 1), or in any of the
listed assets (columns 2-8), and Panel B gives the amount in dollars. Cognitive ability (knowledge and ability) is measured by tests given around age 17. Additional
controls included are sibling-group fixed effects, family income in every year with dummies proxying for missing data and fixed effects for age, gender, survey year, birth
order and birth year. Standard errors are clustered at the individual level. (Numbers with *, **, or *** indicate significance at the 10-, 5-, or 1-percent level, respectively.)
Dependent Variable: Any $ in
Asset
Money Left
Stocks, Bonds
& Mutual
Funds
Savings
IRAs &
Keogh
Tax-Deferred
Accounts
Years
1990 - 2004
1988 - 2000
1985 - 2000
1994 - 2000
1994 - 2000
1994 - 2000
1988 - 2000
1988 - 2000
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
Panel A: Any Investment Income
Knowledge
Ability
Years of Education
Num of Observations
R-Squared
F-stat of Knowledge & Ability
Panel B: Amount
Knowledge
Ability
Years of Education
Num of Observations
R-Squared
F-stat of Knowledge & Ability
2.559 **
(1.287)
3.632 ***
(1.226)
0.173
(0.372)
25993
0.38
0.00
2394.04
(4551.32)
-430.40
(4047.50)
1568.23
(1349.05)
15089
0.48
0.82
3.390 ***
(1.134)
1.832 *
(0.942)
1.508 ***
(0.329)
1.073
(0.935)
5.183 ***
(0.938)
1.653 ***
(0.256)
1.167
(1.515)
1.926
(1.360)
1.055 **
(0.468)
34663
0.33
0.00
44006
0.42
0.00
14220
0.42
0.07
674.29
(607.41)
342.06
(473.52)
314.63 *
(183.30)
127.19
(660.78)
1065.39 *
(564.26)
300.68 *
(181.25)
736.83
(835.33)
-457.13
(751.17)
598.55 **
(299.95)
33455
0.19
0.06
44432
0.26
0.02
13645
0.34
0.68
0.538
(1.709)
5.333 ***
(1.512)
1.262 **
(0.512)
CDs, Loans, Rights to Estate, Income from Other
Mortgage
Sources (Interest,
Investment
Assets
Rent, Dividends, etc.)
Trust
-0.002
(0.844)
1.180 *
(0.706)
0.290
(0.232)
-0.610
(0.424)
0.115
(0.419)
-0.111
(0.140)
14195
0.39
0.00
14239
0.24
0.10
34696
0.22
0.28
1936.09
(1303.73)
1258.74
(1218.08)
630.97
(448.32)
-590.09
(439.07)
581.39
(449.94)
36.03
(128.83)
-326.16
(1002.24)
-464.75
(778.82)
-83.11
(269.48)
13174
0.38
0.01
14103
0.16
0.36
34118
0.12
0.58
5.317 ***
(0.907)
4.092 ***
(0.844)
1.477 ***
(0.225)
76372
0.39
0.00
30.58
(56.95)
-18.81
(45.02)
90.20 ***
(17.97)
74277
0.11
0.86
Table VIII
2SLS Estimates of the Effect of Schooling on Possible Mechanisms from Various Sources
The sample comprises individuals reported in the 5% samples of the 1980, 1990, and 2000 census, except column (2) which only uses 1990 and 2000 data and column (5)
which uses data only from 1980 due to changes in the variables reported. The sample includes individuals who are aged 18-75 and born before 1965. The independent
variable, years of schooling, is instrumented using compulsory schooling laws. Other controls include fixed effects for gender, race, 3-year age groups, 10-year birth cohorts,
state of birth, state of residence, census year and a cubic polynomial in earned income. Regressions also include state of residence fixed effects interacted with a dummy
variable for being born in the South and turning age 14 in 1958 or later to account for the impact of Brown v. Board of Education for blacks. In column 1, the dependent
variable is the share of individuals aged 65-70 in 1970 in the same occupation and state who report receiving retirement income. In column 2, the dependent variable is the
share of neighbors in the current year aged 65 and above who report receiving retirement income other than Social Security. (Numbers with *** indicate significance at the 1percent level.)
Share of Employees in
Moved in Past Five Years
Has a First
Has a Second
Share of Neighbors
Dependent Variable:
Occupation with
Mortgage
Mortgage
with Pension
City
House
Pension
(1)
(3)
(4)
(5)
(6)
(2)
Years of schooling
Num of Observations
F-stat of excluded instruments
0.985 ***
(0.318)
13,013,023
48.4
0.921 ***
(0.299)
9,400,737
5.1
1.102
(1.318)
10,240,086
52.2
-2.640 ***
(0.726)
8,231,828
45.7
0.878
(0.827)
3,146,775
53.9
-1.696 ***
(0.652)
14,119,649
38.6
Table IX
Estimates of the Effect of Not Feeling in Control on Savings, NLSY
Data are from the National Longitudinal Survey of Youth. Panel A reports whether the individual has any money left at the end of the month (column 1), or in
any of the listed assets (columns 2-8), and Panel B gives the amount in dollars. The independent variable of interest is a measure of whether the respondent
agrees with the statement “I have little control over the things that happen to me,” with 1 indicating strong disagreement and 4 indicating strong agreement.
Additional controls included are sibling-group fixed-effects, family income in every year with dummies proxying for missing data and fixed effects for age,
gender, survey year, birth order and birth year. Standard errors are clustered at the individual level. (Numbers with *, **, or *** indicate significance at the 10-,
5-, or 1-percent level, respectively.)
Dependent Variable:
Panel A: Any Money
Little Control
Num of Observations
R-Squared
Panel B: Amount
Little Control
Num of Observations
R-Squared
Money Left
Stocks, Bonds
& Mutual
Funds
(1)
(2)
-2.82 ***
(0.95)
0.01
(0.79)
Savings
IRAs &
Keogh
Tax-Deferred
Accounts
CDs, Loans,
Mortgage
Assets
Rights to
Estate,
Investment
Trust
Income from
Other Sources
(Interest, Rent,
Dividends, etc.)
(3)
(4)
(5)
(6)
(7)
(8)
-2.09 **
(0.89)
-1.20
(0.92)
-1.54
(1.08)
-0.26
(0.48)
0.10
(0.34)
-1.44 *
(0.77)
21,229
0.41
21,261
0.38
17,593
0.49
13,831
0.42
13,807
0.40
13,851
0.24
21,292
0.26
24,503
0.45
-10323.00 ***
(3090.18)
27.73
(527.41)
186.16
(612.54)
-221.19
(510.84)
-1276.97
(835.30)
3.88
(198.17)
-85.34
(743.49)
-14.40
(67.58)
12,575
0.50
20,399
0.24
18,720
0.37
13,277
0.35
12,823
0.38
13,721
0.16
20,924
0.17
23,558
0.21
9
Not for Publication: On-Line Data Appendix
9.1
Comparison of Census and Survey of Consumer Finances Data
Census data have not been used much to track investment income, and one may naturally
have concerns about the reliability of the data, as well as comparability with more standard
data sources, such as the Survey of Consumer Finances (SCF). In this appendix, we compare
the means and distributions of the variables of interest, and describe the relationship between
investment income and …nancial wealth. In the census data, we use the variable “INCINVST”
as a measure of investment income, and the variable “INCRETIR” for retirement income (see
Ruggles et. al, 2004). For the SCF, we use the sum of non-taxable investment income (x5706),
other interest income (x5708), dividends (x5710), and income from net rent, trusts, or royalties
(x5714). In both the census and the SCF, reported numbers appear to be pre-tax income,
though the census …gures are less precise. Neither the SCF nor census measure includes capital
gains. (The income portion of the questionnaire for the census is reproduced below). Retirement
income is measured in the SCF as the sum of current account-type pension bene…ts and nonaccount-type bene…ts.31
Table A2 presents the means, standard deviations, ranges, and percentiles for the investment
income and retirement income variables. Analysis is limited to a sample of households aged 3675, who earn investment income below $50,000. (This is the same sample used to evaluate the
e¤ect of education on investment income.) Relative to the SCF, census respondents appear to
underreport both investment and retirement income. The mean investment income is 17 percent
lower, at $1,264, compared to the SCF average …gure of $1,515. A nearly identical percent fewer
report receiving any invesment income: the …gure is 33% in the SCF, and 27% in the Census.
We speculate that the reason for this is that the survey of consumer …nances is much more
detailed than the census, and that the SCF is done in person. Nonetheless, the distributions
appear to be comparable, with a median of zero in both datasets, and similar 75th, , 90th , and
99th percentiles.
The apparent underreporting of retirement income in the U.S. Census is more severe: the
average reported in the census is approximately 30 percent lower than the average in the SCF,
and approximately 20 percent fewer individuals report any retirement income in the U.S. Census:
22 percent, against 27 percent in the SCF. Nonetheless, again the two distributions appear to
track one another reasonably closely.
The results suggest that the dollar …gures estimated from the census may not be precisely
correct. Nevertheless, the two data sources are not strikingly di¤erent, and the e¤ect on estimated coe¢ cients is likely relatively small.
An alternative check of the comparability of the two datasets is to regress the dependent
variables used in our main paper on individual characteristics, such as age, income, race,and
education level. The coe¢ cients obtained from the SCF and Census are quite similar. Indeed,
equality cannot be rejected for 35 out of 36 demographic variables. (Results not reported.)
Appendix Table A3 provides a detailed breakdown of …nancial market participation, using
the 2001 SCF. Each row reports the fraction of households that use a variety of …nancial services,
for a given range of investment income. The ranges were chosen so that there would be at least
31
The former are, x6464, x6469, x6474, x6479, x6484, and x6489, and the latter are x5326, x5326,x5334, x5418,
x5426, x5434. All values are converted to annual …gures, in 2000 dollars.
30 observations in each range. For example, 61% of households report earning no investment
income. Among this population, 88% have a transaction account (checkings, savings, or moneymarket fund). As reported investment income increases, …nancial market participation generally
increases.
A second potential concern with the use of census data is that information is available on investment income, not …nancial wealth. In particular, if the relationship between …nancial wealth
and investment income is highly non-linear, results using one measure may not translate well
to the other. Figure A1 plots the relationship between investment income and …nancial wealth,
from a Fan local linear regression, using data from the 2001 Survey of Consumer Finances. While
visual inspection reveals a slight increase in slope around the point of $25,000 (consistent with
evidence from Calvet, Campbell, and Sodini, 2007, that investors with higher income achieve
higher risk-adjusted returns), to a …rst approximation, the relationship is linear.
9.2
Census Income Questions
We reproduce here the questions on income from the 2000 Census “long form.”
31. INCOME IN 1999 - Mark [X] the “Yes" box for each income source received during
1999 and enter the total amount received during 1999 to a maximum of $999,999. Mark
[X] the “No" box if the income source was not received. If net income was a loss, enter the
amount and mark [X] the “Loss" box next to the dollar amount.
For income received jointly, report, if possible, the appropriate share for each person;
otherwise, report the whole amount for only one person and mark the “No" box for the
other person. If exact amount is not known, please give best estimate.
a. Wages, salary, commissions, bonuses, or tips from all jobs - Report amount before
deductions for taxes, bonds, dues, or other items.
O Yes
Annual amount - Dollars
$[ ][ ][ ],[ ][ ][ ].00
O No
b. Self-employment income from own nonfarm businesses or farm businesses, including
proprietorships and partnerships - Report NET income after business expenses.
OYes
Annual amount - Dollars
$[ ][ ][ ],[ ][ ][ ].00
O No
O Loss
c. Interest, dividends, net rental income, royalty income, or income from estates and
trusts - Report even small amounts credited to an account.
O Yes
Annual amount - Dollars
$[ ][ ][ ],[ ][ ][ ].00
O No
d. Social Security or Railroad Retirement
O Yes
Annual amount - Dollars
$[ ][ ][ ],[ ][ ][ ].00
O No
e. Supplemental Security Income (SSI)
O Yes
Annual amount - Dollars
$[ ][ ][ ],[ ][ ][ ].00
O No
f. Any public assistance or welfare payments from the state or local welfare o¢ ce
O Yes
Annual amount - Dollars
$[ ][ ][ ],[ ][ ][ ].00
O No
g. Retirement, survivor, or disability pensions - Do NOT include Social Security.
O Yes
Annual amount - Dollars
$[ ][ ][ ],[ ][ ][ ].00
O No
h. Any other sources of income received regularly such as Veterans’(VA) payments, unemployment compensation, child support, or alimony - Do NOT include lump-sum payments
such as money from an inheritance or sale of a home.
O Yes
Annual amount - Dollars
$[ ][ ][ ],[ ][ ][ ].00
O No
32. What was this person’s total income in 1999? Add entries in questions 31 a-31 h;
subtract any losses. If net income was a loss, enter the amount and mark [X] the “Loss" box
next to the dollar amount.
O None OR Annual amount - Dollars
$[ ][ ][ ],[ ][ ][ ].00
O Loss
9.3
National Longitudinal Survey of Youth 1979
The NLSY79 cohort is a nationally representative sample of young people aged 14-22 when the
survey began in 1979. Respondents were interviewed annually until 1994 and then biennially
since 1996. While each survey contains di¤erent questions and often special sets of questions
on topics such as military participation, time-use or alcohol and substance abuse, each survey
contains a core set of questions on respondents’labor force experience, labor market attachment,
investments in education and training. Summary statistics on the variables used in this paper
are available in Appendix Table A1.
From these questions, sta¤ at the Center for Human Resource Research create consistently
coded variables on a number of demographic characteristics. Two such variables are used in the
estimates above. Information on educational attainment and enrollment has been used to create
a variable for highest grade completed as of May 1 of the survey year. Separate questions on
income from various sources have been used to create a consistent estimate of “total net family
income". This variable summarizes all income received in the household, and does not account
for taxes or other adjustments. From 1979 to 1986 total net family income was calculated from
a Household Interview administered to parents for respondents who lived with their parents.
While 19 sources of income are asked separately (such as wages, military income, farm income,
business income, inheritance and gifts), income from investments is included in the “Other
Income" category: “Aside from the things you have already told me about, during 19XX, did
you (or your spouse/partner) receive any money from any other sources such as the ones on
this card? For example, things like interest on savings, payments from social security, net rental
income, or any other regular or periodic sources of income?" Questions on di¤erent types of
assets, such as IRAs and Keogh accounts or 401Ks and pre-tax annuities, di¤er slightly across
years, resulting in question-speci…c sample periods.
In 1980, respondents in the NLSY79 sample were adminstered the Armed Services Vocational
Aptitude Battery (ASVAB) in a joint e¤ort of the U.S. Departments of Defense and Military
Services to update the ASVAB norms. In total, 11,914 NLSY79 respondents (94% of the sample) participated in the test. The ASVAB measures di¤erent aspects of ability, knowledge and
skill in 10 tests, each in one of the following areas: general science, arithmetic reasoning, word
knowledge, paragraph comprehension, numerical operations, coding speed, auto and shop information, mathematics knowledge, mechanical comprehension and electronics information. Scores
on these tests are used to estimate each respondent’s percentile score in the Armed Forces Qualifying Test (AFQT), as well as our measures of knowledge and ability. The AFQT score is a
function of the individual’s score on tests in arithmetic reasoning, word knowledge, paragraph
comprehension and numerical operations. Our measure of innate ability uses these tests plus a
test in coding speed, while our measure of acquired knowledge includes tests in general science,
auto and shop information, mathematics knowledge, mechanical comprehension and electronics
information. Our results are robust to slightly di¤erent decompositions.
Online Appendix Table A1
Summary Statistics, NLSY
Summary statistics from the National Longitudinal Survey of Youth. The survey was conducted annually from 1979-1984,
and biennially since 1996. Column (1) indicates the years from which data are used.
Variable
Year(s)
Mean
St. Dev.
Min
Max
(1)
(2)
(3)
(4)
(5)
Knowledge
1981
58.71
22.47
0
119
Ability
1981
117.82
40.74
0
213
Cognitive Ability
1981
36.81
28.19
1
99
Years of Education
1985 - 2004
12.94
2.33
0
20
Age (in year 1979)
1979
17.16
2.06
14
22
Male
1979
0.52
0.50
0
1
Total Net Family Income
1979 - 2004
51197.32
85407.59
0
1637987
Pearlin "Little Control Over Life"
1992
1.80
0.67
1
4
Any …
1990 - 2004
65.60
47.50
0
100
Money Left
1985 - 2000
63.66
48.10
0
100
Savings
1988 - 2000
16.16
36.81
0
100
Stocks, Bonds & Mutual Funds
1994 - 2000
19.48
39.60
0
100
IRAs & Keogh
1994 - 2000
32.51
46.84
0
100
Tax-Deferred Accounts
1994 - 2000
4.96
21.72
0
100
CDs, Loans, Mortgage Assets
1988 - 2000
3.22
17.65
0
100
Rights to Estate, Investment Trust
1979 - 2002
29.90
45.78
0
100
Income from Other Sources
Amount …
1990 - 2004
65147
103348
1
989100
Money Left
1985 - 2000
6699
24144
0
835000
Savings
1988 - 2000
2556
19379
0
989100
Stocks, Bonds & Mutual Funds
1994 - 2000
4671
23928
0
549500
IRAs & Keogh
1994 - 2000
9513
33984
0
549500
Tax-Deferred Accounts
1994 - 2000
1014
11397
0
549500
CDs, Loans, Mortgage Assets
1988 - 2000
2236
40288
0
3114800
Rights to Estate, Investment Trust
1979 - 2002
500
3373
0
168780
Income from Other Sources
Percentile …
1990 - 2004
42
25
0
89
Money Left
1985 - 2000
37
29
0
87
Savings
1988 - 2000
11
27
0
82
Stocks, Bonds & Mutual Funds
1994 - 2000
13
28
0
81
IRAs & Keogh
1994 - 2000
21
32
0
81
Tax-Deferred Accounts
1994 - 2000
3
16
0
80
CDs, Loans, Mortgage Assets
1979 - 2002
21
32
0
86
Income from Other Sources
Online Appendix Table A2
Comparison of Data from 2001 SCF and 2000 Census
Note: This table compares the means, standard deviations, and percentiles for the key variables, using data from both the Census and the Survey of Consumer
Finances. The Census data are from the 2000 census, while the SCF data are from the 2001 survey of consumer finances. The sample for investment income
variables in both surveys is adults aged 36-75 who report investment income below $50,000. For retirement income, the sample is individuals aged 50 to 75 who
report retirement income less than or equal to $52,000. N indicates the number of unique individuals used to estimate numbers; for the SCF, appropriate weights
were used.
Investment Income
Any Investment Income
Retirement Income
Any Retirement Income
SCF
Census
SCF
Census
SCF
Census
SCF
Census
Mean
Standard Deviation
Min
Max
1515
5089
0
49800
1264
4543
0
49900
0.33
0.47
0
1
0.27
0.45
0
1
4114
9545
0
50400
2866
7477
0
51000
0.27
0.44
0
1
0.22
0.41
0
1
Percentiles
1%
10%
25%
Median
75%
90%
99%
0
0
0
0
200
3500
29000
0
0
0
0
50
2900
25000
0
0
0
0
1
1
1
0
0
0
0
1
1
1
0
0
0
0
1320
15600
43200
0
0
0
0
0
11165
36539
0
0
0
0
1
1
1
0
0
0
0
0
1
1
N
2,720
5,427,616
2,735
5,450,827
1,739
2,883,474
1,739
2,883,474
Online Appendix Table A3: Relationship Between Reported Investment Income and Financial Market Participation
This table reports the fraction of individuals in the 2001 Survey of Consumer Finances who report having various financial assets, for given ranges of financial income. For example, 88% of those who report $0 in investment
income posess transaction accounts, while 100% of those whose investment income is between $1 and $100 have transaction accounts. N reports the number of observations for the sample, while the subsequent column gives the
fraction of the population falling within this bracket. The sample is restricted to individuals aged 37-75, and includes all states except Alaska.
Investment Income
Between
0
1
101
201
301
401
501
751
1,001
2,001
3,001
4,001
5,001
7,001
9,001
11,001
13,001
15,001
20,001
30,001
40,001
Entire Sample
and
0
100
200
300
400
500
750
1,000
2,000
3,000
4,000
5,000
7,000
9,000
11,000
13,000
15,000
20,000
30,000
40,000
50,000
N
1485
74
58
52
39
31
67
41
101
92
60
45
67
57
39
42
40
67
85
63
39
Share of
Population
61.18%
3.20%
2.38%
2.09%
1.47%
1.31%
2.53%
1.98%
3.78%
3.19%
1.97%
1.46%
1.72%
1.74%
0.94%
0.78%
0.86%
1.50%
1.47%
1.26%
0.53%
3219
100%
Average
Household
Investment
Income
Income
0
56,905.05
55,211.74
43.28
70,446.57
137.12
83,534.93
244.60
81,405.99
339.81
87,256.52
454.26
87,493.63
651.68
131,784.36
870.34
108,518.24
1,405.45
99,132.36
2,460.72
77,221.07
3,432.61
143,430.14
4,548.36
131,413.99
5,889.38
136,458.55
7,841.64
127,695.19
9,833.59
186,370.45
11,868.38
140,045.60
14,033.14
186,290.65
17,251.65
226,698.89
24,153.53
195,245.78
34,065.13
784,627.31
44,570.31
95,717.47
6,450.95
Transaction
Account
88%
100%
100%
100%
100%
100%
100%
97%
99%
100%
100%
99%
99%
100%
100%
99%
100%
100%
99%
100%
100%
Checking
Account
78%
92%
98%
91%
94%
88%
91%
90%
90%
90%
88%
90%
85%
88%
98%
94%
84%
96%
85%
88%
85%
92%
83%
Percentage of Individuals with:
Savings
Individual
Account
CD
Stocks
51%
10%
10%
79%
6%
15%
83%
18%
33%
87%
11%
18%
69%
37%
24%
69%
24%
25%
73%
26%
39%
70%
33%
45%
66%
27%
43%
73%
25%
40%
57%
33%
43%
72%
40%
49%
59%
36%
59%
62%
34%
48%
58%
19%
48%
70%
27%
61%
43%
41%
40%
70%
42%
50%
52%
37%
68%
48%
41%
71%
49%
14%
84%
57%
16%
23%
Mutual
Fund
12%
13%
22%
26%
21%
28%
27%
29%
47%
41%
46%
48%
60%
59%
51%
48%
49%
86%
69%
57%
83%
Retirement
Account
48%
57%
72%
85%
70%
80%
70%
77%
85%
76%
77%
87%
85%
72%
72%
90%
87%
94%
91%
89%
95%
24%
60%
Appendix Table A4
Estimates of the Effect of Education on Cognitive Ability, Knowledge, Ability and Being
in Control, NLSY
This table uses variation within families to estimate the relationship between educational
attainment and cognitive ability, and the relationship between feeling in control and cognitive
ability and education. Data are from the National Longitudinal Survey of Youth. Cognitive
ability is measured using the AFQT test but is also decomposed into knowledge and innate
ability. The dependent variable in column (4) is a measure of whether the respondent agrees
with the statement “I have little control over the things that happen to me,” with 1 indicating
strong disagreement and 4 indicating strong agreement. Other controls include sibling-group
fixed effects, log family income in every year with dummies proxying for missing data and
fixed effects for age, gender, race, survey year, birth order and birth year. Standard errors are
clustered at the individual level. (Numbers with *, or *** indicate significance at the 10-, or 1percent level, respectively.)
Dependent Variable:
Year
Panel A:
Years of Education
Num of Observations
R-Squared
Panel B:
Cognitive Ability
Num of Observations
R-Squared
Cognitive
Ability (AFQT)
1981
(1)
5.193 ***
(0.303)
4594
0.9
Cognitive
Ability
(Knowledge)
1981
(2)
4.312 ***
(0.264)
4604
0.9
Cognitive
Ability
(Ability)
LittleControl
1981
(3)
1993
(4)
6.428 ***
(0.435)
4604
0.8
-0.033 ***
(0.010)
3807
0.5
-0.0017 *
(0.0010)
3682
0.5