Sociology Mind
2011. Vol.1, No.4, 192-205
Copyright © 2011 SciRes. DOI:10.4236/sm.2011.14025
Toward a Children’s Savings and College-Bound Identity
Intervention for Raising College Attendance Rates:
A Multilevel Propensity Score Analysis
William Elliott1*, Gina Chowa2, Vernon Loke3
1School of Social Work; University of Pittsburgh, Pittsburgh, USA;
2School of Social Work, University of North Carolina, Chapel Hill, USA;
3School of Social Work, Eastern Washington University, Cheney, USA.
Email: *welliott@ku.edu
Received May 24th, 2011; revised July 3rd, 2011; accepted August 13th, 2011.
It has been suggested that children’s savings programs will be more effective if they are combined with strate-
gies to build children’s college-bound identities. In this study we use a multi-level treatment approach to pro-
pensity score analysis to test this proposition. Findings suggest that children who have savings and are certain
they will graduate from a four-year college are more likely to attend college than their counterparts. Given this,
we suggest that children’s savings policies designed to increase college attendance rates will be more effective if
they include strategies for building children’s college-bound identity and college-bound identity programs will
be more effective if they are linked to children’s savings programs.
Keywords: Wealth, Assets, College Attendance, Identity-Based Motivation, Savings, Child Development
Accounts (CDAs), College Expectations, PSID
Introduction1
In 2008, 55% of children who graduated high school and
were from the lowest family-income quintile enrolled in college
compared to 80% of children who graduated high school and
were from the highest-income quintile, a gap of 25% (Baum,
Ma, & Payea, 2010). A well-recognized barrier to college ac-
cess among low- and moderate-income children is high college
costs (ACSFA, 2010). In recent years, the federal government
has increasingly relied on loans such as the Federal Stafford
and PLUS loan programs as a way to combat high costs. How-
ever, emphasis on loans has led to a growing number of chil-
dren leaving college burdened with high amounts of debt (Col-
lege Board, 2009).
Finding new and innovative ways to increase college atten-
dance rates among low- and moderate-income children is a
priority in today’s global, high tech economy. Researchers have
identified a number of factors, including social capital (Porfeli,
Wang, Audette, McColl, & Algozzine, 2009), human capital
(Paulsen, 2001), and economic capital (Coleman, 1988) as be-
ing key predictors of college attendance. In this study, we focus
on economic capital. According to Sirin (2005), economic
capital is perhaps the most widely applied contextual variable
in research on education. Research shows that, as family re-
sources available to youth increase, their educational perform-
ance, high school graduation, and college attendance rates im-
prove (Coleman, Campbell, Hobson, McPartland, Mood, &
Weingeld, 1966). However, it is not merely the amount of re-
sources but the diversity of the resources that leads to greater
academic achievement. As Coleman et al. (1996) posit, children
from families of higher SES do better because they are exposed
to a wider set of resources that they can tap into to promote
learning. While education research has given considerable at-
tention to income as a form of economic capital, assets have
been largely overlooked, particularly children’s financial assets.
In the last decade, Child Development Accounts (CDAs)
have been proposed as a potentially novel and promising
mechanism for helping to build children’s assets and helping
them pay for college (Sherraden 1991). An example of a CDA
policy is the America Saving for Personal Investment, Retire-
ment, and Education (ASPIRE) Act. ASPIRE would create
“KIDS Accounts,” or a savings account for every newborn,
with an initial $500 deposit, along with opportunities for finan-
cial education.2 Children living in households with incomes
below the national median would be eligible for an additional
contribution of up to $500 at birth and a savings incentive of
$500 per year in matching funds for amounts saved in accounts.
When account holders turn 18, they would be permitted to
make tax-free withdrawals for costs associated with post-sec-
ondary education, first-time home purchase, and re-tirement
security.
However, it is desirable to conduct advance tests of large
scale children’s savings policies like the ASPIRE act prior to
passing them into legislation. Over the last five years, research-
ers have conducted a number of tests of CDAs using a variety
of proxies. Most of this research has focused on household
assets (e.g., Conley, 2001; Destin, 2009; Haveman & Wolff,
2005; Nam & Huang, 2009; Williams Shanks & Destin, 2009).
Household assets are most commonly defined as net worth (i.e.,
total family assets minus debt), liquid assets (i.e., easily con-
verted into cash), and illiquid assets (i.e., hard to convert into
cash). It is beyond the scope of this article to provide a com-
prehensive review of the research in this area.3 Briefly, some
researchers find that children who live in high wealth house-
1This publication is part of the College Savings Initiative, a research and
policy design collaboration between the Center for Social Development at
Washington University in St. Louis and the New America Foundation in
Washington, DC. The College Savings Initiative is supported by the Lumina
Foundation for Education and the Bill & Melinda Gates Foundation.
2At this writing, the ASPIRE Act remains on the Congressional agenda
(http://www.newamerica.net/publications/policy/aspire_act_bill_summary).
3For a comprehensive review of research on household assets and chil-
dren’s education see Elliott, Destin, and Friedline
(
2011
)
.
W. ELLIOTT ET AL. 193
holds are more likely to have higher math and reading scores
(Zhan, 2006), higher high school graduation rates (Nam &
Huang, 2009), higher college attendance rates (Conley, 2001),
and higher college graduation rates (Zhan & Sherraden, 2009)
than children from low wealth households.
Alongside research on household assets, a less developed
body of research has emerged in recent years focusing on when
children have savings of their own. Researchers studying chil-
dren’s savings posit that ownership has unique qualities. This is
in line with consumer research findings. From a consumer re-
search perspective, ownership instills in people (including chil-
dren as young as age five) a greater sense of perceived control
and sense of self (e.g., Belk, 1988; Furby, 1980). According to
Belk (1988), it is through the process of ownership that items
such as money, other people, and pets can become part of the
self. The greater exercise of power a child has over a possession,
such as money, the more closely identified with the self it be-
comes (Furby, 1978). What makes ownership important is what
children perceive that ownership gives them control over – for
e.g., a stake in financing college. In a study of 51 fourth grade
children in a college savings program, Elliott, Sherraden, John-
son, and Guo (2010) find that children who are in the school
savings program are statistically more likely to perceive that
saving is a way to help pay for college than children in a com-
parison group.
The unique effect of ownership may provide low- and mod-
erate-income children with a means to overcome everyday nega-
tive signals that result from a lack of family assets. Children’s
savings make future identities particularly salient, as children are
actively involved in the process that is linked to their college
goals. For example, in addition to saving for college, it may be
that children’s savings increases the likelihood that children will
actively develop strategies to confront costs beyond saving, such
as supplementing their savings with loans and financial aid.
In this study, we build on research on children’s savings and
their educational outcomes by examining whether children’s
asset-building programs are stronger when they are designed to
build children’s expectations for graduating from college along
with building their savings. More specifically, we examine
whether children who have savings and who are certain they
will graduate from a four-year college are more likely to attend
college than if they have no savings and are uncertain they will
graduate, if they have savings but are uncertain they will gradu-
ate, or if they are certain they will graduate but have no savings.
Review of Research
Over the past several years asset researchers have been in-
vestigating the effects of children’s savings on children’s edu-
cational outcomes using the Panel Study of Income Dynamics
(PSID) and its supplements the Child Development Supplement
(CDS) and the Transition into Adulthood (TA) supplement. To
conserve space, only the studies that address the relationship
between children’s savings and college attendance will be re-
viewed below. For a complete review of research on children’s
savings and children’s educational outcomes please see Elliott,
Destin, and Friedline (2011). We also review relevant research
on the relationship between assets and children’s college ex-
pectations.
Research on Children’s Savings and College
Attendance and Graduation
In regards to children’s savings and children’s college out-
comes, Elliott and colleagues have conducted four studies
(Elliott & Beverly, 2011a-b; Elliott, Constance-Huggins, &
Song, 2011; Elliott & Nam, in press). In study one, Elliott and
Beverly (2011a) examine children’s savings effects using an
aggregate sample (N = 1003) of Black and White children ages
17 to 23. They find that children who have designated a portion
of their own savings for school purposes are approximately two
times more likely to be currently attending college or already
graduated. As is the case for all four studies, they account for
missing data by using list-wise deletion and test whether miss-
ing data are missing completely at random (MCAR) using
chi-square and t-tests. Missing data can limit the generalizabil-
ity of these studies. In this study approximately 280 (28%)
cases are deleted due to missing data.
In study two, Elliott, Constance-Huggins, and Song (2011)
examine whether children’s savings effects vary by income
level. To do this, they use separate samples of low-to-moderate-
income (below $50,000; N = 495) children and high-income
($50,000 or more; N = 508) children. Due to missing data, 160
(32%) cases from the low- to moderate-income sample and 157
(31%) cases from the high-income sample are deleted. They
find that, among low-to-moderate-income children, those who
have savings designated for school are about two times more
likely to be currently enrolled in college or to have already
graduated. In the case of high-income children, children’s sav-
ings is not statistically significant. The study authors suggest
that this non-significance may support the proposition that
having children’s savings no longer matters above a certain
income threshold. That is, above this threshold, income might
be high enough that children cannot reasonably doubt that they
will be unable to afford college.
In study three, Elliott and Nam (in press) examine whether
children’s savings effects vary by race. They use separate sam-
ples of Black (N = 469) and White (N = 534) children. In the
sample of Black children, 167 (36%) cases are deleted due to
missing data, and 183 (34%) cases in the White sample are
deleted using list-wise deletion. Findings suggest that children
who have designated a portion of their savings for school are
two times more likely to be attending college or have graduated
from college among both samples of Black and White children.
In the final study, Elliott and Beverly (2011b) restrict the
sample to children who are certain they will graduate from a
four-year college (N = 333). In this sample, 33 (10%) of cases
are deleted in list wise deletion. By restricting the sample, the
researchers are able to determine the amount of “wilt” that oc-
curs and whether children’s savings helps to reduce it. “Wilt” is
the percent of children who expect to graduate from a four-year
college prior to leaving high school but do not attend college
between the ages of 17 and 23; in other words, wilt describes
children who expected to attend college but have not attended
in the years immediately following high school graduation. The
study finds that more than half of children (55%) who do not
have savings of their own experience wilt. However, among
children who expect to graduate from a four-year college, hav-
ing basic savings is associated with children being approxi-
mately six times more likely to attend college, while children
who have designated a portion of their basic savings for school
are approximately three times more likely to attend college.
In sum, study four raises some questions about whether posi-
tive results associated with children’s savings and children’s
educational outcomes in the other three studies may be being
driven by children who have both savings and positive college
expectations. On the whole, there is evidence to suggest that
children’s savings may be positively associated with children’s
W. ELLIOTT ET AL.
194
college attendance.
Research on Childrens Savings and College Expectations
Elliott conducts four studies that examine the relationship
between children’s savings and children’s college expectations
(Elliott, 2009; Elliott & Beverly, 2011a; Elliott, Choi, Destin, &
Kim, 2011; Elliott, Kim, Jung, & Zhan, 2010). These four
studies use data from the PSID and its supplements. Among
these four studies, Elliott and Beverly (2011a) and Elliott, Choi,
et al. (2011) are the only studies to use the TA supplement; the
other two studies use data from the CDS. Further, all but one
study uses list wise deletion to account for missing data. To
date, Elliott (2009) is the only asset study focused on household
or children’s savings, to use multiple imputations to complete
missing data. With the exception of one of the four studies
(Elliott, Choi, et al., 2011), children’s savings studies have
focused on children’s college expectations as a way to explain
the relationship between children’s savings and educational
outcomes.
In study one, the only study to use multiple imputations,
Elliott (2009) finds that children’s school savings is a signifi-
cant predictor of children’s math scores when children’s college
expectations are not included in the model. Second, he finds
that children’s school savings are a significant predictor of
children’s expectations. Third, he finds that expectations are a
significant predictor of math scores when children’s school
savings are not included in the model. Finally, when children’s
college expectations and children’s school savings are included
in the same model, children’s school savings remain signifi-
cantly related to math scores but the effect is reduced. Accord-
ing to the Baron and Kenny (1986) method of testing mediation,
this suggests that children’s expectations act as a partial media-
tor between children’s school savings and children’s math
scores. He also uses bootstrapping and Sobel’s test to further
test whether indirect effects occur. Both methods confirm that
children’s school savings have indirect effects on children’s
math scores that occur through children’s college expectations.
Elliott (2009) also examines the relationship between the
amount of children’s school savings and math achievement. He
finds that amount is not significant.
In study two, discussed above in the review of college atten-
dance and graduation research, Elliott and Beverly (2011a) also
examine the relationship between children’s school savings and
children’s college expectations. According to the Baron and
Kenny method, they find that children’s college expectations
partially mediate the relationship between children’s school
savings and college progress (i.e., currently attending or already
graduated). Bootstrapping confirms this finding.
In study three, Elliott, Kim, Jung, and Zhan (2010) use
PSID/CDS data (N = 1063) to test whether mediation effects
vary by race (White/Black). Separate samples of White (N =
576) and Black (N = 487) youth are analyzed. They correct for
missing data with the Yuan and Bentler (2000) correction for
non-normality data with missing data. The Jamshidian and
Bentler (1999) method allows a model to be estimated without
imputation and loss of subjects. The Yuan and Bentler (2000)
correction is similar to Satorra and Bentler (1994) with com-
plete data. Using Structural Equation Modeling (SEM) and
bootstrapping, they find that school savings is significantly
related to expectations for both White and Black youth. In the
case of math, they find that net worth and school savings have
indirect effects through college expectations for White youth
only (i.e., expectations mediate the relationship between assets
and math achievement). In the case of reading, they find that
there are no indirect effects regardless of race.
In study four, Elliott, Choi, et al. (2011) conduct a simulta-
neous test of whether children’s savings predict children’s col-
lege expectations or college expectations predict children’s
savings. They correct for missing data in a similar fashion as
Elliott, Kim, Jung, and Zhan (2010) do. They find that chil-
dren’s savings has a slightly stronger relationship with chil-
dren’s expectations than children’s expectations has with sav-
ings. However, they suggest that the best interpretation of the
data is that two-way causation exists.
In sum, the potential for multiple effects may make policies
that seek to build assets among children particularly alluring.
Further, findings of two-way causation suggest that asset-
building policies that seek to build both children’s savings
along with children’s college-bound identity may be most ef-
fective at increasing the number of children who have savings
and their college outcomes.
Conceptual Framework
To understand how a college-bound identity is formed, rein-
forced, and influences outcomes, we use Elliott, Choi et al.’s
(2011) theory of asset effects. Their theory is grounded in an
Identity-Based Motivation (IBM) theory of children’s motiva-
tion and behavior (for more information on IBM, see Oyserman
& Destin, 2010). Using the IBM framework, Elliott, Choi, et al.
(2011) propose that three principal components explain the
relation between assets, college-bound identity and motivation:
1) identity salience, 2) congruence with group identity, and 3)
interpretation of difficulty. These principles have been shown
to be important predictors of children’s school behaviors (Oy-
serman & Destin, 2010).
Salience captures the idea that children are more likely to
work toward a goal when images of their own future are on
their mind. People pay attention to things that they believe are
the causes of things that matter to them. As mentioned in the
introduction, Elliott, Sherraden, et al. (2010) find that children
see savings as a way to pay for college. Another way of stating
this finding is that owning savings may be seen as a cause of
being able to attend college. As such, owning savings may help
make college more salient.
Another important factor in the connection between context,
college-bound identity, and behavior is a link to group identity.
Congruence with group identity occurs when an image of the
self feels tied to ideas about relevant social groups such as
friends, classmates, family, and cultural groups. When this
occurs, the congruent personal identity is reinforced. Elliott,
Choi, et al. (2011) point out that assets are almost always con-
nected to the family. For example, when children open an ac-
count they are supported by parents or other family members.
Further, parents are often a primary source of children’s income
through gifts or allowances, for example. As Elliott, Choi, et al.
(2011) state, “When children and their families save money for
college, the meta-message asserts ‘we save,’ ‘we go to college,’
reinforcing the college-bound identity through its congruence
with the actions and goals of the larger group” (p. 1105).
Finally, IBM highlights the importance of having a means
for positively interpreting and overcoming difficulty. From this
perspective, in order for children to sustain effort and work
towards an image of themselves as being college-bound, the
context must provide a way to address inevitable obstacles to
the goal of attending college, such as being able to pay for col-
lege. It is clear how having savings provides children with a
strategy for paying for college.
This paper builds on previous research in several important
W. ELLIOTT ET AL. 195
ways. A way that it builds on previous research is by using
propensity score analyses (PSA). PSA is a relatively new statis-
tical method for testing causal inferences using survey data
(Rubin, 1996). PSA allows researchers to balance potential bias
between those children, for example, who are exposed to hav-
ing savings and those who are not based on known covariates
(Rosenbaum & Rubin, 1983). While most previous research
examining the relationship between assets and children’s edu-
cational outcomes uses survey data, no study uses PSA. Until
recently, propensity score methods have been limited to
two-group situations such as a single treatment and a compare-
son group. However, Imbens (2000) extends the method to
multi-group situations (also see Guo & Fraser, 2010).
Because of these methodological advancements, we are able
to build on past research by examining whether asset-building
interventions that seek to build both children’s savings and a
positive college-bound identity, are more likely to be related to
children attending college than interventions that only build
children’s savings or those that only build children’s college-
bound identities. Specifically we hypothesize that the combined
treatment of school savings and positive college-bound identity
(i.e., having school savings and being certain they would
graduate from a four-year college) is more likely to be associ-
ated with children being on course than school savings only
treatment (having school savings and being uncertain they
would graduate from a four-year college) or college-bound
identity-only treatment (being certain they would graduate from
a four-year college and not having school savings).
Methods
Data
This study uses longitudinal data from the PSID and its sup-
plements, the Child Development Supplement (CDS) and the
Transition into Adulthood supplement (TA). The PSID is a
nationally representative longitudinal survey of U.S. individu-
als and families that began in 1968. The PSID collects data on
such things as employment, income, and assets. Our independ-
ent variables related to households and parents are taken from
1999, 2001, and 2002 PSID data.
The CDS was administered to 3563 PSID respondents in
1997 to collect a wide range of data on parents and their chil-
dren, aged birth to 12 years. Questions cover a broad range of
developmental outcomes across the domains of health, psycho-
logical well-being, social relationships, cognitive development,
achievement, motivation, and education. Follow-up surveys are
administered in 2002 and 2007. For this study, independent
variables for young adults are taken from the 2002 CDS be-
cause this is the first year data are collected on parents’ school
savings for youth and youth’s school savings. Age 12 is the
first year that youth are asked questions about savings and col-
lege expectations. The TA supplement, administered in 2005
and 2007, measures outcomes for young adults who partici-
pated in earlier waves of the CDS and are no longer in high
school. Our outcome variables are taken from the 2007 TA.
The three data sets are linked using PSID, CDS, and TA map
files containing family and personal ID numbers. The linked
data sets provide a rich opportunity for analyses in which data
collected at one point in time (2001 or earlier) can be used to
predict outcomes at a later point in time (2007), and stable
background characteristics can be used as covariates. Because
the PSID initially oversampled low-income families, both the
descriptive and multivariate analyses are weighted using the
last observed weight variable as recommended by the PSID
manual (Gouskova, 2001).
Variables
There are two variables of interest in this study—children’s
college expectations and children’s savings. Children’s college
expectations are created using 2002 CDS data. In the CDS,
children were asked what they thought the chances were that
they would graduate from a four-year college. Children re-
sponded by saying no chance, some chance (about 50:50),
pretty likely, or it will happen. Children who responded that
their chances of graduating from a four-year college were 50%
or less were defined as “uncertain.”
The children’s savings variable is also created using 2002
CDS data. The CDS asks children between the ages of 12 to 18
whether they had a savings or bank account in their name. The
children’s savings variable divides children into two categories:
1) those who had an account in 2002, and 2) those who did not
have account. There are several important differences between
the accounts examined in this study and CDA accounts pro-
posed in the ASPIRE act and other popular education ac-
counts such as Coverdell Education Savings Accounts, Uni-
form Gifts to Minors Act (UGMAs), 529 College Savings Plans
run by States, and Roth Individual Retirement Arrangements
(IRAs). Popular educational accounts offer their owner’s pro-
tection from taxation and in some cases an infrastructure that
provides such things as direct deposit and match savings to
encourage and promote savings. In order not to be taxed, how-
ever, savings in these accounts typically cannot be withdrawn
without penalty until youth reach college age and it must be
spent on college related expenses. As a result, these accounts
can more aptly be defined as being non-liquid in nature. Unlike
in these popular education accounts, children can easily with-
draw money from the accounts in this study and use that money
without penalty but they do not benefit from tax breaks or other
incentives that are common components of CDAs (such as
initial deposits or match saving where for every dollar a child
saves the federal government or other agency matches it with
an additional dollar).
Using children’s expectations and children’s savings vari-
ables, we create four treatment groups or doses similar to Im-
bens (2000) multiple dose treatment approach. The first dose is
children who have no savings and are uncertain whether they
will graduate from a four-year college. The comparison group
in this sample is children with savings only, children who are
certain only, and children who both have savings and are cer-
tain. The second dose is children who have saving only. The
comparison group is children with no savings and who are un-
certain, children who are certain only, and children who both
have savings and are certain. The third dose is children who are
certain only. The comparison group is children with no savings
and who are uncertain, children with savings only, and children
who both have savings and are certain. The final dose is chil-
dren who both have savings and are certain they will graduate
from a four-year college. The comparison group is children
with no savings and who are uncertain, children with savings
only, and children who are certain only.
Outcome Variable. The outcome variable combines two
variables from the TA. First, youth were asked if they had ever
attended college. If they answered yes, they were asked whether
they attend or had attended a 2-year college, a four-year college,
or graduate school. We created a dichotomous variable indicat-
ing whether youth had ever attended a four-year college. These
data were collected in 2007.
W. ELLIOTT ET AL.
196
Control Variables. There are 15 control variables. They
may be thought of as falling into three broad areas: household
characteristics, child characteristics, and neighborhood charac-
teristics. Household characteristics consist of family income;
net worth; household size; head’s education level; head’s mari-
tal status; and home scale. Child characteristics consist of race,
gender, academic achievement, 2002 age, self-efficacy, and
self-concept. Neighborhood characteristics consist of urbanicity,
private school attendance, and peer expectations.
Family income is calculated by averaging family income for
1993, 1997, and 2002. Income averaged over multiple years
provides the best estimate of “permanent income” (Blau, 1999;
Mayer, 1997). Income is inflated to 2002 price levels using the
Consumer Price Index for 1993 and 1997. Because family in-
come is positively skewed, the log of family income is used in
regression analyses.
Net worth. Net worth in the PSID is a continuous variable
that sums separate household values for a business, checking or
savings accounts, home equity, real estate, stocks, and other
assets, and subtracts out credit card and other debt. Net worth is
averaged for 1994, 1999, and 2001; each year of net worth is
inflated to 2002 price levels. Because net worth is positively
skewed, the log form of net worth is used for regression analy-
ses. Since some individuals have a negative value on the net
worth variable, it is necessary to make adjustments to these
numbers so that the natural log of net worth could be calculated.
All net worth values that are less than or equal to 0 are re-coded
as 1 so that the natural log could be ascertained (e.g., Henretta
& Campbell, 1978; Orr, 2003).
Household size, head’s marital status, and head’s education
all came from the 2001 PSID. Household size is a continuous
variable. Head’s marital status is a categorical variable (married
or unmarried). Head’s education is a continuous variable (1 to
16), with each number representing a year of completed
schooling. We also use a categorical variable, dividing heads
into three groups: those with a high school degree or less, those
with some college, and those with a four-year degree or more.
Home scale. The Home Observation for Measurement of the
Environment-Short Form from the Caldwell and Bradley
HOME Inventory (Caldwell & Bradley, 1984) is used as a
measure of the cognitive stimulation and emotional support
parents provide to children. The particular items used in the
PSID Child Development Supplement were taken directly from
the National Longitudinal Survey of Youth, Mother-Child Sup-
plement so that the scales would be as similar as possible
(Baker, Keck, Mott, & Quinlan, 1993). It includes the follow-
ing items: how often the child gets out of the house, how many
books the child has read, how often the mother takes the child
to grocery store, how many cuddly or role-playing toys the
child has, the mother’s belief about how the child learns best,
how many push or pull toys the child has, how often the mother
talks to the child while working, and how often the mother
reads to the child.
Children’s race (White or Black), gender (male or female),
age, self-efficacy, self-concept, academic achievement, peer
expectations, and private school attendance come from 2002
CDS data.
Academic achievement. This variable is continuous, a com-
bination of math and reading scores. The Woodcock Johnson
(WJ-R), a well-respected measure, is used by the CDS to assess
math and reading ability (Mainieri, 2006). In descriptive analy-
sis, we use a dichotomous variable indicating whether children
have average or above-average achievement or below-average
achievement. Average or above average is coded 1 and below
average is coded as 0. Age in 2002 is a continuous variable. In
the descriptive analysis, we use a dichotomous variable indi-
cating whether children were below or at age 16 or older in 2002.
Below or at age 16 is coded 0 and older than 16 is coded 1.
Childrens self-efficacy. This variable is measured using
Pearlin’s self-efficacy scale (for information, see Pearlin, Me-
naghan, Lieberman, & Mullan, 1981). According to Mainieri
(2006), the children’s self-efficacy scale measures the amount
of control children perceive they have over their life in the
PSID/CDS. For descriptive purposes, the mean score is used to
create a dichotomous variable (average or above average
self-efficacy and below average self-efficacy).
Childrens self-concept. This is a continuous variable meas-
ured using Rosenberg’s self-esteem scale (for information, see
Rosenberg 1986). According to Mainieri (2006), children’s
self-concept measures the degree of satisfaction one has with
herself in the PSID/CDS. For descriptive purposes, the mean
score is used to create a dichotomous variable (average or
above-average self-esteem and below-average self-esteem).
Urbanicity. The 2003 Rural-urban Continuum Codes form a
classification scheme that distinguishes metropolitan counties
by size and nonmetropolitan counties by degree of urbanization
and proximity to metro areas.4 The PSID increases the codes by
one (ranges from 1 to 10). We collapse it into a three-level
variable indicating whether children live in a metropolitan,
urban, or rural area.
Childrens peer expectations. Children are asked how many
of their friends planned to attend college: 1) None, 2) a few, 3)
some, 4) many, and 5) almost all or all. Peer expectations are
recoded into a dichotomous variable. If children respond by
selecting 1, 2, or 3, they are coded as 0; if they select options 4
or 5, they are coded as 1.
Private school attendance. In 2002, children are asked
whether they have ever attended private school (yes or no).
Study Sampl e
The 2007 TA sample consisted of 1118 participants. The
sample in this study is restricted to Black and White youth be-
cause only small numbers of other racial groups exist in the TA.
Our final sample consisted of 1,003 children and their families.
Four separate samples are created based on the dosage being
investigated. Table 1 provides the percent in each treatment
dose and its comparison group before and after weighting the
sample using propensity score weighting. After weighting the
treatment and comparison groups were more balanced between
the treatment doses (i.e., no savings/uncertain, savings only,
certain only, and combined). For example, prior to weighting
there was a 26% gap between the non-savings/uncertain to
graduate treatment dose and the combination treatment dose;
however, after weighting, there was a 7% gap between the
lowest and highest dose.
Source: Weighted data from the Panel Study of Income Dy-
namics and its supplements.
Notes: Estimates are propensity score-adjusted using the
weighting scheme in Guo & Fraser, 2010 (also see Foster, 2003;
Imbens, 2000). The propensity score weights are based on the
propensity scores (or predicted probabilities) calculated using the
results of the multinomial logit model. The term certain identifies
children who expected to graduate from a four-year college prior
to leaving high school.
4For more information on the scale go to: http://www.ers.usda.gov/
Briefing/Rurality/RuralUrbCon/.
W. ELLIOTT ET AL. 197
Table 1.
The percent of the sample in each trea tme nt do se and comparison group before and after adjusting for propensity score weight.
Group % Before Adjustment % After Adjustment
No savings/uncertain to graduate dose 17 22
Comparison (savings only; certain only and combined) 83 78
Savings only dose 17 24
Comparison (no savings/uncertain; certain only and combined) 83 76
Certain to graduate only dose 24 25
Comparison (no savings/uncertain; savings only and combined) 76 75
Combination dose 43 29
Comparison (no savings/uncertain; savings only and certain only) 57 71
Analysis Plan
Primarily analyses were conducted using STATA version 10
(STATA Corp, College Station, TX).5 There were four stages
of analysis conducted in this study. In stage one of the analyses;
we completed missing data using multiple imputations. Missing
data can lead to inaccurate parameter estimates and biased
standard errors and population means, resulting in researchers
inaccurately reporting statistical significance or insignificance
(Graham, Taylor, & Cumsille, 2001). In this study, there was
less than 1% missing data on college attendance. However, data
for some control variables were also missing. For example,
both children’s savings and children’s college expectations had
about 20% missing. A rule of thumb for how much data can be
imputed is about 20% (Little & Rubin, 2002). We used the
chained equation method of multiple imputations (Royston,
2004, 2005a, b; van Buuren et al., 2006) to create five inde-
pendent data sets that included all continuous and categorical
variables with no missing data. This method allowed us to
specify the multivariate structure as a series of imputation
models based on other variables. Logistic regression was used
to impute the incomplete categorical variables, and linear re-
gression was used to impute continuous variables. Each missing
value was filled in with a set of plausible values that were cre-
ated with information from other values of a variable and some
associated variables.
In stage two, we conducted propensity score weighting with
multi-treatments/dosages in order to balance selection bias
between those children; for example, those who were exposed
to having savings and those who were not based on known
covariates (Guo & Fraser, 2010; Imbens, 2000). More specifi-
cally, in stage two of the analysis, we created four groups: 1)
had no school savings and uncertain they would graduate from
a four-year college prior to leaving high school; 2) had school
savings and were uncertain they would graduate from a
four-year college prior to leaving high school; 3) certain they
would graduate from a four-year college and had no school
savings prior to leaving high school; and 4) had school savings
and were certain they would graduate from a four-year college
prior to leaving high school. Next we estimated a multinomial
logit regression predicting multi-group membership using 13 of
the 15 covariates included in this paper. The only variables
included in the multinomial logit regression were those posi-
tively correlated following Guo and Fraser’s (2010) approach.
Household size and urbanicity were not included in creating the
propensity score weight or further analyses because they did
not have a significant association with college attendance at the
bivariate level. The resulting coefficient estimates were used to
calculate propensity scores for each group. The inverse of that
probability was used to create the propensity score weight.
In stage three, we test covariate imbalance after weighting.
Since propensity score weighting does not use matching, we
run a weighted simple logistic regression using college atten-
dance (i.e., a dichotomous covariate) as the dependent variable,
and one of the dichotomous dosage variables (i.e., either no
savings/uncertain; savings only, certain only, or combined) as
the single independent variable (Guo & Fraser, 2010). Results
from simple logistic regressions are reported in Appendix A
and B along with frequencies or means depending on whether it
is a continuous covariate or categorical covariate. Appendix A
reports information on all covariates before matching, Appen-
dix B reports information after weighting.
In stage four, we used logistic regression as the primary ana-
lytic tool to assess statistical significance for the overall rela-
tionship between each dose separately and college attendance
without and with propensity score weights included. Moreover,
we provided measures of predictive accuracy through the
McFadden’s pseudo R2 (not equivalent to the variance ex-
plained in multiple regression model, but closer to 1 is also
positive). We also reported odds ratios (OR) for easier inter-
pretation. The odds ratio is a measure of effect size, describing
the strength of association. Identical analyses were repeated for
all five imputed data sets with no missing data, and the results
were pooled (i.e., calculated an average for the five imputed
data sets) to yield less biased parameter estimates in the overall
regression model. This method is superior to other kinds of
imputation methods, such as mean substitution, hot decking,
regression imputation, and single imputation, in terms of pro-
ducing more accurate estimates (Little & Rubin, 2000; Rubin,
1996).
Results
In the first part of this section we discuss findings from the
covariate balance checks. Then we report logistic regression
results for each treatment group.
Bivariate Results from Covariate Balance Checks
Results from the balance checks are presented in see Appen-
dix A, B. In the unadjusted sample, almost all covariates
showed significant group differences regardless of the treat-
ment dose. Once propensity score weighting was conducted,
group differences were no longer significant in almost all cases.
This suggests that weighting was successful in reducing bias
among observed covariates in almost all cases.
5Covariate balance checking is conducted using SAS (SAS Institute
Inc., 2008).
W. ELLIOTT ET AL.
198
Logit Results for No Savings/Uncertain Treatment
Group
Table 2 provides information on unadjusted and adjusted
logit models examining the relationship between the no sav-
ings/uncertain treatment group and whether children had ever
attended college by 2007. Approximately 26% of the variance
in college attendance is explained in the adjusted model. Head’s
education level, household net worth, and children’s academic
achievement are all statistically significant positive predictors
of college attendance among the no savings/uncertain sample of
children in the adjusted model.
The no savings/uncertain treatment group is not statistically
significant when compared to children with savings only, chil-
dren who are certain only, and children who both have savings
and are certain while controlling for all other variables.
Logit Results for the Savings Only Treatment Group
Table 3 provides information on unadjusted and adjusted
logit models examining the relationship between the sav-
ings-only treatment group and whether children have ever at-
tended college. The adjusted model explains about 25% of the
variance in college attendance. Head’s education level, parent’s
college expectations, household net worth, and children’s aca-
demic achievement are positive statistically significant predict-
tors among the sample of children with savings only. Self-ef-
ficacy is a negative predictor of college attendance.
Table 2.
Logit examining the relationship between no savings and no expectations treatment group and college attendance i n 2007 (N = 1003).
Unadjusted
No Savings/Uncertain Treatment
Adjusted
No Savings/Uncertain Treatment
Covariates
B S.E. B S.E.
No savings/uncertain treatment –0.409 0.230 –0.738 0.357
Home scale 0.970 0.517 0.545 0.830
Head’s education level 0.161** 0.046 0.285* 0.080
Head’s marital status 0.661** 0.195 0.522 0.286
Parent’s college expectations for their child 1.605*** 0.190 1.447 0.255
Household net worth 0.071** 0.023 0.076* 0.029
Household income –0.043 0.028 –0.050 0.043
Child’s race 0.305 0.212 0.367 0.320
Child’s age in 2002 0.033 0.054 0.154 0.087
Child’s self-efficacy –0.261 0.171 –0.558 0.241
Child’s self-concept 0.363 0.236 0.152 0.315
Child’s academic achievement 0.023*** 0.004 0.014* 0.006
Child’s college expectations for peers 0.597** 0.191 0.265 0.256
Whether child ever attended private school 0.560 0.449 –1.067 0.919
McFadden pseudo R2 .35 .26
Source: Weighted data from the Panel Study of Income Dynamics and its supplements. Notes: Data imputed using multiple imputations. S.E. = robust standard error. O.R.
= odds ratio. Estimates are propensity score-adjusted using the weighting scheme in Guo & Fraser, 2010 (also see Foster, 2003 and Imbens, 2000). The propensity score
weights are based on the propensity scores (or predicted probabilities) calculated using the results of the multinomial logit model. * p < .05; **p < .01; ***p < .001.
Table 3.
Logit examining the rela t i onship between the savin gs on ly treatment group and college attendance in 2007 (N = 1003).
Unadjusted
Savings Only Treatment
Adjusted
Savings Only Treatment
Covariates
B S.E. B S.E.
Savings only treatment –0.535* 0.231 –0.600 0.322
Home scale 1.121* 0.515 0.887 0.838
Head’s education level 0.168*** 0.046 0.285** 0.076
Head’s marital status 0.679** 0.195 0.423 0.305
Parent’s college expectations for their child 1.613*** 0.190 1.487*** 0.253
Household net worth 0.072** 0.023 0.072* 0.030
Household income –0.040 0.028 –0.040 0.047
Child’s race 0.203 0.216 0.263 0.346
Child’s age in 2002 0.060 0.055 0.137 0.086
Child’s self-efficacy –0.255 0.170 –0.600* 0.252
Child’s self-concept 0.401 0.235 0.270 0.305
Child’s academic achievement 0.023*** 0.004 0.014* 0.006
Child’s college expectations for peers 0.614** 0.189 0.189 0.268
Whether child ever attended private school 0.573 0.454 –0.845 0.870
McFadden pseudo R2 .35.25
Source: Weighted data from the Panel Study of Income Dynamics and its supplements. Notes: Data imputed using multiple imputations. S.E. = robust standard error. O.R.
= odds ratio. Estimates are propensity score-adjusted using the weighting scheme in Guo & Fraser, 2010 (also see Foster, 2003 and Imbens, 2000). The propensity score
weights are based on the propensity scores (or predicted probabilities) calculated using the results of the multinomial logit model. * p < .05; **p < .01; ***p < .001.
W. ELLIOTT ET AL. 199
The savings only treatment group is not statistically signifi-
cant when compared to children with no savings and who are
uncertain, children who are certain only, and children who both
have savings and are certain while controlling for all other
variables.
Logit R esults for the Certain Only Treatment Group
Table 4 provides information on the unadjusted and adjusted
logit models examining the relationship between the certain-
only treatment group and college attendance. The certain-only
treatment group sample consists of children who expected to
graduate from a four-year college prior to leaving high school.
The adjusted model explains about 27% of the variance in col-
lege attendance. Similar to Table 3, head’s education level,
parent’s college expectations, household net worth, and chil-
dren’s academic achievement are all positive statistically sig-
nificant predictors of children in the certain-only treatment
group while self-efficacy is a negative predictor.
The certain-only treatment group is not statistically signifi-
cant when compared to children with no savings and who are
uncertain, children with savings only, and children who both
have savings and are certain while controlling for all other
variables.
Logit R esults for the Combined Treatment Group
Table 5 provides information on unadjusted and adjusted
logit models examining the relationship between the combined
treatment group (i.e., have savings and are certain they will
graduate from a four-year college) and whether children have
ever attended college by 2007. Approximately 26% of the
variance in college attendance is explained in the adjusted
model. Head’s education level, parent’s college expectations,
household net worth, and children’s academic achievement are
all statistically significant positive predictors of children with
savings and who are certain they will graduate from a four-year
college, while self-efficacy is a negative predictor of college
attendance.
Unlike the other treatment groups, the combined treatment
Table 4.
Logit examining the re l ationship between the certain only treatment group and college attendance in 2007 (N = 1003).
Unadjusted
Certain Only Treatment
Adjusted
Certain Only Treatment
Covariates
B S.E. B S.E.
Certain only treatment –0.001 0.102 0.174 0.127
Home scale 1.091* 0.512 0.730 0.824
Head’s education level 0.163** 0.046 0.285** 0.076
Head’s marital status 0.663** 0.195 0.454 0.300
Parent’s college expectations for their child 1.641*** 0.189 1.444*** 0.250
Household net worth 0.075** 0.022 0.071* 0.030
Household income –0.042 0.028 –0.043 0.047
Child’s race 0.311 0.216 0.325 0.338
Child’s age in 2002 0.036 0.054 0.141 0.085
Child’s self-efficacy –0.267 0.170 –0.619* 0.248
Child’s self-concept 0.388 0.234 0.261 0.299
Child’s academic achievement 0.024*** 0.004 0.014* 0.006
Child’s college expectations for peers 0.645** 0.188 0.213 0.267
Whether child ever attended private school 0.605 0.450 –0.960 0.904
McFadden pseudo R2 .36 .27
Source: Weighted data from the Panel Study of Income Dynamics and its supplements. Notes: Data imputed using multiple imputations. S.E. = robust standard error. O.R.
= odds ratio. Estimates are propensity score-adjusted using the weighting scheme in Guo & Fraser, 2010 (also see Foster, 2003 and Imbens, 2000). The propensity score
weights are based on the propensity scores (or predicted probabilities) calculated using the results of the multinomial logit model. * p < .05; **p < .01; ***p < .001.
Table 5.
Logit examining the impa c t relationship between the combin ed t re a tm en t g ro up and college attendance in 2007 (N = 1003).
Unadjusted
Combined Treatment
Adjusted
Combined Treatment
Covariates
B S.E. B S.E.
Combined treatment 0.251*** 0.072 0.276** 0.081
Home scale 0.922 0.589 0.687 0.824
Head’s education level 0.151** 0.052 0.285** 0.081
Head’s marital status 0.662** 0.218 0.485 0.301
Parent’s college expectations for their child 1.565*** 0.203 1.503*** 0.257
Household net worth 0.065** 0.024 0.079* 0.029
Household income –0.043 0.031 –0.047 0.044
Child’s race 0.313 0.238 0.270 0.340
Child’s age in 2002 0.032 0.060 0.152 0.088
Child’s self-efficacy –0.244 0.185 –0.523* 0.247
Child’s self-concept 0.369 0.253 0.169 0.311
Child’s academic achievement 0.022*** 0.004 0.014* 0.006
Child’s college expectations for peers 0.547* 0.206 0.237 0.266
Whether child ever attended private school 0.459 0.545 –0.932 0.934
McFadden pseudo R2 .35 .26
Source: Weighted data from the Panel Study of Income Dynamics and its supplements. Notes: Data imputed using multiple imputations. S.E. = robust standard error. O.R.
= odds ratio. Estimates are propensity score-adjusted using the weighting scheme in Guo & Fraser, 2010 (also see Foster, 2003 and Imbens, 2000). The propensity score
weights are based on the propensity scores (or predicted probabilities) calculated using the results of the multinomial logit model. * p < .05; **p < .01; ***p < .001.
W. ELLIOTT ET AL.
200
group has a statistically significant positive association with
college attendance when compared to children with no savings
and who are uncertain, children with savings only, and children
who are certain only after controlling for all other covariates.
In sum, head’s education level, parent’s college expectations,
household net worth, children’s academic achievement, and
child’s self-efficacy are all significant predictors of college
attendance among all treatment groups except for the no sav-
ings/uncertain treatment group. Only self-efficacy is a negative
predictor of college attendance. Further, among the four treat-
ment groups, only the combined treatment group is a significant
predictor of college attendance.
Discussion
Rising college costs have led to fewer low- and moder-
ate-income children attending and graduating from college in
an era when college graduation is increasingly important to
career success. CDAs have been proposed as a potentially novel
and promising mechanism for reducing college costs and en-
couraging college attendance. Previous research suggests that
children’s asset-building programs may be enhanced if they are
combined with strategies to build children’s college-bound
identity (e.g., Elliott & Beverly, 2011b; Elliott, Choi, et al.
2011). In this study we examine whether children who have
savings and who are certain they will graduate from a four-year
college (i.e., a proxy for college-bound identity) are more likely
to attend college shortly after graduating high school than chil-
dren who do not have savings and who are uncertain, children
who have savings but are uncertain, and children who are cer-
tain but do not have savings.
Similar to Elliott and Beverly (2011b) who find that chil-
dren’s savings is associated with college attendance among a
sample of children who expect to graduate from a four-year
college, we find that the combined treatment group is a signifi-
cant predictor of college attendance. In regards to household net
worth, previous research suggests that net worth is a significant
predictor of college attendance but only when children’s aca-
demic achievement or children’s cognitive ability is not in-
cluded (e.g., Conley, 2001; Jez, 2008; Nam & Huang, 2009). In
contrast, controlling for children’s academic achievement, we
find that net worth is a significant predictor in all four treatment
groups. The reason for the different findings may be because
we use very different samples than previous research.
Consistent with previous research (e.g., Stratton, O’Toole, &
Wetzel, 2007), parent’s education level and children’s aca-
demic achievement remain statistically significant predictors of
whether children attend college. Parent’s college expectations
for their child are also a significant positive predictor of college
attendance for all treatment groups except for the no sav-
ings/uncertain treatment group. This too is consistent with pre-
vious research on college attendance (e.g., William Shanks &
Destin, 2009).
Surprisingly, self-efficacy is a negative predictor in all treat-
ment groups except for the no savings/uncertain treatment group
where it is not significant. Self-efficacy is usually thought of as
being a positive predictor of children’s academic outcomes (e.g.,
Bandura, 1997). A reason for the negative findings may be
because the self-efficacy scale used in this study measures chil-
dren’s global self-efficacy (e.g., “I can make things happen”).
However, Bandura (1997) suggests that in order to accurately
predict academic outcomes, self-efficacy “… beliefs should be
measured in terms of particularized judgments of capability” (p.
42). An example of a more particularized judgment is “I can
make things happen in reading class.”
Limitations
Propensity score analyses have two clear disadvantages rela-
tive to randomized trials. One is the need to assume conditional
independence (i.e., selection bias is eliminated by controlling
for observed covariates). This may not be true as it is impossi-
ble to know all the covariates that may influence the choice to
participate in treatment. The precision of controlling for treat-
ment choice goes as far as the covariates included in the study.
In randomized trials, the researcher can be confident that the
treatment group and the control group are similar on both ob-
served and unobserved characteristics. Second, whereas pro-
pensity score analyses can only estimate treatment effects
where there is overlap between exposed group (e.g., group that
has savings) and unexposed group (e.g., does not have savings),
random assignment ensures that there is common support
across the whole sample. These considerations make experi-
mental techniques superior to propensity score analyses in a
number of important ways.
However, randomization also has its limitations. A practical
limitation is cost. Cost is a major concern with designing ran-
dom control trials to test CDAs. Providing children with initial
deposits and matches (i.e., every one dollar saved is matched
with an additional dollar) can be expensive. Another concern is
time. CDAs that are in the state’s name with the youth as the
beneficiary are being tested in a large experiment in Oklahoma
called SEED for Oklahoma Kids (SEED OK).6 However, be-
cause the accounts were issued at birth in 2004, it will be a
number of years before researchers can test this design as it
relates to college progress. In the meantime, CDAs have al-
ready been proposed in Congress and policymakers are forced
to make decisions sooner rather than later on their value as a
mechanism for increasing rates of college progress. Providing
policymakers with information now is of the utmost importance,
as is using the best available data and methods.
Lastly, Destin and Oyserman (2010) suggest that educa-
tion-dependent identity (i.e., children expect to get a job that
requires a degree) may be a better predictor of children’s true
college expectations than the measure used in this study (i.e.,
no chance, some chance (about 50:50), pretty likely, or it will
happen). Future research may want to examine whether educa-
tion-dependent identity is a better measure of children’s college
expectations.
Implications
Overall, findings from this study suggest that a way to in-
crease college attendance rates is to create education policies
that will both increase children’s savings along with a more
positive college-bound identity. CDAs are a policy mechanism
for promoting children’s savings. However, CDA programs
proposed in the ASPIRE Act have not attempted to incorporate
strategies for building a more positive college-bound identity.
We suggest that a way that CDAs can be designed to help build
children’s college-bound identities is by incorporating Identity
Based Motivation (IBM) strategies into the financial education
component of CDA policies. IBM assumes that people’s per-
ceptions of their possible selves are dynamically constructed in
context.7
6For more information on SEED OK, see
http://csd.wustl.edu/AssetBuilding/SEEDOK/
7See Elliott, Choi, et al. (2011) for more information on how IBM can
be used to explain asset effects. Possible selves programs can be designed to increase student
W. ELLIOTT ET AL. 201
motivation by having students examine their future and think
about goals that are important to them for attending college.
There are many ways that the financial education curricula in
existing CDA policies could be adapted to have students exam-
ine their future and think about goals that are important to them
for attending college. For example, financial education curricu-
lums could be designed to also teach children about the cost of
college, about financial aid, and the role savings can play in
meeting college costs. In this manner, children are being taught
strategies to overcome a perceived difficulty (i.e., ability to pay
for college) related to college attendance. According to IBM, in
order to sustain and work towards an image of a future self
(such as a college-bound identity), the context must provide a
way to address inevitable obstacles to the goal such as paying
for college (Elliott, Choi, et al. 2011). Further, they could be
taught about how much they can expect to save by earning
incentives, initial deposits, match savings (i.e., for every dollar
saved an additional dollar is placed in the child’s account up to
a certain amount each year), and interest, for example.
Policies that seek to increase parent’s college expectations
for their child may also be a particularly promising strategy for
increasing children’s college attendance rates. When elements
of a family’s environment contain cues about assets, like par-
ents’ savings accounts, the presence of such resources can bol-
ster parents’ expectations for their children, influencing both
their own interactions with children and children’s own college
expectations and school-related behaviors. Children’s savings
as proposed in the ASPIRE Act may be a way to positively
influence parent expectations and build congruence with group
identity, a key component of an IBM theory of college-bound
identity. IBM suggests that when an image of the self feels tied
to ideas about relevant social groups (such as, friends, class-
mates, family and cultural groups), the congruent personal
identity becomes reinforced. As Loke and Sherraden (2009)
suggest, CDAs may have a multiplier effect by engaging the
larger family in the asset-accumulation process. A way that this
may happen in CDAs is by allowing parents to make voluntary
after-tax contributions into children’s accounts.
Conclusion
A clear implication of this study is that when children have
savings and have a positive college-bound identity they are
more likely to attend college than both children who have sav-
ings but are uncertain that they will graduate from college, and
children who are certain they will graduate from college but
have no savings. Given this, policies that seek to build both
children’s savings and more positive college-bound identities
are likely to be more successful at increasing college attendance
than those that only promote savings or only promote a positive
college-bound identity.
References
ACSFA (2010). The rising price of inequality: How inadequate grant
aid limits college access and persistence. Washington, DC: Advisory
Committee on Student Financial Assistance.
Baker, P., Keck, C., Mott, F., & Quinlan, S. (1993). NLSY child hand-
book revised edition: A guide to the 1986-1990 national longitudinal
survey of youth child data. Columbus, OH: Center for Human Re-
source Research.
Bandura, A. (1997). Self-efficacy: The exercise of control. New York:
W.H. Freeman.
Baron, R. M., & Kenny, D. A. (1986). The moderator-mediator variable
distinction in social psychological research: Conceptual, strategic,
and statistical considerations. Journal of Personality and Social
Psychology, 51, 1173-1182. doi:10.1037/0022-3514.51.6.1173
Baum, S., Ma, J., & Payea, K. (2010). Education pays 2010: The bene-
fits of higher education for individual and society. New York, NY:
CollegeBoard.
Belk, R. W. (1988). Possessions and the extended self. The Journal of
Consumer Research, 15, 139-168. doi:10.1086/209154
Blau, D. M. (1999). The effect of income on child development. The
Review of Economics and Statistics, 81, 261-276.
doi:10.1162/003465399558067
Brown, W. G., Chingos, M., et al. (2009). Crossing the finish line:
Completing college at America’s public universities. Princeton, NJ,
Princeton University Press.
Caldwell, B. M., & Bradley, R. H. (1984). Home observation for
measurement of the environment. Little Rock, AR: University of Ar-
kansas.
Carnevale, A. P., Smith, N., & Strohl, J. (2010). Help wanted: Projec-
tions of jobs and education requirements through 2018. Washington,
DC: Georgetown University Center on Education and the Workforce.
Coleman, J., Campbell, E., Hobson, C., McPartland, J., Mood, A., &
Weinfeld, F. D. (1966). Equality of educational opportunity. Wash-
ington, DC: Department of Health, Education and Welfare.
Coleman, J. (1988). Social capital in the creation of human capital. The
American Journal of Sociology, 94, S95-S120.
doi:10.1086/228943
College Board. (2009). Trends in student aid. Trends in higher educa-
tion series. Washington, DC: College Board.
Conley, D. (2001). Capital for college: Parental assets and postsecond-
ary schooling. Sociology of Education, 74, 59-72.
doi:10.2307/2673145
Cook, T. D., Church, M. B., et al. (1996). The development of occupa-
tional aspirations and expectations among inner-city boys. Child De-
velopment, 67, 3368-3385. doi:10.2307/1131783
Dauphin, A., El Lahga, A.-R., et al. (2008). Are children deci-
sion-makers within the household (IZA Discussion Papers 3728).
Bonn, Germany: Institute for the Study of Labor (IZA).
Destin, M. (2009). Assets, inequality, and the transition to adulthood:
An analysis of the Panel Study of Income Dynamics (Issue Brief).
New York: The Aspen Institute, Initiative on Financial Security.
Destin, M., & Oyserman, D. (2009). From assets to school outcomes:
How finances shape children’s perceived possibilities and intentions.
Psychological Science, 20, 414-418.
Destin, M., & Oyserman, D. (2010). Incentivizing education: Seeing
schoolwork as an investment, not a chore. Journal of Experimental
Social Psychology. Online First.
Elliott, W. (2009). Children’s college aspirations and expectations: The
potential role of college
development accounts (CDAs). Children and Youth Services Review,
31, 274-283. doi:10.1016/j.childyouth.2008.07.020
Elliott, W. and Beverly, S. (2011). Staying on course: The effects of
savings and assets on the college progress of young adults. American
Journal of Education, 117), 343-374. doi:10.1086/659211
Elliott, W. and Beverly, S. (2011b). The role of savings and wealth in
reducing “wilt” between expectations and college attendance. Jour-
nal of Children & Poverty, 17, 165-185.
Elliott, W., Choi, E. H., Destin, M. and Kim, K. (2011). The age old
question, which comes first? A simultaneous test of children’s sav-
ings and children’s college-bound identity. Children & Youth Ser-
vices Review, 33, 1101-1111.
doi:10.1016/j.childyouth.2011.02.001
Elliott, W., Constance-Huggins, M., & Song, H.-a. (2011). Factors
predicting college progress among low- to moderate-income and
high-income young adults: Assets as an understudied form of eco-
nomic capital (CSD Working Paper 11-15). St. Louis, MO: Wash-
ington University, Center for Social Development.
Elliott, W., Destin, M, and Friedline, T. (2011). Taking stock of ten
years of research on the relationship between assets and children’s
educational outcomes: Implications for theory, policy and interven-
tion. Children and Youth S ervic es Review, 33, 2312-2328.
doi:10.1016/j.childyouth.2011.08.001
Elliott, W., Kim, K., Jung, H., & Zhan, M. (2010). Asset holding and
educational attainment among African American and Caucasian
children. Children and Youth Services Review, 32, 1497-1507.
doi:10.1016/j.childyouth.2010.03.019
W. ELLIOTT ET AL.
202
Elliott, W., & Nam, I. (in press). Direct effects of assets and savings on
the college progress of black young adults. Educational Evaluation
and Policy Analysis.
Elliott, W., Sherraden, M., Johnson, L., & Guo, B. (2010). Young chil-
dren’s perceptions of college and saving: The potential role of child
development accounts. Children and Youth Services Review, 32,
1577-1584. doi:10.1016/j.childyouth.2010.03.018
Executive Office of the President. Office of Management and Budget
(2009). Simplifying student aid: The case for an easier, faster, and
more accurate FAFSA. Washington, DC: Office of Management and
Budget.
Friedline, T., & Elliott, W. (in press). Predicting savings from adoles-
cence to young adulthood. Journal of the Society for Social Work
and Research.
Furby, L. (1978). Possession in humans: An exploratory study of its
meaning and motivation. Social Behavior and Personality, 6, 49-64.
doi:10.2224/sbp.1978.6.1.49
Furby, L. (1980). The origins and early development of possessive
behavior. Political Psychology, 2, 30-42. doi:10.2307/3790969
Gouskova, E. (2001). The 2002 PSID child development supplement
(CDS-II) weights. URL (Retrieved 7 August 2010)
http://psidonline.isr.umich.edu/CDS/questionnaires/cdsiiweights.pdf
Graham, J. W., Taylor, B. J., & Cumsille, P. E. (2001). Planned missing
data designs in analysis of change. In L. M. Collins, & A. Sayer
(Eds.), New methods for the analysis of change (pp. 335-353). Wash-
ington, DC: American Psychological Association.
doi:10.1037/10409-011
Guo, S., & Fraser, W. M. (2010). Propensity score analysis: Statistical
methods and applications. Thousand Oaks, CA, Sage Publications,
Inc
Haskins, R. (2008). Education and economic mobility. In J. B. Isaacs, I.
V. Sawhill, & R. Haskins, Getting ahead or losing ground: Eco-
nomic mobility in America (pp. 91-104). Washington, DC: The
Brookings Institution.
Haveman, R., & Wolff, E. N. (2005). Who are the asset poor? Levels,
trends, and composition, 1983-1998. In M. Sherraden (Ed.), Inclu-
sion in the American dream: Assets, poverty, and public policy (pp.
61-86). New York: Oxford University Press.
Heckman, J. J., & Smith, J. A. (1995). Assessing the case for social
experiments. American E conomic Association, 9, 85-110.
Henretta, J. C., & Campbell, R. T. (1978). Net worth as an aspect of
status. American Journal of Sociology, 83, 1204-1223.
doi:10.1086/226679
Huang, J., Guo, B., et al. (2010). Parental income, assets, borrowing
constraints and children’s post-secondary education. Children and
Youth Services Review, 32, 585-594.
doi:10.1016/j.childyouth.2009.12.005
Imbens, G. W. (2000). The role of the propensity score in estimating
dose-response functions. Biometrika, 87, 706-710.
doi:10.1093/biomet/87.3.706
Jamshidian, M., & Bentler, P. M. (1999). ML estimation of mean and
covariance structures with missing data using complete data rou-
tines." Journal of Edu ca tio nal and Behavioral Statistic s, 24, 21-41.
Jez, S. J. (2008). The influence of wealth and race in four-year college
attendance University of California, Berkeley, Center for Studies in
Higher Education.
Levison, D. (2000). Children as economic agents. Feminist Economics,
6, 125-134. doi:10.1080/135457000337732
Little, R. J. A., & Rubin, D. B. (2002). Statistical analysis with missing
data. New York, NY: John Wiley and Sons.
Loke, V., & Sherraden, M. (2009). Building assets from birth: A global
comparison of child development account policies. International
Journal of Social Welf ar e , 18, 119-129.
doi:10.1111/j.1468-2397.2008.00605.x
Lu, B., Zanutto, E., Hornik, R., & Rosenbaum, P. R. (2001). Matching
with doses in an observational study of a media campaign against
drug abuse. Journal of American Statistical Association, 96, 1245-
1253. doi:10.1198/016214501753381896
Mainieri, T. (2006). The panel study of income dynamics child devel-
opment supplement: User guide for CDS-II. Ann Arbor, MI: Univer-
sity of Michigan, Institute for Social Research.
Marjoribanks, K. (1984). Ethnicity, family environment and adoles-
cents’ aspirations: A follow-up study. Journal of Educational Re-
search, 77, 166-171.
Mau, W. C. (1995). Educational planning and academic achievement of
middle school students: A racial and cultural comparison. Journal of
Counseling and Develop me nt , 73, 518-526.
Mayer, S. (1997). What money can’t buy: family income and children’s
life chances. Cambridge, MA: Harvard University Press.
Meeks, C. B. (1998). Factors influencing adolescents’ income and
expenditures. Journal of Family and Economic Issue s, 19, 131-149.
doi:10.1023/A:1022948623549
Mickelson, R. A. (1990). The attitude-achievement paradox among
black adolescents. Sociology of Education, 63, 44-61.
doi:10.2307/2112896
Nam, Y., & Huang, J. (2009). Equal opportunity for all? Parental eco-
nomic resources and children’s educational achievement. Children
and Youth Services Review, 31, 625-634.
doi:10.1016/j.childyouth.2008.12.002
Orr, A. (2003). Black-white differences in achievement: The impor-
tance of wealth. Sociolo gy of Education, 76, 281-304.
doi:10.2307/1519867
Oyserman, D., & Destin, M. (2010). Identity-based motivation: Impli-
cations for intervention. The Couseling Psychologist, 38, 1001-1043.
Paulsen, M. B. (2001). The economics of human capital and investment
in higher education. In M. B. Paulsen & J. C. Smart (Eds.), The fi-
nance of higher education: Theory, rese a r c h , policy and practice, pp.
55-94. New York, Agathon Press.
Pearlin, L. J., Menaghan, E. G., et al. (1981). The stress process. Jour-
nal of Health and Social Behavior, 22, 337-356.
doi:10.2307/2136676
Porfeli, E., Wang, C., et al. (2009). Influence of social and community
capital on student achievement in a large urban school district. Edu-
cation and Urban Society, 42, 72-95.
doi:10.1177/0013124509343373
Reynolds, J. R. & Pemberton, J. (2001). Rising college expectations
among youth in the United States: A comparison of the 1979 and
1997 NLSY. The Journal of Human R esources, 36, 703-726.
doi:10.2307/3069639
Rosenbaum, P. R., & Rubin, D. B. (1983). Reducing bias in observa-
tional studies using subclassification on the propensity score. Journal
of the American Statistical Association, 79, 516-524.
doi:10.2307/2288398
Rosenberg, M. (1986). Conceiving the self. New York, NY, Basic
Books.
Royston, P. (2004). Multiple imputation of missing values. Stata Jour-
nal, 4, 227-241
Royston, P. (2005a). Multiple imputation of missing values: Update.
Stata Journal 5, 188-201.
Royston P. (2005b). Multiple imputation of missing values: Update of
ice. Stata Journal, 5, 527-536.
Rubin, D. B. (1996). Multiple imputation after 18+ years. Journal of
the American Statistical Association, 91, 474-489.
doi:10.2307/2291635
Rubin, D. B. (2005). Causal inference using potential outcomes: Design,
modeling, decisions. Journal of the American Statistical Association,
100, 322-331. doi:10.1198/016214504000001880
SAS Institute Inc. (2008). SAS/STAT® 9.2 User’s Guide. URL (Ac-
cessed on 5 July 2009)
http://support.sas.com/documentation/cdl/en/statugsurveylogistic/618
36/PDF/default/statugsurveylogistic.pdf.
Sherraden, M. (1990). Stakeholding: notes on the theory of welfare
based assets. Social Service Review, 64, 581-601.
doi:10.1086/603797
Sherraden, M. (1991). Assets and the poor: A new American welfare
policy. Armonk, NY: M.E. Sharpe.
Sirin, S. R. (2005). Socioeconomic status and academic achievement: A
meta-analytic review of research. Review of Educational Research,
75, 417-453. doi:10.3102/00346543075003417
STATA. (2009). Survey—Introduction to survey command. URL (Ac-
cessed on 9 November 2010)
http://www.stata.com/bookstore/pdf/svy_survey.pdf
Stratton, L., O’Toole, D., & Wetzel, J. (2007). Are the factors affecting
dropout behavior related to initial enrollment intensity for college
undergraduates? Research in Higher Education, 48, 453-485.
doi:10.1007/s11162-006-9033-4
van Buuren, S., Brand, J. P. L., Groothiuis-Oudshoorn, C. G., & Rubin,
D. B. (2006). Fully conditional specification in multivariate imputa-
W. ELLIOTT ET AL. 203
tion. Journal of Statistical Computation and Simulation, 76, 1049-
1064. doi:10.1080/10629360600810434
Williams Shanks, T., & Destin, M. (2009). Parental expectations and
educational outcomes for young African American adults: Do
household assets matter? Race and Social Problems, 1, 27-35.
doi:10.1007/s12552-009-9001-7
Yilmazer, T. (2008). Saving for children’s college education: An em-
pirical analysis of the trade-off between the quality and quantity of
children. Journal of Family and Economic Issues, 29, 307-324.
Zhan, M. (2006). Assets, parental expectations and involvement, and
children’s educational performance. Children and Youth Services
Review, 28, 961-975.
Zhan, M., & Sherraden, M. (2009). Assets and liabilities, educational
expectations, and children’s college degree attainment (Working
Paper No. 09-60). St. Louis, MO: Washington University, Center for
Social Development.
W. ELLIOTT ET AL.
204
Appendix A
Table 1.
Covariate balance in 1003 matched pairs of a no savings/uncertain dose, savings only dose, certain to graduate dose only and combination dose
before adjusting for p r o p e ns i t y s c o r e w e i g h t .
Balance Balance
Covariates No Savings / Uncertain
(% or X̄)
Comparison
(% or X̄)
Β Robust
SE
No Sav-
ings /
Uncertain
(% or X̄)
Comparison
(% or X̄)
Β Robust
SE
Home scale 0.96 1.11 –0.534*** 0.057 1.03 1.09 –0.233*0.112
Head’s education level 11.66 13.25 –0.042*** 0.005 12.45 12.97 –0.017 0.035
Head is married 48 68 –0.849** 0.172 66 63 0.125 0.292
Head is not married 52 32 34 37
Parent expects child to
graduate 19 63 –1.98*** 0.21 46 54 –0.318 0.329
Parent does not expect child
to graduate 81 37 54 46
Household net worth 7.15 9.58 –0.019*** 0.003 8.71 8.76 0.000 0.005
Household income 8.97 9.88 –0.012** 0.004 9.39 9.65 –0.004 0.007
Child’s race Black 63 43 0.805*** 0.178 53 47 0.265 0.320
Child’s race White 37 57 47 53
Child’s age in 2002 16.4 16.31 0.005 0.007 16.65 16.20 0.027* 0.013
Child’s self-efficacy 2.96 3.08 –0.045* 0.02 3.05 3.04 0.003 0.189
Child’s self-concept 3.36 3.44 –0.064* 0.028 3.25 3.46 –0.157*0.058
Child’s academic achieve-
ment 187.62 209.28 –0.003*** 0.003 199.20 206.15 –0.001 0.004
Expects most/all peers to
graduate from college 45 78 –1.481*** 0.177 67 71 –3.039 0.277
Does not expect most/all
peers to graduate from
college
55 22 33 29
Attended private school 1 8 –2.01** 0.723 1 11 –4.404*0.663
Did not attend private
school 99 92 99 89
Balance Balance
Covariates Savings only
(% or X̄)
Comparison
(% or X̄) β Robust
SE
Savings
only
(% or X̄)
Comparison
(% or X̄) β Robust
SE
Home scale 1.06 1.09 –0.607 0.4 1.16 1.05 2.633* 1.026
Head’s education level 12.85 13.01 –0.031 0.034 13.12 12.78 0.066 0.061
Head is married 68 64 0.163 0.179 69 62 0.334 0.337
Head is not married 32 36 31 38
Parent expects child to
graduate 38 60 –0.876** 0.173 66 47 0.754 0.316
Parent does not expect child
to graduate from college 62 40 35 52
Household net worth 8.83 9.25 –0.023 0.019 9.60 8.68 0.281 0.044
Household income 9.7 9.73 –0.003 0.025 9.80 9.52 0.035 0.049
Child’s race Black 32 50 –0.757*** 0.179 33 53 –0.858*0.334
Child’s race White 68 50 67 47
Child’s age in 2002 16.98 16.19 0.307*** 0.056 15.85 16.43 –0.208 0.136
Child’s self-efficacy 2.99 3.08 –0.246 0.123 3.20 2.99 0.520 0.349
Child’s self-concept 3.42 3.43 –0.06 0.189 3.49 3.39 0.600 0.441
Child’s academic achieve-
ment 200.2 206.81 –0.001* 0 215.26 201.37 0.002 0.001
Child expects most/all peers
to graduate from college 57 76 –0.875*** 0.175 75 70 0.264 0.305
Attended private school 43 24 25 30
Did not attend private
school 3 8 19 5 1.350 0.739
97 92 –1.06 0.472 81 95
Source: Weighted data from the Panel Study of Income Dynamics and its supplements. Notes: Data imputed using multiple imputations. Estimates are propensity
score-adjusted using the weighting scheme in Guo & Fraser, 2010 (also see Foster, 2003 and Imbens, 2000). The propensity score weights are based on the propensity
scores (or predicted probabilities) calculated using the results of the multinomial logit model. Comparison groups consist of all children not in the dose category. The term
certain identifies children who expected to graduate from a four-year college prior to leaving high school.* p < .05; **p < .01; ***p < .001.
W. ELLIOTT ET AL. 205
Appendix B
Table 2.
Covariate balance in 1003 matched pairs of a no savings/uncertain dose, savings only dose, certain to graduate dose only and combination dose after
adjusting for prop ensity score weight.
Balance Balance
Covariates Certain Only
(% or X̄)
Comparison
(% or X̄) β Robust
SE
Certain Only
(% or X̄)
Comparison
(% or X̄) β Robust
SE
Home scale 1.05 1.09 –0.906* 0.356 1.05 1.09 –0.541 0.668
Head’s education level 12.33 13.19 –0.168***0.033 13.02 12.80 0.044 0.159
Head is married 54 68 –0.611***0.152 67 63 0.210 0.232
Head is not married 46 32 33 37
Parent expects child to graduate 53 57 –0.179 0.149 53 51 0.079 0.236
Parent does not expect child to
graduate 47 43 47 49
Household net worth 8.26 9.47 –0.063***0.016 9.10 8.63 0.024 0.077
Household income 9.3 9.86 –0.049* 0.02 9.51 9.62 –0.009 0.036
Child’s race Black 73 39 1.457*** 0.164 47 49 –0.104 0.232
Child’s race White 27 61 53 51
Child’s age in 2002 15.95 16.44 –0.189***0.047 16.44 16.26 0.060 0.076
Child’s self-efficacy 3.03 3.07 –0.118 0.124 3.01 3.06 –0.077 0.188
Child’s self-concept 3.455 3.42 0.174 0.164 3.45 3.40 0.182 0.248
Child’s academic achievement 196.37 208.57 –0.004***0.001 204.63 204.55 0.000 0.003
Child expects most/all peers to
graduate 76 72 0.179 0.171 67 71 –0.012 0.250
Attended private school 24 28 36 29
Did not attend private school 5 8 –0.363 0.316 9 8 0.124 0.649
Balance Balance
Covariates Combined
(% or X̄)
Comparison
(% or X̄) β Robust
SE
Combined
(% or X̄)
Comparison
(% or X̄) β Robust
SE
Home scale 1.05 1.09 –0.906* 0.356 1.06 1.08 –0.519 0.721
Head’s education level 12.33 13.19 –0.168***0.033 12.81 12.88 –0.014 0.063
Head is married 54 68 –0.611***0.152 55 67 –0.559*0.262
Head is not married 46 32 45 33
Parent expects child to graduate 53 57 –0.179 0.149 44 55 –0.445 0.242
Parent does not expect child to
graduate 47 43 56 45
Household net worth 8.26 9.47 –0.063***0.016 7.80 9.14 –0.065 0.032
Household income 9.3 9.86 –0.049* 0.02 9.64 9.57 0.006 0.037
Child’s race Black 73 39 1.457*** 0.164 59 44 0.594* 0.239
Child’s race White 27 61 41 56
Child’s age in 2002 15.95 16.44 –0.189***0.047 16.28 16.31 –0.009 0.076
Child’s self-efficacy 3.03 3.07 –0.118 0.124 2.94 3.09 –0.355 0.214
Child’s self-concept 3.455 3.42 0.174 0.164 3.45 3.40 0.238 0.277
Child’s academic achievement 196.37 208.57 –0.004***0.001 200.03 206.58 –0.003 0.002
Child expects most/all peers to
graduate from college 76 72 0.179 0.171 68 72 –0.224 0.241
Attended private school 24 28 32 28
Did not attend private school 5 8 –0.363 0.316 6 10 –0.616 0.522
Source: Weighted data from the Panel Study of Income Dynamics and its supplements. Notes: Data imputed using multiple imputations. The weights (adjusted) are based
on the propensity scores (or predicted probabilities) calculated using the results of the multinomial logit model. Comparison groups consist of all children not in the dose
category. Estimates are propensity score-adjusted using the weighting scheme in Guo & Fraser, 2010 (also see Foster, 2003 and Imbens, 2000). The propensity score
weights are based on the propensity scores (or predicted probabilities) calculated using the results of the multinomial logit model. The term certain identifies children who
expected to graduate from a four-year college prior to leaving high school.* p < .05; **p < .01; ***p< .001.