Economics of Education Review xxx (xxxx) xxx–xxx

Contents lists available at ScienceDirect

Economics of Education Review

journal homepage: www.elsevier.com/locate/econedurev

Assessing the impact of financial education programs: A quantitative

model☆
Annamaria Lusardia, , Pierre-Carl Michaudb, Olivia S. Mitchellc
⁎

a
The George Washington University School of Business & NBER, United States
b
HEC Montreal, CIRANO and NBER, Canada
c
Wharton School, The University of Pennsylvania & NBER, United States

ABSTRACT

Prior studies disagree regarding the effectiveness of financial education programs, especially those offered in the workplace. To explain such measurement differences
in evaluation and outcomes, we employ a stochastic life cycle model with endogenous financial knowledge accumulation and investigate how financial education
programs optimally shape key economic outcomes. This approach permits us to measure how such programs shape wealth accumulation, financial knowledge, and
participation in sophisticated assets (e.g. stocks) across heterogeneous consumers. We apply conventional program evaluation econometric techniques to simulated
data, distinguishing selection and treatment effects. We show that the more effective programs provide follow-up in order to sustain the knowledge acquired by
employees via the program; in such an instance, financial education delivered to employees around the age of 40 can raise savings at retirement by close to 10%. By
contrast, one-time education programs do produce short-term but few long-term effects. We also measure how accounting for selection affects estimates of program
effectiveness for those who participate. Comparisons of participants and non-participants can be misleading, even using a difference-in-difference strategy when the
common-trend assumption is unlikely to hold. Random program assignment is needed to evaluate program effects on those who participate.

1. Introduction and decumulation plans, all of which require expertise in dealing with
financial markets, and that they have the capacity to undertake com-
Employees and their families are increasingly responsible for se- plex economic calculations. Yet as Lusardi and Mitchell (2008, 2011a,
curing their own financial well-being. Prior to the 1980s, U.S. workers 2011b, 2017) have reported, few people possess the financial knowl-
relied mainly on Social Security and employer-sponsored defined ben- edge adequate to make and execute complex financial plans. Moreover,
efit (DB) pension plans for their retirement security. Today, by contrast, acquiring such knowledge is likely to come at a cost. Previously, we
Baby Boomers are increasingly relying on defined contribution (DC) built and calibrated a stochastic life cycle model featuring uncertainty
plans and Individual Retirement Accounts (IRAs) to finance their in income, longevity, capital market returns, and medical expenditures;
golden years. The transition to a DC retirement saving model has the that study also incorporated an endogenous knowledge accumulation
advantage of permitting more worker flexibility and labor mobility than process and a sophisticated saving technology (Lusardi, Michaud, &
in the past, yet it imposes greater responsibility on individuals to save, Mitchell, 2017; henceforth LMM). In the model, financial knowledge
invest, and decumulate their retirement wealth sensibly. At the same provided consumers with access to sophisticated financial products that
time, financial markets have become more complex, offering products boosted their expected return on financial assets. Naturally, those
that are often difficult to understand. Whether individuals—in parti- seeking to transfer resources over time by saving benefited most from
cular, older individuals—are equipped to deal with this new financial financial knowledge.
landscape is an important question that has implications for families, The contribution of the present paper is to show how our stochastic
society, and policymakers. life cycle model incorporating endogenous human capital acquisition
Traditional economic models of saving and consumption decisions can be used to help evaluate financial literacy programs. Specifically,
implicitly assume that people are able to formulate and execute saving since knowledge is at the core of the model, the approach permits us to

The research reported herein was performed pursuant to a grant from TIAA-CREF. The authors also acknowledge support provided by Netspar, the Pension
☆

Research Council and Boettner Center at The Wharton School of the University of Pennsylvania, and the RAND Corporation. Michaud acknowledges additional
support from the Fond Quebecois de Recherche sur la Societe et la Culture (FQRSC #145848). Useful comments were provided by Robert Clark. Opinions and
conclusions expressed herein are solely those of the authors and do not represent the opinions or policy of any institution with which the authors are affiliated.
©2019 Lusardi, Michaud, and Mitchell. All rights reserved.
⁎
Corresponding author.
E-mail addresses: [email protected] (A. Lusardi), [email protected] (P.-C. Michaud), [email protected] (O.S. Mitchell).

https://doi.org/10.1016/j.econedurev.2019.05.006
Received 17 August 2018; Received in revised form 23 April 2019; Accepted 14 May 2019
0272-7757/ © 2019 Elsevier Ltd. All rights reserved.

Please cite this article as: Annamaria Lusardi, Pierre-Carl Michaud and Olivia S. Mitchell, Economics of Education Review,
https://doi.org/10.1016/j.econedurev.2019.05.006
A. Lusardi, et al. Economics of Education Review xxx (xxxx) xxx–xxx

evaluate how financial education interventions can influence saving evaluation studies that only 40.5% made use of RCTs). The modern
and investment decisions. Several prior studies have sought to measure program evaluation literature has identified three commonly-used
how financial training programs change behavior, but few have ex- metrics for such comparisons: an Intent to Treat (ITT) measure, an
perimental data needed to capture precisely the impact of the inter- Average Treatment Effect on the Treated (ATET) measure, and a Local
ventions. Using our model, we evaluate the effectiveness of efforts to Average Treatment (LATE) measure. In our context, the ITT compares
build workplace financial education using econometric methods com- outcomes of those who were versus were not offered the program, ir-
monly used to estimate the effect of such programs. We do this in a respective of whether and which people actually elected the program
simulation framework. Inasmuch as all counterfactuals are known in when offered. The ATET measures the effect for the treated, not the
the context of our model, this allows us to compare ”true” outcomes average effect of moving someone into treatment, and hence it is often
with estimates commonly generated by conventional program evalua- the only way to estimate program effects when selection is present; that
tion techniques. We show that it is frequently optimal for individuals to is, one may not be able to evaluate a program’s average treatment effect
fail to invest in knowledge, as it is expensive to acquire and will not when those who do participate differ endogenously from those who do
benefit everyone. Nevertheless, providing employees with financial not. Kaiser and Menkhoff (2017) report that evaluation studies gen-
knowledge can be valuable, depending on when it is offered and what erally report larger ATET estimates than ITT estimates.1 Finally, the
reinforcement is provided. To this end, we use conventional program LATE measure defined by Imbens and Angrist (1994) captures the effect
evaluation econometric techniques and simulated data to take into of the program for those who would participate in the program only if it
account selection and treatment effects: this allows us to measure how is offered.2 Randomization of eligibility is a key ingredient for the re-
such programs shape wealth accumulation, financial knowledge, and covery of LATE by instrumental variables regression.
participation in sophisticated assets (e.g. stocks) across heterogeneous In the context of financial education programs, some authors
consumers. Relatively more effective programs are those which embed seeking to evaluate the impact of the programs have estimated ITT
follow-up or are continued over time, so as to help employees retain effects by comparing outcomes for people who were and were not ex-
knowledge acquired via the program. In such cases, financial education posed to the training, given the option to undertake the programs. Good
delivered to employees around the age of 40 will optimally enhance examples include studies of high school financial literacy mandates at
savings at retirement by close to 10%. By contrast, programs that different times across states (c.f., Bayer, Bernheim, & Scholz, 2009;
provide one-time education can generate short-term but few long-term Bernheim, Garrett, & Maki, 2001). Yet other researchers have estimated
effects. Finally, we evaluate how important it is to account for selection the effect of participating in a program which may include both treat-
in program participation. We conclude that comparing participants and ment and selection effects; numerous examples are cited in Lusardi and
non-participants, even in a difference-in-difference framework, can Mitchell (2014). Finally, several researchers have sought to estimate
deliver misleading estimates of program effectiveness. program effectiveness using instrumental variables estimation, seeking
The paper has several parts. Next, we briefly summarize prior stu- to control on potential unobserved factors driving program participa-
dies, and then, we describe our model and outline our calibration ap- tion and thus recovering the LATE (Lusardi & Mitchell, 2014). Our
proach. We then present a series of scenarios where we evaluate the general conclusion is that much remains to be learned about how fi-
simulated impacts of alternative financial education programs. In turn, nancial education affects key outcomes of interest. Without a well-de-
we use the resulting datasets to examine various econometric models fined control group selected via randomized assignment, it is typically
conventionally used to evaluate such programs. Our paper concludes difficult to measure the effect of financial education programs, since
with a short discussion of the insights that policy as well as the finance assumptions needed to estimate what adopters would have done in the
and pension industries can gain from this work. absence of the program (the counterfactual) are probably too strong.
To remedy this problem, below we show how we can use our model
2. Prior literature (LMM 2017) to help clarify what can happen when a financial educa-
tion program evaluation lacks a guiding theoretical framework. Most
In the wake of the financial crisis and ensuing Great Recession, in- importantly, given individual heterogeneity and the costs and benefits
terest has burgeoned in programs seeking to enhance financial literacy. of financial literacy, not everyone will gain from financial education.
For instance, the Organization for Economic Cooperation and Accordingly, one should not expect a 100% participation rate in every
Development OECD (2005) has published a long list of reports on the financial education program. Moreover, according to our model, fi-
importance of financial literacy and financial education programs. nancial education programs may not always boost savings, and in fact
Several U.S. education programs focus on educational interventions for they might not increase savings at all for some. Therefore, it is in-
young people before they enter the labor market (Mandell, 2008; accurate to conclude that lack of saving means that financial education
Richardson & Seligman, 2014; Walstad, Rebeck, & MacDonald, 2010), is ineffective. Instead, lack of saving can actually be optimal behavior
while others examine programs offered to working-age adults, often for some, and financial education would not be expected to change that
initiated by employers who seek to enhance employee appreciation of behavior. In this respect, our framework helps explain who is likely to
and investment in their workplace-based financial literacy programs participate in such programs, what behavioral outcomes can result, and
(Bernheim & Garrett, 2003; Clark, Ambrosio, McDermed, & Sawant, whether lack of impact is proof of program ineffectiveness.
2006; Clark, Morrill, & Allen, 2012; Collins & Urban, 2016; Lusardi,
Keller, & Keller, 2008).
Despite the widespread popularity of such programs, our recent 3. The model and calibration
literature review (Lusardi & Mitchell, 2014) as well as Collins and
Rourke (2010) argued that relatively little could be learned from most 3.1. Model
existing evaluations. This is because analysts have typically not fol-
lowed the protocol required by ‘gold standard’ randomized controlled In what follows, we focus on workplace financial education pro-
trials (RCTs), enabling researchers to extrapolate from observed results. grams of the sort most often offered by employers with defined
More specifically, a good evaluation will compare outcomes for a ran-
domly selected ‘treatment’ versus ‘control’ group, where the former will 1
In some cases, however, if a proper counterfactual can be identified, the
be exposed to a well-defined financial literacy program, while the latter average treatment effect can be estimated.
will not (Imbens, 2010; Imbens & Wooldridge, 2009). For example, 2
In a randomized control trial with one-sided non-compliance (individuals
Kaiser and Menkhoff (2017) report in a meta-analysis of 126 impact not assigned to treatment cannot receive it), the LATE estimate may coincide
with the ATET effect.

2
A. Lusardi, et al. Economics of Education Review xxx (xxxx) xxx–xxx

contribution pensions.3 We consider employees who can elect to take ft + 1 = (1 ) ft + it

advantage of such programs, which for the present purposes can be
where δ is a depreciation rate and it is gross investment. Depreciation
conceptualized as financial education of one year’s duration, delivered
exists both because consumer financial knowledge may decay, and also
to employees who have not previously anticipated getting such an offer.
because some knowledge may become obsolete as new financial pro-
We characterize each program in terms of three key parameters: an
ducts are developed. Alternatively, financial education can be modeled
eligibility rule, a program cost, and the program’s effectiveness. We
as a permanent boost to knowledge if the depreciation rate were to
assume eligibility is assigned randomly to all employees of a given age,
become smaller or even zero.
which we vary across experimental settings (more on this below). The
The consumer is also eligible for a government transfer trt which
impact of the financial education program is to reduce the employee’s
guarantees a minimum consumption floor of cmin (as in
cost of investing in knowledge. When a program is of high quality, it
Hubbard, Skinner, & Zeldes, 1995). This consumption floor can lower
provides an incentive to acquire more knowledge, and individual em-
the expected variance of future consumption, which diminishes the
ployees will then decide whether to participate in the program. Costs
precautionary motive for saving. Transfers are defined as
matter as well: for instance, if the program is free, all workers will
trt = max(cmin x t , 0) where cash on hand is:
participate (or at best they will be indifferent). In order to capture the
time/money costs of participating in the program, we model the par- x t = at + yt oopt
ticipation cost for the program as a fixed variable; a more general fra-
mework could depend on income or education, but for the present Here yt is net household income and oopt represents out-of-pocket
purposes we keep it fixed. medical expenditures. Both variables are stochastic over and above a
Following our prior work (LMM 2017), each individual is posited to deterministic trend. The sophisticated technology cannot be purchased
select his consumption stream by maximizing expected discounted if x t cd < cmin (that is, the government will not pay for costs of ob-
utility, where utility flows are discounted by β. Utility is assumed to be taining the technology). End-of-period assets are given by:
strictly concave in consumption and defined as ntu(ct/nt), where nt is an at + 1 = R (ft + 1 )(xt + trt ct (it ) cd I ( > 0))
t
equivalence scale capturing (known) differences in consumption pat-
terns across demographic groups (Scholz, Seshadri, & Khitatrakun, where R (ft + 1 ) = (1 t ) R + t R (ft ) . We impose a borrowing con-
2006). Each person’s faces a stochastic mortality risk (in addition to straint on the model such that assets at + 1 must be non-negative.
income and medical expenditure risk), and decisions are made from Following the literature, the individual’s net income (in logs) during
time t=0 (age 25) to age T (or as long as the individual is still alive; his worklife is given by a deterministic component which depends on
T=100). We examine people of three different education profiles (High education, age, and an AR(1) stochastic process; retirement occurs at
School dropouts or <HS; High school graduates or HS; and those with age 65. After retirement, the individual receives retirement income
at least some college, whom we call the College+). It is important to which is a function of pre-retirement income; a similar stochastic AR(1)
allow for heterogeneity in earnings because different groups receive process is assumed for post-retirement out-of-pocket medical ex-
different rewards from the progressive social insurance system, as de- penditures.4 Finally, we allow for mortality risk at all ages, denoting pe,t
scribed in LMM (2017), and they face differential patterns of income, as the one-year survival probability. Mortality risk is allowed to differ
mortality, demographics, and out-of-pocket medical expenditure risk. across education groups, as in LMM (2017).
We also posit that each individual can invest resources using two The state-space in period t is defined as st = ( y, t , o, t , e, ft , at ) where
different investment technologies. One is a basic technology (for ex- ηy,t and ηo,t are shocks to income and medical spending. The consumer’s
ample, a checking account) which yields a certain (low) return r decisions are given by (ct, it, κt). Accordingly, there are three continuous
(R = 1 + r ). This represents the expected return to consumers without control variables, consumption, investment, and the share of invest-
any financial know-how. The other is a more sophisticated technology ment in the technology, and a discrete one, participation. There are five
which enables the consumer to receive a higher expected return, which state variables. We represent the problem as a series of Bellman equa-
increases in financial knowledge f but comes at a cost. Specifically, the tions such that, at each age, the value function has the following form:
consumer must pay a direct cost (fee) to use the technology, cd, and he
V (st )=max ne, t u (ct / ne, t )
must also invest time and money in acquiring the knowledge to gen- ct , it , t
erate a sufficiently high excess return. Obtaining knowledge in the form
+ pe, t V (st + 1) dFe ( o) dFe ( y ) dF ( ) at + 1
of investment it thus has a cost of πi(it); we assume that this cost y o

function is convex, reflecting decreasing returns in the production of =R (ft + 1 )(at + ye, t + trt ct (i t ) cd I ( > 0)),
t
knowledge. We remain agnostic about whether the average cost of in-
vesting in additional knowledge is higher or lower for more educated at + 1 0ft + 1 =(1 ) ft + it R (ft + 1 ) = (1 t)R + tR (ft ).
households; rather, we assume initially that all households face the
We index variables by e where education differences are assumed to be
same cost function. The rate of return to the sophisticated technology is
present. The model is solved by backward recursion after discretizing
stochastic, with an expected return that depends on the individual’s
the continuous state variables.5
level of financial knowledge at the end of t, R (ft + 1 ) . Thus, the stochastic
return function is log-normally distributed with
3.2. Calibration
log R (ft + 1 ) = r + r (ft ) + t where σε is the standard deviation of a
normally distributed shock εt. The function r (ft + 1 ) is increasing in ft + 1
and can be interpreted as an excess return function. Since the variance To explore the impact of financial education on employee behavior,
is assumed fixed, this implies that individuals with higher financial for calibration purposes we assume that u(ct/nt) is of the CRRA form
knowledge obtain a higher Sharpe ratio (higher risk-adjusted returns) with relative risk aversion σ. Here we assume = 1.6, close to the value
on their investments. We denote by κt the fraction of wealth that the estimated by Attanasio, Banks, Meghir, and Weber (1999) using con-
consumer invests in the sophisticated technology in period t. sumption data. Following SSK (2006), we define an equivalence scale
Financial knowledge evolves according to the following equation: that accounts for consumption differences in household size by

4
Because these expenditures are generally low prior to retirement (and to
save on computation time), we allow only for medical expenditure risk after
3
See for instance, (Bayer et al., 2009; Bernheim & Garrett, 2003; Clark et al., retirement (as in HSZ 1995).
5
2006), and Clark, Morrill, and Allen (2014). For additional details on the solution method see LMM (2017).

3
A. Lusardi, et al. Economics of Education Review xxx (xxxx) xxx–xxx

education group and changes in demographics over the life cycle. As- 4. Simulating the impact of financial education programs
suming that z (j, k ) = (j + 0.7k )0.75 where j refer to the number of adults
in the household and k is the number of children under age 18, we then 4.1. The programs
define ne, t = z (je, t , ke, t )/ z (2, 1) where je,t and ke,t are the average number
of adults and children in the household by age and education group. We Given the model described above and the parameters of interest, we
use data from the PSID to estimate the time series of average equiva- can evaluate the impact of employer-provided financial education
lence scales by education group. The age profile of those scales is hump- programs on a variety of outcomes, including whether and which em-
shaped and more amplified for less-educated households. For the base ployees elect to participate, how much they invest in financial knowl-
case, we use a discount factor of 0.96 (as in SSK, 2006, and edge, and how they use their sophisticated technology to invest. We let
Campbell and Viceira (2002). The annual minimum consumption floor eligibility for a particular financial education program offered in a
is set at $10,000 for a couple with one child. given year be expressed using a binary variable dit, and in what follows
Post-retirement income is defined to be a function of pre-retirement we assume eligibility is assigned randomly to employees of that age
income, estimated from fixed-effect regressions on net household in- (which we vary across experimental settings). We model the financial
come on age and a retirement dummy as in LMM (2017), analyzed education program as reducing the employee’s cost of investing in
separately by education level. This produces replacement rates of 0.75 knowledge. A financial education program can increase welfare be-
for high school dropouts, 0.74 for high school graduates, and 0.63 for cause knowledge acquisition is costly. If providing education is cheaper
the College+, close to those based on total retirement income in the than self-education, then it could be welfare improving to provide fi-
literature (e.g. Aon Consulting, 2008). Following retirement, we let nancial education in the workplace. (Although we do not do a full
income decline at the rate estimated with PSID data controlling for welfare calculus with the simulations, one could easily use the model to
education and cohort effects; this pattern is mostly due to changes in do this.) We express p (it ) = (it ) , where ϑ < 1 captures the efficiency
household composition (e.g. widowhood). of the program. If the program is high quality, it provides an incentive
Turning to the financial market variables, we posit a safe asset re- to acquire more knowledge and more employees will then decide to
turn of r =2% (as in Campbell & Viceira, 2002). As the excess return participate in the program. Costs matter as well. For instance, if the
function has not been previously established, we note that the range of program is free, all workers will participate (or be indifferent). In order
risk-adjusted excess portfolio returns reported by to capture the fixed time and money costs of participating in the pro-
von Gaudecker (2015), for example, is −0.017 (5th percentile) to 0.054 gram, we define ψ as the participation cost for the program.
(95th percentile). Using Euler equations, Jappelli and Padula (2013) If the employee is eligible for a program, we define Vp(st) ( p = 0, 1)
estimate that each point of financial literacy is associated with an ex- as the value (indirect utility) of not participating versus participating,
pected increase in the return to saving from 0.2 to 1%. Clark, Lusardi, respectively. The individual participates if v (st ) = V1 (st ) V0 (st ) is
and Mitchell (2017) use administrative data on 401(k) participants and greater than zero. We add a zero mean disturbance to this difference,
find that there is about a one percentage point difference in returns ζit ∼ N(0, σv). Hence, participation is given by:
between those who have the lowest financial literacy score and those
pit = I (v (st ) + > 0).
that have the highest. We therefore use a linear function by setting it

rmax = r (fmax ) = 0.04 and rmin = r (fmin ) = 0, where 0.04 is chosen to In order for ζit to have the correct scale, we fix σv to the standard de-
match the equity premium used in the portfolio literature. Below, we viation of the simulated utility differences (0.001).
choose a convex cost function for investing in financial knowledge, The simulations to follow explore a number of different programs.
which therefore embodies decreasing returns to producing knowledge. First, program eligibility is a function of age, so we evaluate how results
We set = 0.16 in the simulations (Campbell & Viceira, 2002).6 change depending on whether the program is provided to employees at
Estimating the price of acquiring financial knowledge is difficult, as age 30, 40, or 50. When a worker is of the targeted age, he or she is
little is known regarding inputs to the production process (time and deemed to be eligible with probability 0.5. We also explore how pro-
expenditures on financial services), along with investments in, as op- gram effectiveness affects outcomes, by varying ϑ ∈ [0.1, 0.5].
posed to, the stock of financial knowledge. As in LMM (2017), we Additionally, we vary the fixed cost of participating (i.e., ψ ∈ [250,
model the process using the function (it ) = 50it1.75 . This form posits 500]). And in a last and very important case, we also allow for the
that the first units of knowledge are inexpensive, while marginal costs program to affect knowledge depreciation. That is, in one alternative,
rise thereafter. To parametrize the participation cost for the sophisti- we posit that the financial education program provides knowledge that
cated technology (cd), we use the median estimate of $750 (in $2004), does not depreciate over time. This last experiment captures the pos-
following Vissing-Jorgensen (2003). We also require an estimate of the sibility that a program could provide employees with financial advisers
depreciation factor for financial knowledge, δ, though little is known on who could be accessed over time. A total of six illustrative scenarios is
the size of this parameter. We use a value of 6% in our baseline cali- considered below.
bration which is consistent with estimates of the depreciation of human
capital. 4.2. Who participates in financial education programs?
Given this calibration, we can find optimal consumption, financial
knowledge investment, and technology participation at each point in To understand who participates and who does not in a workplace
the state-space and at each age. Having done so, we then use our de- financial education program of the sort described here, we first explore
cision rules to simulate 2500 individuals moving through their life employees’ participation patterns across various scenarios. Table 1 re-
cycles. We draw income, out-of-pocket medical expenditure, and rate of ports how participation rates in the programs vary given (randomly-
return shocks, and we use these to simulate the life cycle paths of all assigned) employee eligibility, where it is clear that participation rates
consumers. These consumers are given the initial conditions for edu- overall (last column) are generally below 100%. We emphasize that this
cation, earnings, and assets derived from the PSID for individuals age is not a sign of program failure; rather, people must incur a cost when
25–30. We initialize financial knowledge at the lowest level (0). A list of investing in knowledge, and knowledge depreciates with time. For both
the baseline parameters and their values is provided in the Appendix. reasons, not everyone will partake of the opportunity to build knowl-
edge. It is also worth noting that program participation rates rise de-
pending when (at which age) the program is offered. This is to be ex-
pected, since employees have little money to manage earlier in life and
6
For information on how we estimate income and medical expenditure people tend to save most between the ages of 40 and 60. Furthermore,
processes as well as mortality risk by education, see LMM (2017). we find that program participation is higher for the better-educated,

4
A. Lusardi, et al. Economics of Education Review xxx (xxxx) xxx–xxx

Table 1 about whether they did or did not optimally take part in each program.
Program participation. We report participation rates in the program by those Using under six different financial education settings, Figs. 1–6 re-
eligible for a series of scenarios and for three education levels. Age refers to the port results from simulations of average profiles of investment in
time at which the program is implemented, ϑ is the relative marginal cost of knowledge, knowledge stocks, changes in wealth (in percent), and
investing in knowledge in the program, and ψ is the fixed cost of participating in
shares of wealth invested in the sophisticated technology. Specifically,
the program.
Figs. 1–3 analyze how results change when the program is offered to a
Age ϑ ψ less HS HS College+ Total worker at age 30, 40, or 50. Fig. 4 reports results for a program offered
to a 40-year old employee with an enhanced efficiency parameter, and
30 0.5 500 0.1471 0.3528 0.4373 0.3594
40 0.5 500 0.32 0.4629 0.4828 0.4509
in Fig. 5 we lower the fixed cost of knowledge (shown in the same order
50 0.5 500 0.4389 0.4855 0.5879 0.5219 as in Tables 1 and 2). Fig. 6 illustrates how results change when fi-
40 0.25 500 0.3657 0.5505 0.6034 0.5466 nancial knowledge depreciation is shut down, as for instance when an
40 0.25 250 0.4629 0.6286 0.682 0.6277 employer maintains the employee’s financial sophistication post-pro-
40 0.1 100 0.6114 0.7371 0.7931 0.743
gram via continued monitoring.
A comparison of the first three figures shows how results change
when we raise the age at which the program is offered. In each case, the
due to the larger gain from investing in knowledge for those in-
upper left-hand panel depicts the impact on investment in financial
dividuals. Conversely, the least-educated are less likely to partake of the
knowledge, while the impact of the program on the stock of financial
program offering. As showed in LMM (2017), the uneducated optimally
knowledge appears in the upper right-hand panel. In the lower left, we
save less, both as a result of their greater reliance on the social safety
report the percentage change in wealth, and in the lower right, the
net and their shorter life expectancies. The final two rows of the table
share of the population using the sophisticated investment technology.
indicate how participation rates for a program offered at a given age,
Each panel includes three lines: the solid line refers to non-enrolled but
say age 40, vary depending on two factors: program efficiency, and the
eligible participants; the dashed line refers to enrolled participants; and
cost of participation. Logically enough, more efficient programs attract
the dotted line indicates how participants would have behaved without
higher participation, whereas higher costs reduce participation.
the program being introduced – a true counterfactual for those who did
Table 2 summarizes the baseline characteristics of those who elect
enroll when they could.
to participate in a financial education program when offered, versus
Fig. 1 shows what happens when the program is made available to
those who do not (conditional on being eligible at a given age). Results
employees age 30. Those who participate in the program do invest
indicate that program participants have higher earnings, more initial
substantially in financial knowledge, which translates into a higher
knowledge, and more wealth, while nonparticipants are poorer, earn
stock of financial knowledge compared to their own (no-program)
less, and have little financial knowledge at baseline. This selection oc-
counterfactual. We also see that those who participate in the program
curs regardless of the age at which the program is offered. Importantly,
cut back on their investment after the program expires. Along with
it implies that an average program effectiveness measure which as-
depreciation in financial knowledge, this leads to a dampening of the
sumes that all programs and nonparticipants could benefit equally will
program’s effect when it is over. After the initial ramp-up in financial
likely be biased.
knowledge, the marginal effect on behavior compared to the proper
The fact that those who optimally elect to undertake the financial
counterfactual is quite small. Conversely, we see that those who do
education program differ systematically from those who do not un-
invest in the financial knowledge program are markedly different from
derscores the fact that a careful program evaluation must take into
those who do not. In other words, both financial knowledge and so-
account the process by which people endogenously select into the
phisticated investment profiles are much higher compared to em-
program. That is, it would be misleading to compare outcomes for
ployees who optimally elect not to participate, underscoring the sample
program participants versus nonparticipants, since each group has
selection concern made earlier. In fact, if one were to compare program
different reasons for its behavior. Moreover, any evaluation program
participants and nonparticipants, one would (erroneously) conclude
that cannot carefully control the sample’s baseline characteristics will
that the program had an enormous impact on the stock of financial
be subject to such selection bias. Of course some of these characteristics
knowledge, producing a 20 percentage point advantage for partici-
– e.g., financial knowledge – may be difficult to measure precisely.
pants. Yet the true counterfactual shows that the net effect of a one-year
Nevertheless, unless randomization is available, modeling the selection
program offered at age 30 is quite small, particularly by the time the
process is critical.
worker attains age 65. Results are similar across Figs. 1–3, though when
the program is implemented on older versus younger workers, the
4.3. The effect of financial education programs over the life cycle
consequences appear slightly larger.
Somewhat larger program effects are evident in Figs. 4 and 5. When
A particularly useful aspect of our simulation approach is that the
the program offered becomes more efficacious for a 40-year old em-
same simulated respondents are observed in different experimental
ployee (Fig. 2 versus 4), the employee experiences a much larger bump-
settings, and in turn, they are offered different financial education
up in knowledge which persists for some time, and savings rise detec-
programs. Accordingly, we can compare life cycle investment, wealth,
tably. Similar results obtain when the cost of knowledge is reduced
and saving profiles for the same individuals, along with information

Table 2
Characteristics of participants and non-participants. We report means of baseline characteristics (income, financial knowledge, and wealth) for participants (p) and
non-participants (np). Age refers to the time at which the program is implemented, ϑ is the relative marginal cost of investing in knowledge in the program, and ψ is
the fixed cost of participating in the program.
Age ϑ ψ Income (np) Fin (np) Wealth (np) Income (p) Fin (p) Wealth (p)

30 0.5 500 35,142 2.701 18,440 56,576 11.9 51,159

40 0.5 500 41,287 15.47 43,827 70,090 44.03 120,307
50 0.5 500 46,653 34.45 142,381 69,423 61.5 207,872
40 0.25 500 36,476 10.7 33,482 69,035 42.98 115,491
40 0.25 250 36,134 11.07 35,028 65,036 38.6 103,988
40 0.1 100 33,593 10.12 38,386 61,426 34.65 92,119

5
A. Lusardi, et al. Economics of Education Review xxx (xxxx) xxx–xxx

Fig. 1. Effects of the financial education program at age 30 over the life-cycle. Intervention at age 30 with = 0.5 and = 500 . We plot the average age profile of
investment in knowledge, stock of knowledge, percent change in wealth, and the share of wealth invested in sophisticated products by participation status. For those
who participated, we also plot the age profile had they not participated in the program.

(Fig. 2 versus 5). Here again, investment in knowledge rises and some so, when. Nevertheless, in some cases, the econometrician may be able
persistence in higher savings can be seen. to observe wealth at one age (e.g., retirement), accompanied with an
A much larger and longer-term impact results from shutting down indicator of whether the person had ever been exposed to such a pro-
the knowledge depreciation parameter, confirmed by a comparison of gram earlier in life.7 This can allow a determination of how offering an
Figs. 2 and 6. The 40-year old employee offered access to a financial educational program affects outcomes of interest. In other cases, one
education program whose effects do not decay will average three times might know which employees elected to take a program, permitting a
more investment in knowledge, which in turn boosts his saving sub- comparison of outcomes between those who participated and those who
stantially. This effect persists until retirement, underscoring the long- did not. Rarely are both available, in practice, and the different out-
term effect of not only building the knowledge, but also extending it comes are not directly comparable unless, as shown above, strong as-
throughout time. In other words, a one-time financial education pro- sumptions hold about the selection process into the program.
gram may have little effect, but the long term effects of a persistent Results in Table 3 illustrate how results differ in our simulated
financial education program can be sizable. setting where we can measure each of the key employee subgroups. For
the six scenarios described earlier, we present four columns of retire-
ment wealth values. The first column summarizes wealth levels for
5. Evaluating financial education programs
participants who elected to take the program when offered. The second
column reports counterfactual wealth for the same people if the pro-
Next we use our simulated data to investigate the effect of the
gram had never been offered. The third column shows wealth levels for
programs of interest using the different metrics employed in the fi-
nonparticipants – those who were offered but declined to participate –
nancial education literature, as described above.
and the final column summarizes average wealth for those never of-
fered the program. As before, each row represents a different policy
5.1. Long-term effects experiment, with a program offered at age 30, 40, or 50 (first three

Frequently, empirical researchers may not know when individuals

in any given survey may have been exposed to or offered some sort of 7
For instance the Health and Retirement Study has asked older individuals if
program. In the present case, for instance, an employee may not recall their employers had offered them workplace-based financial education pro-
whether his employer ever offered a financial education program and if grams (Lusardi, 2004).

6
A. Lusardi, et al. Economics of Education Review xxx (xxxx) xxx–xxx

Fig. 2. Effects of the financial education program at age 40 over the life-cycle. Intervention at age 40 with = 0.5 and = 500 . We plot the average age profile of
investment in knowledge, stock of knowledge, percent change in wealth, and the share of wealth invested in sophisticated products by participation status. For those
who participated, we also plot the age profile had they not participated in the program.

rows), or at age 40 and three sets of other parameters comparable to participated. Since program eligibility is random in our simulation,
those developed in Figs. 1–5. everyone who was eligible to elect the program is in the ITT group.
Turning to the first row of that Table, program participants held From Table 1, we know that 36% of those offered the program parti-
mean wealth at retirement of $544,045. Had they not participated in cipated, which when combined with data in Table 3, yields an average
the program, the same people’s mean wealth would have been about wealth level of $407,647 for the eligible, versus $392,069 among the
0.5% lower (the difference is statistically insignificant). This is the ineligible. Hence, the intent-to-treat estimate of being eligible for the
properly-measured program effect on those who participated, con- program is (4.2%).
sistent with Fig. 1. In other words, the program did boost both financial We can recover the effect of the program on participants by com-
knowledge and wealth at the time the employees were offered the paring program participants and non participants. To do so, Imbens and
program, but by retirement, the effect virtually disappeared.This echoes Angrist (1994) suggest using the Wald estimator:
findings by Kaiser and Menkhoff (2017) from a meta-analysis of 126
E [wi,65 |di = 1] E [wi,65 |di = 0]
impact evaluation studies, who concluded that effects tend to decrease =
with time after exposure, consistent with our simulations. E [pi |di = 1] E [pi |di = 0]
In the real world, of course, we typically cannot observe the ideal
where wi, 65 is wealth of respondent i at age 65, di denotes eligibility,
counterfactual; instead, we must find ways to identify a counterfactual
and pi participation. The expectation operator is E[]. Under certain
and therefore the average effect of the program on the treated. If one
assumptions, Imbens and Angrist (1994) show that this Local Average
could reasonably assume that program participation were independent
Treatment Effect (LATE) captures the effect for a group of individuals
of wealth, then nonparticipants could be used to measure the coun-
who comply with the treatment being offered. Since the ineligible
terfactual: the estimated program effect would be to raise retirement
cannot participate, E [pi |di = 0] = 0, we have one-sided non-compliance
wealth by 60% ($554,045/$330,924-1).
and therefore the effect becomes:
These numbers would lead one to conclude that the program was
extremely effective in boosting saving. However, as demonstrated ear- E [wi,65 |di = 1] E [wi,65 |di = 0]
= .
lier, this is a severely upward-biased metric because participation is E [pi |di = 1]
correlated with wealth at baseline. Alternatively, we could investigate
the effect of offering the program without conditioning on those who This delivers the average effect of the program on the treated, or the
ATET (Imbens & Angrist, 1994). Cole, Sampson, and Zia (2011) and

7
A. Lusardi, et al. Economics of Education Review xxx (xxxx) xxx–xxx

Fig. 3. Effects of the financial education program at age 50 over the life-cycle. Intervention at age 50 with = 0.5 and = 500 . We plot the average age profile of
investment in knowledge, stock of knowledge, percent change in wealth, and the share of wealth invested in sophisticated products by participation status. For those
who participated, we also plot the age profile had they not participated in the program.

Bruhn, Ibarra, and McKenzie (2014) also follow this strategy. For our is found in the final row of Table 3, where depreciation has been shut
first scenario in Table 3, this yields a change of (0.042/0.36 = 0.116), down. This program does increase retirement wealth substantially, by
or a 11.6% increase, in percentage terms. The effect for those induced 7.7%, yet this is much smaller than the 2 times wealth increment re-
to participate by being eligible is therefore worth a 11.6% increase in sulting from (incorrectly) using the nonparticipant pool as the com-
wealth. The finding that ATET effects are larger than ITT effects has parator group.
been reported as a regularity in the meta-analysis of Kaiser and These effects are likely to be heterogeneous. Hence in Table 4 we
Menkhoff (2017). report estimates of average wealth at retirement among those who
Continuing down the rows in Table 3, it is interesting to note that participated and the wealth they would have accumulated without the
the largest bias generated by comparing participants and non- intervention (the counterfactual). We do this by quartile of lifetime
participants occurs when the program is offered to employees at age 40. income and education level. Both for income and education, the effects
At earlier ages, selection is less strong since participants and non- for low socio-economic status person (1st quartile of income or high
participants are more similar and wealth is lower. Later in life, how- schooldropouts) are much smaller, in line with many studies which
ever, the saving motive switches from precautionary to retirement have found smaller effects among disadvantaged groups (Kaiser &
preparation, and behavioral differences are exacerbated. After age 50, Menkhoff, 2017). The effects are largest in the middle of the distribu-
these differences again diminish. Since most financial education in the tion while they are also smaller at the top of the socio-economic ladder,
workplace occurs mid-career (around the age of 40), our model suggest perhaps because the latter respondents have plenty of time and re-
that selection can be a major threat to the evaluation of such programs. sources to invest in financial knowledge.
It is also of interest that the largest effects occur for most efficient We also assess how these estimates depend on some of the key
programs provided at low cost. For example, the next-to-final row in parameters of the calibrated model. In particular, the calibration of δ
Table 3 (where = 0.25, and = 250) shows that the true program and π0 does not rest on a large literature documenting their respective
effect boosts retirement wealth by $37,013 or 7.7%. Comparing the values. Hence Table 5 reports estimates similar to those in Table 3 for
ineligible with the eligible groups, we obtain a difference in wealth of two levels of δ (0.03 and 0.09 compared to 0.06 in the baseline) and
$28,967 or 7.5%. The Wald estimator of the effect for those who two levels of π0 (25 and 75 compared to 50 in the baseline). We re-run
comply with the offer of the program yields an estimated 11.9% effect the scenario with an intervention at age 40, with = 0.1 and = 100
of (7.5% / 0.628). The largest effect of all program scenarios evaluated but preventing the impact of the intervention from depreciating. We

8
A. Lusardi, et al. Economics of Education Review xxx (xxxx) xxx–xxx

Fig. 4. Effects of thefinancial education program at age 40 over the life-cycle, with lower marginal cost. Intervention at age 40 with = 0.25 and = 500 . We plot the
average age profile of investment in knowledge, stock of knowledge, percent change in wealth, and the share of wealth invested in sophisticated products by
participation status. For those who participated, we also plot the age profile had they not participated in the program.

find that estimates of the participation rate in the program are rather of 21.9% with a standard error of 11%. Controlling for covariates yields
stable across scenarios, being higher in the low depreciation scenario an even larger ITT estimate (in Table 4, the ITT estimate is 13.8%).
and lower with a high depreciation rate. Turning to wealth outcomes,
the intervention raises wealth by 2.94% in the baseline compared to
5.1.2. OLS on program participation
2.6% with = 0.03 and 3.4% with = 0.09 . As for π0, the intervention
Thus far, we have argued that, due to selection bias, comparisons of
increases wealth at retirement by 1.4% if 0 = 25 and 4.3% if 0 = 75.
participants and nonparticipants do not identify the effect of the pro-
Hence, our effects are relatively stable for different values of these two
gram on outcomes. But one might wonder whether this could be re-
parameters.
medied by controlling for factors observed early in life, say at age 25.
Since financial knowledge is likely to be zero at age 25, there are two
5.1.1. Intent-to-treat exogenous outcomes on which we could condition: wealth at age 25,
As noted above, the intent-to-treat measure in our setting compares and average lifetime income (in addition to the education dummies). To
outcomes of those who were program-eligible to those who were not, evaluate this, we run the following OLS regression:
assuming that program eligibility is exogenous. To test this with our
log wi,65 = x i + pi + i .
simulated data, we implement the following regression which controls
for education and average lifetime income: This delivers the average effect of the program on the treated, if ϵi⊥pi|xi.
log wi,65 = x i + di + i . Table 7 reports the new point estimates of △ along with their standard
errors.
Under random assignment, we have ϵi⊥di. Results in Table 7 show that when a financial education program is
Table 6 reports, for each of our six scenarios, the point estimate of offered early in life, such as at age 30, baseline controls can sufficiently
△ along with its standard error. In five of the six cases, the program correct for selection since estimated effects are close to zero. At older
effects are small and statistically insignificant, ranging from −0.091 to ages, however, the controls and functional form are insufficient to
0.1235. This confirms the unconditional level estimates reported in control for biases imparted by endogenous selection. In other words,
Table 3. By contrast, the program effect is positive and statistically the estimated effect of participating in the program becomes large and
significant for the final experiment, where financial knowledge is pre- statistically significant when the program is offered to older workers.
served through time (no depreciation). The estimate suggests an effect This is mainly due to the fact that incentives to save and acquire

9
A. Lusardi, et al. Economics of Education Review xxx (xxxx) xxx–xxx

Fig. 5. Effects of the financial education program at age 40 overthe life-cycle, with lower program and marginal costs. Intervention at age 40 with = 0.25 and
= 250 . We plot the average age profile of investment in knowledge, stock of knowledge, percent change in wealth, and the share of wealth invested in sophisticated
products by participation status. For those who participated, we also plot the age profile had they not participated in the program.

knowledge are a function of the growth of income, rather than simply difference (DD) strategy of the following form:
its level.
log wit = µi + t + z it + xit + it ,

5.1.3. Local average treatment effects

for z it = (pit , dit ) . Identification of the average effect posits that
The Wald estimator can also be implemented as an instrumental
ϵit⊥zit|xit, λt, μi. The common-trend assumption imposes that, in the
variables (IV) regression (Imbens & Angrist, 1994). In our case, the first-
absence of the program, the average change in wealth of those who
stage regression for participation is:
participate (z it = 1) would have been the same as for those not parti-
pi = x i + di + i cipating (z it = 0 ). We can estimate this equation using fixed-effect re-
gression using either pit or dit. As described above, estimates of △
assuming that eligibility is independent of ϵi. Results are reported in capture both the ATET and the ITT effects.
Table 8 along with standard errors. Our findings confirm that programs To implement this approach in our simulated data, we consider two
which do not affect depreciation have little effect on retirement-age periods: one year prior to the program, and five years after the program.
wealth levels. Although the point estimates are generally positive, the Since we can directly compute the average effect of the program on
standard errors are often large. Only in the final scenario where the those who participated (using the true counterfactual), we also report
average effect on the treated is positive does the LATE estimator pick up this estimate in column 4 of Table 9. We find that the true effect of the
the effect and the estimate becomes statistically significant. Accord- program on those who participate is generally small, except when the
ingly, this IV estimator is a proper estimator of the average treatment program is highly effective. Using non-participants as the counter-
effect on the treated (ATET) when eligibility or assignment to treatment factual (hence implementing DD with pit) yields generally large and
is random. positive effects. The key explanation for why these estimates are biased
is that the common-trend assumption does not hold for participants and
5.2. Contemporaneous effects non-participants. That is, participants in financial education programs
in our scenario would save more in the absence of the programs,
Several evaluations of financial education programs compare the compared to non-participants. For this reason, using the trend on
same individuals prior to and after receiving the training. When the wealth of nonparticipants as a counterfactual grossly overestimates the
same is done for a control group, one can implement a difference-in- effect of the programs. Implementing DD with eligibility yields

10
A. Lusardi, et al. Economics of Education Review xxx (xxxx) xxx–xxx

Fig. 6. Effects of the financial education program at age 40 over the life-cycle, with no depreciation. Intervention at age 40 with = 0.1 and = 100 and no
depreciation of knowledge among participants to the program. We plot the average age profile of investment in knowledge, stock of knowledge, percent change in
wealth, and the share of wealth invested in sophisticated products by participation status. For those who participated, we also plot the age profile had they not
participated in the program.

Table 3 Table 4
Wealth at retirement by groups. We report mean wealth at retirement (age 65) Heterogeneity in the effect of the intervention at retirement. We report mean
for those who participate in the program, mean wealth for those who partici- wealth at retirement (age 65) for those who participate in the program, mean
pate had they not participated (counterfactual), non-participants among those wealth for those who participate had they not participated (counterfactual),
eligible, and those not eligible. Age refers to the time at which the program is and the mean difference in percentage points. Intervention at age 40, with
implemented, ϑ is the relative marginal cost of investing in knowledge in the = 0.1 and = 100.
program, and ψ is the fixed cost of participating in the program.
Participants Counterfactual Effect (%)
Age ϑ ψ Participants Counterfactual Non-participant Non-eligible
Quartile lifetime income
30 0.5 500 544,045 541,444 330,924 390,971 1st 19,298 19,048 1.31
40 0.5 500 559,223 556,719 291,247 385,994 2nd 150,038 140,005 7.16
50 0.5 500 531,500 529,369 281,307 387,816 3rd 430,359 405,129 6.3
40 0.25 500 554,244 547,139 246,738 385,994 4th 776,854 763,225 1.8
40 0.25 250 515,331 508,604 245,737 385,994 Education level
40 0.1 100 511,730 474,717 230,563 385,994 less HS 181,287 188,624 -3.9
HS 299,799 267,386 12.2
College+ 450,404 442,410 1.81
Total 375,853 357,687 5.1
relatively smaller biases, compared to using participation.

6. Discussion and conclusions rationally elect not to invest in knowledge as it is expensive to acquire
and does not benefit everyone (LMM 2017). The present paper goes
In previous research, we have demonstrated that important seg- farther by using our theoretical model to evaluate the impacts of well-
ments of the population are financially unsophisticated and do not specified financial education programs that could be offered by em-
understand simple interest, inflation, and risk diversification (Lusardi & ployers to workers of different ages. In particular, we use our stochastic
Mitchell, 2008; 2011a; 2011b). We have also shown that it is actually life cycle model incorporating endogenous knowledge accumulation to
optimal for many people to be unsophisticated, in that some people will evaluate six different financial literacy program scenarios. This is useful

11
A. Lusardi, et al. Economics of Education Review xxx (xxxx) xxx–xxx

Table 5
Effect of varying parameters in wealth at retirement by groups. We report mean wealth at retirement (age 65) for those who participate in the program, mean wealth
for those who participate had they not participated (counterfactual), non-participants among those eligible, and those not eligible. We do this for alternative values of
δ (the depreciation rate of financial knowledge) and π0 (the marginal cost of investing in financial knowledge). In the baseline specification, = 0.06 and 0 = 50 .
Intervention at age 40, with = 0.1 and = 100.

Participants Counterfactual Nonparticipant Noneligible Participation (%)

Baseline 488,690 474,717 230,563 385,994 74.6

= 0.03 462,589 450,855 336,727 399,842 77.3
= 0.09 481,752 465,739 188,464 356,388 68.9
0 = 25 488,187 481,275 359,259 422,069 72.3
0 = 75 484,719 464,689 172,703 357,403 71.6

Table 6 Table 9
Effect of offering financial education program on wealth at retirement (Intent- Difference-in-difference effect of financial education program on wealth. We
to-Treat). We report for each program the intent-to-treat estimate of the pro- report estimates of the effect of the financial education program on wealth (in
gram along with its standard error. This estimate is obtained by regressing log percent) 5 years after the program, relative to one year prior to the program.
wealth at retirement on eligibility for the program, and controls for education This is done using 3 counterfactuals. The first uses outcomes of those treated
and average lifetime income. had they not participated (average effect on the treated). The second and third
columns use different counterfactuals: the second uses non-participants who
Age ϑ ψ ITT SE
were eligible, and the third uses those ineligible.
30 0.5 500 −0.09088 0.1133 Age ϑ ψ Counterfactual Non-participant Non-eligible
40 0.5 500 0.09512 0.1135
50 0.5 500 0.03485 0.1136 30 0.5 500 0 .01375 0.6978 0.6141
40 0.25 500 0.1205 0.1129 40 0.5 500 0.005544 0.2721 0.2155
40 0.25 250 0.1235 0.1129 50 0.5 500 0.00279 0.08075 0.0414
40 0.1 100 0.2188 0.113 40 0.25 500 0.02559 0.3563 0.2377
40 0.25 250 0.02453 0.3748 0.2163
40 0.1 100 0.08895 0.437 0.2407
Table 7
Effect offinancial education program participation on wealth at retirement
(OLS). We report for each program the estimate of the effect of the program financial education program evaluation. First, we show that low par-
along with its standard error. This estimate is obtained by regressing log wealth ticipation rates in such programs can be rational, once we recognize
at retirement on participation in the program and controls for education, that improving financial literacy does not benefit everyone and ac-
average lifetime income, and initial wealth (at age 25). quiring knowledge is costly. In particular, the low-income and less-
Age ϑ ψ OLS SE educated have less to gain from participating in such programs. For this
reason, it is incorrect to conclude that financial education programs are
30 0.5 500 0.08816 0.1515 not valued and ”preach only to the converted.” Rather, the decision to
40 0.5 500 0.3362 0.1393
invest in financial education depends on its costs and benefits, factors
50 0.5 500 0.7225 0.1318
40 0.25 500 0.3755 0.1309 which differ across individuals. Second, our model emphasizes the role
40 0.25 250 0.366 0.1243 of self-selection in financial education, particularly at older ages.
40 0.1 100 0.4725 0.1184 Accordingly, great care is required to rigorously evaluate the effec-
tiveness of financial education in non-experimental settings, where self-
selection tends to occur. Third, prior studies have taken too narrow a
Table 8 focus by overlooking the crucial role of knowledge retention, once the
Effect of financial education program participation on wealth at retirement
financial education is obtained. That is, financial education delivered to
(LATE-IV). We report for each program the estimate of the local average
employees around the age of 40 can raise savings at retirement by close
treatment effect along with its standard error. This estimate is obtained by
to 10%, if the knowledge gained can be maintained. Fourth, and re-
instrumental variables regression of log wealth at retirement on participation in
the program and controls for education and average lifetime income. The in- latedly, we show that short-term financial education programs are un-
strumental variable is eligibility for the program. likely to dramatically alter saving, especially when offered to young
people. They are more effective when targeted at peak saving years
Age LATE SE
(e.g., post-age 40).
ϑ ψ

30 0.5 500 −0.2531 0.3155 A final important lesson from our work is that measures of financial
40 0.5 500 0.2089 0.2487 education program effectiveness shape outcomes across heterogeneous
50 0.5 500 0.06684 0.2173 individuals. Therefore evaluators must build several key elements into
40 0.25 500 0.219 0.2048
their study designs. First, it is essential to have accurate measures of
40 0.25 250 0.195 0.1779
40 0.1 100 0.294 0.1512 what information the program delivers and what sort of follow-up is
provided. Second, the researcher must measure baseline features of the
eligible sample including wealth, income, and financial literacy. Third,
since no empirical studies have the kind of information needed to it is necessary to randomize eligibility for the treatment. And fourth,
capture precisely the impact of the interventions. In our case, we know longer-term follow-up is crucial. Intent-to-Treat methods, which make
all relevant counterfactuals to compare ”true” outcomes with program effective use of randomization, deliver powerful information about the
effectiveness estimates generated by conventional econometric techni- effectiveness of the program. If one exploits randomization further, it is
ques. Financial education programs could be welfare enhancing if the possible to estimate the average treatment effect on the treated using
cost of providing education is lower than the cost for individuals to randomization as an IV. As noted by Kaiser and Menkhoff (2017), these
acquire knowledge by themselves. methods remain under-exploited in the evaluation of financial educa-
Our approach provides several important insights regarding tion programs. One avenue for future research is to investigate these
methods both within the context of actual field experiments and

12
A. Lusardi, et al. Economics of Education Review xxx (xxxx) xxx–xxx

hypothetical choice experiments, to measure how financial education 324–347.

may improve outcomes in an unambiguous direction. Clark, R., Morrill, M., & Allen, S. (2012). Effectiveness of employer-provided financial
information: Hiring to retiring. American Economic Review May: 314-31
Cole, S., Sampson, T., & Zia, B. (2011). Prices or knowledge? What drives demand for
financial services in emerging markets. Journal of Finance, 66(4), 1933–1967.
Table A.1
Collins, J. M., & Rourke, C. M. O. (2010). Financial education and counseling: Still
Baseline Parameter Values. Baseline values holding promise. Journal of Consumer Affairs, 44(3), 483–489.
are as follows: relative risk aversion ( = 1.6 ), Collins, J. M., & Urban, C. (2016). The role of information on retirement planning:
financial knowledge depreciation rate Evidence from a field study,. Economic Inquiry, 54(4), 45–67.
( = 0.06 ), investment production function von Gaudecker, H. M. (2015). How does household portfolio diversification vary with
financial sophistication and advice? Journal of Finance, 70(2), 489–507.
( (i ) = 50i1.75 ), participation cost (cd = 750 ),
Hubbard, R. G., Skinner, J., & Zeldes, S. (1995). Precautionary savings and social in-
discount factor ( = 0.96 ). The cost of in- surance. Journal of Political Economy, 103(2), 360–399.
vesting in knowledge takes the form Imbens, G. (2010). Better LATE than nothing: Some comments on Deaton (2009) and
(i ) = 0 i 1. See text. Heckman and Urzua (2009). Journal of Economic Literature, 48(2), 399–423.
Imbens, G., & Angrist, J. (1994). Identification and estimation of local average treatment
Parameter Value effects. Econometrica, 62, 467–475.
Imbens, G., & Wooldridge, J. M. (2009). Recent developments in the econometrics of
σ 1.6 program evaluation. Journal of Economic Literature. 47(1), 5–86.
β 0.96 Jappelli, T., & Padula, M. (2013). Investment in financial knowledge and saving deci-
r 0.02 sions. Journal of Banking and Finance, 37, 2779–2792.
r(fmax) 0.04 Kaiser, T., & Menkhoff, L. (2017). Does financial education impact financial literacy and
π0 50 financial behavior, and if so, when? The World Bank Economic Review, 31(3),
π1 1.75 611–630.
cd 750 Lusardi, A. (2004). Saving and the effectiveness of financial education. In O. S. Mitchell, &
S. Utkus (Eds.). Pension design and structure: New lessons from behavioral finance (pp.
δ 0.06
157–184). Oxford: Oxford University Press.
cmin 10,000
Lusardi, A., Keller, P., & Keller, A. (2008). New ways to make people save: A social
marketing approach,. In A. Lusardi (Ed.). Overcoming the saving slump: How to increase
the effectiveness of financial education and saving programs (pp. 209–236). Chicago:
Supplementary material University of Chicago Press.
Lusardi, A., Michaud, P.-C., & Mitchell, O. S. (2017). Optimal financial literacy and saving
for retirement. Journal of Political Economy, 125(2), 431–477.
Supplementary material associated with this article can be found, in Lusardi, A., & Mitchell, O. S. (2008). Planning and financial literacy: How do women
the online version, at doi:10.1016/j.econedurev.2019.05.006. fare? American Economic Review, 98, 413–417.
Lusardi, A., & Mitchell, O. S. (Mitchell, 2011a). Financial literacy and planning: Implications
for retirement wellbeing. Financial literacy: Implications for retirement security and the
References financial marketplace. Oxford: Oxford University Press17–39.
Lusardi, A., & Mitchell, O. S. (Mitchell, 2011b). Financial literacy around the world: An
Aon Consulting (2008). Replacement ratio study: A measurement tool for retirement planning. overview. Journal of Pension Economics and Finance, 10(4), 497–508.
Global corporate marketing and communications www.aon.com/retire Lusardi, A., & Mitchell, O. S. (2014). The economic importance of financial literacy:
Attanasio, O., Banks, J., Meghir, C., & Weber, G. (1999). Humps and bumps in lifetime Theory and evidence. Journal of Economic Literature, 52(1), 5–44.
consumption. Journal of Economic and Business Statistics, 17(1), 23–35. Lusardi, A., & Mitchell, O. S. (2017). How ordinary consumers make complex economic
Bayer, P., Bernheim, D., & Scholz, J. K. (2009). The effects of financial education in the decisions: Financial literacy and retirement readiness. Quarterly Journal of
workplace: Evidence from a survey of employers. Economic Inquiry, 47(4) 605–24 Finance, 7(3).
Bernheim, D., & Garrett, D. (2003). The effects of financial education in the workplace: Mandell, L. (2008). Financial education in high school. In A. Lusardi (Ed.). Overcoming the
Evidence from a survey of households. Journal of Public Economics, 87, 1487–1519. saving slump: How to increase the effectiveness of financial education and saving programs
Bernheim, D., Garrett, D., & Maki, D. (2001). Education and saving: The long-term effects (pp. 257–279). Chicago, IL: University of Chicago Press.
of high school financial curriculum mandates. Journal of Public Economics, 80, Organization for Economic Cooperation and Development(OECD) (2005). Improving fi-
435–565. nancial literacy. Analysis of issues and policies. Geneva: OECD Publications.
Bruhn, M., Ibarra, G. L., & McKenzie, D. (2014). The minimal impact of a large-scale Richardson, D., & Seligman, J. (2014). Financial literacy among graduate students. A case
financial education program in mexico city. Journal of Development Economics, 108, study from the Indiana University system. TIAA-CREF Institute Unpublished paper
184–189. Scholz, J., Seshadri, A., & Khitatrakun, S. (2006). Are americans saving optimally for
Campbell, J. Y., & Viceira, L. (2002). Strategic asset allocation. Oxford: Oxford University retirement? Journal of Political Economy, 114(4), 607–643.
Press. Vissing-Jorgensen, A. (2003). Perspectives on behavioral finance: Does ‘irrationality’
Clark, R., Ambrosio, M.d., McDermed, A., & Sawant, K. (2006). Retirement plans and disappear with wealth? evidence from expectations and actions. NBER
saving decisions: The role of information and education,. Journal of Pension Economics Macroeconomics Annual, 18(1), 139–194.
and Finance, 45–67. Walstad, W., Rebeck, K., & MacDonald, R. (2010). The effects of financial education on
Clark, R., Lusardi, A., & Mitchell, O. S. (2017). Financial knowledge and 401(k) invest- the financial knowledge of high school students. Journal of Consumer Affairs, 44(2),
ment performance: A case study. Journal of Pension Economics and Finance, 16(3), 336–357.

13