Content uploaded by Tim Kaiser
Author content
All content in this area was uploaded by Tim Kaiser on Jan 08, 2018
Content may be subject to copyright.
Content uploaded by Tim Kaiser
Author content
All content in this area was uploaded by Tim Kaiser on Jan 08, 2018
Content may be subject to copyright.
P R W P 8161
Does Financial Education Impact Financial
Literacy and Financial Behavior,
and If So, When?
Tim Kaiser
Lukas Menkho
Development Economics Vice Presidency
Strategy and Operations Team
August 2017
WPS8161
Public Disclosure AuthorizedPublic Disclosure AuthorizedPublic Disclosure AuthorizedPublic Disclosure Authorized
Produced by the Research Support Team
Abstract
e Policy Research Working Paper Series disseminates the ndings of work in progress to encourage the exchange of ideas about development
issues. An objective of the series is to get the ndings out quickly, even if the presentations are less than fully polished. e papers carry the
names of the authors and should be cited accordingly. e ndings, interpretations, and conclusions expressed in this paper are entirely those
of the authors. ey do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and
its aliated organizations, or those of the Executive Directors of the World Bank or the governments they represent.
P R W P 8161
is paper is a product of the Strategy and Operations Team, Development Economics Vice Presidency. It is part of a
larger eort by the World Bank to provide open access to its research and make a contribution to development policy
discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org.
e authors may be contacted at lmenkho@diw.de.
A meta-analysis of 126 impact evaluation studies nds that
nancial education signicantly impacts nancial behavior
and, to an even larger extent, nancial literacy. ese results
also hold for the subsample of randomized experiments
(RCTs). However, intervention impacts are highly hetero-
geneous: nancial education is less eective for low-income
clients as well as in low- and lower-middle income economies.
Specic behaviors, such as the handling of debt, are more
dicult to inuence and mandatory nancial education
tentatively appears to be less eective. us, intervention
success depends crucially on increasing education intensity
and oering nancial education at a “teachable moment.”
Does Financial Education Impact Financial Literacy and Financial
Behavior, and If So, When?
Tim Kaiser and Lukas Menkhoff
JEL classification: D 14 (personal finance), I 21 (analysis of education)
Key words: financial education, financial literacy, financial behavior, meta-analysis, meta-
regression, impact evaluation
Tim Kaiser is a research associate at the University of Kiel, Germany and the German Institute
for Economic Research (DIW Berlin); his email address is tkaiser@diw.de.
Lukas Menkhoff (corresponding author) is the head of department of International Economics
at the German Institute for Economic Research (DIW Berlin) and Professor of Economics at
the Humboldt-University of Berlin; his email address is lmenkhoff@diw.de.
Acknowledgements: We thank the authors who responded to our requests to provide their
datasets or further details about their studies for their kind cooperation. Moreover, we
appreciate valuable comments from participants at the Research in Behavioral Finance
Conference 2016 in Amsterdam, the Meta-Analysis in Economics Research Network
Colloquium 2016 in Conway, What Works Global Summit 2016 in London, the Conference in
Behavioral Economics and Financial Literacy 2016 in Barcelona, and seminar participants in
Berlin, Halle, Hamburg, Kampala, Kiel, and Vienna. In particular, we thank the editor (Eric
Edmonds), three anonymous referees, Martin Brown, Nathan Fiala, Greg Fisher, Antonia
Grohmann, Roy Kouwenberg, Jochen Kluve, Andreas Lutter, Christian Martin, Olivia
Mitchell, Bob Reed, Anna Sokolova, Tom Stanley, Bertil Tungodden, Ludger Wössmann, and
Dean Yang. Research assistance by Melanie Krüger and Iven Lützen, and financial support by
DFG through CRC TRR 190 are gratefully acknowledged.
2
I. INTRODUCTION
The financial behavior of consumers and small-scale entrepreneurs is receiving increased interest.
Evidence suggests a remarkable incidence of suboptimal individual financial decisions despite the fact
that these decisions are highly relevant for individual welfare. The most prominent case of such an
important financial decision in advanced economies is the amount and kind of retirement savings (cf.
Duflo and Saez 2003). Studies show that undersaving is prevalent in many advanced economies and
that households tend to save in inefficient ways, indicating that many may be unable to cope with the
increasingly complex financial markets (e.g., Lusardi and Mitchell 2007; Choi et al. 2011; Behrman
et al. 2012; van Rooij et al. 2012). This kind of behavior also stretches across other areas, including
portfolio composition (Campbell 2006;Choi et al. 2010;Bucher-Koenen and Ziegelmeyer 2014;von
Gaudecker 2015), excessive and overly expensive borrowing (Stango and Zinman 2009; Gathergood
2012; Agarwal and Mazumder 2013; Gerardi et al. 2013; Zinman 2015), as well as participation in
financial markets in general (van Rooij et al. 2011). Related problems arise in developing countries
often with even more serious consequences as people are exposed to heavy shocks without having
sufficient insurance or mitigation instruments (e.g., Cole et al. 2011; Drexler et al. 2014; Gibson et al.
2014; Sayinzoga et al. 2016). All this strongly motivates providing financial education to foster
financial behavior.
In surprising contrast to this obvious motivation for financial education stands the lack of
compelling evidence that providing financial education is an effective policy for targeting individual
financial behavior (Hastings et al. 2013; Zinman 2015). Narrative literature reviews are inconclusive,
either emphasizing the effectiveness of education measures (e.g., Fox et al. 2005; Lusardi and Mitchell
2014) or emphasizing the opposite (e.g., Willis 2011). Further, the two available meta-analyses of this
issue do not converge in their findings: Fernandes et al. (2014) summarize overall unreliable effects of
financial education, whereas Miller et al. (2015) show that education can be effective in targeting
specific financial behaviors. Given this inconclusive evidence on a most important issue, what can we
learn in order to explain the heterogeneity in findings and to make financial education more effective?
We go beyond the extant literature and systematically code the circumstances of financial
education for our meta-analysis. This allows us to examine the determinants of a positive impact of
education. Another unique characteristic of our analysis is the focus on both objectives of financial
education (i.e., improvements in financial literacy and financial behavior). Hence, we investigate the
role of financial literacy for financial behavior in a unified setting. Finally, our study benefits from a
rapidly rising field (see figure S1.1 in the supplemental appendix S1).
We follow the established procedures for the meta-analysis approach (e.g., Lipsey and Wilson
2001). The result is a sample of 126 studies reporting 539 effect sizes. Studies targeting entrepreneurs
and exclusively measuring business outcomes (such as revenues) are omitted by design. We only
3
consider studies reporting about interventions, such as trainings and counseling efforts. Thus, we focus
strictly on exogenous variation in financial education and neglect works exclusively analyzing the
possible impact of cross-sectional (baseline) differences in financial literacy on financial behavior.
Finally, we carefully code interventions as we examine in detail how financial education was delivered
to the target groups.
Our meta-analysis results in six principle findings: (i) increasing financial literacy helps. Financial
education has a strong positive impact on financial literacy with an effect size of 0.26 (i.e., above the
threshold value of 0.20 that characterizes “small” statistical effect sizes [see Cohen 1977]). Moreover,
effects on financial literacy are positively correlated with effects on financial behavior; (ii) financial
education has a positive, measurable impact on financial behavior with an effect size of 0.09. An effect
size of 0.08 is still found under rigorous randomized experiments (RCTs); (iii) effects of financial
education depend on the target group. First, teaching low-income participants (relative to the country
mean) and target groups in low- and lower-middle income economies has less impact, which is an
obvious challenge for policymakers targeting the poor. Second, it appears to be challenging to impact
financial behavior as country incomes and mean years of schooling increase, probably because high
baseline levels of general education and financial literacy cause diminishing marginal returns to
additional financial education; (iv) success of financial education depends on the type of financial
behavior targeted. We provide evidence that borrowing behavior may be more difficult to impact than
saving behavior by conventional financial education; (v) increasing intensity supports the effect of
financial education; and (vi) the characteristics of financial education can make a difference. Making
financial education mandatory is associated with deflated effect sizes. By contrast, a positive effect is
associated with providing financial education at a “teachable moment” (i.e., when teaching is directly
linked to decisions of immediate relevance to the target group (cf. Miller et al. 2015:13).
Complementing these findings, the meta-analysis also provides interesting non-results because
several characteristics of financial education are without systematic impact on financial behavior.
These include the age and gender of participants, the setting, or the choice of intervention channel
through which financial education is delivered.
The findings reported above clearly motivate to implement financial education because it can
positively affect financial literacy and financial behavior. However, its limited effectiveness raises two
additional problems for policymakers: First, what can be done to make financial education generally
more effective? Second, as a particularly obstinate aspect of the general question raised before, how
can one reach those people who do not participate voluntarily? Problematic groups in this respect
include low-income individuals, residents of low-income countries, and all those who do not self-select
into education measures, as indicated by negative effects from mandatory courses and RCTs. For these
groups, it appears that financial education needs an improved approach to be successful. More research
4
and experience is necessary to better identify the determinants of successful financial education (e.g.,
Hastings et al. 2013).
Our study follows several earlier survey studies about financial education. Most of these studies
have a narrative character, among them widely cited works such as Fox et al. (2005), Willis (2011),
Hastings et al. (2013), and Lusardi and Mitchell (2014). This gives the authors some flexibility about
selecting and interpreting the most relevant studies. A quantitative meta-analysis is more rigid in
approach but has the advantages that transparent rules of procedure ensure replicable results and that
quantitative relations can be derived. Overall, narrative surveys and meta-analyses complement each
other.
We perform a meta-analysis because there are just two earlier systematic accounts of the financial
education literature that leave much room for more research. The study by Miller et al. (2015) covers
only 19 papers due to its extremely restrictive selection criteria, requiring interventions on identical
outcomes. This limits the sample sizes to about five studies and estimates per subsample, which does
not allow investigating the sources of heterogeneity.
Thus, the most similar study to our work is Fernandes et al. (2014), which covers 90 effect sizes
from financial education reported in 77 papers. Despite an overlap of 44 percent with their sample of
studies, our research differs in four crucial ways, which explains our new results: (i) most important is
that we analyze determinants of program effectiveness in a broader way by applying respective coding;
(ii) we consider various outcomes per study (on average about four per study) and their respective
effect sizes; moreover, (iii) we cover recent and mostly randomized experiments providing evidence
of effective interventions; and (iv) we cover additional studies focusing exclusively on financial
literacy as the outcome variable.
This paper is structured in seven further sections. Section 2 introduces our meta-analytic approach.
Section 3 describes our data. Section 4 provides first results of the meta-analysis, while section 5 uses
these results to explain heterogeneity of financial education treatment effects. Robustness tests are
mentioned in section 6, and section 7 concludes with policy considerations and venues for future
research.
II. META-ANALYTIC METHOD
Meta-analysis is a quantitative method to synthesize findings from multiple empirical studies on the
same empirical research question. In a meta-analysis, the dependent variable is comprised of a
summary statistics reported in the primary research reports, while the explanatory variables may
include characteristics of the research design, the sample studied, or, in case of impact evaluations, the
policy intervention itself (cf. Stanley 2001: 131). Meta-analyses can provide answers to two specific
5
questions (cf. Muller 2015; Pritchett and Sandefur 2015; Vivalt 2015). First, is the combined
(statistical) effect across all studies reporting effects of similar interventions on similar outcomes
significantly different from zero? And, second, what explains heterogeneity in the reported findings?
In order to be able to aggregate summary statistics reported across heterogeneous studies, one must
standardize these statistics into a common metric. If all studies would operationalize and measure
outcomes in the same unit, meta-analysis could be performed directly using economic effect sizes (e.g.,
elasticities or marginal effects) in contrast to statistical effect sizes (cf. Stanley and Doucouliagos
2012: 23). This, however, is rarely the case in a large sample of heterogeneous (quasi-) experimental
impact evaluations.
Thus, we use a standard approach of coding a variable capturing intervention success and impact.
Our impact measure (effect size) is the standardized mean difference (SMD) for each treatment effect
estimate. We use the bias corrected standardized mean difference (Hedges’) as our effect size
measure, which is defined as the mean difference in outcomes between the treatment (M) and control
(M) (i.e., the treatment effect) groups as a proportion of the pooled standard deviation (SD) of the
dependent variable:
(1) =
with
(2) =()()
.
where n and are the sample size and standard deviation of the treatment group, and and
are for the control group. Additionally, we capture the standard error of each standardized mean
difference (), which is defined as:
(3) =
+
()
Hedges’ informs about the size and direction of an effect in scale-free standard deviation units.
This metric is only slightly different from other popular effect size measures in experimental impact
evaluations, such as Cohen’s d and Glass∆ (see, e.g., Banerjee et al. 2015). Hedges’ , however,
introduces minor corrections that reduce bias in the effect size estimate in cases with small sample
sizes and when the sample sizes of treatment and control groups are unequally distributed. Results are
qualitatively robust to using alternative measures or relying on (partial) correlations (cf. Lipsey and
Wilson 2001).
6
As a rule of thumb, Cohen (1977) suggests that effect sizes smaller than 0.20 should be considered
as a “small effect”; effect sizes around 0.50 indicate a “medium effect”; while effect sizes greater than
0.80 constitute “large effects.” Where pure mean comparisons, standard deviations, and sample sizes
for each experimental outcome are not reported directly we exhaust all possibilities to calculate or
estimate effect sizes () and its corresponding standard error from the range of available statistical
data (cf. Lipsey and Wilson 2001).
In the estimation of summary effects of the literature, our main approach follows a full pooling
least squares meta-regression framework (e.g., Card et al. 2015). Accordingly, the financial education
treatment effect () can be explained by exogenous, observable characteristics, the impact on an
outcome i, reported in study j is expressed as a linear function
(4) =+
+
where is a vector of observable (exogenous) study-level covariates, such as intensity of
intervention, α is an intercept, and denotes an error-term independent from . We estimate our
models using multiple effect sizes per study and account for heteroscedasticity by clustering standard
errors at the study-level. Reassuringly, results are not sensitive to a set of changes in estimation strategy
and accounting for publication selection bias (see section 6 and supplemental appendix S3).
III. SAMPLE DESCRIPTION
This section describes the selection of studies, the extraction of effect sizes and study-level covariates,
and types of financial education programs.
Selection of Studies
We follow the established meta-analytical protocol (cf. Lipsey and Wilson 2001: 23, Stanley 2001:
143). This starts with systematically searching the relevant databases, including working papers, for
the following keywords: (i) financial literacy; (ii) financial knowledge; (iii) financial education; (iv)
financial capability; and (v) combinations of these keywords with “intervention.” Moreover, we
consider all records from meta-analyses (Fernandes et al. 2014; Miller et al. 2015) and narrative
literature reviews (Fox et al. 2005; Collins and O’Rourke 2010; Willis 2011; Xu and Zia 2012;
Hastings et al. 2013; Blue et al. 2014; Lusardi and Mitchell 2014). This search resulted in over 500
potentially relevant published journal articles and over 600 results from working paper databases with
some apparent overlap. We stopped collecting studies in October 2016 (see appendix S1).
From this collection, we drop studies that do not meet our three criteria for inclusion: (i) reporting
on impacts of an exogenous educational intervention on financial literacy and/or financial behavior;
7
(ii) providing a quantitative assessment of intervention impact that allows coding an effect size statistic
() and its standard error; and (iii) relying on an observed counterfactual in the estimation of
intervention impacts. This selection process leads to a final sample of 126 independent intervention
studies that report 539 effect sizes (further details in tables S1.1 and S1.2 in the supplemental appendix
S1). Of these, 90 studies report 349 effect sizes on financial behavior, and 67 studies report 190 effect
sizes on financial literacy. Among these 90 plus 67 studies, there are 31 studies reporting effect sizes
on both financial literacy and behavior.
RCTs are rare in the early years of the literature, but their share has risen dramatically, with the
majority of studies conducted from 2011 onward being randomized evaluations (see figure 1). This
development in the literature is very favorable for meta-analyses, since it ensures a high internal
validity of research findings reported in the primary studies and helps to clearly distinguish between
selection and treatment effects.
Figure 1. Number of Studies in Our Sample by Research Design per Year
Source: Authors’ calculations based on the data source discussed in the text.
8
Extraction of Effect Size Estimates and Study Descriptors
As the next step, we code the effect of financial education on financial literacy (i.e., a measure of
performance on a financial knowledge test), since knowledge development is the primary goal of
financial education (Hastings et al. 2013; Lusardi and Mitchell 2014). Moreover, we code treatment
effects of financial education on several financial behaviors (see table S1.2 in the supplemental
appendix S1), such as an increase in savings after the treatment. Multiple estimates per study are
considered if multiple outcomes, time-points, or treatments are reported; however, results are robust
to aggregating all effects per study into one synthetic effect size. Further details about this process are
described in supplemental appendix S1.
Types of Financial Education Programs
Our dataset includes four main types of financial education programs. First, and most frequent, are
evaluations of classroom financial education (approximately 83 percent of all estimates) in various
settings, such as schools, universities, the workplace, or specific sites such as savings groups or
microfinance institutions. These studies are quasi-experiments or RCTs, in which the researcher has
control over content, intensity, and survey design in order to measure specific outcomes. There is an
increasing interest in the literature in multiple-treatment and cross-over designs to investigate optimal
delivery strategies and potential causal mechanisms (i.e., Drexler et al. 2014; Carpena et al. 2015;
Skimmyhorn 2016). These studies have high internal validity but may report site-specific effects that
causally interact with unobserved features of the specific sites (cf. Muller 2015). Additionally,
measurement of outcomes is typically in the short or medium run (approx. 65 percent), since long time
series are usually not available. A different strand of the literature evaluating this type of program
looks at classroom financial education utilizing (plausibly exogenous) variation in (mandatory) school
financial education mandates (e.g., Tennyson and Nguyen 2001; Brown et al. 2016). These studies are
typically quasi-experimental in nature, and, while possibly weaker in internal validity, possess high
external validity, since they typically have large sample sizes and measure relatively long-run effects
on behavioral outcomes, such as savings.
A second type of intervention is online financial education (approx. 8 percent of estimates). While
similar in research design to experiments on classroom financial education, these studies usually
estimate the effect of certain online modules on financial literacy and behavior and typically evaluate
instructional videos or interactive applications.
The third type of financial education treatments evaluated in the literature are individualized
counseling interventions (two percent of estimates). These have been mainly studied in the US and
typically study outcomes related to the handling of (mortgage) debt.
9
As a fourth and last type, we identify informational and behavioral nudges, such as information
fairs at the workplace and informational brochures (seven percent of estimates). These studies typically
evaluate behavioral change in response to these low-intensity treatments. There is one study in our
sample that studies the effect of a behavioral nudge in the form of “financial edutainment” in mass-
media (cf. Berg and Zia 2013). This is an intervention designed to impact financial behaviors through
a non-cognitive channel (as opposed to increasing financial knowledge), and the included study
evaluates the impact of financial messages inserted into episodes of a popular television series in South
Africa.
IV. RESULTS FROM META-ANALYSIS
We report the mean effects for all studies (section 4.1) and then for subsamples: financial literacy and
financial behavior (section 4.2), types of financial education programs (section 4.3), research designs
(section 4.4), and different country groups (section 4.5).
Summary Effects of Financial Education
Here we discuss the average effects of financial education on financial literacy and financial behavior.
Based thereon, we study the relation between these two outcomes. As a starting point, we note that the
summary effect of financial education on all kinds of reported outcomes is estimated to be g = 0.148
(p = .000, n = 539). However, heterogeneity in effect sizes is high, indicating that outcomes could be
disaggregated for meaningful analyses.
Financial behavior. We find that the average impact of educational interventions on financial
behaviors is statistically highly significant (g = 0.086) (see table S1.3 in the supplemental appendix
S1). The main reason that we get a more favorable result than Fernandes et al. (2014) is that we profit
from a moderate, positive time trend (more details in supplemental appendix S2). To compare the
magnitude of this effect size to results from health promotion on behavioral change (e.g., weight loss
and nutrition in obesity studies), Portnoy et al. (2008) report in their meta-analysis of 75 RCTs an
average effect size of about 0.1.
Financial literacy. The average impact of financial education on financial literacy is substantially
higher (g = 0.263, p = .000, n = 190) than the one on financial behavior (see figure S1.2 and table S1.3
in the supplemental appendix S1). Moreover, financial education explains 1.7 percent of the variance
in financial knowledge and, thus, appears to be only slightly less effective than educational
interventions in other domains such as math and science instruction (cf. Fernandes et al. 2014: 1867).
To put this effect size in perspective: the meta-analysis of 225 studies by Freeman et al. (2014) reports
an average effect size of around 0.47 for studies evaluating student performance in response to
10
alternatives to lecturing in undergraduate science education; however, these interventions occur in a
university context and last for a full semester.
Relationship between financial literacy and behavior. The intuition is that increases in financial
literacy scores are an important intermediate result in a causal chain expected to lead to behavior
change (e.g., Grohmann et al. 2015; Fort et al. 2016). Indeed, for a sample of 31 studies, we find in a
regression with standard errors clustered at the study-level that the effect size on financial literacy is a
statistically significant predictor of effect size on financial behavior (b = 0.230, p =.022). Thus, an
increase of one standard deviation unit in financial literacy scores is related to an average increase of
0.23 standard deviation units of the financial behaviors studied. However, the non-overlapping
confidence intervals of these effect sizes also indicate that these two elements of the causal chain
should be analyzed separately when attempting to explain the heterogeneity in effect sizes.
Effect Sizes by Type of Financial Behavior
Figure 2 shows the average effect size for the seven categories of financial behaviors targeted by the
educational interventions in our sample.
Figure 2. Forest Plot of Effect Sizes by Type of Financial Behavior Studied
Source: Authors’ calculations based on the data source discussed in the text.
11
Average effect sizes for three out of seven categories of outcomes are clearly positive and highly
statistically significant at the one percent level. Additionally, all confidence intervals for the different
types of financial behaviors overlap each other, indicating that there are no extreme differences in
impacts depending on the specific form of financial behavior targeted. In detail: (i) the average effect
size on “budgeting” appears to be higher than those on downstream behaviors; and (ii) effect sizes
related to saving and retirement saving appear to be higher than the average effect size of financial
education on borrowing behavior; (iii) this latter average effect size is small (g = 0.02) and insignificant
from zero; (iv) similarly, the average effect sizes for “insurance” (g = 0.05), “remittances” (g = 0.03),
and “bank account behavior” (g = 0.00) are estimated to be small and insignificant from zero, although
based on a few studies per category only. Thus, debt-related financial behaviors may be the most
challenging to target through financial education (see Miller et al. 2015: 238). Overall, these findings
correspond to the results provided by Fernandes et al. (2014) and Miller et al. (2015) and extend to our
much larger sample.
Effect Sizes by Type of Financial Education Intervention
We form subsamples by the main types of financial education interventions, as discussed in section
3.3. First, we compare classroom financial education to three types of non-classroom delivery channels
(online financial education, counseling, and informational/behavioral nudges). Second, we distinguish
between financial education at school and two non-school settings (workplace and other settings).
Panel A of table 1 shows results split by outcomes on financial literacy and financial behavior. While
in-person classroom trainings appear to be (unconditionally) more effective than non-classroom
delivery channels in increasing financial knowledge, we observe no statistically significant difference
regarding impacts on financial behavior. Turning to the intervention setting, it appears that
interventions in schools are more effective at increasing financial literacy but yield marginally
significant smaller treatment effects on financial behavior. However, we note that these relations are
obviously partially confounded with several other relevant variables (e.g., the age of the participants,
the delay in measurement, and research design), which indicates the importance of an examination in
a multivariate setting (cf. section 5).
12
Table 1. Effect Sizes of Financial Education by Intervention Type, Research Design, and Country
Groups
Outcome Type Studies Obs. ES (g) SEg p-value Diff. (t-value)
A Effect sizes by intervention channel & setting
Fin. literacy Classroom 58 135 0.294 0.054 0.000 0.106**
(2.015)
Non-classroom 9 55 0.188 0.039 0.001
- Online 5 41 0.217 0.060 0.018
- Counseling 0
- Nudge 4 14 0.103 0.045 0.108
Fin. behavior Classroom 70 317 0.084 0.013 0.000 -0.014
Non-classroom 20 32 0.098 0.020 0.000 (0.452)
- Online 11 18 0.085 0.034 0.031
- Counseling 7 8 0.095 0.030 0.020
- Nudge 2 6 0.140 0.007 0.031
Fin. literacy School 35 62 0.373 0.076 0.000 0.163***
Non-school 32 128 0.210 0.035 0.000 (3.273)
- Workplace 1 1 0.164 0.063
- Other 31 127 0 210 0.035 0.000
Fin. behavior School 27 90 0.057 0.014 0.000 -0.039*
Non-school 63 259 0.096 0.014 0.000 (1.96)
- Workplace 17 47 0.121 0.049 0.023
- Other 46 212 0.090 0.015 0.000
B Effect sizes by research design
Fin. literacy RCTs 33 135 0.209 0.033 0.000 -0.185***
Quasi-exp. 34 55 0.394 0.083 0.000 (-3.638)
Fin. behavior RCTs 40 227 0.081 0.015 0.000 -0.012
Quasi-exp. 50 122 0.093 0.022 0.000 (-0.661)
C Effect sizes by country group
Fin. literacy High income 53 123 0.328 0.058 0.000 0.183***
Developing 14 67 0.145 0.031 0.000 (3.787)
- Low 3 6 0.219 0.069 0.086
- Lower-middle 6 44 0.155 0.047 0.023
- Upper-middle 5 17 0.092 0.023 0.017
Fin. behavior High income 66 168 0.071 0.019 0.000 -0.027
Developing 24 181 0.098 0.014 0.000 (-1.512)
- Low 6 39 0.161 0.038 0.009
- Lower-middle 12 90 0.091 0.008 0.000
- Upper-middle 6 52 0.06 0.023 0.045
Notes: Average effect sizes (g) estimated via OLS regressions of effect sizes fitting only an intercept. Sample is split by an
indicator of intervention type, research design, or country group. “Channel” is a categorical variable operationalized in the
form of four dummy variables: Classroom, Counseling, Online, and “Nudge” where “Nudge” is the default (omitted)
category in the regressions. “Setting” is a categorical variable operationalized through three dummy variables: School,
Workplace and Other where Other is the omitted category in the meta-regression analyses. Country groups are based on
the World Bank Atlas method and refer to 2015 data on GNI per capita. Low-income economies are defined as those with
a GNI per capita of $1,025 or less in 2015, lower-middle income economies are defined by a GNI per capita between
$1,026 and $4,035, upper-middle income economies are those with a GNI per capita between $4,036 and $12,475, and
high-income economies are defined by a GNI per capita greater than $12,475. Standard errors are clustered at the study
level. ***, ** and * denote significance at the one percent, five percent and ten percent level.
Source: Authors’ analysis based on data sources discussed in the text.
13
Effect Sizes by Research Design
Regarding research design, Fernandes et al. (2014: 1865) find that weaker research designs lead to
inflated effect sizes. Thus panel B of table 1 compares average effect sizes as a function of research
design. When we focus on financial behaviors as outcomes, RCTs show statistically highly significant
(unconditional) effect sizes of 0.081. These are only slightly smaller than for quasi-experiments with
0.093, indicating that the small but positive significant effects of financial education exist even under
the most rigorous empirical standards. RCTs also provide a significant positive effect of financial
education on financial literacy with 0.209. Here the difference to other designs (effect size of 0.394)
is significant at the one percent level.
Effect Sizes by Country Groups
To investigate another potential source of heterogeneity, we disaggregate our data by country groups.
Panel C of table 1 shows effect sizes by country groups as classified by the World Bank based on 2015
GNI per capita. We find that effect sizes on financial literacy are significantly higher in developed
(high income) economies (g = 0.328) than in developing economies (low income, lower- and upper-
middle income economies, g = 0.145). Turning to effect sizes on financial behavior, this difference is
statistically insignificant in this unconditional comparison but differences between country groups
become more nuanced and statistically significant when controlling for other relevant variables (see
section 5.2).
V. EXPLAINING HETEROGENEITY IN FINANCIAL EDUCATION TREATMENT EFFECTS
Section 4 shows that the average effect size of financial education is accompanied by large
heterogeneity. Thus, we examine whether there are factors explaining this heterogeneity. This will also
suggest directions that future financial education policies might take in order to increase their impact
on financial behavior.
Potential Correlates of Effect Size
The effectiveness of financial education is potentially influenced by the peculiarities of the specific
intervention. Based on prior literature, we group these characteristics into four categories: (i) the
research design; (ii) the intensity of education; (iii) the target group of education; and (iv) the
characteristics of the education program.
(i) Regarding the research design of a financial education study, we expect the method of
investigation (i.e., RCT vs. less rigorous designs) to be relevant. Second, the concrete measurement of
an effect will influence the estimated size of impact. It is known that focusing on treatment on the
14
treated (TOT) (i.e., measuring a treatment effect on the population who actually received or attended
the treatment) generally results in higher effect sizes than focusing on the intention to treat (ITT) effect
(i.e., the population who was in principle assigned to treatment). However, ITT may be more relevant
for policy (cf. Imbens and Wooldridge 2009: 15; Gertler et al. 2011: 73). Third, the delay between
financial education treatment and measurement of the effect may negatively influence the effect size
since effects of the intervention may decay over time (cf. Fernandes et al. 2014: 1867). Additionally,
we control for the precision of effect size estimation by the inverse standard error (or the [squared]
standard error, see supplemental appendix S3). All these variables are defined in table 2, which also
provides descriptive statistics.
(ii) A core variable of financial education interventions is the intensity of education (i.e., the
number of hours taught). It is expected that higher intensity will support the effect. However, the time-
frame over which the financial education intervention is delivered to the target group may also be of
importance. We expect differences between high intensity and low intensity relative to the duration.
Thus, we code the hours of financial education per week (i.e., intensity per week) and the duration of
the intervention in weeks to investigate this issue.
(iii) The expectation regarding a possible relation between the target group of education and
effectiveness of financial education is as follows. Generally, learning is easier for younger people,
younger people may be more open to new concepts and their baseline financial literacy scores are low
(e.g., Lusardi and Mitchell 2014), meaning that the age of the target group may be negatively related
to the effect size of financial education. Second, a gender gap in financial literacy is treated as a stylized
fact in the literature (cf. Lusardi and Mitchell 2014) which may also translate to gender differences in
treatment effects. Thus we include the percentage of women in the sample. Third, it is expected that
the acquaintance of the target group with an educational environment may be helpful. As a proxy for
such openness to education, we take the income of the target group relative to the overall population.
Fourth, we expect that the overall institutional level of education should support domain-specific
educational efforts (Jappelli 2010). As a proxy for this potential relationship, we take a country’s
population mean years of schooling as reported by the United Nations Development Program Human
Development Reports. Additionally, we augment our data with country-level financial literacy data
from a 2015 global financial literacy survey (Klapper et al. 2015). We hypothesize that financial
education interventions may yield higher effects when the population baseline financial literacy is
lower, indicating more room for improvement through education. Finally, as a control variable we
code the country of intervention according to the World Bank country group classifications.
(iv) Regarding the characteristics of the education program, it seems interesting whether the
channel (i.e., classroom, online, individual counseling, etc.) is important in explaining education
effectiveness, since these formats come with different trainer to participant ratios and may rely on
15
different pedagogical approaches to financial education. It may be that willingness to learn and change
financial behavior is lower when financial education is mandatory (cf. Collins 2013) or motivation to
participate in financial education is not intrinsic but driven by incentives provided by the offering
institution. Lastly, these characteristics may be correlated with specific settings (i.e., at school or at the
workplace).
Next, and going further in this direction, it is coded whether participants are educated at a teachable
moment (i.e., that they have the possibility to apply their knowledge in a concrete case of interest to
them, e.g., Doi et al. 2014; Miller et al. 2015). Thus, we capture whether the education addressed
immediate financial issues (such as borrowers already in default, or migrants confronted with deciding
through which channel remittances are sent). Alternatively, financial education was generic and
offered at an unspecific moment, as is often the case in large scale financial education programs (e.g.,
Bruhn et al. 2014).
Table 2. Summary Statistics
Variable Obs. Mean Std. Dev. Min. Max.
A Descriptive statistics at the study-level
RCT 126 0.405 0.493 0.000 1.000
TOT 115 0.452 0.500 0.000 1.000
Delay 93 82.231 273.613 0.000 1566
1/SE 126 57.535 210.450 2.480 1636.712
Intensity 87 11.211 14.929 0.100 87.000
Duration 76 7.341 14.150 1.000 103.000
Age 109 30.717 14.120 9.000 63.870
Percent female 123 54.011 18.493 0.000 100.000
Low income clients 102 0.529 0.502 0.000 1.000
Years of schooling 126 11.270 2.843 3.200 13.600
FL in population 124 50.419 11.658 24.000 66.000
Mandatory 96 0.292 0.457 0.000 1.000
Incentivized 86 0.314 0.467 0.000 1.000
Teachable moment 126 0.397 0.491 0.000 1.000
B Descriptive statistics at the estimate-level
RCT 539 0.672 0.470 0.000 1.000
TOT 510 0.282 0.451 0.000 1.000
Delay 463 93.742 292.025 0.000 1566.000
1/SE 539 41.260 124.389 2.740 957.167
Intensity 451 15.384 23.444 0.100 144.000
Duration 434 7.908 14.236 1.000 103.000
Age 494 31.814 11.720 9.000 63.870
-continued-
16
Percent female 525 52.923 18.200 0.000 100.000
Low income clients 451 0.681 0.467 0.000 1.000
Years of schooling 539 9.890 3.463 3.200 13.600
FL in population 523 44.170 14.668 24.000 66.000
Mandatory 480 0.240 0.427 0.000 1.000
Incentivized 445 0.247 0.432 0.000 1.000
Teachable moment 539 0.479 0.500 0.000 1.000
Notes: “RCT” is a dummy variable with “1” if selection into treatment was conducted through randomization and “0”
otherwise (such as matched designs). “TOT” is a dummy variable with “1” if the effect size estimate is derived from the
treatment effect on the treated and “0” if it is derived from the ITT estimate. “Delay” is a continuous variable indicating
the delay between treatment and measurement of outcomes in weeks. “1/SE” is the inverse standard error for each effect
size estimate. “Intensity” is the total number of hours of financial education exposure to the treated. “Duration” indicated
the time-frame of financial education in weeks. “Age” is the mean age of the sample in years. “Percent Female” is the
relative frequency of female participants in the sample in percent. “Low income” is a dummy variable with “1” if the mean
annual income per capita of the sample is below the country average income per capita. “Mandatory” is a dummy variable
with “1” indicating mandatory participation in financial education and “0” voluntary participation. “Incentivized” is a
dummy variable with “1” when incentives to participate where provided and “0” if participation was unconditional on
incentives. “Teachable moment” is a dummy variable indicating whether the financial education intervention was offered
at a teachable moment.
Source: Authors’ analysis based on data sources discussed in the text.
Meta-regression Models Explaining Intervention Impacts
This section examines determinants of financial education effectiveness using a multivariate meta-
regression framework including the above discussed potential correlates as right-hand side variables.
Our procedure is motivated by economic and econometric considerations. From an economic point of
view, we aim for including all variables that have a substantial theoretical foundation. From an
econometric viewpoint, the specification should be parsimonious, especially in the presence of a
relatively small sample size of studies.
Thus, we start with a specification where we include all reasonable and available variables (table
3, column 1). In order to keep the number of studies considered high, we impute average or default
values for missing observations (we show in supplemental appendix S3 that our main results are
insensitive to imputation). The discussion considers groups of variables in four blocks, following their
introduction in section 5.1.
17
Table 3. Explaining Heterogeneity in Effect Sizes on Financial Behavior
Notes: Non-standardized coefficients from OLS regressions. Dependent variable in columns (1) and (2) is effect size
(Hedges’ g) on financial behavior in the full sample of studies reporting on financial behavior as an outcome. Column (3)
shows results for RCTs only. Column (4) and (5) show results for financial behavior split by country groups. Column (6)
limits the sample to classroom trainings only. Robust standard errors clustered at the study-level in parentheses. ***, **
and * denote significance at the one percent, five percent and ten percent level.
Source: Authors’ analysis based on data sources discussed in the text.
(1)
All
(2)
All
(3)
RCTs
(4) Low
inc. econ
(5) High /
middle inc.
econ
(6) Low
income
clients
RCT -0.070** -0.068** -0.209** -0.079** -0.066**
(0.027) (0.028) (0.091) (0.036) (0.032)
TOT 0.079*** 0.068** 0.012 -0.016 0.076** 0.031
(0.027) (0.027) (0.040) (0.066) (0.035) (0.032)
Delay 0.000 0.000 -0.001** -0.001** 0.000 -0.000
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
1/SE -0.000 -0.000 0.000 -0.003 -0.000 0.000
(0.000) (0.000) (0.001) (0.002) (0.000) (0.000)
Intensity / week 0.004** 0.004*** 0.007*** 0.004** 0.003 0.004***
(0.002) (0.001) (0.001) (0.002) (0.003) (0.001)
Duration -0.000 -0.000 -0.001 -0.001 -0.000 0.000
(0.000) (0.000) (0.001) (0.001) (0.001) (0.000)
Age -0.001
(0.001)
Percent female -0.000
(0.001)
Low income clients -0.065*** -0.055*** -0.074*** -0.042** -0.048**
(0.020) (0.017) (0.024) (0.019) (0.021)
Years of schooling -0.016*** -0.019*** -0.016** -0.026*** -0.025*** -0.011*
(0.006) (0.006) (0.006) (0.009) (0.009) (0.006)
FL in population -0.003
(0.002)
Country group
a) Low/lower-mid. inc. econ. -0.129* -0.093** -0.092** -0.059
(0.073) (0.036) (0.042) (0.042)
b) Upper-mid. inc. econ. 0.000
(0.060)
Channel
a) Classroom -0.003
(0.028)
b) Counseling -0.018
(0.033)
c) Online -0.028
(0.028)
Setting
a) School 0.022
(0.023)
b) Workplace 0.041
(0.036)
Mandatory -0.074*** -0.051** -0.078* -0.015 -0.065** -0.052
(0.024) (0.023) (0.044) (0.042) (0.025) (0.033)
Incentivized -0.012
(0.029)
Teachable moment 0.079*** 0.064** 0.016 0.025 0.069** 0.072**
(0.021) (0.026) (0.035) (0.026) (0.029) (0.032)
Constant 0.477*** 0.332*** 0.338*** 0.514*** 0.406*** 0.188*
(0.157) (0.079) (0.095) (0.110) (0.114) (0.095)
R2 0.210 0.183 0.149 0.170 0.204 0.109
n (Studies) 90 90 40 18 72 44
n (Effect sizes) 349 349 227 129 220 234
18
Research design. Starting with the research design of the underlying primary studies, we find that
RCTs report—ceteris paribus—slightly smaller effect sizes than non-RCTs, which is in line with
earlier presumptions (see table 1, panel B). However, now this difference is statistically significant
(see column 1 of table 3). As expected, the operationalization of treatment effects as TOT-estimates
leads to inflated effect size estimates. Apparently, the delay between intervention and measurement of
outcomes does not seem to be systematically related to effect sizes in this estimation (cf. supplemental
appendix S3 for an alternative approach and investigation of heterogeneous treatment effects
depending on delay in measurement). In addition, estimates with large inverse standard errors are
associated with smaller effect sizes, indicating that larger and more precise studies report smaller effect
sizes overall. However, this coefficient is small in size and insignificant.
Intensity. Turning to the relationship between intensity per week and duration, column 1 of table 3
shows that intensity has a significant positive effect on treatment effects on financial behavior. Thus,
an increase of one hour of financial education per week leads to a 0.004 standard deviation unit increase
in the impact on financial behaviors studied. Considering that the average weekly duration is in this
subsample is roughly nine weeks and weekly intensity is only about four hours, doubling the weekly
intensity to eight hours while keeping everything else constant at the mean, would lead to an average
treatment effect around 14 percent higher than the empirical mean predicted treatment effect in this
fully specified model.
Target group. Among participant characteristics, age and gender are not significant explanatory
variables. However, the coefficient on “low income clients” is highly significant and negative,
indicating that these individuals are more difficult to educate. Regarding increasing mean years of
schooling at the country level, returns to additional financial education appear to diminish. This is in
line with results from two studies in very different contexts (Europe and India) that report higher
treatment effects for lower-educated individuals and diminishing returns to financial education upon
higher baseline levels of education (cf. Cole et al. 2011; Fort et al. 2016). Similarly, the coefficient for
baseline financial literacy in the population is also negative, albeit statistically insignificant. While
these results suggest declining marginal returns to financial education, the negative effect for low- and
lower-middle income economies—and also the above-mentioned coefficient on low-income clients—
shows a countervailing influence from challenging groups or country circumstances.
Characteristics of education. Regarding the channel variables, column 1 shows that no alternative
channel appears to be generally more or less effective than financial education in classroom settings
or informational nudges (omitted category). The same is true for the setting of the intervention where
school and workplace settings are not systematically different from other settings. However,
mandatory financial education and implementing financial education at a “teachable moment” appear
to be important. Specifically, we find, that making financial education mandatory decreases effect sizes
19
by 0.074 standard deviation units: The predicted value for effect size on financial behavior in
mandatory formats with everything else kept equal at the (empirical) mean would be only g = 0.030
(SE = 0.020, p = .134); thus, economically small and statistically insignificant from zero. In contrast,
offering financial education at a teachable moment increases effect sizes by 0.079 standard deviation
units. Thus, the predicted value for effect size on financial behavior would be ceteris paribus g = 0.124
(SE = 0.014, p = .000) (i.e., statistically highly significant), roughly 48 percent larger than the
unconditional average effect size found in the sampleand about 45 percent larger conditional on the
empirical means for all other covariates in this full model.
Parsimonious specification. We reduce the above discussed fully specified model by keeping the
variables on research design and intensity but otherwise eliminating the insignificant variables.
Column 2 of table 3 describes the resulting reduced model that confirms the fully specified regression
results from column 1. There are just some smaller changes in the estimated standard errors that occur
at a few variables. This indicates that it is justified to rely on the parsimonious specification, in
particular when we analyze subsamples with a much smaller number of observations in the following.
Meta-Regression Models for Subsamples
Given the large degree of heterogeneity across the 90 studies and their underlying financial education
programs, we move to an analysis of more homogenous subsamples.
RCTs only. Many will agree that RCTs fulfill the most rigorous requirements, implying that results
limited to this subsample of studies are indeed reliable. We do not prefer this procedure because many
observations are lost. Nevertheless, it is reassuring that results qualitatively hold, as shown in column
3 of table 3 for the subsample of 40 RCTs covering 227 effect sizes. However, while the negative
coefficient for mandatory courses remains to be large in magnitude and statistically (marginally)
significant, the coefficient for teachable moment loses explanatory power in this estimation.
Interventions in low and lower-middle income economies. This subsample covers 18 studies that
report 129 effect sizes (see column 4 of table 3). Again, all coefficients have the same sign and similar
magnitude as in our parsimonious specification (column 2 in table 3), but differences in standard errors
arise. While intensity of the intervention remains a strong predictor and low-income clients in low-
income economies also benefit significantly less from financial education, mandatory formats and
timing in the sense of offering financial education at a teachable moment appear less predictive of
treatment effects.
Interventions in upper-middle and high-income economies. Turning to the 72 studies that examine
financial education in more affluent economies (column 5 of table 3), we find that results again are
qualitatively very similar to the pooled analysis in column 2. Here, the opposing coefficients for
mandatory formats and offering financial education at a teachable moment are statistically significant
20
at the five percent -level, indicating that these effects may be primarily driven by interventions in
middle or high income economies.
Interventions for low-income individuals. Examining the subsample of 44 studies focusing on low-
income individuals results in a similar picture arising. Effects appear to be higher with increased
training intensity and offering financial education at a teachable moment. However, country-level
years of schooling and country income are now only marginally significant and insignificant
covariates, respectively. Additionally, the coefficient for mandatory courses still has the same sign and
similar magnitude, but is estimated with a larger standard error.
Disaggregating financial behaviors and financial behaviors by target group. As discussed in
section 4.2, it appears to be easier to affect financial behaviors in terms of (retirement-) savings and
budgeting compared to borrowing behavior. Thus, we disaggregate the sample into three categories of
financial behaviors and search for potentially heterogeneous effects of our main explanatory variables.
We reduce the choice of variables for some subsamples to avoid problems with degrees of freedom
due to relative few observations.
Column 1 of table 4 shows results for the subsample of 32 studies reporting effect sizes on
borrowing behavior. This result matches our main results of the aggregated sample of effect sizes
(column 2 of table 3) with significant positive effects from increased intensity, negative effects for
low-income target groups, and countries, negative effects from making financial education mandatory
and positive effects from offering financial education at a teachable moment. Column 2 of table 4
shows results for the subsample of 20 studies that focus on borrowing as the outcome and have low-
income clients as the target group. Again, results are nearly identical. However, the delay in
measurement is now a marginally significant predictor: effect sizes in this sample seem to diminish as
time between intervention and measurement of outcomes increases. Hence, treatment effects on debt
related behaviors among low-income individuals may be shorter-lived.
Turning to effect sizes reported in 67 studies on (retirement-) saving (column 3 of table 4), we
observe that the relevant variables from our benchmark model (column 2 of table 3) remain significant
predictors. However, voluntary versus mandatory formats seem to be unrelated to effectiveness.
Column 4 of table 4 shows the results on savings and retirement savings for low-income individuals
reported in 31 studies. Signs and magnitude are similar to the benchmark estimation, but the only
coefficients estimated with a small standard error are intensity per week and the teachable moment.
Thus, qualitative results hold, but effect sizes on saving behavior for low-income individuals may be
difficult to impact through the considered covariates.
21
Table 4. Explaining Heterogeneity in Effect Sizes for Subsamples by Type of Financial Behavior and
Target Group
Notes: Non-standardized coefficients from OLS regressions with clustered standard errors at the study-level in parentheses.
We only include right hand side variables where differential information from at least two studies is available in the
regressions. ***, ** and * denote significance at the one percent, five percent and ten percent level.
Source: Authors’ analysis based on data sources discussed in the text.
Turning to the subsample of 20 studies on budgeting and record keeping behavior (column 5 of
table 4), on which financial education yields the largest effects, we find that intensity is not
significantly related to effect size. Additionally, all of the other signs and relative magnitudes of the
coefficients remain similar to our benchmark estimation; however, with increased standard errors due
to only 20 studies and 40 observations. Completing this exercise, we now examine determinants of
treatment effects for the subsample of studies reporting on budgeting outcomes for low-income clients
(column 6 of table 4). There are 11 studies in this subsample reporting 27 estimates. Again, qualitative
results are similar and intensity now, again, is a marginally significant predictor of effect sizes on
budgeting behavior.
Overall, we find that the positive effects from increased intensity appear to be driven by
interventions focused on (retirement-) saving and borrowing behavior, whereas the timing and
voluntary participation matter, especially for borrowing behavior. Thus, the financial behavior that is
(1)
Borrow
(2)
Borrow
×
low inc.
clients
(3)
Save
(4)
Save
×
low inc.
clients
(5)
Budget
(6)
Budget
×
low inc.
clients
RCT -0.136*** -0.100*** -0.002 -0.035
(0.022) (0.026) (0.045) (0.058)
TOT 0.089** 0.106** 0.090 0.074
(0.033) (0.039) (0.054) (0.079)
Delay -0.000 -0.000* 0.000 -0.000 -0.001 -0.019
(0.000) (0.000) (0.000) (0.000) (0.002) (0.012)
1/SE 0.000 0.001** -0.000 0.000 -0.003* -0.007
(0.000) (0.000) (0.000) (0.000) (0.002) (0.005)
Intensity / week 0.003** 0.003** 0.003* 0.004** 0.037 0.595*
(0.001) (0.001) (0.002) (0.002) (0.031) (0.308)
Duration -0.000 -0.000 -0.001 0.000 -0.000 0.017
(0.000) (0.000) (0.001) (0.001) (0.003) (0.014)
Low income clients -0.043** -0.050**
(0.019) (0.022)
Years of schooling -0.023*** -0.023*** -0.018*** -0.011 -0.020* 0.017
(0.006) (0.008) (0.007) (0.011) (0.011) (0.022)
Low/lower-mid. inc. econ. -0.178*** -0.199*** -0.142*** -0.102
(0.052) (0.067) (0.045) (0.066)
Mandatory -0.069** -0.120*** -0.025 -0.010
(0.032) (0.039) (0.031) (0.049)
Teachable moment 0.100*** 0.087*** 0.084** 0.114*
(0.025) (0.026) (0.036) (0.065)
Constant 0.375*** 0.326** 0.305*** 0.147 0.361** -0.685
(0.087) (0.114) (0.091) (0.165) (0.134) (0.524)
R2 0.473 0.394 0.194 0.147 0.206 0.359
n (Studies) 32 20 67 31 20 11
n (Effect sizes) 100 73 166 91 40 27
22
hardest to impact (borrowing) needs special effort in the sense of increased intensity and timing the
financial education intervention at a teachable moment.
VI. ROBUSTNESS
The robustness tests cover eight different aspects and are reported in full in supplemental appendix S3.
All of them confirm our qualitative findings. Here, we just mention these tests: (i) testing the average
treatment effect with several alternative meta-regression models; (ii) repeating the parsimonious
benchmark model without imputing missing values; (iii) running this model for studies about the US
only; (iv) running this benchmark model with classroom studies only; (v) running this model with
equal weight per study by either calculating one synthetic effect size per study or weighting effect sizes
accordingly; (vi) running the benchmark specification with different empirical approaches; (vii)
analyzing the influence of delay on effects; and (viii) testing a different definition of training intensity.
Additionally, we further examine publication bias and possible heterogeneity in study quality in
supplemental appendix S4 and use alternative econometric techniques that account for publication
selection bias in supplemental appendix S3.
VII. CONCLUDING POLICY DISCUSSION
This meta-analysis covers studies that potentially contribute to realizing policy objectives, such as
improved financial literacy and changes in individual financial behavior. Due to this close link to
economic policy, we discuss insights that have potential policy relevance in three steps:
General policy lessons: (i) the most important policy lesson from our research is that financial
education can be effective. However, the field of financial education is not developed enough that
established standards could be followed “blindly,” rather the process of designing interventions needs
careful attention due to large heterogeneity across program types and individual studies;(ii)
interventions targeting improvements in financial literacy are quite successful as they achieve
effectiveness similar to comparable education interventions in other domains. As financial literacy
education basically aims at improving financial knowledge and awareness, it seems evidentiary that it
works well in the classroom and at school (see e.g. Bruhn et al. 2016).
Improved financial literacy also has an indirect positive effect on financial behavior, although this
indirect effect is small so that changes in financial behavior should also be addressed directly.
(iii) Education interventions targeting financial behavior have desired effects on average. Although
these effects are economically rather small, they are statistically robust. Impacts on financial behavior
23
are higher if the intensity of education is increased and if financial education is offered at a teachable
moment. The effects are smaller if “problematic” groups are addressed, such as low-income clients.
Policy lessons for subgroups. As the universe of studies covers widely diverse financial education
interventions, we draw three lessons for more homogeneous groups: (i) regarding the country groups,
education effects seem to be somewhat lower in low and lower middle-income countries. This is
probably due to the disadvantageous institutional circumstances in these countries. A relative
advantage in these countries, however, is that the general level of education (mean years of schooling
in the population) is comparatively low so that marginal returns to additional domain-specific
education are high. The lower opportunity costs of education may be a reason why mandatory
participation conditions, such as school based programs, are less problematic and offering financial
education at teachable moment appears to be of lesser importance in these countries;(ii) while
problematic target groups, such as low-income clients, are more difficult to educate in general, the
determinants of effective financial education are not different from the general population. If there is
a difference, it appears that a teachable moment is relatively important, indicating that there is a
particular need to get the attention of this target group; (iii) regarding the outcomes of financial
education, improving debt related behavior is, on average, hardly successful. At the same time,
mistakes can be rather consequential and the structure of many significant determinants is the same as
for other financial behaviors, such that the general lessons may translate to this specific case; however,
it needs much more input to reach economically significant results. Moreover, there is variation across
studies revealing clear success cases, which suggests that it is useful to go down to the study level and
learn from best practices. The effects on improving savings or budgeting behavior are much larger in
magnitude than on borrowing.
Research on open policy issues. In order to improve financial education policies in the future we
see three areas of urgent research: (i) we need quite generally more reliable evidence on the
effectiveness of financial education interventions. Almost two-thirds of the evidence comes from the
US, indicating that there are large gaps of evaluation elsewhere;(ii) regarding the documentation of
impact evaluations within published reports, it would be very desirable to provide more information
about study and program characteristics (see Miller et al. 2015). A straight-forward example is the
quality of teacher training or implementation, which can make a crucial difference but is unknown in
almost all studies (Brown et al. 2016). The same applies to the ways in which the curriculum is
structured and implemented (see Drexler et al. 2014 as a notable exception); (iii) finally, in order to
come closer to welfare assessments, information in two directions is needed: first, information about
program costs is frequently missing. Thus, in terms of welfare, positive education effects could be
balanced with the true costs of the intervention (see also Lusardi et al. 2016). Second, the discussion
of effectiveness of financial education policy should also consider principal alternatives to financial
24
education in general. Such alternatives include limiting the kind of available products (choices),
altering the choice architecture (e.g., Carroll et al. 2009), working with nudges (e.g., Thaler and
Benartzi 2004; Willis 2011), considering the promotion of commitment devices (e.g., Brune et al.
2016), offering incentives (e.g., Saez 2009), or implementing more rigid consumer financial protection
policies (cf. Campbell et al. 2011).
There are two arguments in favor of implementing financial education. First, the small average
effect comes with low average intensity. More than 70 percent of our considered studies invest no
more than two days in education, indicating that these measures may have only small effects, but also
low costs. Second, the average small effect of financial education is accompanied by large
heterogeneity, indicating that those offering financial education measures can still learn from best
practice experiences, a development that is ongoing as evidenced by time trend of slowly increasing
effectiveness documented in rigorous impact evaluation studies.
25
REFERENCES
Agarwal, S., and B. Mazumder. 2013. “Cognitive Abilities and Household Financial Decision
Making.” American Economic Journal: Applied Economics 5 (1): 193–207.
[CrossRef][10.1257/app.5.1.193]
Banerjee, A., E. Duflo, N. Goldberg, D. Karlan, R. Osei, W. Parienté, J. Shapiro, B. Thuysbaert, and
C. Udry. 2015. “A Multifaceted Program Causes Lasting Progress for the Very Poor: Evidence
from Six Countries.” Science 348 (6236): 1260799.
Behrman, J.R., O.S. Mitchell, C.K. Soo, and D. Bravo. 2012. “How Financial Literacy Affects
Household Wealth Accumulation.” American Economic Review: Papers and Proceedings 102
(3): 300–4.
Berg, G., and B. Zia. 2013. “Harnessing Emotional Connections to Improve Financial Decisions.
Evaluating the Impact of Financial Education in Mainstream Media.” World Bank Policy
Research Working Paper 6407.
Blue, L., P. Grootenboer, and M. Brimble. 2014. “Financial Literacy Education in the Curriculum:
Making the Grade or Missing the Mark?” International Review of Economics Education 16 (Part
A): 51–62.
Brown, M., J. Grigsby, W. van der Klaauw, J. Wen, and B. Zafar. 2016. “Financial Education and
the Debt Behavior of the Young.” Review of Financial Studies 29 (9): 2490–522.
Brune, L., X. Giné, J. Goldberg, and D. Yang. 2016. “Facilitating Savings for Agriculture: Field
Experimental Evidence from Malawi.” Economic Development and Cultural Change 64: 187–
220.
Bruhn, M., L. de Souza Leao, A. Legovini, R. Marchetti, and B. Zia. 2016. “The Impact of High
School Financial Education: Evidence from a Large-Scale Evaluation in Brazil.” American
Economic Journal: Applied Economics 8 (4): 256–95.
Bruhn, M., G.L. Ibarra, and D. McKenzie. 2014. “The Minimal Impact of a Large-Scale Financial
Education Program in Mexico City.” Journal of Development Economics 108: 184–9.
Bucher-Koenen, T., and M. Ziegelmeyer. 2014. “Once Burned, Twice Shy? Financial Literacy and
Wealth Losses During the Financial Crisis.” Review of Finance 18 (6): 2215–46.
Campbell, J.Y. 2006. “Household Finance.” Journal of Finance 61 (4): 1553–604.
Campbell, J.Y., H.E. Jackson, B.C. Madrian, and P. Tufano. 2011. “Consumer Financial Protection.”
Journal of Economic Perspectives 25 (1): 91–114.
Card, D., J. Kluve, and A. Weber. 2010. “Active Labour Market Policy Evaluations: A Meta-
Analysis.” Economic Journal 120 (548): 452–77.
Card, D., J. Kluve, and A. Weber. 2015. “What Works? A Meta-Analysis of Recent Active Labor
Market Program Evaluations.” NBER Working Paper 21431.
Carpena, F., S. Cole, J. Shapiro, and B. Zia. 2015. “The ABCs of Financial Education. Experimental
Evidence on Attitudes, Behavior, and Cognitive Biases.” World Bank Policy Research Working
Paper 7413.
Carroll, G.D., J.J. Choi, D. Laibson, B.C. Madrian, and A. Metrick. 2009. “Optimal Defaults and
Active Decisions.” Quarterly Journal of Economics 124 (4): 1639–74.
Choi, J.J., D. Laibson, and B.C. Madrian. 2010. “Why Does the Law of One Price Fail? An
Experiment on Index Mutual Funds.” Review of Financial Studies 23 (4): 1405–32.
———. 2011. “$100 Bills on the Sidewalk: Suboptimal Investment in 401(k) Plans.” Review of
Economics and Statistics 93 (3): 748–63.
Cohen, J. (1977). “Statistical power analysis for the behavioral sciences” (Rev. Ed.). Hillsdale, NJ:
Lawrence Erlbaum Associates.
Cole, S., T. Sampson, and B. Zia. 2011. “Prices or Knowledge? What Drives Demand for Financial
Services in Emerging Markets?” Journal of Finance 66 (6): 1933–67.
Collins, J.M. 2013. “The Impacts of Mandatory Financial Education: Evidence from a Randomized
Field Study.” Journal of Economic Behavior and Organization 95: 146–58.
Collins, J.M., and C.M. O’Rourke. 2010. “Financial Education and Counseling—Still Holding
Promise.” Journal of Consumer Affairs 44 (3): 483–98.
26
Doi, Y., D. McKenzie, and B. Zia. 2014. “Who You Train Matters: Identifying Combined Effects of
Financial Education on Migrant Households.” Journal of Development Economics 109: 39–55.
Drexler, A., G. Fischer, and A. Schoar. 2014. “Keeping It Simple: Financial Literacy and Rules of
Thumb.” American Economic Journal: Applied Economics 6 (2): 1–31.
Duflo, E., and E. Saez. 2003. “The Role of Information and Social Interactions in Retirement Plan
Decisions: Evidence from a Randomized Experiment.” Quarterly Journal of Economics 118 (3):
815–42.
Fernandes, D., J.G. Lynch Jr, and R.G. Netemeyer. 2014. “Financial Literacy, Financial Education,
and Downstream Financial Behaviors.” Management Science 60 (8): 1861–83.
Freeman, S., S.L. Eddy, M. McDonough, M.K. Smith, N. Okoroafor, H. Jordt, and M.P. Wenderoth.
2014. “Active Learning Increases Student Performance in Science, Engineering, and
Mathematics.” Proceedings of the National Academy of Sciences 111 (23): 8410–5.
Fort, M., F. Manaresi, and S. Trucchi. 2016. “Adult Financial Literacy and Households’ Financial
Assets: The Role of Bank Information Policies.” Economic Policy 31 (88): 743–82.
Fox, J., S. Bartholomae, and J. Lee. 2005. “Building the Case for Financial Education.” Journal of
Consumer Affairs 39 (1): 195–214.
Gathergood, J. 2012. “Self-Control, Financial Literacy and Consumer Over-Indebtedness.” Journal
of Economic Psychology 33 (3): 590–602.
Gerardi, K., L. Goette, and S. Meier. 2013. “Numerical Ability Predicts Mortgage Default.”
Proceedings of the National Academy of Sciences 110 (28): 11267–71.
Gertler, P.J., S. Martinez, P. Premand, L.B. Rawlings, and C.M. Vermeersch. 2011. Impact
Evaluation in Practice. Washington DC: World Bank Publications.
Gibson, J., D. McKenzie, and B. Zia. 2014. “The Impact of Financial Literacy Training for
Migrants.” World Bank Economic Review 28 (1): 130–61.
Grohmann, A., R. Kouwenberg, and L. Menkhoff. 2015. “Childhood Roots of Financial Literacy.”
Journal of Economic Psychology 51: 114–33.
Hastings, J.S., B.C. Madrian, and W.L. Skimmyhorn. 2013. “Financial Literacy, Financial
Education, and Economic Outcomes.” Annual Review of Economics 5: 347–73.
Imbens, G.W., and J.M. Wooldridge. 2009. “Recent Developments in the Econometrics of Program
Evaluation.” Journal of Economic Literature 47 (1): 5–86.
Jappelli, T. 2010. “Economic Literacy: An International Comparison.” Economic Journal 120 (548):
F429–51.
Klapper, L., A. Lusardi, and P. van Oudheusden. 2015. “Financial Literacy Around the World:
Insights from the Standard and Poor’s Rating Services Global Financial Literacy Survey.”
http://gflec.org/initiatives/sp-global-finlit-survey/ ; last checked 07 August 2017.
Lipsey, M.W., and D.B. Wilson. 2001. Practical Meta-Analysis. Thousand Oaks, CA: Sage.
Annamaria Lusardi, Pierre-Carl Michaud, and Olivia S. Mitchell, "Optimal Financial Knowledge and
Wealth Inequality", Journal of Political Economy 125, no. 2 (April 2017): 431-477.
———. 2014. “The Economic Importance of Financial Literacy: Theory and Evidence.” Journal of
Economic Literature 52 (1): 5–44.
Lusardi, A., P.C. Michaud, and O.S. Mitchell. 2016. “Optimal Financial Knowledge and Wealth
Inequality.” Journal of Political Economy.
Miller, M., J. Reichelstein, C. Salas, and B. Zia. 2015. “Can You Help Someone Become Financially
Capable? A Meta-Analysis of the Literature.” World Bank Research Observer 30 (2): 220–46.
Muller, S.M. 2015. “Causal Interaction and External Validity: Obstacles to the Policy Relevance of
Randomized Evaluations.” World Bank Economic Review 29: S217–25.
Portnoy, D.B., L.A. Scott-Sheldon, B.T. Johnson, and M.P. Carey. 2008. “Computer-Delivered
Interventions for Health Promotion and Behavioral Risk Reduction: A Meta-Analysis of 75
Randomized Controlled Trials, 1988–2007.” Preventive Medicine 47 (1): 3–16.
Pritchett, L., and J. Sandefur. 2015. “Learning from Experiments When Context Matters.” American
Economic Review: Papers and Proceedings 105 (5): 471–5.
27
Saez, E. 2009. “Details Matter: The Impact of Presentation and Information on the Take-Up of
Financial Incentives for Retirement Saving.” American Economic Journal: Economic Policy 1
(1): 204–28.
Sayinzoga, A., E.H. Bulte, and R. Lensink. 2016. “Financial Literacy and Financial Behaviour:
Experimental Evidence from Rural Rwanda.” Economic Journal 126 (594): 1571–99.
Skimmyhorn, W. 2016. “Assessing Financial Education: Evidence from Boot Camp.” American
Economic Journal: Economic Policy 8 (2): 322–43.
Stango, V., and J. Zinman. 2009. “Exponential Growth Bias and Household Finance.” Journal of
Finance 64 (6): 2807–49.
Stanley, T.D. 2001. “Wheat from Chaff: Meta-Analysis As Quantitative Literature Review.” Journal
of Economic Perspectives 15 (3): 131–50.
Stanley, T.D., and H. Doucouliagos. 2012. Meta-Regression Analysis in Economics and Business.
New York: Routledge.
Tennyson, S., and C. Nguyen. 2001. “State Curriculum Mandates and Student Knowledge of
Personal Finance.” Journal of Consumer Affairs 35 (2): 241–62.
Thaler, R.H., and S. Benartzi. 2004. “Save More Tomorrow: Using Behavioral Economics to
Increase Employee Saving.” Journal of Political Economy 112: 164–87.
van Rooij, M., A. Lusardi, and R. Alessie. 2011. “Financial Literacy and Stock Market
Participation.” Journal of Financial Economics 101 (2): 449–72.
———. 2012. “Financial Literacy, Retirement Planning and Household Wealth.” Economic Journal
122 (560): 449–78.
Vivalt, E. 2015. “Heterogeneous Treatment Effects in Impact Evaluation.” American Economic
Review: Papers and Proceedings 105 (5): 467–70.
von Gaudecker, H.-M. 2015. “How Does Household Portfolio Diversification Vary with Financial
Literacy and Financial Advice?” Journal of Finance 70 (2): 489–507.
Willis, L.E. 2011. “The Financial Education Fallacy.” American Economic Review: Papers and
Proceedings 101 (3): 429–34.
Xu, L., and B. Zia. 2012. “Financial Literacy Around the World: An Overview of the Evidence with
Practical Suggestions for the Way Forward.” World Bank Policy Research Working Paper 6107.
Zinman, J. 2015. “Household Debt: Facts, Puzzles, Theories, and Policies.” Annual Review of
Economics 7: 251–76.
28
Appendix
to accompany
“Does financial education impact financial literacy and financial behavior,
and if so, when?”
Appendix S1: Supplementary material
Appendix S2: Comparison of our dataset and results to previous meta-analyses
Appendix S3: Robustness checks
Appendix S4: Publication bias and heterogeneity of study quality
Appendix S5: Overview of studies included in the statistical meta-analysis
Appendix S6: References for studies included in the statistical meta-analysis
29
Appendix S1: Supplementary material
This Appendix S1 contains two kinds of information: First, there are three tables (Table
S1.1 to Table S1.3) and two figures (Figure S1.1 and Figure S1.2), which are referred to in the
main text, mainly in the earlier sections.
Second, there is a longer documentation about “Additional information on selection of
studies and extraction of effect size estimates and study descriptors.” This documentation
provides deeper information that complements Section 3.1 (Selection of studies) and Section
3.2 (Extraction of effect size estimates and study descriptors) of the main text.
30
Table S1.1: Summary of financial education studies by publication date and country
Number of studies Percent of sample
(1) (2)
A By publication date
1999 2 1.59
2000 0 0.00
2001 5 3.97
2002 1 0.79
2003 4 3.17
2004 3 2.38
2005 6 4.76
2006 5 3.97
2007 6 4.76
2008 6 4.76
2009 8 6.35
2010 10 7.94
2011 7 5.56
2012 15 11.9
2013 9 7.14
2014 11 8.73
2015 15 11.9
2016 13 10.32
B By country of intervention Income
Australia 2 1.59 High
Bosnia and Herzegovina 1 0.79 Upper-middle
Brazil 1 0.79 Upper-middle
China 1 0.79 Upper-middle
Dominican Republic 1 0.79 Upper-middle
Germany 1 0.79 High
Ghana 1 0.79 Lower-middle
Hong Kong, China 1 0.79 High
India 8 6.35 Lower-middle
Indonesia 2 1.59 Lower-middle
Italy 7 5.56 High
Kenya 1 0.79 Lower-middle
Malawi 1 0.79 Low
Mexico 1 0.79 Upper-middle
Mozambique 1 0.79 Low
New Zealand 2 1.59 High
Pakistan 1 0.79 Lower-middle
Qatar 1 0.79 High
Rwanda 1 0.79 Low
Singapore 1 0.79 High
South Africa 1 0.79 Upper-middle
Spain 1 0.79 High
Sri Lanka 1 0.79 Lower-middle
Tanzania 2 1.59 Low
USA 83 65.87 High
Uganda 2 1.59 Low
Low inc. econ. 7 5.5
Lower-middle inc. econ. 14 11.11
Upper-middle inc. econ. 6 4.76
High inc. econ. 99 78.57
Total 126 100
Notes: Country group classifications refer to 2015 World Bank data on GNI per capita (Atlas method).
31
Table S1.2: Overview of coded outcomes and definitions
Outcome category Definition Freq.
Financial literacy (190 estimates)
A Financial knowledge (+) Raw score on financial knowledge test 190
Indicator of scoring above a defined threshold (100%)
Indicator of solving an item correctly
Financial behaviors (349 estimates)
B Borrowing & debt management behavior 100
(28.65%)
1) Reduction of loan default
within a certain time-frame
(+)
2) Reduction of delinquencies
within certain time frame (+)
3) Better credit score (+)
Binary indicator
Binary indicator
Continuous measure of credit score
4) Reduction in informal
borrowings (+)
5) Lower cost of credit / interest
rate (+)
Binary indicator of informal loan or reduction
in number of informal loans
Sum of real interest amount or interest rate
and (if applicable) cost of fees
6) Any debt (-) / (+) (depending
on intervention goal)
7) Any formal loan (+)
8) Total amount borrowed (-) /
(+) (depending on intervention
goal)
Binary indicator
Binary indicator
Continuous measure of borrowed amount
9) Total outstanding debt (-) / (+)
(depending on intervention
goal)
Continuous measure of total debt
10) Better borrowing index (+) Study-specific index of survey items to
measure borrowing amount, frequency, and
repayment
11) Uses credit card up to limit (-)
Binary indicator
C Budgeting & planning behavior 40
(11.46%)
1) Having a written budget (+) Binary indicator
2) Positive sentiment toward
budgeting (+)
Binary indicator
3) Having a financial plan (+) Binary indicator
4) Keeping separate records for
business and household (+)
Binary indicator
5) Seeking information before
making financial decisions (+)
Binary indicator
6) Self-rating of adherence to
budget (+)
Study-specific scale
D Saving & retirement saving behavior 166
(47.56%)
1) Total savings held (+)
2) Savings rate or savings within
timeframe (+)
3) Savings index (+)
4) Any savings (+)
Continuous measure of savings amount or
categorical variable indicating amount within
range
Savings relative to income
Amount over defined time-frame
Study-specific index of survey items designed
to measure savings amount and frequency
Binary indicator
-continued-
32
5) Has formal bank (savings)
account (+)
Binary indicator
6) Investments into own or other
business (stocks) (+)
7) Holds any stocks or bonds (+)
Continuous measure of amount invested
Binary indicator
8) Has any retirement savings (+)
9) Participates in retirement
savings plan (e.g. 401k) (+)
10) Amount of retirement savings
(+)
Binary indicator
Binary indicator
Continuous measure of retirement savings
amount
11) Retirement savings rate (+)
12) Positive sentiment towards
investing funds (+)
Retirement savings relative to income
Binary indicator
13) Reduction of excess risk in
retirement fund (+)
Continuous measure of retirement savings
amount allocated to risky assets
14) Reduction of cost of savings
product (fees paid) (+)
Continuous measure of fee amount paid
15) Increase in contribution rate to
retirement savings plan (+)
Indicator of increase or continuous measure of
amount increase
16) Net wealth (+)
Continuous measure of net wealth
E Insurance & risk mitigation behavior 16
(4.59%)
1) Any formal insurance (+)
2) Having a diversified portfolio
(+)
Binary indicator
Numbers of assets in portfolio; Standard
deviation of returns in portfolio
F Remittance behavior 16
(4.59%)
1) Lower cost of remittance
product (+)
2) Lower remittance frequency
and higher amount (lower
cost) (+)
3) More control over remitted
funds (+)
Continuous measure of cost or binary choice
of lower cost product
Measure of remittance frequency within
timeframe and continuous amount remitted
Study-specific scale to measure control over
remitted amount
G Bank account behavior 11
(3.15%)
1) Has formal bank (checking)
account (+)
2) Opens formal account within
certain time frame
3) Uses formal bank account
Binary indicator
Binary indicator
Binary indicator
Notes: When necessary, outcomes are reverse-coded so that positive signs reflect positive financial education
treatment effects (i.e. when the dependent variable is coded as the probability of default, we transform this to the
reduction in probability of default in order to be able to assign a positive sign).
33
Table S1.3: Summary of estimated financial education impacts
Outcome Significance at 5% Significance at 10% Average
effect
size
Negative Insig. Positive Negative Insig. Positive (SE)
A Effects on financial literacy
Fin.
literacy
1
(0.53%)
72
(37.89%)
117
(61.58%)
2
(1.05%)
62
(32.63%)
126
(66.32)
0.263***
(0.414)
B Effects on financial behavior
Fin.
behavior
8
(2.29%)
215
(61.60%)
126
(36.10%)
18
(5.16%)
181
(51.86%)
150
(42.98%)
0.086***
(0.012)
Borrowing 5 80 15 10 70 20 0.023
(5.00%) (80.00%) (15.00%) (10.00%) (70.00%) (20.00%) (0.014)
Budgeting
& planning
0
(0.00%)
15
(37.5%)
25
(62.50%)
1
(2.50%)
10
(25.00%)
29
(72.50%)
0.207***
(0.053)
Saving 2
(1.67%)
61
(50.83%)
57
(47.50%)
6
(5.00%)
49
(40.83%)
65
(54.17%)
0.108***
(0.017)
Retirement
Saving
0
(0.00%)
22
(47.83%)
24
(52.17%)
0
(0.00%)
17
(36.96%)
29
(63.04%)
0.108***
(0.034)
Insurance 0 13 3 0 12 4 0.045
(0.00%) (81.25%) (18.75%) (0.00%) (75.00%) (25.00%) (0.024)
Bank
account
behavior
0
(0.00%)
10
(90.91%)
1
(9.09%)
0
(0.00%)
10
(90.91%)
1
(9.09%)
0.003
(0.027)
Remittance
behavior
1
(6.25%)
14
(87.50%)
1
(6.25%)
1
(6.25%)
13
(81.25%)
2
(12.50%)
0.035
(0.046)
Notes: Average effect sizes are estimated via OLS with standard errors clustered at the study-level in
parentheses. ***, ** and * denote significance at the 1%, 5% and 10% level.
34
Figure S1.1: Citations of published items with the keyword financial literacy per year, source: SSCI
Figure S1.2: Kernel-density estimates of effect sizes by outcome (for Hedge’s g<1)
35
Additional information on selection of studies and extraction of effect sizes estimates
and study descriptors.
Selection of studies. We follow the established meta-analytical protocol (cf. Lipsey and
Wilson 2001, p.23; Stanley 2001, p.143; Stanley and Doucouliagos 2012; Stanley et al. 2013).
This starts with systematically searching the relevant databases for the most common keywords
in order to aggregate a large sample of potentially eligible studies to be included in our meta-
analysis. Keywords are (i) financial literacy; (ii) financial knowledge; (iii) financial education;
(iv) financial capability; and (v) combinations of these keywords with “intervention.” To
minimize publication bias and capture the broadest sample of studies possible, we
systematically search not only the relevant databases for published records (e.g. ISI, Business
Source Premier via EBSCO Host, JStor) but also for registered trials, working papers, and
informal research reports (e.g. AEA RCT-registry, SSRN, Fin. Lit. E-Journal, RePEC, NBER,
Worldbank eLibrary). All records from recent systematic accounts of the literature (Fernandes
et al. 2014; Miller et al. 2015) are included in our initial pool of studies. In addition, we screen
the references of narrative literature reviews (Fox et al. 2005; Collins and O’Rourke 2010;
Willis 2011; Xu and Zia 2012; Hastings et al. 2013; Blue et al. 2014; Lusardi and Mitchell
2014).
This search resulted in over 500 potentially relevant published journal-articles and over
600 results from working paper databases with some apparent overlap. We stopped collecting
articles from these databases in October 2016.
From this collection, we drop studies that do not meet our three criteria of inclusion: (i)
Reporting on impacts of an exogenous educational intervention designed to strengthen the
participants’ financial literacy and/or leading to behavioral change in the area of personal
finance; (ii) providing a quantitative assessment of intervention impact that allows coding an
effect size statistic () and its standard error; and (iii) relying on an observed counterfactual in
36
the estimation of intervention impacts. Consequently, we only include experimental studies
with sufficient information on intervention outcomes in our analysis, i.e. RCTs, quasi-
experiments, and natural experiments (see below for coding of studies). Where necessary
information was partially missing, we consulted additional online resources related to the article
or contacted the authors of the primary studies via e-mail.
This selection-process results in a final sample of 126 independent intervention studies
that report 539 effect sizes. Of these, 90 studies report 349 effect sizes on financial behavior,
and 67 studies report 190 effect sizes on financial literacy. Among these 90 plus 67 studies,
there are 31 studies reporting effect sizes for both financial literacy and behavior. Our selection
of studies covers 126 independent interventions from 1999 through 2016. Table S1.1 shows the
composition of our sample of studies by the date of publication (Panel A) and the country in
which the intervention took place (Panel B). While most interventions took place in the U.S.
and other OECD countries, 21.4% of studies were conducted in low- or middle-income
countries. The sample is comprised of 51 RCTs and 75 quasi-experiments. RCTs are rare in the
early years of the literature, but the share has risen dramatically, with the majority of studies
conducted from 2011 onward being randomized evaluations (see Figure 1 in the main text).
Extraction of estimates. The next step in our meta-analytic process is to extract effect
size estimates from the statistical data reported in the primary studies. Our analysis aggregates
treatment effects of financial education interventions on two main categories of outcomes. First,
we code the effect of financial education on financial literacy (i.e. a measure of performance
on a financial knowledge test) since knowledge development is the primary goal of financial
education (Hastings et al. 2013; Lusardi and Mitchell 2014). We do not include self-
assessments of changes in financial knowledge as an outcome.
Second, we code treatment effects of financial education on financial behaviors. These
behaviors can be further disaggregated into the following categories: Borrowing, savings and
37
retirement saving, budgeting and planning, insurance, as well as remittances. Table S1.2
provides an overview of the categories and definitions of effect size estimates by outcome type.
We code all available effect sizes per study on cognitive (financial knowledge) and
behavioral outcomes. We include multiple estimates per study if multiple outcomes, time-
points, or treatments are reported. We only extract main (average treatment) effects reported in
the papers. Thus, we do not code estimates reported in the “heterogeneity-of-treatment-effects-
section” within papers, such as sample splits or interaction-effects of binary indicators (e.g.
gender, income, ability, …) with the treatment indicators. If results are only reported in a
disaggregated manner (only effects on subsamples), we perform a within study (random-
effects) meta-analysis (DerSimonian and Laird 1986) to generate an inverse-variance-weighted
average effect size to proxy the main effect. Additionally, we aim to capture only non-redundant
effect sizes per paper (i.e. we do not include effect sizes for the same intervention on the same
outcome reported in the robustness-section). The number of coded estimates per study ranges
from 1 to 87 estimates. We show in the Appendix S3 (robustness checks) that giving each study
equal weight by creating a single synthetic effect size per study through a within-study meta-
analysis or alternatively by weighting each observation by the inverse number of effect-size
estimates contributed by each study yields similar results.
In addition to the coding of all possible estimates of effect sizes () and their standard
errors of financial education treatment on financial literacy or financial behavior (cf. Section
2), we develop a coding protocol to extract potentially relevant information about the study
(study descriptors) that may serve as predictor variables explaining the variability in effect
sizes. Specifically, we aim at extracting data on (i) research design and measurement of
dependent variables; (ii) the intensity of education; (iii) the sample/target group of the
intervention; and (iv) the details of the intervention itself, such as channel, setting, and
participation conditions. Coding of the included study reports was completed by the authors of
38
this paper and two research assistants who were trained using the guidelines by Lipsey and
Wilson (2001, p.88). Overall intercoder reliability is high and data collection for most of the
variables concerning the setting, participants, and research design of the primary studies was
straightforward. However, key details of the underlying educational intervention are often
missing or underreported in the research reports. If information is only partially missing authors
were asked to provide these details via e-mail.
References in Appendix S1
Blue, L., Grootenboer, P., and Brimble, M. (2014). Financial literacy education in the
curriculum: Making the grade or missing the mark? International Review of Economics
Education, 16, Part A(0): 51–62.
Collins, J.M. and O’Rourke, C.M. (2010). Financial education and counseling—still holding
promise. Journal of Consumer Affairs, 44(3): 483–498.
DerSimonian, R. and Laird, N. (1986). Meta-analysis in clinical trials. Controlled Clinical
Trials, 7(3): 177–188.
Fox, J., Bartholomae, S., and Lee, J. (2005). Building the case for financial education.
Journal of Consumer Affairs, 39(1): 195–214.
Fernandes, D., Lynch, Jr., J.G., and Netemeyer, R.G. (2014). Financial literacy, financial
education, and downstream financial behaviors. Management Science, 60(8): 1861–1883.
Hastings, J.S., Madrian, B.C., and Skimmyhorn, W.L. (2013). Financial literacy, financial
education, and economic outcomes. Annual Review of Economics, 5: 347–373.
Lusardi, A. and Mitchell, O.S. (2014). The economic importance of financial literacy: theory
and evidence. Journal of Economic Literature, 52(1): 5–44. Lipsey, M.W. and Wilson, D.B.
(2001). Practical meta-analysis. Sage, Thousand Oaks, CA.
Miller, M., Reichelstein, J., Salas, C., and Zia, B. (2015). Can you help someone become
financially capable? A meta-analysis of the literature. World Bank Research Observer, 30(2):
220–246.
Stanley, T. D. (2001). Wheat from chaff: Meta-analysis as quantitative literature review.
Journal of Economic Perspectives, 15(3): 131–150.
Stanley, T. D. and Doucouliagos, H. (2012). Meta-regression analysis in economics and
business, Routledge, New York, NY.
Stanley, T., Doucouliagos, H., Giles, M., Heckemeyer, J. H., Johnston, R. J., Laroche, P.,
Nelson, J. P., Paldam, M., Poot, J., Pugh, G., Rosenberger, R. S., and Rost, K. (2013). Meta-
analysis of economics research reporting guidelines. Journal of Economic Surveys, 27(2):390–
394.
39
Willis, L.E. (2011). The financial education fallacy. American Economic Review: Papers and
Proceedings, 101(3): 429–434.
Xu, L., and Zia, B. (2012). Financial literacy around the world: an overview of the evidence
with practical suggestions for the way forward. World Bank Policy Research Working Paper
6107.
40
Appendix S2: Comparison of our dataset and results to previous meta-
analyses
There are two earlier meta-analyses about financial education: The study by Miller et al.
(2015) covers 19 papers due to its extremely restrictive selection criteria. Thus, most similar to
our work is the study by Fernandes et al. (2014), which covers 90 effect sizes from financial
education reported in 77 papers. Despite an overlap of 44% with their sample of studies, our
research differs in four ways which explains our new results: (i) most important is that we
analyze determinants of program effectiveness in a broader way by applying respective coding.
(ii) Then we code the various outcomes per study and their respective effectiveness. Moreover,
(iii) we cover recent and mostly randomized experiments providing evidence of effective
interventions; and (iv) we cover additional studies focusing exclusively on financial literacy as
the outcome variable. We aim to elaborate on these comparisons in this part of the Appendix.
Comparison of studied samples. Our selection-process (see Appendix S1) led us to a
final sample of 126 independent intervention studies that report 539 effect sizes. Of these, 90
studies report 349 effect sizes on financial behavior, and 67 studies report 190 effect sizes on
financial literacy. Among these 90 plus 67 studies, there are 31 studies reporting effect sizes on
both financial literacy and behavior. The sample is comprised of 51 RCTs and 75 quasi-
experiments.
As mentioned, Miller et al. (2015) select 19 intervention-studies for their statistical meta-
analysis. Their main inclusion criterion is that interventions report on identical outcomes. This
limits their analysis to sample sizes of four to six studies (and estimates) per outcome. While
informative of magnitude and significance of effect sizes on identical outcomes, such an
approach prevents a detailed investigation into the sources of heterogeneity, given the very
limited number of studies available. However, we note that the results for size, direction, and
significance of the main behaviors studied in Miller et al. 2015 are in line with our results (see
Figure 2 in the main text).
41
Fernandes et al. (2014), with 77 papers selected, cover 90 effect sizes (15 RCTs and 75
quasi-experiments) of “manipulated financial literacy” (cf. Fernandes et al. 2014, p.1863). Of
their 77 papers, 55 are also part of our sample. We exclude 22 single-group pre-posttest and
quasi-experimental papers because they either do not analyze education interventions (but other
personal finance related programs, e.g. match incentives), report only aggregate measures of
self-reported financial behavior, wellbeing or self-efficacy, or because it is not feasible to
calculate a meaningful effect size statistic. In addition, we include 35 recent studies that were
not previously available. Moreover, we consider another 36 studies examining the impact of
financial education on financial literacy but neglecting possible impacts on financial behavior.
These differences explain the mentioned overlap of 44% regarding studies.
Comparison of estimation results. We estimate the average treatment effect of
educational interventions on financial behaviors to be statistically highly significant (g=0.086,
p=0.000, n=349). Although the average treatment effect of 0.086 is small in magnitude, there
exists a measurable and robust impact of financial education on various kinds of financial
behavior. In comparison, Fernandes et al. (2014) estimate the summary effect of financial
education on financial behavior to be roughly g=0.066. However, the authors use averaged
effect sizes per paper and weight each observation with its average inverse variance. In order
to obtain a better comparison with that study, we exactly apply their method (random effects
meta-regression) with synthetic effect sizes per study to our sample of studies. This provides
an average (weighted) effect size of g=0.079 (p=0.000, n=90) (see Table S3.1 in Appendix S3).
Thus, our estimate of a summary effect for the literature is not too different from theirs.
To investigate the potential source of this difference, we estimate the weighted average
effect size among those recent studies that are not included in Fernandes et al. (2014). Indeed,
we find that there is a larger average effect of financial education on financial behavior in this
sample (g=0.13). This indicates that the new studies covered in our meta-analysis are the main
42
source of difference. Diving deeper into this issue, we find that Fernandes et al. (2014) estimate
extremely small average effect sizes for their sample of 15 RCTs. Our broader sample of
randomized experiments, however, leads to a much more positive assessment. In line with this
observation, the effect size of financial education on financial behavior documented in RCTs
seems to increase over time, indicating a positive time trend in effect sizes: a regression of
effect size on year of study publication results in a statistically highly significant coefficient
(b=0.014, SE=0.004). This moderate, positive time-trend is an important element in explaining
our positive result about the effect of financial education on financial behavior.
Turning to the result concerning the treatment effect of financial education on financial
literacy (measured through knowledge assessments), we estimate the average impact of
financial education on financial literacy to be g=0.263 (p=0.000, n=190). Thus, our analysis of
a comprehensive sample of studies (n=67) leads to a positive assessment of the effectiveness
of financial education on financial literacy. This education explains 1.7% of the variance in
financial knowledge and, thus, appears only slightly less effective than educational
interventions in other domains, such as math and science instruction (cf. Fernandes et al. 2014,
p.1867). Our positive result is in remarkable contrast to Fernandes et al. (2014, p.1867), who
find that financial education only explains 0.4% of the variance in financial literacy and state
accordingly that, “financial education yields surprisingly weak changes in financial knowledge
presumed to cause financial behavior.” However, this result seems a bit fragile as it is based on
only 12 studies and cannot, obviously, be replicated in our larger sample of studies (cf.
Fernandes et al. 2014, p. 1867).
References in Appendix S2
Fernandes, D., Lynch, Jr., J.G., and Netemeyer, R.G. (2014). Financial literacy, financial
education, and downstream financial behaviors. Management Science, 60(8): 1861–1883.
Miller, M., Reichelstein, J., Salas, C., and Zia, B. (2015). Can you help someone become
financially capable? A meta-analysis of the literature. World Bank Research Observer, 30(2):
220–246.
43
Appendix S3: Robustness checks
Appendix S3 contains eight kinds of robustness checks: (i) we estimate the (weighted)
average treatment effect of financial education on financial behavior using five alternative
meta-regression models for continuous effect sizes; (ii) we show results without imputing
missing values; (iii) we run our benchmark analysis with the subsample of studies conducted
in the USA only; (iv) we run our benchmark analysis with the subsample of classroom financial
education studies only; (v) we give each study the same weight in the analysis by creating one
synthetic effect size per study or, alternatively, assigning a weight of the inverse number of
observations contributed by each study to each estimate within a given study; (vi) we re-
estimate our multivariate analysis using eleven alternative meta-regression models; (vii) we
look for heterogeneous impacts depending on the delay in measurement of outcomes; and,
lastly, (viii) we test a different operationalization of training intensity.
(i) Summary of treatment effects on financial behavior under various models. Table
S3.1 shows the estimated (weighted) average effect size of financial education treatment on
financial behavior outcomes for six alternative models. We first perform an analysis on the full
sample (Panel A) and disaggregate our sample further into RCTs only (Panel B) and a
subsample containing only quasi and natural experiments (Panel C).
<Table S3.1 about here>
Column (1) repeats the OLS results, while Column (2) shows results with a single
synthetic (weighted average) effect size per study. Column (3) shows results for random effects
meta-regression (DerSimonian and Laird 1986) with inverse variance weights, synthetic effect
sizes per study, and Knapp and Hartung (2003) adjusted standard errors. This is common in
meta-analyses in other disciplines (such as clinical trials) and thus serves as a further check of
the sensitivity of our results to the estimation strategy. This approach assigns weights for each
44
study based on the inverse variance of the within study measurement error plus the between
study variance (tau squared) (=
). Thus we define our meta-analytic model as
=
++ (6)
where
∼(0,τ) (7)
and
∼(0,) (8)
Here is defined as the effect size estimate of study i, is the corresponding standard
error,is the between study variance in true effects, and is a vector of study level
covariates (including an intercept). We estimate this model using either method of moments
(DerSimonian and Laird 1986) or alternatively restricted maximum likelihood or empirical
bayes.
Column (4) reports estimations based on a GLS random-effects model. If one assumes
that the between-study heterogeneity cannot readily be explained by the observable
characteristics included, (i.e. due to unobserved heterogeneity in implementation quality),
one has to incorporate unobservable characteristics through random effects into the model (cf.
Cho and Honorati 2014). Thus, including an effect capturing unobservable characteristics of
the study, the meta-analytic model is defined as:
=
+
+ (5)
where is the impact (continuous effect size) of a financial education intervention on
outcome i reported in study j, is a vector of observable covariates, is a random effect
of unobservable study characteristics and is an error term independent of and.
Column (5) shows results for full pooling unrestricted weighted least squares using the
inverse standard error (precision) as weights (cf. Stanley and Doucouliagos 2012, 2015).
45
Finally, Column (6) shows results from robust variance meta-regression with dependent effect
sizes (see Tanner-Smith and Tipton 2014).
Reassuringly the direction is positive and statistical significance is found for all of the
considered models and sample splits. Additionally, the magnitude of the coefficient is similar;
however differences in detail do exist: The most common meta-analysis model is presented in
Column (3), which is also the model that Fernandes et al. (2014) and Miller et al. (2015) use
for their analyses. These models compare favorably to our main results discussed in the paper
relying on unrestricted ordinary least squares using multiple estimates per study and clustering
the standard errors at the study-level. In contrast, unweighted random effects GLS leads to a
higher estimate of the average treatment effect (Column 4). This approach is used previously
by Cho and Honorati (2014). The smallest estimate is reported in Column (5): By relying on
unrestricted weighted least squares, very large studies with extremely small standard errors,
which are most often quasi-experimental, receive extreme weight in the calculation of the
summary effect. From our point of view, it does not seem ideal to discount comparatively
smaller studies (which often still have sample sizes of over 1000 individuals) with high internal
validity (RCTs) as strongly as this approach does. Thus, if one incorporates weights based on
the standard error or variance of estimates, it seems advisable to account for between study
heterogeneity through random effects as discussed above and presented in columns (3) and (6).
Finally, column (6) presents results applying a recently developed method that accounts for
dependency among effect sizes (multiple, correlated estimates per study) (see Hedges et al.
2010; Tanner-Smith and Tipton 2014). Again, results are in line with our main results, although
with deflated expectations about the average effect in the whole sample of studies. This estimate
is also in line with the magnitude of the result presented in Fernandes et al. (2014), however,
our assessment about the effectiveness of 40 RCTs on financial behavior remains to be
strikingly different to the evidence synthesized by Fernandes et al. (2014).
46
(ii) Conservative handling of missing data. Next, we turn to estimations of complete
cases only, in order to test the robustness of our results using imputed default categories or
mean values for missing observations. Column (1) in Table S3.2 reports OLS meta-regression
results for complete cases only. These results correspond to the results presented in Table 3 of
the manuscript but show larger standard errors for some of the variables, however, turning none
of the main explanatory variables insignificant. This result strongly supports the conclusions
drawn from estimations with a large number of studies in the sample.
<Table S3.2 about here>
(iii) US only subsample. Then we consider only studies conducted in the U.S., since
these account for 65.87% of the studies and 42.67% of the effect size estimates in our sample
(column 2 of Table S3.2). Again, our results are near identical to the estimation in Table 3.
However, the standard error for the covariate for low-income clients increases and turns this
result insignificant while maintaining its magnitude and sign.
(iv) Classroom trainings only. Further, we consider only studies reporting on classroom
trainings as interventions (column 3). Again, our results are near identical to the estimation in
Table 3. However, the standard error for the covariate for mandatory courses increases.
(v) Equal study weights. Much of the meta-analysis literature in other fields than
economics uses effect size models where each study contributes only one synthetic effect size
to the meta-regression analysis. This procedure assures that the assumption of independent
estimates is not violated. There are different options to provide such a single effect. Some
suggest only using the most robust results in a primary study (cf. Cho and Honorati 2014, p.
119). The textbook literature on meta-analysis, however, tends to recommend creating a
synthetic effect size per study by using the average (or weighted average) effect across multiple
outcomes (cf. Lipsey and Wilson 2001).
47
We follow this approach here for the purpose of robustness exercises, but we point at the
major disadvantage that effects heading in opposite directions within one study may be
cancelling each other out. Column (4) of Table S3.2 shows results for such an approach. The
signs and magnitudes of our coefficients are very similar to the model with multiple non-
synthetic effect sizes per study and standard errors clustered at the study-level. However, in the
estimation based on this sample, the standard errors increase, thus leading to insignificant
covariates in three cases: RCT, intensity per week, and low-income clients. Since this approach
works with much less information than would be otherwise available, we conclude that
qualitatively this check also confirms our main findings derived from the larger sample of
available effect sizes.
Finally, in column (5) we give each study equal weight by assigning the inverse number
of estimates per study as weights for each effect size observation within a study. This yields
very similar results to the approach in column (4).
(vi) Alternative meta-regression models. Here we discuss the use of alternative
statistical regression models in the estimation of predictors of intervention impact.
(Ordered) probit models for sign and significance. In column (1) of Table S3.3 we apply
a probit-regression on an indicator variable of statistically significant effect estimates (at the
5%-level). This is a departure from earlier analyses because we now neglect the size of effects
and only consider their statistical significance. Following the approach applied by Card et al.
(2010, 2015) and Cho and Honorati (2014), we code the sign and significance for each impact
estimate reported in the primary studies. This indicator of intervention success has the
advantage that it is easily interpretable and neutral to the unit of the outcome variable. However,
it only captures the direction and significance of an effect, unlike the standardized mean
difference which preserves its magnitude (cf. Stanley and Doucouliagos 2012, p. 6). Using this
approach, we construct a binary dependent variable taking the value 0 if the primary study
48
impact estimate t-statistic is smaller than 1.96 and taking the value 1 if t ≥ 1.96. Additionally,
we extend this approach and construct an ordered categorical variable that can take three values
of -1 if t ≤ -1.64, 0 if t ≥ - 1.64 and t ≤ 1.64, and 1 if t ≥ 1.64. Thus, we distinguish between
significant negative, insignificant, and significant positive estimates at the 10%-level because
there are hardly negative estimates at the 5%-level (see Table S1.3 in the Appendix S1).
<Table S3.3 about here>
We observe that mostly the sign and significance of the logged odds correspond with the
model using a continuous measure of effect size reported in Table 3, column (2). However,
estimated standard errors differ, as the coefficients for TOT, intensity, and mandatory are now
insignificant – probably resulting from reduced variance in the dependent variable in
comparison to the use of continuous effect sizes.
In column (2) we extend this approach and estimate an ordered probit model where the
dependent variable consists of three ordered categories that distinguish between significant
negative, insignificant and significant positive estimates at the 10%-level of financial education
impact. This leads to a very similar assessment of predictor sign and magnitude as in our
benchmark model in Table 3, column (2), but again slightly different estimates for the standard
errors, with intensity, however, being a significant predictor in this estimation again.
GLS random effects regression. Next, we check whether controlling for unobservables
affects our results. The results in column (3) show coefficients from a GLS random effects
regression based on the assumptions discussed in equation 5. This estimation almost entirely
matches the results of the benchmark model shown in Table 3, column (2) with the exception
of an increased standard error for mandatory financial education.
Unrestricted weighted least squares. Next, we turn to an alternative unrestricted weighted
least squares approach. In column (4) we weight each effect size with its inverse standard error
(1/SE) and account for publication selection bias by including the standard error (SE) of each
49
estimate as a covariate (as suggested by Stanley and Doucouliagos 2012). The results show that
our results, again, largely match the results of the ordinary least squares estimations, however,
the predictor for mandatory courses is now insignificant. In column (5) we redo this analysis
and use the inverse variance as weights and include the variance as a covariate in the analysis
to account for publication selection bias. This estimation, while qualitatively similar, shows no
negative effects (due to increased estimated standard errors) for low-income clients, and
mandatory courses.
Random effects meta-regression (DerSimonian and Laird 1986). Table S3.4 shows our
preferred specification for three different estimators of random-effects meta-regression models
with and without Knapp and Hartung (2003) corrected standard errors, respectively. Using
method of moments (columns 1 and 2), we find that our results are similar to our benchmark
model using OLS in Table 3, column (2), with the exception of increased standard errors,
especially when applying the correction suggested by Knapp and Hartung (2003), for the
coefficients for low-income economies, low-income individuals and intensity per week, which
are now statistically insignificant. Turning to the alternative estimators (restricted maximum
likelihood, and empirical bayes), we find that these results are again nearly identical. Overall,
we conclude that the pattern in sign and magnitude (including most standard errors) of our main
explanatory variables are confirmed under various random effect meta-regression models,
however with a more positive assessment of the intervention impact in low and lower-middle
income economies and for low-income individuals, as well as a positive but insignificant
estimate of intensity per week.
<Table S3.4 about here>
(vii) Heterogeneous impacts depending on delay in measurement. In order to check
for heterogeneous impacts depending on the considered time-frame, we conduct two tests. First,
we model the relationship between delay in measurement and effect size on financial behavior
50
outcomes in a non-linear fashion by creating a categorical variable that distinguishes between
short term (less than one month, approx. 12% of estimates), medium term (less than one year,
approx. 41% of estimates), and long-term (longer than one year, approx. 47% of estimates)
effects on financial behavior. Column (1) of Table S3.5 shows that short term effects tend to be
higher than medium- or long-term effects on financial behavior, which is in line with the present
literature (cf. Fernandes et al. 2014; Lusardi et al. 2015b). Splitting the sample according to
these three time-frames, we observe that most predictors are similar in sign and magnitude in
all subsamples, with some differences regarding signs and significance of predictors. It seems
noteworthy, and reassuring for our results, that the subsample comprising the longer-term
treatment effects appears to be driving our main results. In particular, intensity appears to matter
for effect sizes to be found after a long delay between treatment and measurement. This is in
line with earlier observations by Fernandes et al. (2014) that intensity may interact with delay
since intervention.
<Table S3.5 about here>
(viii) Intensity. Since the intensity of financial education supports its effectiveness, we
check which aspect of intensity of education drives our results. Using only the total number of
hours taught as a linear predictor of effect size (and neglecting the duration of the intervention),
we find that intensity does not predict effect sizes on financial behavior (available on request).
This result remains the same in several variants of variable and model specifications (e.g.
including polynomial forms of intensity, interaction effects between delay and intensity, and
centering) and holds when effect sizes on financial literacy are regressed on this linear predictor.
Thus, the intensity relative to the duration of the intervention appears to matter most for the
impact on financial behavior. This finding seems to have practical implications, since it favors
education with higher relative intensity, i.e. trainings with relatively more hours per week.
51
References in Appendix S3
Card, D., Kluve, J., and Weber, A. (2010). Active labour market policy evaluations: A meta-
analysis. Economic Journal, 120(548): F452–F477.
Card, D., Kluve, J., and Weber, A. (2015). What works? A meta analysis of recent active
labor market program evaluations. NBER Working Paper 21431.
Cho, Y. and Honorati, M. (2014). Entrepreneurship programs in developing countries: A meta
regression analysis. Labour Economics, 28: 110–130.
DerSimonian, R. and Laird, N. (1986). Meta-analysis in clinical trials. Controlled Clinical
Trials, 7(3): 177–188.
Fernandes, D., Lynch, Jr., J.G., and Netemeyer, R.G. (2014). Financial literacy, financial
education, and downstream financial behaviors. Management Science, 60(8): 1861–1883.
Harbord, R. M., Higgins, J. P., et al. (2008). Meta-regression in Stata. Stata Journal, 8(4):493–
519.
Hedges, L.V., Tipton, E., and Johnson, M. C. (2010). Robust variance estimation in meta-
regression with dependent effect size estimate, Research Synthesis Methods 1(1): 39-65.
Knapp, G. and Hartung, J. (2003). Improved tests for a random effects meta-regression with a
single covariate. Statistics in Medicine, 22(17): 2693–2710.
Lipsey, M.W. and Wilson, D.B. (2001). Practical meta-analysis. Sage, Thousand Oaks, CA.
Miller, M., Reichelstein, J., Salas, C., and Zia, B. (2015). Can you help someone become
financially capable? A meta-analysis of the literature. World Bank Research Observer, 30(2):
220–246.
Stanley, T. D. (2001). Wheat from chaff: Meta-analysis as quantitative literature review.
Journal of Economic Perspectives, 15(3): 131–150.
Stanley, T. D. (2008). Meta-regression methods for detecting and estimating empirical effects
in the presence of publication selection. Oxford Bulletin of Economics and Statistics,
70(1):103–127.
Stanley, T. D. and Doucouliagos, H. (2012). Meta-regression analysis in economics and
business, Routledge, New York, NY.
Stanley, T. D. and Doucouliagos, H. (2015). Neither fixed nor random: weighted least squares
meta-analysis. Statistics in Medicine 34(13): 2115–2127.
Tanner-Smith, E. E. and Tipton, E. (2014). Robust variance estimation with dependent effect
sizes: practical considerations including a software tutorial in Stata and SPSS. Research
Synthesis Methods, 5(1):13–30.
52
Table S3.1: Financial education treatment effect on financial behavior under various models
Outcome (1)
OLS
Full pooling
(2)
OLS
Synthetic ES
(3)
RE-Metareg
(4)
RE
GLS
(5)
WLS
1/ SEg
(6)
Robumeta
Panel A : All
Fin. Behavior 0.086*** 0.102*** 0.079*** 0.093*** 0.026** 0.064***
(0.012) (0.013) (0.009) (0.012) (0.011) (0.008)
n(Studies)
n(Effect sizes)
90
349
90
90
90
90
90
349
90
349
90
349
Panel B: RCTs
Fin. behavior 0.082*** 0.102*** 0.075*** 0.089*** 0.067*** 0.078***
(0.014) (0.023) (0.013) (0.021) (0.013) (0.012)
n(Studies)
n(Effect sizes
40
227
40
40
40
40
40
227
40
227
40
227
Panel C: Quasi exp.
Fin. behavior 0.093*** 0.102*** 0.083*** 0.100*** 0.015* 0.059***
(0.022) (0.015) (0.012) (0.015) (0.008) (0.010)
n(Studies) 50 50 50 50 50 50
n(Effect sizes) 122 50 50 122 122 122
Notes: Column (1) shows the average effect size on fin. behavior estimated via OLS with standard errors clustered by Study
ID. Column (2) shows the average effect using only one synthetic (weighted average) effect size per study. Synthetic effect
sizes are estimated via within-study random effects meta-regression (DerSimonian and Laird 1986). Column (3) shows the
average weighted treatment effect estimated via random effects meta-regression (DerSimonian and Laird 1986) and Knapp
Hartung (2003) adjusted standard errors. The Stata command is “metareg”. Column (4) shows the average treatment effect of
fin. edu on fin. behavior utilizing a study random-effects GLS model. Column (5) presents results using unrestricted
weighted least squares where a weight of the respective inverse standard error is assigned to each observation. Column (6)
presents results from robust variance meta-regression with dependent effect size estimates (Tanner-Smith and Tipton 2014).
The Stata command is “robumeta”. Standard errors (clustered at the study-level for Columns (1), (4), (5), and (6)) in
parentheses. ***, ** and * denote significance at the 1%, 5% and 10% level.
53
Table S3.2: Missing data, subsamples and giving each study equal weight
Notes: Column (1) reports results for complete cases only. Columns (2) present results for the sample split of
USA studies only. These splits include only variables for which differential information from at least two studies
are available. Column (3) presents results using one synthetic effect size (weighted within-study average effect
size across all outcomes) per study. Column (4) shows results by weighting each observation by the inverse
number of observations of the study the observation is nested in. Standard errors (clustered at the study-level for
all Columns but (4)) in parentheses. ***, ** and * denote significance at the 1%, 5% and 10% level.
(1)
No
imputations
(2)
US only
(3)
Classroom
only
(4)
Synthetic ES
OLS
(5)
Equal study
weights
RCT -0.052* -0.097** -0.080*** -0.052 -0.042
(0.027) (0.042) (0.028) (0.033) (0.031)
TOT 0.057 0.114*** 0.065** 0.107*** 0.105***
(0.041) (0.040) (0.028) (0.028) (0.035)
Delay -0.000* 0.000 0.000 -0.000 0.000
(0.000) (0.000) (0.000) (0.000) (0.000)
1/SE 0.001 0.000 -0.000 0.000 -0.000
(0.000) (0.000) (0.000) (0.000) (0.000)
Intensity / week 0.005*** 0.006* 0.004*** 0.001 0.001
(0.001) (0.003) (0.001) (0.002) (0.002)
Duration -0.000 -0.001 -0.000 -0.001 -0.000
(0.001) (0.001) (0.001) (0.001) (0.001)
Low income clients -0.047** -0.003 -0.054*** -0.043 -0.049**
(0.020) (0.025) (0.017) (0.027) (0.022)
Years of schooling -0.022*** -0.021*** -0.022** -0.020**
(0.007) (0.006) (0.009) (0.009)
Low/lower-mid .econ -0.113** -0.108** -0.113* -0.108*
(0.044) (0.041) (0.061) (0.059)
Mandatory -0.086* -0.097*** -0.043 -0.097** -0.095***
(0.049) (0.033) (0.028) (0.038) (0.029)
Teachable moment 0.058 0.129*** 0.075** 0.058** 0.058**
(0.052) (0.035) (0.033) (0.028) (0.026)
Constant 0.359*** 0.042 0.364*** 0.364*** 0.344***
(0.097) (0.031) (0.095) (0.118) (0.119)
R2 0.125 0.340 0.177 0.297 0.206
n (Studies) 35 55 70 90 90
n (Effect sizes) 24 135 317 90 349
54
Table S3.3: Alternative meta-regression models
(1) (2) (3) (4) (5)
Probit
5%
Ordered
probit
10%
RE
GLS
WLS
1/SE(g)
weights
WLS
1/Var(g)
weights
RCT -0.794*** -0.802*** -0.087*** -0.086*** -0.044**
(0.225) (0.196) (0.024) (0.020) (0.022)
TOT 0.052 0.002 0.049** 0.038** 0.058***
(0.189) (0.176) (0.023) (0.016) (0.015)
Delay -0.001** -0.000 -0.000 -0.000 0.000
(0.000) (0.000) (0.000) (0.000) (0.000)
1/SE 0.001 -0.000
(0.001) (0.001)
SE
g
0.486*** 0.611**
(0.173) (0.272)
SE
g
2 3.147**
(1.496)
Intensity /week 0.018 0.027* 0.003** 0.003* 0.006**
(0.014) (0.015) (0.002) (0.002) (0.003)
Duration 0.008* -0.000 0.000 0.000 0.000
(0.004) (0.005) (0.001) (0.000) (0.000)
Low inc. clients -0.566*** -0.561*** -0.060*** -0.014* -0.000
(0.160) (0.148) (0.019) (0.007) (0.002)
Years of schooling -0.154*** -0.136*** -0.024*** -0.022*** -0.018***
(0.058) (0.044) (0.006) (0.005) (0.006)
Low/lower-mid. econ. -0.872** -0.792** -0.105** -0.086*** -0.076*
(0.392) (0.314) (0.042) (0.032) (0.045)
Mandatory 0.172 0.130 -0.030 -0.026 -0.017
(0.245) (0.272) (0.026) (0.020) (0.018)
Teach. moment 0.326 0.404** 0.063*** 0.042** 0.068***
(0.219) (0.192) (0.024) (0.017) (0.015)
Constant cut 1 -3.977***
(0.636)
Constant cut 2 -1.999***
(0.594)
Constant 2.009** 0.356*** 0.304*** 0.210***
(0.783) (0.079) (0.066) (0.079)
R
2 0.197 0.301 0.336
Pseudo R2 0.109 0.084
n (Studies) 90 90 90 90 90
n (Effect Sizes) 349 349 349 349 349
Notes: Dependent variable in columns (1) and (2) is a categorical indicator of sign and significance of intervention impact.
Dependent variable in columns (3) and (4) is effect size (Hedges’ g) on financial behavior. Column (1) reports results from
probit-regression with a binary outcome indicating whether financial education had a significantly positive effect on financial
behavior at the 5%-level. Column (2) provides results for ordered probit regression with a dependent categorical variable
taking the value “-1” if financial education had a significantly negative impact on financial behavior, “0” if financial
education had an insignificant effect on financial behavior, and “1” if financial education had a significant positive effect on
financial behavior at the 10%-level. Column (3) reports results from GLS random-effects regression. Column (4) reports
results of weighted least squares estimation with inverse variance weights. Standard errors clustered at the study-level in
parentheses. ***, ** and * denote significance at the 1%, 5% and 10% level.
55
Table S3.4: Random effects meta-regression on synthetic effect sizes with inverse variance weights
(1) (2) (3) (4) (5) (6)
MM MM REML REML EB EB
RCT -0.066*** -0.066*** -0.065*** -0.065*** -0.066*** -0.066***
(0.021) (0.020) (0.021) (0.020) (0.022) (0.022)
TOT 0.061*** 0.061*** 0.061*** 0.061*** 0.063*** 0.063***
(0.020) (0.019) (0.020) (0.019) (0.020) (0.020)
Delay -0.000 -0.000 -0.000 -0.000 -0.000 -0.000
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Intensity /week 0.002 0.002 0.002 0.002 0.002 0.002
(0.002) (0.002) (0.002) (0.001) (0.002) (0.002)
Duration 0.000 0.000 0.000 0.000 0.000 0.000
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001)
Low inc. clients -0.024 -0.024 -0.024 -0.024 -0.025 -0.025
(0.016) (0.016) (0.016) (0.015) (0.017) (0.017)
Years of schooling -0.015** -0.015*** -0.014** -0.014*** -0.015** -0.015**
(0.006) (0.006) (0.006) (0.005) (0.006) (0.006)
Low/lower inc. econ. -0.044 -0.044 -0.043 -0.043 -0.046 -0.046
(0.039) (0.038) (0.039) (0.037) (0.040) (0.040)
Mandatory -0.052** -0.052*** -0.051** -0.051*** -0.053** -0.053**
(0.020) (0.019) (0.020) (0.019) (0.021) (0.021)
Teach. moment 0.053*** 0.053*** 0.052*** 0.052*** 0.053*** 0.053***
(0.018) (0.017) (0.018) (0.017) (0.018) (0.018)
Constant 0.251*** 0.251*** 0.249*** 0.249*** 0.256*** 0.256***
(0.075) (0.073) (0.075) (0.071) (0.076) (0.076)
I2 81.14% 81.14% 81.14% 81.14% 81.14% 81.14%
Adj. R2 - - 0.442 0.442 0.474 0.474
n (Studies) 90 90 90 90 90 90
n (Effect Sizes) 90 90 90 90 90 90
Adjusted errors yes no yes no yes no
Notes: Results from random-effects meta-regression (DerSimonian and Larid 1986) with and without Knapp and Hartung
(2003) adjusted standard errors, respectively. Dependent variable is effect size (Hedges’ g) on financial behavior weighted by
its inverse variance. Columns (1) and (2) show results for method of moments (MM) estimates. Columns (3) and (4) show
results for restricted maximum likelihood (REML) estimates. Columns (4) and (5) show results from empirical bayes
estimates. The Stata command is metareg (Hardbord and Higgins 2008). Standard errors in parentheses. ***, ** and * denote
significance at the 1%, 5% and 10% level.
56
Table S3.5: Effect sizes on financial behavior and heterogeneity of treatment effects by delay in
measurement of treatment effects
(1) (2) (3) (4)
Financial
behavior
Short term
subsample
Medium term
subsample
Long term
subsample
RCT -0.061** 0.148 -0.085*** -0.073*
(0.026) (0.102) (0.027) (0.038)
TOT 0.043* -0.221** 0.043 0.062
(0.025) (0.078) (0.032) (0.049)
Short term 0.089**
(0.039)
Medium term -0.006
(0.018)
1/SE -0.000 -0.005** -0.000 0.000
(0.000) (0.002) (0.000) (0.000)
Intensity /week 0.004*** 0.006 0.002 0.004***
(0.001) (0.007) (0.002) (0.001)
Duration -0.000 0.010** 0.000 0.000
(0.001) (0.005) (0.001) (0.001)
Low inc. clients -0.044*** -0.046 -0.041** -0.045**
(0.014) (0.087) (0.020) (0.019)
Years of schooling -0.021*** -0.103** -0.011 -0.021**
(0.005) (0.047) (0.008) (0.009)
Low/lower inc. econ. -0.122*** -1.127*** 0.034 -0.156***
(0.041) (0.318) (0.055) (0.058)
Mandatory -0.041** -0.076 0.003 -0.056***
(0.019) (0.097) (0.047) (0.021)
Teach. moment 0.090*** 0.202* 0.009 0.109***
(0.028) (0.108) (0.032) (0.024)
Constant 0.332*** 1.634** 0.235** 0.332***
(0.077) (0.624) (0.101) (0.119)
R2 0.204 0.457 0.073 0.319
n (Studies) 90 18 24 53
n (Effect Sizes) 349 42 143 164
Notes: Results from OLS meta-regression with robust standard errors clustered at the study-level. Dependent variable is
effect size (Hedges’ g) on financial behavior. Standard errors in parentheses. ***, ** and * denote significance at the 1%, 5%
and 10% level.
57
Appendix S4: Publication bias and heterogeneity of study quality
We show examinations of conventional visual tests for publication bias in order to address
the so-called file drawer problem (cf. Stanley and Doucouliagos 2012, p. 73) and examine the
sample of studies for heterogeneous results depending on study quality. Note that we also use
formal econometric methods, (i.e. alternative regression approaches) in Appendix S3 that are
in principle capable of generating unbiased estimates in the presence of publication selection
(see table S3.4 columns 4 and 5).
Publication bias. We conduct visual tests for overall publication bias (funnel
asymmetry), so-called funnel plots (cf. Figures S4.1 and S4.2). Precision of the estimated
treatment effect should increase in larger studies. Thus, we scatter effect sizes (multiple effects
per study) against the standard errors of the effect size estimates (inverted y-axis). Effect
estimates from small studies (larger sampling errors) should scatter more widely at the bottom
of the graph, with the spread decreasing as standard errors decrease. In the absence of bias, the
plot resembles a symmetrical inverted funnel. Therefore, asymmetry indicates a publication
bias in the sense that negative or non-results are under-represented (i.e. not published at all).
Inspecting the two plots indicates that symmetry is higher for effect sizes on financial behavior
than for effect sizes on literacy but both outcomes may be affected by publication biases in the
sense that the overall treatment effect may suffer from a slight upward bias. This conclusion,
however, requires the assumption that non-results are not published at all (i.e. the file drawer
problem).
<Figure S4.1 and Figure S4.2 about here>
This assumption may be more plausible for quasi- and natural experiments than for RCTs,
as results from rigorous randomized experiments are likely to be published irrespective of their
results. Therefore, we perform the same visual check on the subsample of RCTs only (cf.
Figures S4.3 and S4.4). Indeed, these plots appear to be more symmetric indicating that
58
publication bias may primarily be an issue within the sample of non-randomized studies. As (i)
nearly 40 percent of our sample is comprised of RCTs; (ii) we control for research design in all
our regressions; and (iii) our main results replicate within the subsample of RCTs, we suggest
that publication biases are not an issue for our analysis. However, we also test the robustness
of our results using weighted least squares and controlling for the standard error (or the squared
standard error, i.e. the variance of the estimate), which is advocated as a robust method in the
presence of publication selection (cf. Stanley and Doucouliagos 2012).
<Figure S4.3 and Figure S4.4 about here>
Publication status and quality. Another concern in any meta-analysis is the issue of
biases arising from the aggregation of results from studies with different publication status and
quality. On the one hand, researchers fear that the tendency of the scientific community to favor
statistically significant positive results over insignificant non-results may lead to biased
estimates favoring the rejection of the null hypothesis of a zero-effect of financial education on
relevant outcomes. The standard solution in the meta-analysis literature is to include as many
unpublished studies (grey literature) as possible to address this potential source of bias a priori.
On the other hand, economists fear that by aggregating studies of different publication
status and quality, the results suffer due to the lack of empirical rigor in grey-literature primary
studies. To shed light on this issue in the financial education literature, we compare average
effect sizes of financial education interventions by different types of publication status and
indicators of quality. Table S4.1 compares average effect sizes on financial literacy and
behavior by publication status in an academic journal. Interestingly, a bias affects only the
effect size estimates on financial literacy, as they appear to be more than twice as high in
published than in unpublished papers (t=3.863). Turning to effect sizes on financial behavior,
however, we observe no significant difference in average effect sizes between published and
unpublished studies.
59
<Table S4.1 about here>
Considering indicators of study quality, we code the article influence score (ISI web of
knowledge) of the respective journal (and year) for every publication and assign a value of 0
for studies available as working papers. Comparing influential (article influence score >1) with
less influential (≤1) publications, we find that the quality bias for financial literacy is now
insignificant (t=0.328): Moreover, influential journals tend to publish studies with 0.04 standard
deviation units smaller effect sizes on behavior (t=-2.189) than non-influential journals. Thus,
more rigorous work reports a slightly smaller average treatment effect than presumably less
rigorous work.
Next, we code the number of citations for each publication as reported in Google Scholar
(as of October 31, 2016). The mean number of citations per article is 53.91and we split the
sample in studies cited above and below this threshold value. Again, we find no significant
differences between highly cited studies and others: If anything, highly cited studies tend to
report smaller average effect sizes