Financial Literacy and Perceived Economic Outcomes

Abstract

We explore the relationship between financial literacy and self-reported, reflective economic outcomes from respondents using survey data from the United States. Our dataset includes a large number of covariates from the National Financial Capability Study (NFCS), widely used by literacy researchers, and we use a new econometric technique developed by Hahn et al., designed specifically for causal inference from observational data, to test whether changes in financial literacy infer meaningful changes in self-perceived economic outcomes. We find a negative treatment parameter on financial literacy consistent with the recent work of Netemeyer et al. and contrary to the presumption in many empirical studies that associate standard financial outcome measures with financial literacy. We conclude with a discussion of heterogeneity of the financial literacy treatment effect on household income, gender, and education level sub-populations. Our findings on the relationship between financial literacy and reflective economic outcomes also raise questions about its importance to an individual’s financial well-being.

KEYWORDS:

• Causal inference
• Financial literacy
• Financial well-being
• Perceived economic outcomes
• Regularization

Disclosure Statement

The authors report there are no competing interests to declare.

Acknowledgments

We are grateful for Annamaria Lusardi, Brigitte Madrian and Carly Urban’s helpful comments on earlier drafts.

Notes

1 As an example of financial education sources the CFPB has a consumer tool tab and the U.S. Treasury department has compiled a list. The Chicago FED lists educational offerings of banks in the system.

2 The life-cycle of savings literature has given insight to why people save, spend and borrow, yet professional personal finance advice is driven by rules of thumb, for example, “‘everybody should save 10% of their periodic earnings for retirement.” See Modigliani and Brumberg (1954) and Friedman (1957).

3 Households and individuals can be financially well without being financially literate. They may rely on CPAs for their taxes, a bank relationship for transactions, or financial advisors for investments and insurance products. Moreover, wealthier households may be able and more interested in off-loading financial matters to others and neither care nor are able to answer successfully the standard set of questions that measure financial literacy.

4 See Hastings, Madrian, and Skimmyhorn (2013), pp. 358.

5 Netemeyer et al. (2018) supplemented their work with a measure of overall well-being developed by Su, Tay, and Diener (2014).

6 As noted by Lusardi and Mitchell (2014), p. 34, “Though it is challenging to establish a causal link between financial literacy and economic behavior, both instrumental variables and experimental approaches suggest that financial literacy plays a role in influencing financial decision making, and the causality goes from knowledge to behavior.” Two noteworthy contributions along this path are Calvet, Campbell, and Sodini (2007) and Agarwal et al. (2009)

7 Indeed, most empirical questions surrounding financial literacy are characterized by these problems.

8 These authors propose jointly modeling the treatment and response $Y,Z|\mathrm{X}$ by first modeling the treatment variable as a function of covariates $Z|\mathrm{X}$, and then modeling the response $Y|Z,\mathrm{X}$. The first likelihood provides information on the propensity of being treated as a function of covariates, and the second uses this information to mitigate endogeneity when estimating the partial effect of Z on Y. Importantly, their procedure provides a way to “shrink-away” irrelevant covariates using Bayesian shrinkage priors.

9 In Appendix A, we visually compare our NFCS sample to the entire survey universe to confirm the sample and universe are similar across several covariates.

10 Odean (1998) defined the disposition effect as the tendency for investors who hold losers to hold them too long, and investors who own winners to have sold them too quickly.

11 Choi, Laibson, and Madrian (2011) sample included employees who were older than 59.5 who were unconstrained by withdrawal penalties: they could have simply withdrew employer contributions but chose not to take advantage of it. Even among a subsequent experiment, the researchers find conclude that low financial literacy and poor choice about a matching contribution are positively related.

12 See Survey of the States, Council for Economic Education.

13 Mangrum (2019) explored student loan repayment behavior after college using university-level data from the College Scorecard database. Universities are populated with students from different states, therefore, different financial education experiences, and Mangrum found that “mandated students” improve financial student loan repayment percentages among first generation and low income borrowers.

14 See Willis (2008), p. 204. Willis interprets policymakers’ promotion of financial literacy as an ineffective substitute for financial regulation that places too high a burden on non-expert consumers.

15 See Willis (2011) p. 429.

16 Willis (2011) notes that empirical work to date is replete with evidence that “biases, heuristics, and other non-rational influences” circumvent good financial decision-making. In this study, we are unable to test a hypothesis related specifically to financial education because the 2015 NFCS study did not survey that question.

17 Lusardi, Michaud, and Mitchell (2017), p. 473.

18 Almenberg and Dreber (2015) linked financial literacy and investing in the stock market with the intent to explore how investing varies between men and women when financial literacy is controlled. The authors’ measure financial literacy by identifying basic and advanced financial skills. While the authors find that men have higher probabilities of investing in the stock market, controlling for financial literacy skills reduces the probability differences between men and women substantially, and makes a “gender gap,” inconsequential.

19 53.4% of Balasubramnian and Brisker’s sample used one of the defined advisers.

20 Generally, the economic outcome for a household at any point in time is its economic net worth; that is, assets including household human capital less debt.

21 Studies were conducted in 2009, 2012, 2015, and 2018. See http://www.usfinancialcapability.org.

22 There are 27,564 and 27,091 observations from the surveys conducted in 2015 and 2018, respectively.

23 We will discuss in the empirical analysis below why this is important for inferring a more accurate causal effect estimate and how regularization provides a greater degree of confidence the estimate.

24 The attributes of the final dataset are available from the authors upon request.

25 The interest rate, inflation and risk questions were designed by Olivia Mitchell and Annamaria Lusardi. See Lusardi and Mitchell (2014). According to the 2015 NFCS national report, the Rule of 72 question was added as an additional interest rate question to “to test the concept of interest compounding in the context of debt.”

26 In the NFCS study, the financial investment question is item B14 where an answer of “yes” is coded as a 1 and an answer of “no” a 2.

27 See http://www.usfinancialcapability.org/results.php?region=US.

28 For example, a person who is not satisfied with her current personal financial condition (Question 1) is also likely to have difficulty paying bills every month (Question 2).

29 Several recent studies develop methods for inferring causal effects using machine learning techniques. Chernozhukov et al. (2016) propose a double machine learning approach. Hill (2011) provide an early method using Bayesian Additive Regression Trees. Taddy et al. (2016) and Hahn, Murray, and Carvalho (2020) develop more recent Bayesian approaches for experimental and observational data, respectively. Wager and Athey (2018) provide theoretical guarantees for an approach based on random forests, but Hahn, Murray, and Carvalho (2020) point out the challenges with their method in practice. Little guidance is given for controlling the level of statistical regularization of the random forests, and these practical issues are manifest for any ML method used for causal inference.

30 Hahn et al. (2018) discussed how naive regularization of γ and β in Model 3 will lead to significant bias in the estimate of α—a phenomena called regularization-induced confounding (RIC). To avoid this, we follow an approach outlined in Hahn, Murray, and Carvalho (2020) to control for a prediction of literacy as a function of the covariates. First, we estimate literacy $\widehat{Z}$ from the selection equation in Model 3 with a shrinkage prior placed on the coefficient vector γ. Second, we control for this prediction $\widehat{Z}$ in the response equation and only regularize the coefficients β, leaving unregularized α and the coefficient on $\widehat{Z}$. The shrinkage priors used are closely related to the horseshoe priors of Carvalho, Polson, and Scott (2010) and estimation is undertaken using a variant of the elliptical slice sampler developed in Hahn, He, and Lopes (2019) and implemented in the R package bayeslm. 100,000 MCMC draws from the posterior distribution are generated after 10,000 draws are generated as burn-in. 95% intervals for the model estimates are computed as the empirical quantiles of this posterior distribution.

31 We use an the R package Bayesian causal forests – bcf – to fit the nonlinear model which gives estimates of the individual treatment effects. The current model is built for a binary treatment variable, so we use our existing literacy variable and create a binary version by dividing at the median literacy value. Thus, all sample observations with measured literacy larger (smaller) than the median will labeled as a 1 (0) in the new binary literacy variable. The quantile at which we divide literacy does not effect the overall heterogeneity results described.

32 There are, of course, many traits of potential interest limited only in number by the variables in the underlying dataset.