Financial Literacy, Financial Education, and Downstream Financial Behaviors

168 papers covering 201 prior studies; manipulated-literacy cells N = 80,014 (low-income, 35 effects) and 271,759 (general population, 55 effects); measured-literacy cells N = 6,975 and 245,248; plus three original studies (n = 543 panel; one probability sample)

All ages and settings, overwhelmingly adults; includes high-school financial education as one of six intervention forms. Mostly US, some developing-country programmes.

The field's founding meta-analysis, and the reason "financial literacy does not work" became a stylised fact. Its decisive move is separating MANIPULATED financial literacy (someone ran an intervention) from MEASURED financial literacy (a survey scored people's knowledge and correlated it with behaviour) — the two give effects differing by roughly an order of magnitude in variance explained. It pools mixed designs within the manipulated set (15 randomised experiments, plus quasi-experiments and pre-post), hence grade C. Two internal gradients matter more than the headline: within the manipulated set, TRUE RANDOMISED experiments give SMALLER effects than quasi-experimental and pre-post designs; and within the measured set, instrumental-variable estimates are smaller than OLS on the same data (0.059 vs 0.091 across the 16 studies with non-weak instruments). Both gradients run the direction that says the apparent effect is design artefact. Kaiser et al. (2020) later re-coded these RCTs and report four coding errors; the Fernandes RCT estimate is built on 15 observations from 13 studies, which is thin. Funded by the National Endowment for Financial Education — an advocacy funder that paid for a null.

Interventions to improve financial literacy explain 0.1% of the variance in the financial behaviours they target (sample-weighted partial r = 0.025 in low-income samples, 0.035 in general population samples). Measured financial literacy correlates far more strongly with the same behaviours (r = 0.113 and 0.134; 1.27% and 1.80% of variance), and that gap is the paper's central claim: the correlation is not a lever. Effects decay with delay — a significant negative linear effect of months since intervention (B = -0.0033, SE 0.0009, p = .002) — and at delays of 18.5 months or more there is no significant effect of even 24 hours of instruction. High-school financial education specifically explained 0.15% of behavioural variance. Financial education also moves KNOWLEDGE weakly: in the 12 papers measuring both, interventions explained 0.44% of variance in financial knowledge, against 2.25% for science and maths instruction in Lipsey & Wilson's benchmark.

High for the measured-literacy half, and the paper demonstrates it rather than assuming it. Financial literacy scores predict behaviour partly because they proxy numeracy, willingness to take risks, confidence and propensity to plan — all substantially heritable — and the partial effects "diminish dramatically" when those traits are controlled or when literacy is instrumented. Low for the manipulated half, which rests on experiments.

Directionally replicated, magnitudinally disputed. Kaiser, Lusardi, Menkhoff & Urban (2020/2022) re-ran the question on 76 RCTs and could not reproduce the near-null: their behaviour estimate is 0.10 SD against Fernandes' implied 0.018 SD, and they report four coding errors in the original RCT classification. What survives is the manipulated-vs-measured gap and the knowledge-vs-behaviour gap; what does not survive is "essentially zero" as the causal estimate, and the rapid-decay claim is also weakened (Kaiser et al. find 0.057 SD at two-plus years, statistically indistinguishable from the short-run estimate).