Financial education affects financial knowledge and downstream behaviors

76 randomised experiments, total sample > 160,000; 673 effect estimates (458 on behaviour from 64 studies, 215 on knowledge from 50 studies). Average study n = 2,136, median 840.

All ages: only 7.5% of estimates concern children under 14, 20% youth aged 14-25, and 72.4% adults over 25. 60.8% of estimates come from developing economies; 72.4% target below-median-income individuals. This is NOT primarily a school literature.

The intended rebuttal to Fernandes et al. (2014), and the strongest evidence financial education has. Restricted to randomised experiments only — an explicit refusal of the quasi-experimental degrees of freedom — and estimated with robust variance estimation under a random-effects model. Numbers here are read from NBER Working Paper 27057 (April 2020); the published Journal of Financial Economics version reports the same design and headline figures. Three things a school builder must hold onto. First, the population: this is an adult, developing-country, low-income literature, and the school-age cell is small. Second, the DOSE is tiny — mean 11.71 hours of instruction, median 7 — so the effects are what seven hours buys, not what a year-long course buys. Third, the authors are the field's principal advocates (Lusardi directs a financial-literacy research centre) and the meta includes their own trials; they report robustness to excluding their own papers, which is the right check, and the estimate survives publication-bias correction, restriction to top economics journals, and restriction to adequately powered studies. Independence is recorded as developer-involved for that reason, not because they built any programme.

In 76 RCTs the random-effects mean treatment effect on financial BEHAVIOUR is 0.1003 SD (95% CI 0.071-0.129) and on financial KNOWLEDGE 0.204 SD (95% CI 0.152-0.255). Under the common-true-effect assumption Fernandes et al. used, the behaviour estimate is 0.065 SD (CI 0.043-0.089) against their 0.018 SD (CI -0.004 to 0.022) from 15 observations in 13 RCTs — so the null is not confirmed on the larger sample. The age split is the school-relevant number and it cuts against schools: behaviour effects are SMALLEST for children under 14 (0.064 SD) versus youth 14-25 (0.1203) and adults (0.1068), while knowledge effects are largest for children (0.2763, but on 15 observations in seven studies, CI 0.0076-0.545). Decay is modest, not catastrophic: 0.0991 SD under six months, 0.071 at six-plus months, 0.0878 at one year, 0.0653 at 1.5 years, 0.0574 at two-plus years (seven studies), none statistically distinguishable from each other. Mean cost $60.40 per participant (median $22.90) for roughly a fifth of an SD.

Low. Randomised assignment throughout; genes cannot differ between arms. The residual concern is generalisation rather than confounding — the samples are volunteers into financial-education studies, largely adults, and 60.8% are in developing economies.

This IS the replication attempt on Fernandes, Lynch & Netemeyer (2014), and it partly fails to reproduce them: the authors could not exactly replicate the original RCT estimate and document four coding errors in it. Direction agrees with Fernandes (knowledge moves more than behaviour; behaviour effects are small); magnitude does not (3.6x under the original modelling assumption, 5.5x under random effects). Kaiser & Menkhoff (2017, 2020) are the same team's earlier passes and agree. No independent team has re-run the coding, which is the outstanding gap: both sides of this dispute have now been adjudicated only by one of the two disputants.