- Original authors / creators
- Shaoran Yu; Zhao Zhang
- Original publication
- Frontiers in Psychology · Frontiers Media SA · 2026-08-10
- License & retained notices
- CC BY 4.0
Copyright © 2026 Yu and Zhang.
Article-specific license evidence ↗Exact license notice
https://creativecommons.org/licenses/by/4.0/This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms. Article-specific license URI: https://creativecommons.org/licenses/by/4.0/ - Edition changes
- Narrative paragraphs and references retained with whitespace normalization; figures, tables and display equations remain in exact retained XML and original source, not the reader edition.
- Evidence limits
- Adjacent research, not a directly sampled family-office study. Applicability to a family office requires separate assessment. Kaiser et al. (2022) noted that long-term follow-up evidence remains relatively sparse, with outcomes measured a median of roughly 6 months after treatment and few studies tracking effects at 1 year or longer. The retention of treatment effects at 6 months is particularly relevant for educational policy because brief curricula are cheap to scale but valuable only insofar as their effects endure. Bover et al. (2026) provide a rigorous randomized demonstration of short-run gains among Spanish high-school students, and the few studies with longer horizons (e.g., Frisancho, 2023, in Peru; Bruhn et al., 2016, in Brazil) document both immediate gains and their partial attenuation over time, but comparable evidence in Chinese vocational settings is absent. Moreover, the differential decay rates of cognitive vs. psychological gains—e.g., whether self-efficacy or knowledge dissipates first—remains an open question with direct curricular design implications. This study has eight limitations that qualify our conclusions and directly inform avenues for future work.Small cluster count (J = 4) constrains cluster-level inference. Although individual-level OLS with HC3 robust SE produced treatment effects with p < 0.05 for FLS, FSES, and BDQ, the enumerated wild cluster bootstrap—our primary cluster-respecting inference—produced p values uniformly at its structural floor (2/16 = 0.125) for all immediate effects. This is a structural feature of any RCT with so few clusters and is not a substantive null result; rather, it reflects the limited information content of cluster-level variation when only four classes are observed. We have reported both inference frameworks transparently. Future replications should aim for at least 8–12 clusters per arm, which would enable conventional cluster-level significance assessment and substantially strengthen causal inference. Our findings should be interpreted as effect-size estimates supported only under individual-level inference; cluster-level inference is bounded by design, and causal claims at the unit of randomization are not supported. Exact cluster-level randomization inference reinforced this reading: the observed allocation produced the largest treatment effect among all six possible class-to-arm assignments for every immediate (T2) outcome, yet—because randomization occurred over only four clusters—the two-sided randomization p value has a structural floor of 1/3 (all six T2 effects: p = 0.333), so conventional cluster-level significance is unattainable by design rather than by absence of effect.Concurrent measurement of mediators and outcome at T2 limits causal mediation interpretation. FSES, CFC, and BDQ were all measured immediately post-intervention, leaving open the possibility of reverse causation (e.g., a high-quality decision experience increasing self-efficacy rather than self-efficacy enabling the decision). While our analysis adjusts for baseline BDQ_T1 and aligns with the predominant cross-sectional mediation tradition in financial education research (cf. Çera et al., 2021), strict causal mediation requires temporal precedence of M before Y. In addition, random assignment identifies only the treatment → mediator paths; the mediator → outcome paths remain observational and vulnerable to unmeasured confounding. The sensitivity analysis reported in Section 4.6 quantifies this vulnerability (indirect effects nullified at ρ = 0.147 for FSES and ρ = 0.051 for CFC), and we accordingly treat H5 as exploratory. Future designs should measure mediators during weeks 2–3 of the intervention—before the T2 outcome assessment—to address this temporal concern directly.Single-instructor design precludes instructor blinding. The PI delivered both treatment and active-control sessions, ensuring instructor-quality consistency but preventing instructor-blinding to study hypothesis. While our active-control design and survey administration by RA mitigate gross expectancy effects, we cannot fully rule out subtle differential investment in the two arms. Expectancy and demand effects therefore remain uncontrolled for all self-report outcomes, and the between-arm difference in perceived course quality (4.16 vs. 3.78) is equally consistent with such effects; this concern compounds the effect-size caution discussed at the opening of this section. Future replications should employ separate instructors per arm with masking to study hypothesis.Single-site sampling. The trial was conducted in a single vocational college in one city, limiting external generalizability. Multi-site replication across regions, urban/rural divides, and college types is needed.The outcome index relies on simulated scenarios. The BDQ index, while showing meaningful convergent validity with self-reported planning behaviors at baseline (mean r = 0.33), reflects decision-making tendencies in simulated scenarios rather than actual financial behavior in the field. Future studies should incorporate administrative data—e.g., actual bank transactions, savings records, credit applications—to strengthen behavioral inference.FSES Chinese translation needs formal psychometric validation. While translate-back-translate procedures and pilot reliability assessment (N = 40, α = 0.81) provide preliminary support, formal CFA validation in a large Chinese vocational sample remains a needed step before broad deployment. Relatedly, the FLS showed modest reliability (KR-20 = 0.68), which attenuates and destabilizes the literacy-outcome estimates specifically; both measurement limitations qualify the corresponding conclusions wherever they appear.Six-month follow-up may be insufficient for durable financial-behavior assessment. While our T3 measurement contributes to the small but growing long-term follow-up literature, ideal financial-behavior research would track participants over 2–5 years to assess durable change in cumulative outcomes. A two-year follow-up extension is planned.Substantial residual direct effect (59% of total) indicates unmeasured mechanisms. Although both FSES and CFC showed indirect effects under individual-level inference, 59% of the total treatment effect on BDQ operates outside these two pathways. Candidates for future investigation include mental accounting (Thaler, 1985), financial socialization (Shim et al., 2009), and peer-effect channels Bursztyn et al. (2014) through which financial literacy training may propagate within social networks.
- Retraction / correction check
- No retracted-publication label in the retrieved Europe PMC record or Crossref update-to/updated-by notice. XML metadata screened for retraction. Publisher article page searched for explicit correction/retraction/concern notices; see verification record. These bounded checks do not cover unlinked or unregistered notices and are not exhaustive independent retraction clearance. Verified 2026-11-05. Automated/source checks are not human peer review.
- Preserved record
- Rights and full text checked 2026-11-05. Included in this volume on 2026-11-05. Original source dates remain unchanged.
Download the preserved source
SHA-256 (retrieved artifact): d278dbec9844a25703102dce628151bd9a41f53acb9b0a4bc41f753a378650e5
SHA-256 (preserved text edition): 737505a0c373d0266989ec09dcfb19125c2d40fc99a3f82d65d6a4fc37d3ad14