Teaching Money, Teaching Minds: The Academic Returns to Financial Education

Students assigned to the treatment group are offered an online financial education course designed specifically for university students by eQwa, an experienced and certified provider of financial education. The course content was developed in collaboration with university student representatives to ensure relevance and relatability. It is delivered through the university’s Moodle platform, where eligible students access it using their institutional credentials. The course is primarily in Italian, with English subtitles to ensure accessibility for international students.

Students in the control group will not have access to the course until November 2026, after all data collection (except for possible long-term effects; see below in Experimental details) is complete. Participants are clearly informed that assignment to the early-access (treatment) or delayed-access (control) group is determined randomly. To assess potential information spillovers, the endline survey will include self-reported measures of exposure to financial education resources and sharing of knowledge by peers.

2025-12-16

2026-06-30

• Academic performance. We measure academic performance using a standardized index (z-score) that combines (i) students’ accumulated ECTS credits and (ii) the average grades for courses passed. This measure combines both the quantity and the quality of academic performance. The index is standardized within the study sample (mean = 0, SD = 1) and reflects academic achievement during the period following the intervention. Specifically, we include credits earned and grades obtained from the start of the intervention until the end of the academic year examination period (October 2026), prior to the new academic year and before the control group gains access to the course. 1 variable. Measured using administrative data.
• Financial literacy. Standardized z-score financial literacy index using the “Big three” (Lusardi and Mitchell 2011) questions. We code correct answers as 1 and every other answer as 0. Also measured at the baseline and added as a control to the regression (see below). 1 variable. Measured using survey data.

Multiple hypothesis testing (primary hypotheses). Since we test two primary hypotheses, we report both conventional (“per-comparison”) p-values and p-values adjusted for multiple hypothesis testing. To address potential concerns about false discoveries arising from testing multiple outcomes, we apply the correction method proposed by Barsbai et al. (2020) using the authors’ mhtreg Stata code. This approach extends the procedure developed by List, Shaikh, and Xu (2019) to allow for multiple-hypothesis correction in multivariate regression settings. The method explicitly accounts for the dependence structure among hypotheses, thereby improving statistical power relative to traditional corrections that assume independence (e.g., Bonferroni (1935); Holm (1979)).

By family:

Academic performance (2 variables)
• Academic performance (disaggregated). Z-score for the primary outcome sub-components (grades and ECTS credits) separately. 2 variables.

Well-being (3 variables)
• Mental health. Standardized z-score mental health index composed of PHQ2 and GAD2 (sometimes referred to as PHQ4 combined). Each question is coded on a scale from 0= not at all to 3=nearly every day. Prefer not to answer is treated as a missing value. Responses coded as standard in the literature. We also consider a broader index that includes all PHQ-8 and GAD-7 questions. Also measured at the baseline and added as a control to the regression (see below). 1 variable.
• Loneliness. Standardized z-score loneliness index composed of UCLA 3-item loneliness scale. Each question is coded on a scale from 0= not at all to 3=nearly every day. Prefer not to answer is treated as a missing value. Responses coded as standard in the literature. Also measured at the baseline and added as a control to the regression (see below). 1 variable.
• Life satisfaction. Numerical 1-10. Responses to a question “All things considered, how satisfied are you with your life as whole these days?” ranging from “completely dissatisfied” to “completely satisfied”. Also measured at the baseline and added as a control to the regression (see below). 1 variable.

Self-control (1 variable)
• Self-control. Standardized z-score self-control index using Falk et al. (2016) willingness to delay gratification using four questions selected. We use the following questions: 1) “I am a person who often acts too hastily.” 2) “I have difficulties resisting temptations.” 3) “I have a tendency to procrastinate on things, even though it would be best to take care of them quickly.” 4) “I am a person who follows my plans.” Each question is coded on a scale from 1 = completely disagree to 10= completely agree. Measured only at endline. 1 variable.

Trust (1 variable)
• Trust. Standardized z-score trust index combining questions on confidence in organizations (government, municipality, National Institute of Social Security, state, banks, other financial operators, charitable or humanitarian organizations, universities). Each question is coded on a scale from 0=not at all to 3=a great deal. Prefer not to answer is treated as a missing value. Also measured at the baseline and added as a control to the regression (see below). 1 variable.

Financial behavior (4 variables)
• Liquidity constraints. Standardized z-score index combining questions on difficulty dealing with unexpected expenditures (on their own and with possible help of others). 1 variable.
• Financial stability. Indicator. Answering “perfectly predictable” or “rather predictable” to a question “Are your earnings predictable from one month to the other?” Also measured at the baseline and added as a control to the regression (see below). 1 variable.
• Savings. Indicator. Answering “Yes” to a question “Are you able to save money in a typical month during the academic year?”. Also measured at the baseline and added as a control to the regression (see below). 1 variable.
• Financial confidence. Indicator. Answering “rather confident” or “very confident” to a question “How confident do you feel in managing your personal finances?” Also measured at the baseline and added as a control to the regression (see below). 1 variable.

All secondary outcomes are measured using survey data.

Multiple hypothesis testing (secondary hypotheses). Alongside conventional (“per-comparison”) p-values, we report p-values adjusted for multiple hypothesis testing to account for the increased likelihood of false discoveries when testing several outcomes within a family of outcomes simultaneously. Specifically, we apply the procedure of List, Shaikh, and Xu (2019), implemented through the mhtreg module of Barsbai et al. (2020), which extends the List-Shaikh-Xu approach to multivariate regression frameworks. We correct for numbers of hypotheses, corresponding to the secondary outcomes within each of the four families (academic performance, well-being, preferences and beliefs, financial behavior). This method accounts for dependence across hypotheses and therefore retains greater power than corrections assuming independence (e.g., Bonferroni 1936; Holm 1979). Researchers with a priori interest in specific outcomes may focus on the corresponding unadjusted p-values, whereas others should rely on the multiple-testing–adjusted results for a more conservative inference.

Sample. We recruit from the entire population of students enrolled at the University of Milan—about 56,000 students with an active student account across all fields and study programs—to participate in a free online financial education course developed by the university in collaboration with a company specializing in financial education. Recruitment is conducted through multiple communication channels, including in-class presentations by research assistants featuring short promotional videos, flyer distribution, posters across university premises, and a coordinated email and social-media campaign. The main advertising campaign took place during October and November 2025. Baseline data were collected during the recruitment process, prior to randomization.

A smaller group of participants (N = 233) was recruited in an earlier wave, when a baseline survey was administered throughout similar channels to all university students in May 2025. The survey promised a possible future offering of a course on financial education, without specifying the exact details. The interest in taking the survey and providing consent for future contact with uncertain outcome was thus lower. Students participating in this wave were asked to register together with the second, major registration wave.

As a further incentive, participants are informed that they can obtain an official certificate of participation and an open badge suitable for inclusion in their résumé. In addition, the first 50 students (25 in each edition) who complete the course are offered a free one-on-one consultation with a certified financial education trainer.

Students are clearly told that access to the course will be staggered: half receive early access (treatment group) and half receive delayed access (control group). This arrangement is presented as a pilot phase necessitated by capacity constraints.

All participants provide explicit informed consent for: (i) participation in surveys; (ii) collection of learning-engagement data from the Moodle platform; and (iii) linkage of these data with administrative academic records provided by the university.

In total, N = 2,237 students registered for the course. Half (N = 1,115) were randomly assigned to the treatment group, and the remainder to the control group. Details on randomization are provided below in the “Randomization Method” section.

Sample selection. To assess selection into the study, we will compare summary statistics from administrative records for the study sample with those of the broader student population who did not register.


Experimental manipulation. See section “Intervention(s)” above.

Timeline.
Participant recruitment: First (small) wave in May 2025, second (major) wave in October/November 2025
Baseline survey data collection: First (small) wave in May 2025, second (major) wave in October/November 2025
Randomization: November 2025
Intervention: December 2025-March 2026
Endline survey data collection: June 2026
Administrative data endline extraction: November 2026

Outcomes. See above for both primary and secondary outcomes.

Other data collected.
We are aware that students may access other types of financial education, being it from classmates or other sources. We capture this by asking students at the endline:
Have you completed a course on financial education in the past 6 months? [Yes / No]
[If yes]: Was it a course offered by the University of Milan? [Yes / No]
Did you discuss the contents of the course on financial education offered by the University of Milan with a fellow student? [Yes, we extensively discussed the details of the contents / Yes, we discussed the contents vaguely / No, we did not discuss the contents with anyone]

Randomization balance. Following the concerns about balance test reporting (e.g., Bruhn and McKenzie 2009), we use an omnibus joint test of orthogonality to test for balance using all baseline data described above. In a single OLS regression, we regress all the variables on the Treatment group indicator. Then we test for joint significance of all the estimated coefficients using an F-test.

Manipulation checks. The survey (both baseline and endline) includes knowledge questions for contents of the course. We measure if participation in the financial education improves knowledge on these questions at the endline. We estimate equation (1) defined below using the knowledge z-standardized index as an outcome. Specifically, we recode questions on (i) knowledge about what to do in case of missing a mortgage or loan payment, on (ii) what the eligibility criteria for obtaining an invalidity pension are, on (iii) comparability of risks of different financial tools, and on (iv) measuring the value of social security contributions to retirement pension. Respondents randomly responded to either (i) or (iii), and (ii) or (iv). Response to each answer is coded as 1 if the reply is correct, while it is 0 if the response is incorrect or the student responds that they do not know the answer.

Using equation (1) below, we also study effects of the treatment on responses on the indicator of “Have you completed a course on financial education (offered by the University of Milan)?”

Standard Errors. Standard errors are clustered at the individual level.

Hypotheses.

P1. Academic performance (z-score index of ECTS + grades)
H0: The mean academic-performance index is the same across arms.
H1: The mean academic-performance index is different in treatment than control.
Rationale: Financial education should reduce finance-related stress and improve planning/time management, freeing cognitive bandwidth and raising both course completion (ECTS) and grades.

P2. Financial literacy (Big Three composite, z-score or 0–3 raw)
H0: The mean financial-literacy score is the same across arms.
H1: The mean financial-literacy score is different in treatment than control.
Rationale: Indirect effect of the course.

S1a. Grades (z-score of average grades, component of primary index)
H0: Mean grades (z) are equal across arms.
H1: Mean grades (z) are different in treatment than control.
Rationale: Same as for P1.

S1b. ECTS credits (z-score of credits earned, component of primary index)
H0: Mean ECTS (z) are equal across arms.
H1: Mean ECTS (z) are different in treatment than control.
Rationale: Rationale: Same as for P1.

S2. Mental health (PHQ-2 + GAD-2 index, standardized; higher = more symptoms)
H0: Mean mental-health index is equal across arms.
H1: Mean mental-health index is different in treatment than control.
Rationale: Improved financial control reduces stress/anxiety.

S3. Loneliness (UCLA-3 index, standardized)
H0: Mean loneliness index is equal across arms.
H1: Mean loneliness index is different in treatment than control.
Rationale: Reduced financial strain can facilitate participation/engagement and reduce isolation.

S4. Self-control (4-item index from Falk et al., standardized; higher = more self-control)
H0: Mean self-control index is equal across arms.
H1: Mean self-control index is different in treatment than control.
Rationale: The course emphasizes planning and resisting temptations, improving intertemporal control.

S5. Trust in organizations (multi-item index, standardized; higher = more trust)
H0: Mean institutional-trust index is equal across arms.
H1: Mean institutional-trust index is different in treatment than control.
Rationale: Greater understanding of financial/welfare systems can increase perceived fairness and competence of institutions.

S6. Liquidity constraints (index on ability to handle unexpected expenses, standardized)
H0: Mean liquidity-constraint index is equal across arms.
H1: Mean liquidity-constraint index is different in treatment than control.
Rationale: Better budgeting and knowledge of support options reduce binding short-run constraints.

S7. Savings (indicator: “able to save in a typical month”)
H0: The proportion reporting they can save is equal across arms.
H1: The proportion is different in treatment than control.
Rationale: Improved budgeting and reduced unnecessary spending increase the likelihood of positive month-end balances.

S8. Financial confidence (indicator: “rather confident” or “very confident” in managing personal finances)
H0: The proportion of students reporting that they feel confident in managing their personal finances is equal across arms.
H1: The proportion of students reporting that they feel confident in managing their personal finances is different in treatment than control.
Rationale: Financial education should increase students’ sense of competence and control over financial matters.


Regression analysis.

We estimate intention-to-treat (ITT) effects of assignment to the treatment group on each primary and secondary outcome. The main specification is a simple linear model:
Y_i=α+β" " T_i+θC_i+γ^' X_i0+δ_(f(i))+ε_i, (1)

where
Y_i is the outcome for student i (academic performance, financial literacy, or a secondary outcome),
T_i is an indicator equal to 1 if student i was assigned to the treatment group (early access), and 0 otherwise (control group, delayed-access),
C_i is an indicator for a one-on-one consultation with a certified financial education trainer for the first 50 completers.
X_i0 is a vector of strata variables and baseline covariates (see below),
δ_(f(i)) are faculty fixed effects, and
ε_i is the error term.
The coefficient of interest, β, captures the ITT effect—the causal impact of assignment to treatment under randomization.

Not available

Randomization. Software-based randomization. Stratified randomization at the individual level.

Stratification. We stratify based on the four dimensions (each with two levels) below as these are most predictive of the outcomes at the baseline:
• Gender (female yes / no) - due to substantial gender differences in baseline financial literacy, we expect stronger effects for female students. Data from administrative records.
• ISEE (household wealth status indicator with lower values indicating lower per-capita wealth) median split. Data from administrative records. Whenever ISEE is not reported (missing value; individuals can withhold this information but cannot claim welfare benefits if ISEE not reported), we treat these individuals as belonging to the above median group of individuals who have a non-missing value for the ISEE as most characteristics for the two groups overlap. This is to be expected as the incentive for reporting ISEE is higher for individuals with lower wealth.
• “Maturità” (final high school exam grade or its foreign equivalent) grade median split. Data from administrative records.
• BA/MA degree student. Data from administrative record.
A small number of students have missing administrative date (due to late enrolment). We create an additional strata for these individuals.

Specifically, using Stata, we re-run the stratified randomization 20,000 times (seed 840801) and let the computer pick an outcome that produces the lowest treatment difference for baseline levels of (i) financial knowledge index, (ii) GAD-2 score, (iii) PHQ-2 score, and (iv) the average of all grades (from administrative data). We define the lowest treatment difference by summing up the absolute values for the four treatment differences and finding the lowest value. We then use this outcome for the final treatment assignment.

Individual students

No

2,237 students

2,237 students

1,115 treatment group
1,122 control group

With N = 2,200 students (1:1 allocation), assuming α = 0.05 and power = 0.80, we can detect a minimum effect size of approximately 0.12 SD on the standardized primary outcomes.