Messaging after a Bank Failure: Messaging Strategies to Stop Contagion

We propose to employ a between-subject survey experiment with depositors in the United States and Canada to test different messaging strategies in the context of a bank failure caused by a deposit run. The survey experiment will be conducted online between April and May 2025, with respondents randomly assigned to one of the four messaging conditions.

The intervention consists of three different government messaging strategies (deposit insurance information-only; conditional commitment to support uninsured deposits; and unconditional commitment to reimburse all depositors regardless of deposit holdings) relative to a control where government is silent ("passive") in the face of a bank failure caused by a deposit run.

2025-04-07

2025-05-16

The primary outcome variables is 'worry.' It captures the extent to which the interventions (different messaging strategies) attenuate or aggravate worry about having access to money relative to the passive (say nothing) control.

The 'worry' outcome will be derived from a question that asks respondents to indicate their level of worry to access to their money on a 0-10 scale, where 0 = not at all and 10 = extremely.

The secondary outcome is withdrawal behaviour - given their level of worry, (a) will respondents withdraw their money from their banking institution and (b) if they choose to move it, where would they move it to?

Another secondary outcome is whether respondents choose to share their decision with others. If they do, we will ask whom they choose to inform and through which channels they communicate their decision.

Deposit runs arise from mass shifts of deposits to other banking establishments and/or cash withdrawals. If government can effectively dampen worry and therefore, presumably, withdrawal behaviour, they can slow or stop a deposit run or bank failure from spreading. The "withdrawal" outcome will be derived from a question that asks respondents to select form one of the three options: leave money where it is, withdraw some of their money, or withdraw all their money. If respondents decide to move some or all their money, a follow up question will be asked whether they would keep it in cash or move it to another financial institution, and if so, the name of the specific institution.

Respondents are also asked whether they would share their decision with others. If they select "yes", they will indicate whom they would inform by choosing from a list of potential recipients (friends, family, co-workers, acquaintances, strangers, or others) and specify their communication method(s) (email, text, phone call, social media, blog post, news media, or others).

The experiment will begin by explaining the concept of a deposit run. It will then asked subjects to imagine that a deposit run caused a bank to fail. We will then randomly assign subjects to four messaging conditions (including one control and three treatments). Each treatment provides one additional piece of information (relative to the control). The messaging conditions map to different government communications strategies in context of a bank failure caused by bank run.

The four information conditions range from passive (no communications, "control") to information on deposit insurance ("DI-Info") to information only to a contingent commitment to reimburse uninsured deposits ("conditional") to a full commitment to reimburse all deposits regardless of amount ("unconditional"). We then ask respondents to indicate their level of worry given the scenario and whether they would leave their money where it is or withdraw some or all their money. If they withdraw, we ask where they would move it to. Additionally, we ask respondents whether they will inform others about their decisions. If they choose to do so, they are asked whom they will inform and how they will communicate.

Randomization by survey firm algorithm.

This is a survey of individuals. Individuals will be randomly allocated to one of four treatments (including control).

No

2400 individuals per jurisdiction (United States, Canada).

2400 individuals in each jurisdiction (United States; Canada)

In each jurisdiction (US, Canada): 600 for passive ("control"); 600 for information-only messaging ("DI-Info"); 600 for conditional support for uninsured depositors ("conditional"); and 600 for unconditional commitment ("unconditional").

To determine the appropriate sample size for our study, we conducted an a priori power analysis using G*Power 3.1 (Faul et al., 2009). Given our study design, we aimed to detect a medium effect size with a power of 0.80 and an alpha level of 0.05. We plan to compare mean differences in worry (0-10 scale) across four treatment conditions, while controlling for covariates (both continuous and categorical, e.g., age, income, gender, knowledge). Given this, we conducted a power analysis for an Analysis of Covariance (ANCOVA) and a Linear Multiple Regression. Power Analysis for ANCOVA We used the “ANCOVA: Fixed effects, main effects and interactions” option in G*Power to estimate the required sample size. The parameters were set as follows: • Effect size (f) = 0.25 (medium, based on Cohen, 1988, 1992) / 0.10 (small effect) • Number of groups = 4 (treatment conditions) • Number of covariates = 6 (e.g., age, income, gender) • Alpha level (α) = 0.05 • Power (1-β) = 0.80 The analysis estimated a required sample size of 179 participants (if medium effect) (or 1,095 if small effect) to achieve sufficient power. Power Analysis for Linear Multiple Regression Given that we also planned to conduct an Ordinary Least Squares (OLS) regression with worry (0-10) as the dependent variable, treatment dummies as independent variables, and the same covariates, we conducted an additional “Linear multiple regression: Fixed model, R² deviation from zero” analysis in G*Power: • Effect size (f²) = 0.15 (medium) / 0.02 (small effect) • Number of predictors = 9 (3 treatment dummies + 6 covariates) • Alpha level (α) = 0.05 • Power (1-β) = 0.80 This analysis suggested a lower sample size of 114 participants (if medium effect) (or 791 if small effect). Final Sample Size Decision This power analysis ensures adequate sensitivity to detect meaningful treatment effects while controlling for key covariates. Given that ANCOVA generally requires a larger sample than linear multiple regression, we adopted the more conservative estimate. The targeted sample size of 2,400 respondents per jurisdiction (United States and Canada) ensures sufficient statistical power for this experimental study. References Cohen, J. (1988). Statistical power analysis for the behavioral sciences. Hillsdale, New Jersey: Lawrence Erlbaum Associates. Cohen, J. (1992). A power primer. Psychological Bulletin, 112(1), 155–159. Faul, F., Erdfelder, E., Buchner, A., & Lang, A.-G. (2009). Statistical power analyses using G*Power 3.1: Tests for correlation and regression analyses. Behavior Research Methods, 41, 1149-1160.