Do saving nudges cause borrowing? Evidence from a mega-study


A vast number of policies aiming to increase savings are currently in place, with many of them involving nudges. These policies are based on the assumption that savings are financed with decreases in consumption. However, when policy makers or researchers evaluate these interventions, they often focus on immediate savings outcomes without looking at where the money comes from, even though a non-trivial fraction of households hold liquid savings while simultaneously carrying credit card debt (the so-called credit card debt puzzle). Understanding whether or not increases in savings are financed with high-interest debt is of critical importance for policy makers and researchers alike because if they are, consumers may end up worse off than when they started. 

In a recent working paper, we investigate whether or not saving nudges lead to increases in credit card borrowing. We use data from a large-scale field experiment paired with comprehensive and accurate panel data on individual bank accounts and credit cards. The data originated from a bank in Mexico, Banorte, which ran a randomized experiment with 3,054,438 customers. Of these, 2,679,545 customers were treated with weekly or bi-weekly ATM and SMS messages encouraging them to save for 7 weeks during fall 2019, while the remaining 374,893 customers received no messages.

We pay particular attention to the borrowing response of individuals with the largest predicted response to the nudge. For them, potential unintended consequences of saving nudges are all the more likely. To identify them, we predict treatment effects at the individual level with a causal forest. The causal forest allows us to avoid the over-fitting problems we would inevitably encounter if we manually searched for subpopulations with large treatment effects. A manual search would misleadingly attribute large treatment effects to subpopulations in which some observations show unusually large savings due to idiosyncratic shocks that could affect borrowing outcomes as well. In contrast, the causal forest is based on a repeated split-sample procedure, in which one sample is used to partition the covariate space and another is used to estimate the corresponding treatment effects. This eliminates the possibility that pre-treatment covariates predict a large treatment effect due to idiosyncratic shocks that could also affect other outcomes, including borrowing decisions.

In figure 1, we split individuals into quartiles of predicted treatment effects resulting from the causal forest and calculate the actual treatment effects for each group. In essence, the forest identifies two groups of individuals: a large first group with a zero treatment effect (quartiles 1 to 3 of the predicted treatment effects), and a smaller second group with positive and significant treatment effects (the top quartile of the predicted treatment effects).