AEA RC T Registr y

The American Economic Association's registry for randomized controlled trials

2015-10-06

2015-10-31

Our key outcome variable will be bank card and credit card spending during October 2015 in the key category (shopping, bars/restaurants, or supermarket) with the largest budget. Because the distribution of this variable will be highly skewed, we will focus on the logarithm of the variable and also plan to winsorize it (at the 99th percentile of its distribution, unless the distribution is so extreme that the 95th percentile is more appropriate). In order to increase statistical precision, we also plan to look at changes in the variable relative to September 2015 (it may not be possible, however, to obtain September 2015 data).

As secondary analyses, we plan to conduct quantile regressions to study differences in the full distribution of the key outcome variable across treatment groups. We also plan to examine secondary outcome variables such as an indicator for whether or not the goal was achieved and actual spending as a proportion of budgeted spending.

Finally, tertiary analyses will examine overall spending (as opposed to spending only within the key category with the largest budget).

The experimental design includes three groups:
1) Broad budget goal - users receive push notifications providing feedback about spending relative to a broad budget goal
2) Narrow budget goal - users receive push notifications providing feedback about spending relative to a narrow budget goal
3) Control - users do not receive push notifications providing feedback on spending relative to budget goals

The experimental design includes three groups:
1) Monthly treatment group (~4000 users) - in the Monthly treatment group will receive information about their spending in the key category since the beginning of the month and will be reminded of their monthly spending goal
2) Weekly treatment group (~4000 users) - users in the Weekly treatment group will receive information about their spending in the key category since the beginning of the week and will be reminded of their weekly spending goal
3) Control group (~1000 users) - users do not receive push notifications providing feedback on spending relative to budget goals

Randomization is done in an office by a computer.

The unit of randomization is the individual user.

No

The sample will include approximately 9000 users (randomization is at the user level).

The sample will include approximately 9000 users (randomization is at the user level).

~4000 in each of two treatment groups and ~1000 in control group

We do not have detailed data with which to perform precise power calculations, but we anticipate that the standard deviation of log winsorized spending in category in the month (in Brazilian reais) will be about about 1. In a comparison of the two treatment groups (~4000 users each), we have ~77% power with alpha=0.05 to detect a change of 0.06 in the variable (a 6% spending change). In a comparison of one of the treatment groups (~4000 users) to the control group (~1000 users), we have ~81% power with alpha=0.05 to detect a change of 0.1 in the variable (a 10% spending change).