For each outcome variable, the PIs estimated two versions of the same general specification. In the general specification, the placebo group is the omitted group. This allows the regression coefficient on the treatment indicator to be interpreted as the effect of the corruption information treatment independent of the effect of receiving a flyer. All estimates include municipal fixed effects. For binary outcome variables, the PIs used a linear probability model. For turnout variables, the PIs calculated outcomes with regard to registered voters only. In addition, the PIs included a baseline poverty index at the precinct level and turnout from the previous election as controls for turnout outcomes.
In the first version, the PIs included linear and quadratic interactions between the treatment indicator and the proportion of public funds spent irregularly or corruptly in that precinct's municipality. In the second, the PIs included interactions between the treatment indicator and dummy variables indicating if the level of corruption that precinct's municipality was low, moderate, or high. Low levels of corruption were defined as 0% to 33% of public funds spent in a corrupt way, moderate levels were defined as 34% to 66% spent in a corrupt way, and high levels were defined as more than 66% spent in a corrupt way. The PIs also re-estimated all models without state capitals and where the failure-to-treat rate was high, in order to deal with spillover effects and noncompliance bias respectively. The results proved robust.
Data collection consisted of electoral results at the precinct level, census baseline demographic characteristics, an endline survey conducted 10 days after the municipal elections, and information on candidates' employment prior to the 2009 elections.