The minimum maximum t-statistic re-randomization method (as described in Bruhn et al., 2009) will be employed. Using the pre-specified stratification variables described below, one thousand draws of treatment assignments will be taken. For each draw, the maximum t-statistic of the t-tests for the stratification variables will be recorded. The one draw with the minimum maximum t-statistic will be chosen as the final treatment assignment.
In the pilot experiment, randomization was stratified by four variables: Hindi literacy, baseline work performance, village population, and subcenter. Literacy was directly measured during the baseline survey through an assessment involving asking the ASHAs to read a Hindi sentence. Baseline performance was operationalized as the number of client visits reported by each ASHA during her first 90 days of using CommCare. Because ASHAs received CommCare training in batches, the dates corresponding to the 90-day interval vary by ASHA, and because some were trained as late as October 2012, the interval was limited to 90 days to avoid censoring the performance data of these late trainees. ASHAs are assigned by the government to cover specific villages, and the village population variable is based on data from the 2001 Indian national census, which was the latest publicly available census at the time. Finally, subcenters are the first-level public health facilities in the local health system. They are staffed by auxiliary-nurse midwives (ANMs), and each subcenter takes as its catchment area a designated, contiguous cluster of surrounding villages. ASHAs are linked to subcenters by virtue of the village that they work in, and they are loosely supervised by the ANM in the subcenter.
In the replication experiment, randomization will be stratified by six variables: Hindi literacy (as described above), intrinsic motivation, extrinsic motivation, prosocial motivation, and number of follow-up visits in the previous 4 and 12 months. The three motivation variables will be based on psychometric scales administered during the baseline survey. Each variable is an unweighted average of the Likert-scale responses to the items constituting each scale. The decision to pre-specify these variables for sub-group analysis is based on findings from the pilot experiment which suggest interaction effects between the treatments and psychometric traits of the health workers. Finally, since one of the main outcome measures is home visits carried out by the health worker, we will stratify on baseline/lagged performance over the previous 4 and 12 months (to capture both short- and medium-term performance).