We selected 28 Private Service Providers (PSPs) and 3 SILC groups per PSP. We then randomly allocated the 28 PSPs (in a computer program) to one of the two arms of the CRT: that is, 14 PSPs were allocated to COFE+SILC and 14 PSPs randomly allocated to SILC alone. The 3 SILC groups within each PSP were therefore allocated to the arm to which that PSP had been allocated.
Regarding the selection of SILC groups within each PSP, we first subset down to only SILC groups with at least 33% of the member households with orphans and vulnerable children (OVC), except when a PSP had fewer than 3 groups with at least 33% OVC. Then we sorted by SILC cycle within each PSP, preferring SILC groups at cycle 1, and kept the three SILC groups with the three largest proportions of OVC at the earliest cycle. Finally, to subset down to 28 PSPs (out of 29 available), we kept the 28 PSPs with the largest number of OVC members.
We then used a constrained randomization routine to allocate the selected 28 PSPs to the two arms of the trial. Our goal was to balance the two arms on important PSP-level characteristics hypothesized to be related to the financial outcomes of interest so as to avoid confounding. We note that randomization is expected to achieve this “balancing” (i.e. to avoid confounding) when a large number of PSPs are randomized but because CRTs usually randomize a small number of clusters (i.e. fewer than 40 total clusters), it is important to use a strategy to help ensure balance is achieved. In this way, we would expect that the “Table 1” for this trial would show comparable characteristics between the COFE+SILC arm and the SILC arm.
The four PSP-level characteristics chosen for balancing were: proportion of group members that are OVC; proportion of group members that are using Edufund (i.e. a shared fund that a SILC group could implement to focus on paying school expenses); PSP experience in months (i.e., months since their first training); and region (Gomba vs. Mityana).
We gave greater weight to OVC and Edufund, since these were determined to be most important of the four. This ensures that these will be the most balanced of the four variables.
In the constrained randomization routine, we used a constrained randomization space cutoff of 0.1, and simulated 50,000 randomization schemes (since the total number of randomization schemes with 28 clusters is >40,000,000). We checked randomization validity, and found acceptable validity given that “a reasonable constrained space should ensure each cluster pair appears in the same arm (and in different arms) for at least 25% of times and at most 75% times”. It fell slightly outside of this range, but not by much (19%).