RDRS is organized administratively into branch offices. Each branch has a set of villages in its catchment area defined by the geographic road) distance to the branch. Branch catchment areas are non-overlapping so each village in the experiment can be allocated to a single branch.
Treatment, defined as the offer of a migration subsidy (incentivization), occurs at the village level. Every eligible household in a treated village is offered the migration subsidy. Our randomization strategy places villages into four categories:
1. Incentivized: Villages in which the migration subsidy offer is made.
2. Spillover: An untreated village geographically in the middle of a group of treated villages.
3. Spillover-control: An untreated village that belongs to a branch that includes treated villages, but is surrounded by other untreated villages.
4. Pure-control: An untreated village that belongs to a branch that has no treated villages.
To achieve this classification, we randomize at two levels. First, we randomly divide branches into treated and control. Branches assigned to be control contain only pure-control villages. Branches assigned to be treated contain the other three types of villages.
Within treated branches, our randomization strategy generates a treated sector (designated as incentivized), a single untreated village within the treated sector (designated as spillover), and an untreated sector (designated as spillover-control). In accordance with the RDRS workplan, the treated sector comprises one third of the villages in a treated branch. For assignment, we identify the centroid of the branch catchment area and then project each village onto a circle around the centroid. We randomly select one village on this circle and designate it as spillover. We then define the incentivized sector as the third of the circle surrounding the spillover village. In effect, we create a “pie slice” (designated as the incentivized sector), with one village in the middle left untreated as spillover.
This strategy stems from the fact that incentivization may generate spillovers onto nearby villages. Spillovers come from three main sources. First, we find in previous work that migrants generally travel in groups and migrants from geographically close sources tend to go to geographically similar destinations. Therefore, inducing migration in one village may lower the returns to migration from nearby villages through the destination labor market. Second, labor markets may be locally integrated. Out-migration from an incentivized village may lower labor supply, raise wages, and induce in-migration from nearby villages. Third, household risk sharing networks may extend beyond village boundaries. An incentivized household may share the benefits of migration with others in nearby villages.
Our randomization strategy creates multiple types of non-incentivized villages to evaluate the geographic extent of these spillovers. The spillover village in a treated branch is on average closest to incentivized villages and therefore most exposed to treatment spillovers. At the other extreme, we believe pure-control villages are sufficiently far from treated regions that their workers are no more exposed to treatment spillovers than workers from anywhere else in the country. Spillover-control villages falls between these extremes and allow us to estimate how quickly the spillovers dissipate with distance.
For evaluation, we plan to survey (record) households in only a subset of incentivized, spillover-control, and pure-control villages. Survey villages are selected as follows:
1. Incentivized: One randomly selected village in the incentivized sector per branch.
2. Spillover: The village in the middle of the incentivized sector, designated as spillover in each treated branch.
3. Spillover-control: The village diametrically opposite the spillover village on the circle projection.
4. Pure-control: One randomly selected village in each untreated branch.
The randomization design generates the four experimental categories while ensuring that the status of a village is uncorrelated with other geographic characteristics. In particular, treatment status is orthogonal to the geographic density of villages and their proximity to a branch’s boundary. The survey design preserves orthogonality between likelihood of being surveyed and geographic characteristic as well. Unfortunately,
in maintaining this orthogonality, we cannot guarantee that spillover villages are closer to the incentivized sector than spillover-control in every treated branch. We do not account for proximity to the centroid in randomization, meaning that a very central spillover-control village may be closer to the treated region than a peripheral spillover village. However, on average, spillover villages are closer to incentivized villages
than spillover-control villages. Similarly, an incentivized village in our sample is on average closer to the incentivized sector than a spillover-control village, but slightly father on average than a spillover village.