Experimental Design Details
Below we explain how our experimental design will allow us to answer our primary research questions. This will also clarify the marginal contribution of each treatment arm.
The experimental variation will allow us to test for the effectiveness of a variety of incentives, in order to identify the mechanisms through which barriers to referrals and treatment are alleviated. To measure the effects of our interventions, we will estimate versions (e.g., probit) of the following econometric model:
yij= α + βE ecnourage + βR reward + βB bonus + βN names +βA anonym + γXi +Tt + εijt
where i denotes the individual patients, j the randomly assigned treatment condition, and t the month of the intervention. yij represents the outcome variable of interest (e.g., yi = 1 if patient i made a referral, and 0 otherwise; or yi = 1 if patient i referred someone who tested positive for TB, and 0 otherwise). The key parameters of interest on the right-hand side of model (1) are the treatment effects βk, with the constant α representing, all else equal, the response in the pure control condition T0. First of all, the estimated coefficient βE will reveal whether encouragement to refer new suspects has an effect, even in the absence of incentives. In fact, an explicit appeal accompanied by referral cards might increase the salience of the request, thereby increasing referrals. Second, comparing βR and βE will determine whether financial rewards are effective at inducing current patients to find and refer new suspects (Q1). Third, comparing βB vs. βR will allow us to assess the extent to which unconditional incentives induce strategic behavior on the part of patients to refer individuals who are not sick only to obtain the reward (Q2a) and whether the current patients have concrete information on the health status of their social contacts (Q2b). Specifically, if patients in the unconditional incentive arm behave opportunistically by disproportionally referring people who are not sick, the number of referrals made will be higher and the share that test positive for TB smaller compared to those for patients in the bonus condition. Such a result will also indicate that patients do have information about the health condition of their social contacts (or an obtain it by exerting more effort). Fourth, comparison between βN and βE will reveal whether outreach by peers (current patients) or outreach by health workers is more effective in enrolling new suspects (Q3). Since current patients may also contact new suspects when they receive incentives based on intake of those new suspects, we can also estimate differences-in-differences specifications that include the interaction between eligibility for a reward, and outreach by peers vs. health workers. Finally, the estimated coefficient βA will inform us about the role of stigma and other “social costs” in making current patients reluctant to make referrals (Q4). For example, if we found that the share of referrals who are TB positive is higher when the referrer remains anonymous, that would indicate that stigma and related social costs are indeed a barrier that is making patients reluctant to approach others in their social circles who are likely also sick. Tt are month fixed effects, included to control for seasonality effects and other unobservable factors associated with the different intervention periods. Individual characteristics (including demographics such as age and gender, and other characteristics such as personality traits, experience with the disease and the clinic, etc.) will be reported in the vector of controls Xi. εijt is the error term. Because the randomization was done at the center level, we adjust the standard errors by allowing for clustering by center.