// UMD ECON-626 // FALL 2019 // PAMELA JAKIELA AND OWEN OZIER // Activity: IV in an RCT, Treatment on the Treated // Data based on Brudevold-Newman, Honorati, Jakiela, and Ozier (2017) // Evaluation of an active labor market program for young women in Nairobi // Three randomly-assigned arms: control, cash grants, and gem // GEM is the IRC's Girls Empowered by Microfranchise training + cash + mentoring program // load data use gemdata.dta // 1. Consider two variables measured at baseline and endline: voced and selfemp // voced is a dummy for having received vocational training // selfemp is a dummy for self-employment // 1a. In the control group, how correlated is b_voced with e_voced? What about // b_selfemp and e_selfemp? corr b_voced e_voced corr b_selfemp e_selfemp // 1b. Regress e_voced on the treatment dummies (gem and grant) controlling for // strata dummies. How do results (and, specifically, t-statistics) compare // when you run specifications with and without the baseline values of // the outcome variable? reg e_voced gem grant i.stratid, r // 1c. Repeat this exercise for e_selfemp. How do results differ? reg e_selfemp gem grant i.stratid, r // 1d. Now estimate the treatment effect on the change in vocational training or // self-employment. gen change_voced = e_voced - b_voced gen change_selfemp = e_selfemp - b_selfemp reg change_voced gem grant i.stratid, r reg change_selfemp gem grant i.stratid, r // 2a. Calculate the expected IV (Wald) estimate of the impact of training // (through the GEM program) on self-employment: use data from the control // group and the GEM treatment arm to estimate the reduced-form and the // first stage in separate regressions, noting the resulting ratio. reg e_selfemp gem if grant==0, r local ivnumerator = _b[gem] reg training gem if grant==0, r local ivdenominator = _b[gem] di `ivnumerator'/`ivdenominator' // 2b. How does this compare to the actual IV estimate of the impact of training // on self-employment (when you use ivregress 2sls)? ivregress 2sls e_selfemp (training = gem) if grant==0, r // 2c. One could also estimate the impact of launching a business (through the // program) on self-employment (using the bizlaunch variable). What // happens when you do this? Should you do this? Why or why not? ** Since assignment to the GEM program might impact those who start training and drop out, ** this will generate a biased estimate of the treatment effect of launching a business // 3. Compare the ITT and TOT estimates of the impact of cash grants on self-employment. tab gotcash if grant==1 // 97 percent of those assigned to the grant arm received the grant reg e_selfemp grant if gem==0, r ivregress 2sls e_selfemp (gotcash = grant) if gem==0, r