Selection into Sickness and Symptoms

Table 6 Effects of Healthcare Choice on Past Schooling Decisions Falsification IV OLS Currently Enrolled Attended School Past Six Months Currently Enrolled Attended School Past Six Months Formal healthcare 0.146 0.0321 0.269 0.285 Days of rainfall × no facility −0.00281 −0.000618 0.00530 0.00554 No facility 0.00337 −0.0250 −0.0291 −0.0322 0.100 0.107 0.0564 0.0609 Days of rainfall −0.0101 −0.00955 −0.0106 −0.00966 0.0238 0.0246 0.0228 0.0246 Observations 1,954 1,954 1,954 1,954 Mean of dependent variable 0.626 0.821 0.626 0.821 Notes: Robust standard errors in parentheses p 0.01, p 0.05, p 0.1. See Table II for additional comments. In Columns 1 and 2 of Table 6, we report second-stage IV results of the effects of formal healthcare on the probability of being currently enrolled and attending school at all in six months prior to survey, respectively. Columns 3 and 4 report results from analogous reduced form regressions of these outcomes on the interaction instrument. We see no evidence of an effect of formal healthcare, as instrumented by the interaction of facility existence and days of rainfall, nor of an effect of the instrument directly on past schooling outcomes. Columns 3 and 4 are also effectively selection equations, whose coefficient esti- mates can be used to test for whether selection into various definitions of enrollment varies with the instrument. The results indicate that the instrument does not signifi- cantly shift the selection probability for either sample.

C. Selection into Sickness and Symptoms

If the instrument affects the probability of reporting an acute illness or the severity and symptoms of illnesses reported, it might have an effect on outcomes outside of its effect on healthcare choice. In particular, due to the self-reported nature of the data on illness, we might be worried that heterogeneity in the threshold level of illness severity required for an individual to report acute illness might be correlated with unobserved characteristics and, in turn, with health, schooling and labor out- comes. This could potentially be a source of bias in our estimates; however, we would only be concerned to the degree that the instrument covaries with this threshold. In Table 7 Selection into Sickness and Symptoms Acute Illness Beginning = 14 Before Survey Fever Symptoms Cough Symptoms Headache Symptoms Days of rainfall × no facility 0.00167 −0.0112 −0.00176 −0.00886 0.00421 0.00711 0.00702 0.00621 No facility −0.0713 0.162 0.0742 −0.104 0.0565 0.109 0.0891 0.0749 Days of rainfall 0.0125 −0.0273 0.0475 0.0215 0.0226 0.0454 0.0370 0.0286 Observations 7,094 1,954 1,954 1,954 Mean of dependent variable 0.276 0.367 0.280 0.266 Notes: Robust standard errors in parentheses p 0.01, p 0.05, p 0.1. See Table II for addi- tional comments. Sample in Column 1 not restricted by acute illness; analysis conducted on all children aged 7–19 inclusive. order to check for this selection into reporting acute illness, we first regress the incidence of sickness using a binary variable for reporting an illness which began in the past 14 days on the full set of regressors including the interaction instrument in a linear probability OLS regression. The results from this regression are reported in Column 1 of Table 7, and verify that the instrument does not predict selection into the sample. In Columns 2–4, we report results from the regressions of binaries for the re- porting of fever symptoms, cough, and headaches, respectively, on the instrument and the usual set of controls. Again, we find no significant effects of the instrument on the probability of reporting specific symptoms. These results provide further evidence that the instrument is orthogonal to illness severity and heterogeneous severity thresholds for reporting illness. Finally, in Table 8 we report first- and second-stage regression results from spec- ifications identical to those reported in Tables II and III, but with the symptom binaries as additional controls. The results appear quite robust to the inclusion of these additional controls. That is, there does not seem to be any evidence that se- lection into sickness or symptoms is driving the results.

D. Instrument Checks