This preview shows page 1. Sign up to view the full content.
Unformatted text preview: nt) relationship between receiving
HIV-positive results on the probability of having sex at follow-up (Table 7, column 8).
Theoretically, if individuals who were more likely to practice safe sex were also more likely to
choose to learn their HIV status and also purchase more condoms at the follow-up, not accounting for selection bias would overestimate the true impact of learning HIV results on later sexual
behavior. However, comparing the OLS estimates to the IV estimates of the impact of learning
HIV results on condom purchases (Table 7) indicates no consistent pattern across each column,
and the differences between the OLS and IV estimates are neither large nor statistically significant for either HIV-negative or HIV-positive individuals (not shown).
Several of the outcome variables in Table 7 are binary, possibly warranting a nonlinear estimation strategy. However, estimation of binary regression models with binary endogenous variables
is difficult and there are often problems with convergence or concavity of the log likelihood surface (David A. Freedman and Jasjeet S. Sekhon 2008; see also Edward Vytlacil and Nese Yildiz
2007; Jacob Nielsen Arendt and Anders Holm 2006). One strategy suggested by Heckman (1978)
is to estimate a bivariate probit model. In order to estimate the effects of learning HIV results
on the likelihood of purchasing any condoms using this nonlinear strategy, it is necessary to
divide the sample by HIV status and estimate the effect of obtaining HIV test results on positive
and negative individuals separately. Because the sample is not pooled among HIV-positives and
HIV-negatives, the nonlinear estimates presented in Table 8 are not directly comparable to those
in Table 7. I therefore present linear OLS and IV estimates as well as the marginal probit and
marginal biprobit (accounting for endogeneity) estimates on the impact of learning HIV results
on condom purchases among HIV-positives and negatives.17
The linear results in column 2, Table 8, do not differ...
View Full Document