Replication data for: Estimating HIV Prevalence: Can Heckman Help?

Estimating national HIV prevalence is challenging due to high rates of HIV test refusal. Barnighausen et al. (2011) proposes the Heckman selection model to account for selection on unobservables, which would result in biased national prevalence estimates. In practice, Heckman models may be difficult to implement because they require a suitable identification variable. Barnighausen et al. (2011) suggests that interviewer identity fulfills the criteria for a valid identification variable. By replicating the paper, we investigate the validity of Heckman estimates of HIV prevalence in Zambia using interviewer identity as the identification variable. Our findings suggest that interviewer identity is endogenous to HIV status under plausible field conditions, and is therefore not a valid identification variable. We find very high predicted prevalence of HIV among men that did not consent to testing (50-54%), and 29-33% among non-consenting men who have never had sex. These surprising results call into question the model's validity. We recommend that DHS evaluate the possibility of altering their data collection procedures to ensure exogeneity of interviewer identity and further evaluate the robustness of the Heckman model before changing HIV prevalence estimation methods.