BACKGROUND: HIV serosurveys have become important sources of HIV prevalence estimates, but these estimates may be biased because of refusals and other forms of non-response. We investigate the effect of the post-test counseling study protocol on bias due to the refusal to be tested. METHODS: Data come from a nine-month prospective study of hospital admissions in Addis Ababa during which patients were approached for an HIV test. Patients had the choice between three consent levels: testing and post-test counseling (including the return of HIV test results), testing without post-test counseling, and total refusal. For all patients, information was collected on basic sociodemographic background characteristics as well as admission diagnosis. The three consent levels are used to mimic refusal bias in serosurveys with different post-test counseling study protocols. We first investigate the covariates of consent for testing. Second, we quantify refusal bias in HIV prevalence estimates using Heckman regression models that account for sample selection. RESULTS: Refusal to be tested positively correlates with admission diagnosis (and thus HIV status), but the magnitude of refusal bias in HIV prevalence surveys depends on the study protocol. Bias is larger when post-test counseling and the return of HIV test results is a prerequisite of study participation (compared to a protocol where test results are not returned to study participants, or, where there is an explicit provision for respondents to forego post-test counseling). We also find that consent for testing increased following the introduction of antiretroviral therapy in Ethiopia. Other covariates of refusal are age (non-linear effect), gender (higher refusal rates in men), marital status (lowest refusal rates in singles), educational status (refusal rate increases with educational attainment), and counselor. CONCLUSION: The protocol for post-test counseling and the return of HIV test results to study participants is an important consideration in HIV prevalence surveys that wish to minimize refusal bias. The availability of ART is likely to reduce refusal rates.