Implications of the HIV testing protocol for refusal bias in seroprevalence surveys

BMC Public Health

Table 5 Comparison of HIV seroprevalence estimates based on standard probit models and models accounting for sample selection under various scenarios (Zewditu Memorial Hospital, Addis Ababa, 2003–04)

	Scenario
	Test of Heckman model^a	Post-test counseling is required	Post-test counseling is optional
E(HIV% – Probit)	17.7 (16.4 – 19.1)	17.8 (16.6–19.1)	21.4 (20.2–22.7)
E(HIV% – Heckman)	23.1 (21.7 – 24.4)	23.4 (22.1–24.7)	23.7 (22.4–25.0)
Observed HIV%	22.2 (19.9 – 24.4)	unknown	unknown
Sample	Consent groups A and B	All consent groups	All consent groups
Assumption	HIV status in consent group B is unobserved	HIV status in consent groups B and C is unobserved	HIV status in consent group C is unobserved
LR test H₀:ρ = 0	p < .01	p < .01	p = .07

Notes: 95%- CI are reported between brackets. Using dummies for admission diagnosis rather than the likelihood of infection in these regressions hardly changes the estimated prevalence rates though one of the selection models did not converge.
^a In the first column, we assume that HIV status in consent group B is unknown, and compare the ordinary Probit and Heckman selection model estimate with the true or observed value of HIV prevalence.

ISSN: 1471-2458