Using interviewer random effects to remove selection bias from HIV prevalence estimates

Table 2 Estimates of HIV prevalence among Men in Zambia (2007)

Model	HIV prevalence	Analytic 95% CI		Bootstrap 95% CI
All Men - Fixed Effects Selection Model	20.1%	19.0%	21.3%
All Men - Random Effects Selection Model	16.3%	15.3%	17.3%	11.0%	18.4%
All Men – Random Effects Bias Correction Selection Model	15.5%	14.5%	16.5%	10.2%	17.9%
Men with Valid HIV Tests	12.1%	11.0%	13.3%
Men with No Contact - Imputation Model	15.3%	14.2%	16.3%
All Men - Imputation Model	12.3%	11.4%	13.2%

In the Heckman-type selection models (rows 1-3), consent to test and HIV status are jointly estimated using a bivariate probit with the following covariates: education, household wealth quintile, type of location, marital status, had a sexually transmitted disease, age at first intercourse, had high risk sex, number of partners, condom use, would care for an HIV-infected relative, knows someone who died of AIDS, previously tested for HIV, smokes, drinks alcohol, language, age group, region, ethnicity and religion. The selection variable which predicts consent but not HIV status is interviewer identity. Full parameter estimates are presented in tables A4-A6 in the appendix (Additional file 1). Analytic standard errors are shown for the fixed effects and random effects models, with bootstrap errors for random effects and random effects bias correction models based on 1,000 replications. Our cluster bootstrap takes account of survey design by drawing a fixed number of clusters (the same as in the original data) from each stratum in each sample. Results from an imputation model are also shown in rows 5–6, along with estimates only using those without missing data (respondents with a valid HIV test). HIV prevalence estimates are weighted. Source: DHS Zambia 2007 (men).

ISSN: 1471-2288