HIV estimation using population-based surveys with non-response: A partial identification approach.

Area of Science:

• Epidemiology
• Biostatistics
• Public Health

Background:

• Estimating HIV prevalence using Demographic and Health Surveys (DHS) data is complicated by non-response and test refusals.
• Standard adjustment methods like imputation assume data are missing at random, which is often not the case.
• Instrumental variable methods can account for non-random missingness but rely heavily on instrument validity.

Purpose of the Study:

• To develop and evaluate a robust method for HIV prevalence estimation that addresses non-response and test refusals in survey data.
• To improve the informativeness and credibility of HIV prevalence estimates derived from population-based surveys.
• To compare the performance of the proposed method against conventional imputation and worst-case bounds.

Main Methods:

• Utilized Manski's partial identification approach to construct instrumental variable bounds for HIV prevalence.
• Employed a pool of candidate instruments, not requiring all to be valid.
• Evaluated the method through a simulation study and applied it to DHS data from Zambia, Malawi, and Kenya.

Main Results:

• Imputation methods produced significantly biased HIV prevalence estimates under mild non-random missingness.
• Worst-case identification bounds were robust but lacked informativeness.
• The proposed union of instrumental variable bounds balanced informativeness and robustness, even with some invalid instruments.

Conclusions:

• Partial identification bounds offer a credible and robust approach to HIV prevalence estimation without stringent assumptions about non-response mechanisms.
• The union of instrumental variable bounds provides a more informative alternative to worst-case bounds while maintaining robustness.
• This method is crucial for accurate HIV surveillance using population-based survey data affected by non-response.