The lay abstract featured today (for HIV estimation using population-based surveys with non-response: A partial identification approach by Oyelola A. Adegboye, Tomoki Fujii, Denis Heng-Yan Leung, Li Siyu) is from Statistics in Medicine with the full Open Access article now available to read here.

Abstract

Estimating HIV prevalence with survey data is a complex task, often hindered by non-response and test refusals. In this context, this research introduces a novel approach that goes beyond traditional methods like imputation, which assumes missing data occurs at random, but this may not always hold true. Instead of point identification, rates are identified within bounds.  Instrumental variables are used to narrow the identification bounds.

The effectiveness of the proposed approach is rigorously assessed through simulation exercises and by analyzing Demographic and Health Survey (DHS) data from Kenya, Malawi and Zambia. The findings are clear–imputation can produce biased results, even with slight non-random missingness. However, the proposed method, which balances informativeness and robustness without making strong assumptions, offers a reliable and effective alternative. It combines instrumental variable bounds from a pool of candidate instruments, ensuring the reliability of the estimation procedures.

Two essential considerations are highlighted. First, including a sufficient number of valid instruments is critical to prevent an undue expansion of the confidence bounds. The objective is to robustly ascertain prevalence while accepting the potential invalidity of some instruments within the instrument pool. Secondly, the careful use of white noise as an instrument is vital as it may lead to wide bounds similar to worst-case scenarios. The strength of an instrument can be tested, which is exemplified in the empirical study, ensuring the reliability of the proposed method.

In summary, this new approach provides a sturdy alternative for estimating HIV prevalence in population-based surveys marked by frequent non-response and limited insight into the reasons for non-response. The new approach would enhance current methodologies substantially; it strikes a balance between robustness and the provision of useful bounds through instrumental variables, thus instilling confidence in the estimation procedures.