How Should We Best Estimate the Mean Recency Duration for the BED Method?

BED estimates of HIV incidence from cross-sectional surveys are obtained by restricting, to fixed time T, the period over which incidence is estimated. The appropriate mean recency duration ([Image: see text]) then refers to the time where BED optical density (OD) is less than a pre-set cut-off C, given the patient has been HIV positive for at most time T. Five methods, tested using data for postpartum women in Zimbabwe, provided similar estimates of [Image: see text] for C = 0.8: i) The ratio (r/s) of the number of BED-recent infections to all seroconversions over T = 365 days: 192 days [95% CI 168–216]. ii) Linear mixed modeling (LMM): 191 days [95% CI 174–208]. iii) Non-linear mixed modeling (NLMM): 196 days [95% CrI 188–204]. iv) Survival analysis (SA): 192 days [95% CI 168–216]. Graphical analysis: 193 days. NLMM estimates of [Image: see text] - based on a biologically more appropriate functional relationship than LMM – resulted in best fits to OD data, the smallest variance in estimates of [Image: see text], and best correspondence between BED and follow-up estimates of HIV incidence, for the same subjects over the same time period. SA and NLMM produced very similar estimates of [Image: see text] but the coefficient of variation of the former was >3 times as high. The r/s method requires uniformly distributed seroconversion events but is useful if data are available only from a single follow-up. The graphical method produces the most variable results, involves unsound methodology and should not be used to provide estimates of [Image: see text]. False-recent rates increased as a quadratic function of C: for incidence estimation C should thus be chosen as small as possible, consistent with an adequate resultant number of recent cases, and accurate estimation of [Image: see text]. Inaccuracies in the estimation of [Image: see text] should not now provide an impediment to incidence estimation.