Estimating the Population Size of Female Sex Workers in Zimbabwe: Comparison of Estimates Obtained Using Different Methods in Twenty Sites and Development of a National-Level Estimate

Background: National-level population size estimates (PSEs) for hidden populations are required for HIV programming and modelling. Various estimation methods are available at the site-level, but it remains unclear which are optimal and how best to obtain national-level estimates. Setting: Zimbabwe. Methods: Using 2015–2017 data from respondent-driven sampling (RDS) surveys among female sex workers (FSW) aged 18+ years, mappings, and program records, we calculated PSEs for each of the 20 sites across Zimbabwe, using up to 3 methods per site (service and unique object multipliers, census, and capture-recapture). We compared estimates from different methods, and calculated site medians. We estimated prevalence of sex work at each site using census data available on the number of 15–49-year-old women, generated a list of all “hotspot” sites for sex work nationally, and matched sites into strata in which the prevalence of sex work from sites with PSEs was applied to those without. Directly and indirectly estimated PSEs for all hotspot sites were summed to provide a national-level PSE, incorporating an adjustment accounting for sex work outside hotspots. Results: Median site PSEs ranged from 12,863 in Harare to 247 in a rural growth-point. Multiplier methods produced the highest PSEs. We identified 55 hotspots estimated to include 95% of all FSW. FSW nationally were estimated to number 40,491, 1.23% of women aged 15–49 years, (plausibility bounds 28,177–58,797, 0.86–1.79%, those under 18 considered sexually exploited minors). Conclusion: There are large numbers of FSW estimated in Zimbabwe. Uncertainty in population size estimation should be reflected in policy-making.

Each population size estimation method makes a number of assumptions, as does respondent driven sampling (1,2). Some of these assumptions can be investigated for possible biases in the estimate of P (proportion attending clinic or receiving a wristband in the RDS surveys) and implications for the PSE qualitatively assessed. For multiplier methods, the two data sources are assumed to be independent. In the case of the service multiplier method (SMM), this means that women attending the programme should not disproportionately have been more likely to be recruited into the survey, and for the unique object multiplier method (UOMM), that the process of distributing wristbands was independent of survey recruitment. While we made efforts to keep these processes separate (e.g. not recruiting seeds based on programme attendance or distributing wristbands directly to RDS seeds) we graphically examined the convergence of the estimate that measured programme attendance or wristband receipt in the RDS survey over sample accumulation to judge whether there was evidence that the final estimate was likely still dependent on seed characteristics or whether the estimate appeared to have stabilized prior to final sample size ('reasonable convergence'). If the estimate had not converged, this could also have indicated that it was too low or high (3) and therefore that our resulting PSE was too high or too low. We also examined recruitment homophily by programme attendance and wristband receipt (the tendency for women to recruit others like themselves on the basis of a given characteristic) and the ratio of the mean network size of those who did attend or receive a wristband to those who did not, a difference that is accounted for in the weighting but which explains a discrepancy between the unweighted and weighted findings.
Capture-recapture methods assume the 'captures' are independent from each other which is difficult to ensure or assess in practice, and assumes limited mobility of FSW to and from the site between captures. The census methods assume that women counted at sites are indeed sex workers.
We used the RDS package (4,5)   • How to summarise the size estimate for the twenty sites with direct PSEs where multiple methods were used • Approach to extrapolating from these 20 sites to obtain a national population size estimate • Develop a list of likely hotspot sites for female sex work around Zimbabwe. These were decided to be all those 36 sites with Sisters clinics, plus any additional sites estimated to have concentrations of female sex workers by the workshop attendees. • Matching hotspots into strata of likely similarity of SW prevalence • Estimate of the proportion of all female sex workers in Zimbabwe who would be found in one of the hotspot sites.

Methods
The process for reaching decisions during the workshop was as follows. First, workshop attendees attended training on methods used to estimate population sizes, their strengths, weaknesses and uncertainties, estimates obtained from 20 sites in Zimbabwe, and methods used to extrapolate site estimates to the national level. Background and contextual information behind each decision was given, followed by an open discussion amongst all workshop attendees. The facilitators then asked participants to put forward suggestionsthese were each discussed and debated, a consensus amongst workshop attendees was reached and participants were reminded of the decisions reached at regular intervals and at the end of the workshop. Particular attention was given to each participant's particular background and expertise.
When considering additional sites that could be hotspots for FSW around Zimbabwe, participants considered each province in turn and reviewed a map. Not all suggestions put forward were necessarily accepted by the group. When creating strata of hotspots considered to be similar to each other with respect to the likely prevalence of sex work among adult women (unknown for those sites without direct PSEs), a table of site classifications was projected so that participants could consult it. This included site names, provinces, primary and secondary classifications (often related to main economic industry: mining, tourism, border site, farming, fishing, army base, mining, rural growth-point) and male and female population from Census 2012. After a first list of groupings was reached, participants were asked to discuss and agree the groupings again before the groupings were finalised.

Attachments:
Slides used during the workshop (excluding those with names sites other than Harare and Bulawayo, decided to be too sensitive to report without anonymising sites) are included in Appendix 4.