Discussion

This study reports on a feasible method to estimate population sizes linking capture–recapture methods to RDS studies that provide representative estimates of the networked population. The difference between the crude and RDS-adjusted proportions in the population estimates for both MSM and FSW highlights the importance of having a representative study as the second data source. As IBBS become more common in countries with concentrated HIV epidemics and use more robust designs to capture representative data, countries should consider including population estimates as part of these activities. The low cost and ease of data collection is all the more reason to conduct these estimations. Due to the close cooperation with NGO working with MSM and FSW in San Salvador the incremental cost to the IBBS for the initial capture was US$500 for each population.

An international meta-analysis of population-based studies that have enquired about male sexual behaviours reported a prevalence of lifetime sex between men in Latin America that varied from 2% to 25%. Sex between men in the past year varied from 1% to 8%.9 A similar study reported the proportion of sex workers among the adult female population in urban centres in Latin America varying from 0.2% to 7%.10 As part of El Salvador's UNAIDS-supported HIV estimation exercise in 2009—which all countries develop biannually—the national estimation group estimated that MSM constitute 2–5% of the urban male population and FSW 0.4–0.8% of the urban female population 15–49 years of age, based on UNAIDS recommendations and available literature. These figures imply a total of 7267–18 167 MSM and 1627–3253 FSW in San Salvador. These figures are very close to our findings for the MSM population. However for FSW, our numbers are almost double the UNAIDS-supported figures and our findings must be interpreted with caution. Based on our study and using the 2009 census urban adult population aged 15–49 years for San Salvador, 3.4% of men in San Salvador are MSM and 1.4% of women are FSW.11

Several assumptions must be met in order to generate a reliable estimate using the capture–recapture. Most important is that the samples should be independent. In order to ensure independence, different sampling strategies were used at each stage; furthermore, provided the RDS estimate is representative of the population, participation in the IBBS is independent of belonging to more visible groups in which the first capture was conducted. The second assumption is that all members in the target population should share equal chance to be selected or that at least the second sample is representative of the population.6 7 The requirement of a simple random sample is unrealistic for research with hidden populations, such as MSM, in which it is challenging to construct a sampling frame. In our study, RDS was used to recruit participants in the second capture; weighted estimates generated from RDS data are considered to be unbiased estimates that are generalisable to the population as a whole. While not providing equal selection probabilities, the RDS-adjusted point estimates are representative of the population, and the adjusted variance is reflected in the CI. The last assumption is that the population is closed. The time between the two phases for this study was short and we believe that in and outmigration was negligible.

We found that using the crude and adjusted estimates on the proportion of people receiving the key chain introduced differences in the results for MSM. It is likely that key chains were preferentially distributed to MSM with larger social networks; thus the recapture fraction is downweighted when applying the RDS adjustment and the estimated number of MSM increases in turn. Using the crude proportion would therefore underestimate the population by failing to correct for dependence between the first and second captures. This dependence occurs because members of the target populations who are more accessible for the key chain distribution would also be more likely to be recruited into the RDS study (because of their relatively large network sizes they have a greater possibility of being invited to participate). For FSW, the differences between the crude and adjusted estimates were not large. Notably, using the adjusted estimates for FSW resulted in a reduction in sampling error in contrast to the increase in error seen for MSM. Interestingly, the homophily score (ie, tendency to recruit within subgroup) among FSW who received the key chain was lower (0.08) than the score among those who did not (0.30), while for MSM the relationship was the reverse (0.13 compared with −0.02, respectively). While not the focus of this work, one possible explanation is that distribution of the unique object to individuals who comprise less-connected versus more-connected subgroups may improve sampling error and should be investigated further.

It is believed that weighted estimates generated from RDS generalise to the population as a whole. However, recent assessments of the RDS sampling methodology have called into question the validity of this assertion and of conventional means of analysing RDS data, given the dependence of RDS on arguably unrealistic assumptions. Computer simulations using US network data suggest that variance may be underestimated, particularly in the presence of tightly linked subgroups that are poorly connected to the rest of the social network (ie, ‘bottlenecks’).12 In this study, for variables such as age, place of work, self-identification and educational level, we observed relatively large numbers of recruitments from each variable category to remaining categories (ie, no zero cells in the recruitment matrix), suggesting that network bottlenecks were not a major concern with respect to these variables (data not shown). Evaluations of the RDS recruitment strategy also suggest that bias may be introduced if too few waves are included or if recruitment is preferential.13 In our study, of the 11 seeds in the RDS study, six resulted in long chains reaching at least five waves of recruitment. The longest recruitment chain included 18 waves. Finally, in RDS, estimates are based on a with-replacement random-walk model, while the actual sampling is without replacement. When a substantial fraction of the target population is sampled this approximation can lead to bias in the resulting estimators.13 Based on our results, the RDS study sample represented less than 5–12% of the estimated population.

This study has several limitations inherent to RDS including those mentioned above. In addition, as in any study of highly stigmatised populations it is possible that members of the MSM and FSW populations may have chosen not to participate for fear of being identified. There have been no other sources of size estimation such as population-based studies in El Salvador to validate these results. Other methods, such as including specific questions in national demographic health surveys, would be useful to triangulate results and improve the estimations. In behavioural surveillance there is always a high potential for social desirability response. Moreover, most methods of size estimation inevitably experience biases that arguably lead to under or overestimation, but such that researchers rarely know which way the biases may go. Therefore, data from a ‘gold standard’ survey in terms of sampling methodology and representativeness, for which there is also wide agreement that the estimates are likely to be conservative, provides a useful lower bound on which all can agree, as a first step forward.

It has been known for many years that the MSM and FSW populations are at greater risk of HIV in El Salvador. Our estimates suggest the size of these populations in San Salvador is considerable and should be targeted for interventions. It is essential that countries with concentrated epidemics have up-to-date estimates of most at-risk populations. Such data will allow the El Salvador Government, national and international agencies and other partners to prepare and budget interventions for these populations.