Introduction
Sex between men is the most frequent mode of HIV transmission in Western Europe, the USA and Australia [Reference Beyrer1]. The number of HIV cases diagnosed among men who have sex with men (MSM) has continued to rise in recent years worldwide [2, Reference Pharris3]. Similar trends have been observed in France [Reference Cazein4, Reference Le Vu5].
Given this dramatic context, monitoring risk behaviours in MSM is crucial to understand the drivers of HIV and other disease epidemics, and to plan and evaluate prevention interventions [Reference Dubois-Arber6]. Because populations at high risk of HIV acquisition are often hard to reach, general population surveys – such as household surveys – typically do not include a large enough sample of the population of interest [Reference McGarrigle7]. With respect to MSM, although some national probability surveys include questions on sexual behaviour and/or identity [Reference Bajos, Bozon and Beltzer8], they usually include relatively small numbers from this population [Reference Prah9]. Obtaining a sufficient sample size is one of the biggest challenges for researchers when conducting behavioural surveillance among MSM [Reference Elford10].
Time-location sampling (TLS), also called time-space sampling or venue-based sampling, is widely used to collect data from hard-to-reach populations who frequent known locations. This technique is especially used in surveys on MSM. The principle is to recruit individuals in physical places at times when they gather there (e.g. gay bars, clubs, backrooms) [Reference Holt11–Reference Snowden13]. Using TLS means that a high proportion of venues attended by a priority population can be included in the sampling frame.
Two main issues of TLS have been discussed in the literature [Reference MacKellar14–Reference Stueve16]: how to take into account the sampling design and how to deal with the frequency of venue attendance (FVA). Accordingly, it is necessary to produce unbiased estimates. Although the use of weights based on individual FVA is recommended and justified [Reference Gustafson17–Reference Xia21], it seems that in reality this practice is not frequently applied. Recently, Léon et al.[Reference Léon, Jauffret-Roustide and Le Strat18] presented TLS in the context of sampling theory and proposed a design-based inference model taking into account sampling weights and the FVA. In their work, an indirect sampling framework and the generalised weight share method (GWSM) [Reference Deville and Lavallée22] were both used in a population of drug users. This method has been never used to produce unbiased estimates of HIV prevalence. To our knowledge, in order to estimate HIV prevalence among MSM attending gay venues, alone Gustafson et al.[Reference Gustafson17] used a methodology which took into account sampling weights and FVA in their inference. The weight calculation was different from GWSM, especially in using of FVA. Concerning inference, Gustafson et al. used a model-based method.
In 2015, Santé publique France, conducted a survey (Prevagay 2015), among MSM attending gay venues in five French metropolitan cities in order to estimate HIV prevalence among MSM attending gay venues. The main objective is to demonstrate the importance of using TLS associated with GWSM. In this paper, we present the methodology of the Prevagay 2015 survey, and describe the implementation of the sampling design. We estimate HIV seroprevalence in MSM attending gay venues in five French metropolitan cities using different sampling weights: no weight, no FVA, GWSM trimming or not extreme values in order to assess the impact of taking into account sampling weights and FVA in inference. Design-based method described by Léon et al.[Reference Léon, Jauffret-Roustide and Le Strat18] was used for inference.
Material and methods
Survey design
The Prevagay 2015 survey was conducted in five French metropolitan cities (Lille, Lyon, Montpellier, Nice and Paris).
The choice of these cities was based both on feasibility constraints and on epidemiological criteria. A minimum number of sufficiently frequented accessible venues was needed. Based on the number of new HIV diagnoses in MSM (French regional HIV monitoring data) [Reference Cazein4], the number of HIV prevalence declaration (Gay and Lesbian press survey 2011) [Reference Velter23], and regional alerts of increasing numbers of STI, we chose four cities (in addition to Paris) with different HIV epidemiological profiles.
In each city, the expected sample size was determined from the expected HIV prevalence and the desired precision of the HIV prevalence estimate. Expected HIV prevalence was based on self-reported HIV status of respondents to the national 2011 Gay and Lesbian press survey, conducted via the internet [Reference Velter23].
Formative research was carried out in collaboration with the ‘Equipe Nationale d'Intervention en Prévention et Santé pour les Entreprises’ (ENIPSE), which is a long established national association that organises, among other activities, disease prevention actions in gay venues. Thanks to this association's long history, we were able to identify gay venues and gain an easier access to them and their managers in order to seek their agreement to participate. We decided to investigate commercial gay venues 4 days a week: bars (with or without backrooms), discos, saunas and sex clubs. The survey took place over 6 weeks in Paris, and 4 weeks in each of the other cities between September and December 2015.
We defined a visit duration as a period of 4 h in venues in Paris and 3 h in other cities. We built a sampling frame of venue-day-times (VDTs). For each venue, we chose two visits a day with a total of eight visits a week.
Once eligible venues were identified and had agreed to participate, we requested ENIPSE staff to collect information on opening hours and the estimated number of attendees eligible for the survey for each visit, in order to build the sampling frame for each venue.
A two-stage TLS design was used. During the first stage in each city, we selected VDTs using simple random sampling without replacement, with a minimum of one visit per venue. The number of sampled visits for a venue was proportional to the average number of MSM attending that venue in all visits over the survey period. In the second stage, for each VDT, MSM were selected using systematic random sampling.
A team of investigators (two to four persons) was created for each city, led by a local ENIPSE staff member. We decided to only use investigators belonging to the MSM community, in order to make contact easier with attendees of gay venues. All investigators were specifically trained to implement the survey.
During visits, each team recruited participants, using flyers and information letters about the survey. They also estimated the number of eligible attendees during the visit and noted the number of refusals to participate in the survey. A form collecting basic sociodemographic information was offered to MSM who refused to participate.
Recruitment for survey and data
MSM were eligible for the survey if they were at least 18 years old, had had sex with men in the previous 12 months, could read and speak French, and agreed to both perform finger-prick blood self-sampling, and answer the questionnaire. Participants responded to questionnaire using electronic tablets and no missing answers were permitted. HIV testing was performed by the National Reference Laboratory for HIV (Tours, France) on dried blood spots (DBS) with a combined immunoassay for detection of both p24 antigen and HIV antibodies (Genscreen ultra HIV Ag-Ab; Biorad), as previously described [Reference Velter24]. HIV-positive specimens were confirmed by a combination of assay of recent infection, serotyping and Western blot [Reference Velter24, Reference Le Vu25].
Data collecting regarding FVA
We presented a list of all the participating gay venues for each city to each participant, asking them the following question: ‘In the last month, how many times did you attend the following venues?’ From this specific question, we estimated the number of FVA during the survey period for each individual by (1) summing the declared number of visits in different venues in the previous month, (2) dividing this number by 30 (mean number of days in a month) and (3) multiplying by the number of sampled days in the participant's city. We assumed that FVA did not vary over the survey period.
Sampling weights
To make inferences from the random sample to the population, a sampling weight was assigned to each participant.
At the first stage, the inclusion probability 
$\pi _k^v $ of a VDT k is equal to the number of sampled visits of the specific gay venue divided by the total number of visits in the sampling frame of the corresponding city. The sampling weight of VDT k is the inverse of its inclusion probability: 
$w_k^v = 1/\pi _k^v $.
In the second stage, the inclusion probability π i|k corresponds to the probability that participant i attends the VDT k. All other things being equal, the probability π i|k is equal to the number of MSM surveyed in VDT k divided by the ENIPSE staff member's estimate of eligible men in k.
Thus, a sampling weight w i for participant i can be equal to:
$$w_i = w_k^v \times w_{i \vert k} $$where
$$w_{i \vert k} = 1/\pi _{i \vert k} $$ However, in TLS, individual FVA is not equal for all participants making w i biased. In this context, we define the GWSM weight, 
$\tilde w_i, \; $ providing from the GWSM taking FVA into account by dividing the simple individual weight w i by the number of visits in participating venues during the survey period (noted nFVAi for participant i). Thus, the GWSM weight is equal to:
$$\tilde w_i = w_i /n{\rm FVA}_i $$Trimmed GWSM sampling weights
Despite thorough formative research, changes from the initial design can occur. For instance, one could initially plan five visits to a gay venue for survey purposes, but only visit once because of the owner's refusal to allow survey staff to return. Another example is that a difference between the expected and the real mean number of MSM attending a venue in a given period could lead to fewer survey participants than initially planned. This can lead to extreme sampling weights which can overrepresent individuals in the estimation of a statistic of interest (e.g. prevalence). The estimation of a statistic could have been biased and its variance overestimated. Accordingly, it was necessary to truncate the largest weights. We decided to replace (trim) the weights exceeding a threshold equal to the median weight plus four times the interquartile range of weights in each city, keeping the same sum of initial weights 
$\sum\nolimits_{i = 1}^n {\tilde w_i} $ where n is the sample size. The estimation of the population size was not modified. Let 
$\tilde w_{{\rm trim\;} i} $ be the trimmed weight of the participant i:
$$\eqalign{\tilde w_{{\rm trim}\; i} = &(\tilde w_i \times 1_{\{ \tilde w_i \le T\}} + T \cr &\times 1_{\{ \tilde w_i \gt T\}} \; ){^\ast}\displaystyle{{\mathop \sum \nolimits_{i = 1}^n \tilde w_i} \over {\mathop \sum \nolimits_{i = 1}^n (\tilde w_i \times 1_{\{ \tilde w_i \le T\}} + T \times 1_{\{ \tilde w_i \gt T\}} )}}\;,} $$where
1{a <b} = 1 if a <b and 0, otherwise and T is the median weight plus four times the interquartile range of weights of the city of participant i.
These trimmed GWSM sampling weights are the weights used for the all analyses of the PREVAGAY survey.
Taking into account weights in inference
In general, the main objective of cross-sectional surveys is to estimate functions of interest in the population, such as a total (e.g. number of MSM frequenting gay venues), a proportion (e.g. prevalence of the HIV-infected in the population) or a mean (e.g. average age of HIV-seropositive men). We used the Horvitz–Thompson estimator [Reference Horvitz and Thompson26] and its variance [Reference Särndal, Swensson and Wretman27] which is widely used in surveys.
Data analysis
We declared survey design using sampling weights and stratification by city and finite correction population (fpc) at each stage. At the first stage, the fpc is equal to the number of sampled VDTs divided by the total number of VDTs. At the second stage, the fpc is equal to the number of interviewed MSM divided by the total number of MSM during visits.
We estimated biological HIV prevalence in each city. We compared HIV prevalence estimates and their 95% confidence intervals using different sampling weights: (1) no weight, (2) no FVA (w i), (3) no trimmed GWSM 
${\rm (}\tilde w_i )$ and (4) trimmed GWSM 
${\rm (}\tilde w_{{\rm trim\;} i} )$. We also estimated the design effect of the estimated HIV prevalence in each city. The design effect is equal to the estimated variance of the estimated HIV prevalence taking into account TLS divided by the estimated variance of the estimated HIV prevalence with simple random sampling.
Results
Profile of respondents
The study recruited 2646 participants in the five cities (with a participation rate of 50%): 478 in Lille, 485 in Lyon, 266 in Montpellier, 328 in Nice and 1089 in Paris. A total of 247 visits took place: 45 in Lille, 42 in Lyon, 45 in Montpellier, 42 in Nice and 73 in Paris. On average, 14 persons in Paris and eight in other cities were included at each visit. The weighted median age of participants was 41 years old. Among them, 64% pursued studies after high school diploma, and 84% defined themselves as homosexuals. The details by city are described in Table 1.
Table 1. Profile of respondents* unweighted** trimmed GWSM weighted
Sampling weight and FVA distribution
The sampling weights before trimming, 
$\tilde w_i $ varied between 0.025 and 200, with 85 participants having weights exceeding the median plus 4 interquartile in their city. The details by city are described in Table 2. The distribution of the trimmed sampling weights 
$\tilde w_{{\rm trim\;} i} {\rm \;} $ for each city and all cities is illustrated in Figure 1. The FVA varied from 1 to 215 with a median of 6 FVA in the five cities: 4 in Lille, 6 in Lyon and Montpellier, 4 in Nice and 8 in Paris. The distribution of FVA is illustrated in Figure 2.
Fig. 1. Distribution of trimmed sampling weights.
Fig. 2. Distribution of FVA (FVA >100 –n = 19– were removed).
Table 2. Minimum and maximum sampling weights before trimming, value of threshold (median plus 4 interquartile of the city) and number of sampling weights greater than the threshold in each city
HIV prevalence estimation
HIV prevalence among MSM attending gay venues in the five cities studied was estimated at 14.3% (95% CI (12.0–16.9)). A weighted logistic regression on HIV status of all participants, adjusted for city, age and education level, concluded there was significant differences in HIV prevalence between all five cities (P < 0.001), particularly between Paris and Lille. However, no significant difference was observed between Paris and the other cities. More specifically, we estimated a prevalence of 7.6% (95% CI (5.1–11.1)) in Lille, 11.4% (95% CI (6.9–18.3)) in Lyon, 16.9% (95% CI (11.2–24.7)) in Montpellier, 17.1% (95% CI (11.8–24.1)) in Nice and 16.1% (95% CI (12.5–20.4)) in Paris.
A weighted logistic regression model used to explain the HIV serological status in each city according to the number of FVA, adjusted for age and education level, showed that FVA had a significant effect on the HIV status in Paris, Lille and Nice: the higher the FVA, the higher the risk of being seropositive for HIV
We compared HIV prevalence estimations using different sampling weights (Fig. 3). The unweighted 95% confidence intervals for HIV prevalence (no weight) overlapped with confidence intervals of estimations using 
$\tilde w_{{\rm trim\;} i} $, 
$\tilde w_i $ and w i. The unweighted prevalence estimates were included in the confidence intervals of estimations using 
$\tilde w_{{\rm trim\;} i}, \tilde w_i {\rm \; and\;} w_i \; $ for all cities but one (Nice). The variance of the unweighted prevalence was narrower than the variances based on GWSM (using 
$\tilde w_i $, and 
$\tilde w_{{\rm trim\;} i} $).
Fig. 3. Estimation of HIV prevalence in each city according to different weights: 
$\tilde w_{{\rm trim\;} i} $ (trimmed GWSM), 
$\tilde w_i $ (no trimmed GWSM), w i (no FVA) and no weight.
We computed the design effect of the estimated HIV prevalence in each city (Table 3). Design effects were different according to cities, with a minimum in Lille and a maximum in Lyon. The design effect of estimated HIV prevalence ranged from 1.2 (Lille) to 4.0 (Lyon).
Table 3. Design effect of the estimated HIV prevalence in each city
Discussion
We applied the GWSM to provide the most accurate estimations for HIV prevalence in MSM attending gay venues. This method took into account the TLS weights and individuals’ FVA. Of all recently published studies about MSM attending gay venues, to our knowledge, only Gustafson et al.[Reference Gustafson17] produced estimates using sampling weights and FVA. However, they used a different estimation method than GWSM. Other studies in Australia and the USA used TLS, but provided estimations without taking into account FVA [Reference Mirandola12] and sometimes without weights [Reference Holt11, Reference Snowden13]. The need to use both sampling weights and FVA in inference of TLS studies has been demonstrated [Reference Gustafson17, Reference Léon, Jauffret-Roustide and Le Strat18]. In the present study, the variance of HIV prevalence estimates was lower than estimations based on GWSM when we did not take into account the sampling design (unweighted estimates). Unweighted estimates, although still commonly used, can incorrectly conclude that significant differences exist in HIV prevalence between cities adjusted for age and education level. We decided to trim extreme sampling weights in order to avoid a variance in estimates which was artificially too large. Although trimming is often used, there seems to be no consensus in the literature on how to trim extreme weights [Reference Alexander, Dahl and Weidman28–Reference Potter31]. Our choice was based on a compromise between as small a change as possible in weights and the greatest reduction of variance of some key statistics (HIV prevalence and negative viral load prevalence).
In this paper, we presented the design effect to help researchers calculate sample sizes when they set up similar survey designs. As the design effect ranged from 1.2 and 4 according to city, we recommend using the maximum value. It will then suffice to multiply the sample size needed from a simple random sampling hypothesis by the design effect.
Our study also showed the impact of the non-use of FVA inference. Indeed, in cities where FVA was positively associated with serological status, prevalence estimates obtained without taking into account FVA were different from those obtained by GWSM. In their simulation work, Léon et al. showed that not considering FVA produced biased estimation [Reference Léon, Jauffret-Roustide and Le Strat18].
With respect to FVA, despite the fact that the GWSM requires only the total number of FVA, we decided to ask participants how frequently they visited specific participating gay venues. We could have asked only one question on frequency attendance in gay venues but it could have been difficult to differentiate between participating venues and non-participating ones. Furthermore, as each city was studied independently, a global question could have led to an overestimation of frequentations for MSM who travelled between cities (i.e., visited participating venues in different cities studied). Moreover, we thought that asking for information on each venue would provide a more precise answer than a general question on all venues.
The participation rate was estimated at 50%. Among people who refused to participate, only 21% agreed to complete a refusal questionnaire. Accordingly, it was difficult to compare respondents with non-respondents. However, it is likely that only the most highly motivated men, for whom prevention is important, agreed to participate [Reference Velter24], and this most probably led to an underestimation of HIV prevalence in our population [Reference Elford10].
The overall HIV infection prevalence was estimated at 14.3% (12.0–16.9). The variations in HIV prevalence we observed across cities (from 7.6% (5.1–11.1) in Lille to 17.1% (11.8–24.1) in Nice) could partly be explained by the differences in the type of recruitment venues where participants were included. Differences in age distributions were also important to explain the differences in HIV prevalence. Despite differences in methodologies, our results regarding HIV prevalence were comparable with those from studies using TLS and conducted throughout cities in Europe [Reference Mirandola12], the USA [Reference Raymond, Chen and McFarland32] and Australia [Reference Holt11].
Limitations
Despite TLS being the current method of choice to conduct surveys among hard-to-reach populations such as MSM, our results cannot be extended to the whole MSM population. Men recruited through TLS were more connected to the gay community than those recruited through Internet sampling, had a greater number of sexual partners, and had more risky sexual behaviours [Reference Zablotska33]. Accordingly, estimates of HIV prevalence in MSM attending gay venues may overestimate HIV prevalence in the MSM population. Respondent-driven sampling and surveys based on online recruitment, represent other alternatives, with the potential of recruiting a broader sample of MSM [Reference Velter23, Reference Zablotska33]. However, neither allows pure random sampling [Reference Beyrer1].
Despite the carefully developed study protocol, individual unforeseen events affected the implementation of the survey and led to modifications in the sampling calendar. The terrorist attacks in Paris in November 2015 created a certain atmosphere of fear, with public disturbances occurring in the city's gay neighbourhood during the subsequent days. We consequently decided to suspend the survey for 1 week and to postpone the end of the data collection. In Nice, serious flooding occurred in October 2015 while the survey was being implemented. The events in Paris and Nice led very probably to a decrease in attendance of gay venues during the survey period. Thus, sampling weights could have been somehow miscalculated and could have introduced some biases in estimations.
Conclusion
Finally, the implementation of TLS and the use of GWSM made it possible, for the first time, to both perform a random survey among MSM in France attending gay venues and to produce reliable statistical results, in particular concerning the estimation of HIV prevalence in that population. In the last decade, a diversification in gay and other MSM social networks has been observed [Reference Zablotska33]. We observed a shift in the way in which MSM meet sexual partners, with a decline in the use of traditional venues (such as bars, saunas and backrooms) and an increase in internet use and apps [Reference Zablotska33]. These changes must be taken into account in MSM behavioural surveillance, for example modifications in recruitment methodologies. This of course creates new challenges as regards issues of representativeness. In order to understand the MSM population in all its diversity, it will be necessary to triangulate data collection methods (population probability surveys, internet convenience sampling and TLS) to implement HIV prevention interventions in line with changes in gay socialising.
Acknowledgements
We thank the PREVAGAY2015 group, composed of A Velter, Alexandre, F Barin, S Chevaliez, D Friboulet, M Jauffret Roustide, F Lot, N Lydié, G Peytavin, O Robineau, L Saboni, C Sauvage, C Sommen. Our thanks also to Lucie Léon, Yann the Strat, Laurence Meyer and Josiane Warzawski for their helpful advice.
We would like to thank all those who agreed to participate in the PREVAGAY2015 study. We also thank the employees of the association ENIPSE who carried out the study on the ground (Sébastien Cambau, Jérôme Derrien, Sylvain Guillet, Loïc Jourdan, Cyril Kaminski, Vivien Lugaz, Cedric Péjou, Erika Thomas Des Chenes, Florian Therond, Richard De Wever) and the associations that supported the study throughout, including AIDES (Vincent Coquelin), Act Up (Hugues Fisher), Le 190 (Michel Oyahon), Sidaction (Sandrine Fournier). We also thank Kevin Babaud, Céline Desouche and Damien Thierry for their excellent technical assistance, and Pascal Chaud, Agnès Lepoutre, Philippe Malfait, Bakhao Ndiaye, Cyril Rousseau, Christine Saura, Yassoungo Silue and Stephanie Vandentorren from the regional units of Santé publique France for their region implication. We would like to extend our warmest thanks to all the institutions that agreed to participate, and all the associations who facilitated the study.
PREVAGAY2015 was financed by the National Agency for Research against AIDS and Viral Hepatitis (ANRS) and the French Health Agencies in the Hauts-de-France, Auvergne-Rhône-Alpes, Ile-de-France, Occitanie and Provence-Alpes-Côte d'Azur regions.
Data collection was carried out by B.V.A.
Declaration of interest
None.
