Epidemiology of HCV infection among drug users (DUs) has been widely studied. Prevalence and sociobehavioural data among DUs are therefore available in most countries but no study has taken into account in the sampling weights one important aspect of the way of life of DUs, namely that they can use one or more specialized services during the study period. In 2004–2005, we conducted a national seroepidemiologic survey of DUs, based on a random sampling design using the Generalised Weight Share Method (GWSM) and on blood testing.
A cross-sectional multicenter survey was done among DUs having injected or snorted drugs at least once in their life. We conducted a two stage random survey of DUs selected to represent the diversity of drug use. The fact that DUs can use more than one structure during the study period has an impact on their inclusion probabilities. To calculate a correct sampling weight, we used the GWSM. A sociobehavioral questionnaire was administered by interviewers. Selected DUs were asked to self-collect a fingerprick blood sample on blotting paper.
Of all DUs selected, 1462 (75%) accepted to participate. HCV seroprevalence was 59.8% [95% CI: 50.7–68.3]. Of DUs under 30 years, 28% were HCV seropositive. Of HCV-infected DUs, 27% were unaware of their status. In the month prior to interview, 13% of DUs shared a syringe, 38% other injection parapharnelia and 81% shared a crack pipe. In multivariate analysis, factors independently associated with HCV seropositivity were age over 30, HIV seropositivity, having ever injected drugs, opiate substitution treatment (OST), crack use, and precarious housing.
This is the first time that blood testing combined to GWSM is applied to a DUs population, which improve the estimate of HCV prevalence. HCV seroprevalence is high, indeed by the youngest DUs. And a large proportion of DUs are not aware of their status. Our multivariate analysis identifies risk factors such as crack consumption and unstable housing.
Drug users (DUs) are at high risk for HIV and HCV infections . Among DUs, HCV, like HIV, is mainly transmitted through needle sharing. However, HCV (Hepatitis C Virus) is more easily transmitted than HIV because its prevalence is greater than for HIV (Human Immunodefiency Virus), it is more resistant to desiccation [2,3] and is also linked to the sharing of drug paraphernalia (water, filter, spoon) [4-8]. HCV infection can, therefore, occur after much less injection and sharing episodes [9-12]. To minimize the risk of HIV and HCV, harm reduction policies have been adopted in many industrialized countries [13-17]. Many studies have studied the epidemiology and risk factors for HCV infection among DUs [18-27] and shown that harm reduction policies have had a major impact on HIV among DUs but a much more limited impact on HCV infection [10,28-36]. HCV prevalence estimates remain high , over 50% in the majority of countries where data are available . In all these studies, prevalence and sociobehavioural data among DUs are available and useful for public health issues. So far, no study has taken into account in the sampling weights one important aspect of the way of life of DUs, namely that they can use one or more specialized services during the study period. This has an impact on their inclusion probabilities, and accordingly on estimates.
In France, most data on HIV and HCV infection among DUs come from notifications or highly selected population samples [39-42] and no prevalence estimates of HIV and HCV infection based on laboratory confirmation were available among DUs. In 2004–2005, we therefore conducted a national survey among DUs who resided in metropolitan France (the ANRS-Coquelicot study) to assess the prevalence of HIV and HCV infections based on serum testing, the frequencies of at risk practices, and the risk factors for HCV prevalence. To improve the estimate of HIV and HCV prevalences, we used a sampling design including the Generalised Weight Share Method (GWSM) described by Ardilly and Lavallee [43,44], taking into account the use of specialized services by each DU during the study period.
In 2004–2005, we conducted a two stage random survey of DUs in five French large cities (Lille, Strasbourg, Paris, Bordeaux and Marseilles). Given the expected size of the sample, these five cities were chosen due to the important number of specialized structures for DUs, and the prevalence of drug use in each city. Prevalence rates of drug addiction were estimated from available data on opiate substitution treatments (OST) . The five cities from five different regions were also selected to represent the diversity of drug addiction (products and modes of consumption) in France. The first stage of the design was constituted of high-and low-threshold care structures for DUs and of general practitioners (GPs) and the second stage by DUs. A DU to be included in the survey was a person >18 years who had injected or snorted drugs "at least once in its life" and accepted to participate in the survey after informed consent.
The sampling of individuals is similar to the time-location sampling . In each city we did an exhaustive inventory of all structures that provide services to DUs: accommodation services including residential centers, hotel rooms, "sleep-in"; drug treatment centers including methadone maintenance or psychotherapy; low threshold services including needle exchange programs and outreach work teams. We then constructed a sampling frame by half-day that structures were open. Pairs (structure/half-day) were selected using an unequal probability sampling. Inclusion probabilities were proportional to the structure's active DUs list to obtain a calendar of half-day visits of structures for DUs selection, inclusion and data collection. In each structure/half day visit, DUs present were selected using a simple random sampling, except for residential centers where all users were included. For GPs we first constituted a random sample of GPs stratified by the volume of substitution treatment prescribed (high versus moderate). Then, inclusion was proposed to all the DUs followed by the selected GPs.
Subject inclusion and data collection
During each structure/half day visit, a number of DUs corresponding to the calculated sample size were asked to participate and included if they consented to the interview and self fingerprick blood sampling on dried blood spots for HIV and HCV testing. The questionnaire was administered over 30 to 40 minutes by professional interviewers independent of the recruitment structures.
Interviewers had been trained for DUs interview in order to minimize the social desirability bias with respect to drug consumption and at risk practices. The questionnaire explored DU'socio-demographic situation, health status, access to HCV screening, knowledge of HCV transmission modes, drug use, and at risk practices. Socio-demographic variables were related to the situation at the time of the study, such as for the level of education, employment, houselhold and living conditions, marital status. Other items investigated different period of DUs' lives such as ever previous jail experience, living with their families or not during adolescence. Health related items included knowledge of current HIV and HCV serological status, medical treatments, circumstances of last HCV or HIV testing and access to HCV treatments. Perception of health status and main health problems were investigated for the 6 months prior to interview. Drug consumed, modes of consumption (injecting, sniffing, smoking) and at-risk practices (sharing syringes, other injection materials, sniffing material or glass pipes) were recorded for the last month using the Injecting Risk Questionnaire (IRQ) . DUs were also questioned on their injection and sniffing history over their whole lives. We also collected information on knowledge of HIV, HCV and HBV transmission modes.
To obtain seroprevalence data, we analyzed blood samples on blotting papers. The dried blood spots (DBS) were cut out with a punch to obtain a circle of about 6 mm diameter, which was placed in 250 μl of 0.01 M phosphate sodium buffer containing 10% bovine serum albumin and 0.05% Tween 20 (PBS-BSA-TW) then incubated at room temperature for one hour in an ultrasonic cleaner. The eluted serum samples were directly used to fill the wells of ELISA microplates (200 μL per well). The above steps were performed strictly as recommended by the manufacturer. The third-generation ELISA kit from Ortho-Clinical Diagnostics (HCV 3.0, Raritan, NJ) was used to detect anti-HCV antibodies. The same procedure was used to detect HIV antibodies using the Ortho HIV1/2 Ab capture Elisa kit. Any anti-HIV positive sample was confirmed by serotyping and/or western blot .
We assessed the validity of the ELISA technique applied to DBS by comparing absorbance values for various serum samples positive for anti-HCV antibodies or anti-HIV antibodies, either tested directly as serum and as DBS eluates. The data showed an excellent concordance of results except for weakly positive sera from patients with resolved HCV for whom HCV-antibodies could be missed. Thus prevalence rates of HCV infection based on the DBS method must be considered as minimal estimates (see appendix - Additional file 1).
Blood samples were collected, after informed consent, by means of fingerprick self-sampling. The individuals were informed that they would not receive the results of the tests made from fingerprick blood sampled on DBS. However, they were informed on the benefits and places of HCV and HIV screening and treatment through leaflets and posters in the survey participating structures. In addition during interviews, interviewers stressed the benefit of screening and directed DUs to near by screening services. The study protocol was approved by an ethical review board (Autorisation Number CCPPRB 02-002) because of its public health added value and the indirect individual benefit for participating DUs.
For this kind of cross-sectional design, we classically assign a sampling weight to each individual to estimate the epidemiological indicators of interest (prevalence, odds ratio etc.) in the target population. The sampling weight is the inverse of the inclusion probability of a DU in the survey sample. The fact that DUs can use more than one structure during the study period has an impact on these inclusion probabilities. To calculate a correct sampling weight, we have to take into account the number of visits in participating structures during the course of the survey for each DU. We are in the general framework of an indirect sampling  in which there are two populations: the DU, noted UA and the structures, noted UB, that are related to one another. We want to produce an estimate for UB using the existing links between the two populations. To estimate an epidemiological indicator in a target population UB using a sample selected from another population UA is a major challenge if the links between the units of the two populations are not one-to-one because it becomes difficult to calculate the inclusion probabilities and consequently the sampling weights. A solution proposed by Ardilly and Lavallee is the generalised weight share method (GWSM) [43,44].
This technique is meant to be a methodological framework to estimate weights, including several weighting techniques used for different designs (longitudinal studies, network sampling, time-location sampling, snowball sampling, etc...). Time-location sampling is more widely described and used in the literature, particularly for sampling hard-to-reach or hidden population . However, we introduce the GWSM as we assert that it is more appropriate than the time-location sampling because (1) the GWSM is more rigorous from a sampling theoretical point of view, while the TLS focuses on the probability of individuals' behaviour (model-based approach) and (2) the GWSM can account for several cases where the relationship between the sampling frame(s) and the target population is more complex.
To our knowledge, this method has not been yet introduced in epidemiology as a rigorous statistical method allowing to (1) calculate sampling weights and (2) encompass several methods more widely used and described in the literature.
This method produces a new weight for each unit in the target population UB which is an average of the sampling weights of the population UA from which the sample is selected. A fictitious sampling example is illustrated in Figure 1. The population UA contains 16 structures and the population UB contains 11 DUs. The shaded structures were sampled pointing to sampled DUs. The arrows indicate the links between the services and the DUs. Individual 1 reported a unique visit during the survey period. Individual 4 reported three visits in a sampled drug treatment center and in two non sampled services (accommodation n°13 and outreach work team n°16) represented by the dotted arrows. Individual 8 represents a fictitious situation in our study as it was almost impossible that a DU was interviewed several times. In practice, three questions were asked to DUs on the number of visits and the identification of structures visited to identify the links between the two populations. The GWSM needs inclusion probabilities only for the selected units j in the sample sA. This is a major advantage in comparison with other methods based on exact calculation of selection probabilities of surveyed units. The classical sampling weights wi = 1/πi are first calculated for each individual i. Then the number of services visited by the individual i during the survey, noted ri and the number of times when the individual i has visited the service k during the survey, noted rik were calculated. Finally a new weight is calculated by . For the illustrative example in figure 1, the new weights are , , .
Figure 1. Observed links (arrows) and declared links (dotted arrows) between the population of services and the population of drug users, ANRS-Coquelicot study, France 2004–2005. The shaded boxes represent the sampled services and the sampled individuals.
All results reported in this article are estimates that take into account the two stage survey design including the GWSM. Statistical analyses were done with Stata 9 software. Comparisons of proportion were made with the chi-square test, with a significance threshold of 0.05. Prevalence ratio (PR) and their 95% confidence intervals (CI) were calculated to analyse univariate association between HCV status and risk factors. A multivariate logistic regression model was then used to identify independent risk factors for HCV seropositivity using a manual stepwise strategy. We also tested interactions between explanatory variables and no interaction was significant. Prevalence odds ratios are presented in our multivariate analysis, as it is more difficult to estimate prevalence ratios in a multivariate analysis taking into account the sampling design.