Setting
This study is a cross-sectional ego-centric network survey of 170 FSWs in Tehran city, conducted between January and June 2017. An ego-centric network survey provides more insight into the effects of the social network around the FSW on her condom use by making an inventory of the personal network of the FSW and collecting relevant individual characteristics of the FSW herself as well as of her sexual partners. This purpose is achieved using a so-called name generator for making the inventory of the members of the personal network, and a name interpreter for determining their relevant characteristics and relevant interactions. In our study, the FSW who is the respondent is the focal individual (ego) in the personal network.
Participants
FSWs are a hard to reach population in Iran, subject to social and official taboos and social stigma. However, in recent years enormous governmental and non-governmental investments and efforts have been made to increase the access of this vulnerable population to HIV prevention programs, such as the creation of drop-in centers (DICs) and consultation centers for at risk women throughout the country. These centers provide free services including basic sexual and reproductive health care, educational programs about sexual transmitted infections (STI) and prevention methods, HIV testing, and counselling [29, 30].
It is virtually impossible to obtain a random sample from a hidden and stigmatized population such as this. Therefore, participants were recruited through snowball, purposeful, and convenience sampling methods as successful alternatives to non-applicable random sampling methods for recruiting ‘hidden’ populations [31,32,33]. Eligibility criteria were being over 16 years old, having had sex for money, drugs, and so on in the last year, identifying oneself as a sex worker, and willingness to participate in the study. For the snowball sample, to initiate the chain referral process, six FSWs who satisfied the inclusion criteria were selected as index participants. After interviewing the index participants, we asked them to introduce some other sex workers whom they had named as their network members. The snowball sample proceeded according to the principles of respondent-driven sampling [31]. We gave the respondents some coupons for their peer friends, which included the ID number, aim of the study, amount of incentive, place of visit for interview, and an expiration date. Each participant was given 100,000 Rails (equal to 3 USD) as a small primary incentive for participating in the study and completing the interview; and a secondary small incentive, of the same amount, if they had introduced their peer friends, who did sex work and named as their network members. This process continued until five waves. In all waves the participants introduced few peer friends. Some introduced only one or two peer friends named as network members, some others introduced other FSWs who had not been named as their network members. It was impossible to access all peer members who were named by the participants. We completed the sample with convenience sampling and purposeful sampling methods. Convenience sampling was used by recruiting participants among FSWs who attended DICs and consultation centers. Recruitment of study participants in DICs was facilitated by DIC staff, who were personally contacted at the center by the first author. The DIC staff identified potential participants and introduced them to the interviewer. To improve diversity from all involved sites, we also used purposeful sampling from outreach spots such as team homes, streets, and parks by a peer outreach worker who collaborated with the study via a DIC. To maintain anonymity of the FSWs, we received verbal informed consent from all participants, because sex work is illegal in Iran. We provided an explanation regarding the study purpose to all participants, and informed them about the confidentiality and their right to withdraw from the study at any time during the interview.
Measurements and variables
The data were gathered using an investigator-constructed questionnaire, after assessing content validity, scalability, and reliability. The questionnaire consisted of two parts, individual and network information. Individual information was collected about the demographic characteristics age, educational level, marital status, and place of living, and about the frequency of sex work in the last month, HIV knowledge, and HIV test. For network information, first, a name generator inventory [34] was applied to indicate the FSW’s sexual network members. The sexual network was defined as the set of nominated persons with whom they had had sex in the past 30 days. We asked the FSWs to nominate up to 5 persons with whom they had any sexual relationships during the past 30 days.
For these nominated persons further information was collected including socio-demographic information, duration of contact, frequency of contact, frequency of condom use, intimacy, social support, and drug or alcohol use before a sexual relationship with him. The data was collected in face-to-face structured interviews by trained peer interviewers, which was helpful to build trust and get honest responses from the participants. Each interview usually lasted around 45 minutes.
Socio-demographic variables
The socio-demographic variables were collected as follows. Age was recorded in years. Educational level was measured in six ordinal categories, coded as 1 to 6: illiterate, able to read and write, primary education, secondary education, high school or diploma, and university education. Marital status was measured in four categories as single, married, divorced, and widowed; for the analysis this was dichotomized into never married and ever married. Place of living was categorized as homeless, living in the home of others, and personal home. The number of people supported by the FSWs was measured as a count; for the analysis it was dichotomized as zero versus more than zero. The variables collected for the network members were age and educational level, measured similarly as for the respondents.
Social network variables
To assess the frequency of contact with the sexual network members, we asked the participants “How many times did you meet him or did you communicate with him in the last month?”. To assess duration of the tie we asked the participants “Since how long have you known him?”. This was coded in months. The intimacy of the relationship between participants and sexual network members was assessed with a five-point Likert response, with categories very close, close, somewhat close, distant, and very distant. Social support was assessed by an investigator-constructed questionnaire with five items. Reliability of the items comprising the social support questionnaire was pilot-tested prior to final implementation in this study. Cronbach’s alpha and intraclass correlation for the scale were .82 and .85, respectively. Mokken scale analysis was used to assess scalability and uni-dimensionality of the social support questionnaire [35]. The Loevinger H-coefficient for the questionnaire was larger than 0.5, characterizing it as strong scale.
Because sexual network members could be nominated for more than one role or interaction (a sexual partner may also have a familial tie and/or be a drug use partner), the multiplexity of their position was determined. This was defined as 1 if they had more than one role in the respondent’s network, and 0 if they had just one role in her network (i.e., only sexual partner). The density of each network was also assessed. To assess density, participants were given a matrix with the names of their mentioned sexual network members on both axes, and then were asked to indicate the pairs of network members who knew each other. The density was calculated as the number of sexual network members who knew each other divided by the maximum number possible, given the size of the sexual network.
The variable of drug and alcohol use before or with sex was dichotomized into “yes=1”, and “no=0”.
Perceived safe sex norm
Perceived safe sex norms were assessed by the perceived norm scale (PNS) of the safe sex norm questionnaire (SSNQ) [16]. The PNS assesses perceptions of the FSW regarding attitudes and behaviors about condom use of their peer friends who do similar sex work. It uses 17 questions each on a five-point Likert scale: “all”, “most”, “about half”, “some” and “none”. These were coded as scores from 1= “none” to 5= “all”. The reliability of the questionnaire was pilot-tested. The attainable score range is 17-85. Higher scores indicate that the perception of FSWs of the attitude and behaviors of their friends is that most of them have a positive attitude about condom use and frequently use it in their sexual relationships [16]. Cronbach’s alpha and the intraclass correlation for this scale were 0.89 and 0.83, respectively. Mokken scale analysis was used to assess scalability and uni-dimensionality of the questionnaire [35]. The Loevinger H-coefficient for the scale was larger than 0.5, characterizing it as strong scale.
HIV Knowledge
Knowledge about HIV transmission was evaluated by a 14-item tool based on an established questionnaire in the Iranian population [36]. The questionnaire was pilot-tested with 28 participants in Tehran. Scalability and uni-dimensionality of the questionnaire were assessed by Mokken scale analysis [35]. Each item and the whole questionnaire had Loevinger H coefficients above 0.4, which is good. Cronbach’s alpha for this scale was 0.86. The sum score was transformed to a scale of 0 to 100, with high scores meaning more knowledge.
Sexual practice and HIV test
For assessing sexual practices, the participants were asked to report about the frequency of sex work in the last month. This was recorded as a count variable. Participants were also asked to report if they ever had a HIV test. This, together with the test result, was recorded as a categorical variable with values “having HIV test & positive result=1”, “having HIV test & negative result or don’t know=2”, and “having no HIV test=3”.
Condom use
Frequency of condom use by participants and the sexual network member was the dependent variable, measured on a five-point ordinal scale. It was defined as Yij = 1 if for network member i of respondent j condom use was reported as ‘never’, Yij = 2 for ‘rarely’, Yij = 3 for ‘sometimes’, Yij = 4 for ‘often’, and Yij = 5 for ‘always’.
Statistical analysis
Pearson correlations were computed to get a basic insight about patterns of association of all variables. The distributions of the variables were assessed carefully for missing values and outliers, as these might unduly affect the results. Frequency of sex work had three values larger than 20 (two values 25, one 60) which were considered outliers; these were truncated to 20.
Taking into account the FSW as well as the sexual partners implies a multilevel structure [37], with the sexual partners, as network members, nested in FSWs as respondent level. The dependent variable, condom use, is a variable at the partner level. As explained below, the analysis proceeds in steps according to a conceptual framework summarizing the role of individual and network characteristics for HIV risk behaviors of FSWs. This framework is based on previous studies of the effects of various factors on HIV risk behaviors [14, 38,39,40] and goes from the more general background characteristics to the social network characteristics.
Responses about network members of the same respondent are likely to be correlated. Therefore, we used multilevel analysis [37, 41]. The number of respondents (called ‘level-2 units’ in the terminology of multilevel analysis) was 170; the total number of sexual network members (‘level-1 units’) was 615. We created a group mean (for groups defined as all sexual network members of a given FSW) for each explanatory variable at the level of the sexual partner, including age of sexual partner, education of sexual partner, frequency of contact, duration of tie, intimacy, social support, and drug or alcohol use before or with sex. This is required to investigate the difference between within-group and between-group regressions [37]. The within-group regression coefficient is the estimated parameter at the partner level, the between-group regression coefficient is the sum of the FSW-level and the partner-level coefficients.
Since the distribution of the dependent variable is highly skewed, with five values, we employed a multilevel ordered logistic regression model for ordered categorical outcomes [37]. This was the multilevel proportional odds model, which can be formulated as threshold model with C ̶ 1 thresholds where C is the number of categories of the outcome variable; here C = 5. The mathematical expression of the model is
$${\displaystyle \begin{array}{c}\mathrm{P}\left\{{Y}_{ij}=c\right\}=\mathrm{P}\left\{{\theta}_{c-1}<{\tilde{Y}}_{ij}\le {\theta}_c\right\}\\ {}{\tilde{Y}}_{ij}={\sum}_h{\beta}_h{x}_{hij}+{U}_{0j}\end{array}}$$
Here Yij is the observed dependent variable for network member j of respondent i; c is an outcome ranging from 1 to 5; and P indicates probability. Ỹij is a hypothetical unobserved auxiliary variable, which can be regarded as an underlying continuous variable that is observed after categorization according to thresholds θ1, θ2, θ3, and θ4. The observed outcome is c when Ỹij is between the two thresholds θc-1 and θc, where the two outer thresholds formally are defined as minus or plus infinity: θ0 = ̶ ∞ and θ5 = +∞. The βh are regression parameters; finally, the xhij are the explanatory variables, which cover characteristics of respondents i as well as of sexual network members j; and U0j is a respondent-level random effect with a standard logistic distribution.
We calculated the intraclass correlation coefficient (ICC), a descriptive statistic that measures the proportion of total variance of an outcome that is accounted for by the group level; in this case, the groups in the data refer to the FSWs. In other words, the ICC measures similarity in condom use between sexual partners of the same FSW. It was calculated according to formula (17.26, page 311) in [37], taking the within-group variance equal to π2/3=3.29 (the variance of the logistic distribution).
The model selection utilizes a conceptual framework based on previous studies of the effects of various factors on HIV risk behaviors [14, 38,39,40] and goes from more general background characteristics to the social network characteristics. This framework distinguishes three groups of independent variables. The first group is composed of the individual background characteristics of the FSW and her sexual partners, as indicated by age, education, and number of supported people. The second group consists of HIV knowledge and the behaviors directly associated with sex work: its frequency, whether it is accompanied by drug or alcohol use, and HIV testing. The third group is composed of social network characteristics and psychosocial mechanisms through which these may affect condom use, regarded as a behavior protecting against HIV risk: personal network density, and the tie characteristics such as duration of the tie, frequency of contact, and intimacy, social support, social norms, and drug or alcohol use before or with sexual relationship. These three groups may be interpreted as reflecting a hypothetical causal ordering, but we use this as a framework guiding the analysis and do not rely on assumptions of causality. In the multilevel ordered logistic regression analyses, a stepwise model selection procedure was employed, in which the groups of variables were entered sequentially. This allowed estimating the effect of social network characteristics on condom use while controlling for individual background characteristics of the FSWs and their sexual partners. It started with the empty model which contained only the dependent variable and the threshold parameters. Micro soft Excel was used for data management (data entry, quality control and cleaning of quantitative data). The data was analyzed by the ordinal package [42] in the R statistical system [43] which allows fitting a variety of mixed effects models for categorical outcomes. p-values less than 5% were regarded as statistically significant.
