Combination methods for HIV prevention in men who have sex with men (MSM)

  • Protocol
  • Intervention



This is the protocol for a review and there is no abstract. The objectives are as follows:

This review seeks to examine the effectiveness of prevention interventions for MSM that combine at least two of the three categories of intervention modality (biomedical, behavioural and structural) compared either to other HIV prevention interventions or to minimal/no HIV prevention.

Below, descriptions of the criteria that will be used to categorise intervention components are provided under the heading ‘Types of interventions’ and the full list of outcomes of interest are provided in the section ‘Types of outcome measures’.

All planned comparisons are detailed in the section ‘Data synthesis’, followed by the planned variables of interest for subgroup and meta-regression analyses.


Description of the condition

Addressing the global HIV/AIDS pandemic is a matter of extreme urgency. At the end of 2011 an estimated 34 million people were living with HIV globally, while 2.5 million became newly infected with the virus and 1.7 million died of AIDS-related causes that year (UNAIDS 2012). In nearly every country for which figures are available, HIV prevalence among men who have sex with men (MSM) exceeds that within the general population, with regional prevalence estimates ranging from 3% in the Middle East and North Africa to 25% in the Caribbean (Beyrer 2012). Myriad biological, social and institutional factors combine to explain the susceptibility of MSM populations to HIV infection. Unprotected receptive anal intercourse (URAI) facilitates the transmission of HIV more readily than unprotected vaginal intercourse; a recent systematic review estimated that transmission probability during URAI is approximately 18 times greater than during unprotected vaginal intercourse (Baggaley 2010). Moreover, the stigma and discrimination faced by MSM in many societies, and the state-sanctioned homophobia of some countries, including the criminalisation of homosexuality, create environments in which MSM are difficult to reach with prevention services, are disproportionately affected by HIV risk factors like untreated sexually transmitted infections (STIs) and mental health problems, and cannot always safely access health services, including HIV testing and treatment (Altman 2012).

Description of the intervention, and how the intervention might work

With no known cure for HIV/AIDS, prevention remains central to the global response to the pandemic. Condom provision and promotion, and behavioural interventions ranging from individual and group counselling to gay venue-based peer education, constitute the traditional approach to prevention.

The efficacy of correct and consistent condom use for reducing HIV transmission is well established (Weller 2002). A review of five cohort studies of MSM and transgender people revealed an overall 64% protective effect of consistent condom use (WHO 2011), while a large prospective cohort study of MSM found that condom use reduced per-contact risk of HIV acquisition among receptive partners by 78% compared to unprotected anal intercourse (UAI; Vitinghoff 1999). However, acceptability of and adherence to condoms is imperfect and varies across populations of MSM (UNAIDS 2012), limiting their success in the production of population-level prevention benefits (Sullivan 2012). Behavioural interventions to combat these barriers to condom use, and to improve other sexual behaviour outcomes like reduced number of partners, have proven only partially effective, and most evaluations of these interventions have been unable to detect an effect on HIV incidence (Herbst 2005; Herbst 2007; Johnson 2008; Lorimer 2013).

Increasing acknowledgement of the ‘structural’ drivers of HIV, that is, the broader social, economic, political and environmental factors that increase vulnerability to HIV infection, such as human rights violations, stigma, and poverty (Auerbach 2011), and the consequent need for structural-level prevention interventions (Blankenship 2000; Gupta 2008), combined with recent advances in biomedical prevention strategies including oral pre-exposure prophylaxis (Grant 2010), and the growing consensus that neither biomedical nor behavioural interventions alone are sufficient to curb the epidemic (Coates 2008; Padian 2008; Sullivan 2012) has led to a shift in popular thinking on HIV prevention in recent years. Combination HIV prevention, in which multilevel packages of prevention components drawn from the three major classes of intervention modality – biomedical, behavioural and structural – are strategically assembled and tailored to the needs of target populations (UNAIDS 2010), is widely considered the best available strategy for achieving population-level decreases in HIV incidence (Chang 2013; Hankins 2010; Kurth 2011; Merson 2008; Piot 2008; Vermund 2013). If assembled appropriately, these tailored intervention combinations are hypothesised to produce synergies between single, partially effective prevention modalities, yielding greater outcome effects than the sum of their constituent parts (Merson 2008; Piot 2008; UNAIDS 2010).

For the purposes of this review, combination HIV prevention interventions will be defined as intervention packages that combine components from at least two of the three classes of intervention modality (biomedical, behavioural and structural). Drawing on descriptions and representative examples outlined in the UNAIDS Discussion Paper on combination HIV prevention (UNAIDS 2010), intervention component categories will be defined as follows. Biomedical intervention components reduce the risk of HIV exposure, transmission, or acquisition either by providing a physical barrier between the virus and susceptible tissues (e.g., condoms) or by establishing physiological conditions that limit the replication and/or survival of the virus in (e.g., oral antiretroviral-based prevention) or on (e.g., antiretroviral microbicides) the body. Behavioural intervention components (e.g., risk reduction counselling) aim to decrease the risk of HIV transmission among MSM by promoting behaviour change within the MSM population through, for example, the enhancement of HIV-related knowledge and attitudes, or the development of protective skills and emotions (Herbst 2007). Finally, structural intervention components (e.g., population-wide homophobia reduction interventions) are designed to alter the broad contextual factors, operating beyond the MSM network and gay community (Adimora 2010), that work to increase the vulnerability of MSM to HIV infection and/or hamper the effectiveness of other HIV prevention interventions (Auerbach 2011).

Mathematical modelling has provided conceptual support for the hypothesis that a strategic combination of HIV prevention interventions, if adequately scaled up, could have considerable population-level effects on HIV incidence in both the general population (Cremin 2013; Hallett 2008) and in MSM populations (Beyrer 2012a; Sullivan 2012; Wirtz 2013). Empirical evidence is required to confirm these findings.

Why it is important to do this review

This review will aim to summarise the existing evidence on multicomponent HIV prevention interventions targeted at MSM populations that combine components from more than one category of intervention modality. While no known trials of combination prevention for MSM have been conducted with the explicit goal of identifying synergies between individual components, it is informative to identify and summarise the results of existing intervention studies that have drawn components from multiple categories.

Our results will illuminate the effects on HIV and STI incidence and risky sexual behaviour of the different intervention component combinations that have been tested to date, and the study-level factors associated with favourable effects. This synthesis may help to inform the design of future combination prevention trials for populations of MSM, both through the mapping of the component combinations that have hitherto produced positive, negative, or null results, and the identification of combinations yet to be evaluated in MSM samples.


This review seeks to examine the effectiveness of prevention interventions for MSM that combine at least two of the three categories of intervention modality (biomedical, behavioural and structural) compared either to other HIV prevention interventions or to minimal/no HIV prevention.

Below, descriptions of the criteria that will be used to categorise intervention components are provided under the heading ‘Types of interventions’ and the full list of outcomes of interest are provided in the section ‘Types of outcome measures’.

All planned comparisons are detailed in the section ‘Data synthesis’, followed by the planned variables of interest for subgroup and meta-regression analyses.


Criteria for considering studies for this review

Types of studies

The review will include randomised controlled trials (RCTs), in which participants were allocated to intervention conditions either individually or by cluster, and non-randomised controlled studies in which observations were made in all groups before and after intervention implementation. Included studies will be limited to those in which a combination HIV prevention intervention was compared against a control group receiving a different HIV prevention intervention (including both single-component and other combination interventions), standard care, or no intervention.

We recognise that the exclusion of less rigorous non-randomised studies (e.g., single-group before-and-after studies) will probably result in the de facto exclusion of a subset of intervention evaluations, particularly those containing structural components, in which the use of control groups was considered not feasible, ethical and/or necessary by the study authors. Given the growing consensus that, in principle, no defining characteristic of structural interventions precludes the use of comparison groups or random allocation in their evaluation (Bonell 2006; Hayes 2010), evidenced in part by the multiple RCTs of combination prevention interventions containing structural components that have been conducted (e.g., Pronyk 2006) or are presently underway (e.g., Vermund 2013), it was determined that requiring included studies to employ comparison groups was reasonable.

Types of participants

Studies specifically targeting MSM, irrespective of age, language, geographic location, ethnicity, or any other sociodemographic indicator, will be included. For the purposes of this review, ‘MSM’ refers to a behavioural category inclusive of all biological males who have sex with biological males, regardless of their sexual or gender identity. MSM encompasses, for example, transgender women, male sex workers, straight-identified MSM, gay and bisexual men, and MSM who identify with a number of other culturally specific sexual and gender identities.

It bears mentioning that the term ‘MSM’, which was coined in the 1990s with the intention of reducing stigma against gay, bisexual, transgender and straight-identified men who have sex with other men, oversimplifies the array of sexual behaviours in which men engage, sacrificing specificity of behaviour in favour of the sensitive inclusion of all biological males who have sex with biological males. The term also fails to differentiate between the variety of sexual and gender identities, and social and behavioural characteristics thereof, represented in populations of ‘homosexually active’ men (Young 2005). While the importance of understanding the social and behavioural diversity of MSM is appreciated, the term ‘MSM’ continues to be widely used in the literature and, in the interest of standardisation across studies, its use was deemed necessary in this review.

If non-MSM participants are included in a study, we will obtain outcome data for the subset of MSM included.

Types of interventions

Any HIV prevention intervention containing at least two of the three major categories of intervention component (biomedical, behavioural and structural) will be eligible for inclusion. For the purposes of this review, components will be defined as below, based largely on the descriptions and representative examples provided in the UNAIDS Discussion Paper on combination HIV prevention (UNAIDS 2010). In order to be classified as biomedical, behavioural and/or structural, an intervention must explicitly aim to reduce HIV transmission risk according to the means described in the definitions below.

Biomedical interventions are those that reduce exposure to, transmission of, or infection with HIV, either by providing a physical barrier between the virus and susceptible tissue or by creating a physiological environment in or on the body that is hostile to viral replication and/or survival. Examples include: condom provision, medical male circumcision, oral pre- and post-exposure prophylaxis, rectal microbicides, antiretroviral ‘treatment as prevention’, and biomedical STI treatment services.

Behavioural interventions are strategies that promote individual behaviour change to reduce the risk of HIV transmission (though they need not be delivered at the individual level). While the categorisation of most behavioural intervention components is uncontroversial, some could arguably be classified as biomedical interventions. HIV testing, for example, is often justifiably considered a biomedical intervention because of its provision by medical practitioners and its effect as a gateway to engagement with medical services for those who test positive. However, for the purposes of this review we will consider HIV testing a behavioural intervention component because the well established preventive effects associated with the receipt of a positive test result (Crepaz 2009; Denison 2008; Marks 2005) operate primarily through behaviour change resulting from knowledge of positive serostatus. Other examples of behavioural interventions include: individual and group counselling, behaviour change communication (including gay venue-based risk reduction and condom promotion), educational interventions targeting MSM, peer education and persuasion, community development and/or mobilisation programs aiming to alter predictors of HIV risk within the gay community (and/or MSM network), and prevention commodity (e.g., condom) social marketing (e.g., via posters and billboards) targeting MSM.

Structural interventions are defined, for the purposes of this review, as interventions that aim to alter environmental or societal-level factors beyond the gay community (and/or MSM network) that may affect vulnerability to HIV infection or that may dilute the effectiveness of other HIV prevention interventions. These factors include, inter alia, population-wide accessibility of prophylactic technologies (Blankenship 2000); HIV/AIDS-related stigma among healthcare practitioners and policy-makers and in the general population (Valdiserri 2002); society-wide gay/MSM-related stigma (i.e., homophobia), which acts as a barrier to health service and prophylactic access for MSM (Fay 2011; Santos 2013) and, through the production of internalised homophobia among individual MSM (Berg 2013), can increase HIV-related risk behaviours like substance abuse (Shoptaw 2009) and unprotected anal sex (DeLonga 2011; Ross 2008; Ross 2013); and the presence of discriminatory policies, including those that criminalise consensual homosexual sexual activity (Beyrer 2011; Itaborahy 2013) and that outlaw or accord lesser status to same-sex domestic partnerships (Klausner 2006). While structural interventions may ultimately result in behaviour change, they are distinguished from behavioural interventions by the targeting of the contextual factors beyond the gay community/MSM network, that nonetheless affect individual or group behaviour, rather than the behaviours themselves (Gupta 2008). For example, while a behavioural intervention may aim to directly change risky behaviour by providing sexual health education to MSM and encouraging them to engage with health services, seek testing, and disclose their HIV/STI status to partners, a structural intervention may aim to decrease homophobia and HIV-related stigma among healthcare providers, creating ‘safe spaces’ the for testing and monitoring of STIs in the MSM population, thus facilitating risk reduction behaviour. Still, some intervention components could arguably be classified either as structural or behavioural interventions. Community development and/or mobilisation programs, for example, may be considered structural interventions because they often work by addressing population- and society-wide determinants of HIV risk like homophobia. The definition of structural interventions that will be used in this review is largely consistent with this classification. However, many community development and/or mobilisation interventions aim specifically to decrease risky sexual behaviour among MSM by changing community norms within the gay (or MSM) community, rather than the population at large. As noted above, this subset of interventions will be categorised as behavioural interventions for the purposes of this review unless it is clear that the community development/mobilisation is primarily geared towards campaigns that are structural in orientation. Examples of structural interventions include: large-scale community development programs, community mobilisation and dialogue programs, and educational media interventions that address determinants of HIV risk operating beyond the MSM population; stigma reduction programs and anti-homophobia campaigns targeted at the general population, policy makers, healthcare providers, or other practitioners or leaders; policy and legal reform; and broad-based programmes designed to improve access to health and HIV prevention services, including prophylactic technologies, across the general population.

Many condom-based prevention interventions, particularly those that both provide and promote the use of condoms, are likely to qualify both as biomedical and behavioural interventions. The distinction between these two types of component deserves elaboration. For an intervention to fulfil the classification criteria for any component category, it must both explicitly aim to reduce HIV transmission risk through the means described in the relevant definitions above and, if applicable, provide the material features of the intervention modality thought to be required to achieve risk reduction. The features involved in condom promotion interventions (behavioural) may include, for example, educational materials, posters, or counselling sessions, while those involved in condom provision (biomedical) are, at minimum, the condoms themselves. A condom promotion intervention that does not involve condom provision would be classified as a single-component behavioural intervention, while a condom provision program implemented without a substantive behavioural condom promotion component would be classified as a single-component biomedical intervention. A combination behavioural-biomedical condom promotion and provision intervention would therefore need to explicitly include the behavioural and biomedical intervention features (i.e., respectively, condom promotion materials or activities, and condoms) required to reduce HIV transmission risk in accordance with the above criteria. It is worth noting that large-scale condom distribution interventions are often classified as structural interventions (Charania 2011), particularly when their aims include the improvement of the accessibility, acceptability and/or availability of condoms (Blankenship 2000). While interventions involving only the distribution of condoms to MSM will be defined as single-component biomedical interventions in this review, those that combine widespread condom distribution with additional components addressing population-wide condom accessibility, acceptability, and/or availability (i.e., structural-level mediators of HIV vulnerability) will be categorised as combination biomedical-structural interventions.

Because of potential disagreement about the definitions of intervention categories, the description of intervention components by study authors as biomedical, behavioural or structural will not be used to categorise interventions for the purposes of this review. Rather, we will examine the content of interventions as described in study reports (or via communication with authors, as needed) to classify intervention components and categorise interventions according to the above criteria.

We anticipate considerable heterogeneity between studies on the basis of the different combinations of intervention components that are evaluated. As discussed below in the ‘Data synthesis’ section we will analyse pooled study results in subgroups on the basis of the presence of different types of intervention component and combinations thereof.

Types of outcome measures

To be eligible for inclusion in the review a study must have assessed at least one of the outcomes listed below. Biological outcomes (especially HIV incidence) are the most accurate indicators of the success of HIV prevention interventions. However, given that only one behavioural intervention trial (Koblin 2004) and no known evaluations of structural interventions for MSM have included HIV incidence as an endpoint, and that many interventions are designed to reduce HIV transmission risk among MSM who are living with HIV (for whom HIV incidence is an irrelevant outcome), it is useful and necessary to record several relevant sexual behaviour outcomes. It may also be informative to record any structural-level pathway variables targeted by interventions, though ex ante determination of these variables is problematic because their relevance as predictors of HIV vulnerability tends to vary depending on the context and the presence of other factors, and evidential support for hypothesised pathways between putative structural factors and HIV risk is limited (Auerbach 2011). Nonetheless, in addition to biological and behavioural outcomes, we will collect data on the broad categories of structural factors thought to be associated with HIV risk that are listed below.

When outcome data are reported at multiple follow-up points, data from the latest endpoint will be recorded in order to capture sustained intervention effects.

Primary outcomes

Biological outcomes:

  • Incidence of HIV

  • Incidence of other STIs

Behavioural outcomes:

  • Number of occasions of unprotected anal intercourse (UAI) during recall period

  • Proportion of participants reporting at least one occasion of UAI during recall period

  • Number of partners for UAI during recall period

Secondary outcomes

Behavioural outcomes:

  • Number of occasions of unprotected receptive anal intercourse (URAI) during recall period

  • Proportion of participants reporting at least one occasion of URAI during recall period

  • Number of partners for URAI during recall period

  • Number of occasions of unprotected insertive anal intercourse (UIAI) during recall period

  • Proportion of participants reporting at least one occasion of UIAI during recall period

  • Number of partners for UIAI during recall period

  • Number of occasions of UAI with serodiscordant partner(s) and/or partner(s) of unknown serostatus during recall period

  • Proportion of participants reporting at least one occasion of UAI with serodiscordant partner(s) and/or partner(s) of unknown serostatus during recall period

  • Number of serodiscordant partners and/or partners of unknown serostatus for UAI during recall period

Structural outcomes:

  • Indicators of societal homophobia and/or anti-MSM stigma, including tolerance of and attitudes towards homosexuality reported by decision-makers, healthcare providers, and members of the general population; frequency of reported anti-gay physical and verbal abuse and human rights violations; and introduction of punitive anti-gay laws and policies

  • Indicators of societal HIV/AIDS-related stigma, including tolerance of and attitudes toward people living with HIV (PLHIV) reported by decision-makers, healthcare providers, and members of the general population; frequency of reported physical and verbal abuse against PLHIV and HIV-related human rights violations; and introduction of coercive or discriminatory legislation against PLHIV

  • Indicators of the successful enforcement of laws and implementation of policies protecting MSM or PLHIV rights or preventing MSM- or HIV/AIDS-related discrimination

  • Potential society-wide economic mediators of HIV risk, including income per capita, economic inequality metrics (e.g., Gini coefficient), and affordability of housing and other basic needs

  • Indicators of the society-wide accessibility and uptake of health services including HIV prevention programming, HIV counselling and testing, STI prevention and treatment services, and HIV prophylactic technologies

  • Educational indicators, including societal-wide educational attainment, and levels of HIV/AIDS- and sexual health-related knowledge

Search methods for identification of studies

It is possible that reporting of relevant studies may not be uniform and that pertinent reports may be published in grey literature, local publications, bulletins of international, regional and national organisations, and conference proceedings. Moreover, relevant reports may be published in languages other than English. Several sources of published and unpublished data (listed below) will therefore be searched without restriction on the basis of publication date, language of report, location in which studies were conducted, or their publication status.

Electronic searches

Comprehensive search strategies, developed with the guidance of the Cochrane Review Group on HIV/AIDS, will be tailored to search a broad range of relevant databases, including the Cochrane Central Register of Controlled Trials (CENTRAL), MedLine, EMBASE, PsycInfo, Global Health, Web of Science, ERIC, the Social Sciences Citation Index, CINAHL Plus, Sociological Abstracts, TRIP, Social Policy and Practice, LILACS, and the WHO Global Health Library. The search strategy will combine the Cochrane HIV/AIDS Review Group HIV string, the Cochrane Collaboration highly sensitive search string for RCTs (altered to capture non-randomised studies), and a population filter for identifying studies involving MSM.

The AIDS Education Global Information System (AEGIS;, the U.S. National Library of Medicine (NLM) Gateway (, and the websites of the International Society for Sexually Transmitted Disease Research (ISSTDR; and international AIDS Society (IAS; will be searched for reports included in various HIV/AIDS conference proceedings

Websites of relevant organisations and agencies, including UNAIDS, WHO, other UN agencies, the World Bank, the U.S. Centres for Disease Control and Prevention (CDC), and the U.K. National Health Service (NHS) Centre for Reviews and Dissemination (CRD) at University of York, will be searched for relevant reports. The WHO International Clinical Trials Registry Platform (ICTRP) and will be searched for ongoing or prospective trials.

Searching other resources

Issues of major HIV-related journals from the past five years, including AIDS, AIDS Care, AIDS & Behavior, Journal of Acquired Immune Deficiency Syndromes, HIV Clinical Trials, AIDS Prevention and Mental Health, AIDS Education and Prevention, and AIDS Patient Care & STDs, will be hand searched in order to identify reports not published in the electronic databases listed above

Reference lists of included study reports will be scanned for mention of other, previously unidentified studies.

Study authors will be contacted by email, as needed, in order to access unpublished data of trials located through the above methods and to identify any other studies not identified through the above methods.

Data collection and analysis

Selection of studies

Two authors (BV & GJMT) will independently screen titles and abstracts for potential relevance. Full text versions of articles deemed potentially relevant by at least one author will be obtained for closer review by both authors, who will independently identify studies meeting the inclusion criteria. Disagreements will be resolved by deferring to the third author (CB). Reasons for exclusion of any full-text articles will be provided in a table of excluded studies.

Data extraction and management

Two authors (BV & GJMT) will independently conduct data extraction using a specialised form designed with the assistance of the Cochrane HIV/AIDS Review Group. Data will be extracted (or derived) on: study characteristics, including design and methodological quality, allocation procedure, blinding and sampling methods; description of the intervention, including its theoretical basis (if applicable), processes involved, delivery method, length of intervention and/or number of sessions, and categorization of intervention components (i.e., biomedical, behavioural and/or structural); study location and context; duration of follow-up; sample size, number of participants in each group, and attrition rates; information about participants (age, sexual orientation, gender identity, and other demographic characteristics); baseline health information; baseline sexual behaviour practice; and all sexual behaviour (e.g., UAI) and HIV/STI incidence outcome data provided. Conflicting assessments will be discussed and resolved between the two authors. The third author (CB) will resolve any persistent disagreements. Missing data will be sought from study authors, as needed.

Data will be entered into, managed and analysed using Review Manager (RevMan) 5.2 software. One author (BV) will enter data into RevMan and another (GJMT) will check all entries for fidelity with the data recorded on the final data extraction forms.

Assessment of risk of bias in included studies

The methodological quality of included studies will be evaluated using the Cochrane Collaboration’s tool for assessing risk of bias, as described in the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2011). Two authors (BV & GJMT) will independently classify studies as presenting either a ‘low risk’ of bias, ‘high risk’ of bias, or ‘unclear risk’ of bias in the following domains: random sequence generation, allocation sequence concealment, blinding of outcome assessment, incomplete outcome data, selective outcome reporting, and other potential biases or threats to validity. Criteria described in Chapter 8 of the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2011) will be used to appraise studies.

Discordant assessments will be resolved by discussion between the two assessing authors. Unresolved disagreements will be settled through consultation with the third author (CB).

Measures of treatment effect

For ease of comparison between interventions and for consistency of analysis it is desirable to express all intervention effect estimates using a consistent metric. All outcomes will therefore be expressed as odds ratios (OR) with 95% confidence intervals (CI).

For studies reporting dichotomous outcomes, such as incidence of UAI, we will calculate ORs and variances using dichotomous event tables of treatment successes and failures and standard OR formulae described in the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2011). For studies reporting means and standard deviations of continuous or interval-level outcomes, such as number of partners for UAI or number of occasions of UAI during the recall period, we will calculate standardized mean differences (and associated variances), and convert these to ORs (and variances) using methods described by Hasselblad 1995 (cited in Johnson 2002).

We will calculate all ORs such that values below one represent a favourable effect of the intervention and values greater than one represent a favourable effect of the comparison group. Values of approximately one represent null effects.

All OR and variance estimates will be converted to the logarithmic scale for meta-analysis. Resultant summary effect and variance estimates in the log format will be converted back to ORs for presentation and reported with 95% CIs.

Unit of analysis issues

For all included cluster RCTs, in which groups or clusters are the unit of randomisation rather than individuals, we will check that reported outcome data have been adjusted by authors to account for intra-cluster correlation. If necessary, authors will be contacted to confirm this. If unadjusted data are presented, we will use methods described in Section 16.3 of the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2011) to adjust data by deriving a ‘design effect’ adjustment factor for each relevant trial using an intra-cluster correlation coefficient (ICC) imputed based on values reported in other included studies. If no included studies report an ICC value, we will use an estimate of 0.005, the value reported in a community-level trial of HIV prevention for MSM (Kelly 1997) and employed in a Cochrane review of behavioural HIV prevention interventions for MSM (Johnson 2008). Where imputation is necessary, sensitivity analysis (discussed below) will be performed to assess the robustness of our results to the imputation of a range of different ICC values.

If studies are included in which multiple active intervention groups in which we are interested are compared to a single control group, we will divide the control participants by the number of intervention groups, as described in Section 16.5 of the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2011), to create separate pair-wise comparisons such that no single participant’s data are analysed more than once.

Dealing with missing data

Any missing data (e.g., outcomes) or statistics (e.g., standard deviations) will be requested from study authors.

Attrition rates for all studies will be noted and we will assess whether analyses have been performed on an intention–to-treat (ITT) basis by study authors. If not, we will perform ITT analyses under the assumption that all missing outcome data were treatment failures in the intervention group and successes in the comparison group. Sensitivity analyses (described below) will be performed to assess the robustness of the review’s results to changes to this ‘worst-case scenario’ assumption.

The likely impact of missing data will be discussed in the interpretation of the review’s results and loss to follow-up will be assessed as a potential source of bias.

Assessment of heterogeneity

Statistical heterogeneity will be assessed visually using the forest plot, and quantitatively using the Chi-squared and I2 statistics (Higgins 2003). Heterogeneity statistics will not, however, be used to determine the appropriateness of a particular model for meta-analysis.

Assessment of reporting biases

The potential risk for publication bias will be explored by generating and examining funnel plots of individual study effect size estimates versus standard errors, and evaluating any asymmetry thereof (Egger 1997). Funnel plots are not informative when few studies are included and/or when included studies have similar sample sizes (Higgins 2011). We will therefore not use funnel plots for outcomes for which ten or fewer studies are included, or for which studies are of similar size.

Data synthesis

If included studies, and/or subgroups of included studies (e.g., studies with similar combinations of the three categories of intervention component), are deemed sufficiently similar in terms of participants, outcomes and intervention components to allow for meaningful quantitative synthesis, standard meta-analytic methods will be employed, as per the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2011). Random-effects models, based on the method of moments (DerSimonian 1986), will be used under the assumption that the true effect sizes estimated by studies are drawn from a distribution of true effects rather than a single true value.

We anticipate considerable heterogeneity between interventions based on the presence of different categories of intervention modality (biomedical, behavioural and structural). To assess this, we will stratify interventions based on the four possible intervention component combinations included in the review (i.e., biomedical + behavioural; behavioural + structural; biomedical + structural; and biomedical + behavioural + structural).

Comparisons specified a priori are:

  • All combination interventions against minimal/no HIV prevention,

  • All combination interventions against other HIV prevention,

  • Biomedical and behavioural against minimal/no HIV prevention,

  • Behavioural and structural against minimal/no HIV prevention,

  • Biomedical and structural against minimal/no HIV prevention,

  • Biomedical, behavioural, and structural against minimal/no HIV prevention,

  • Biomedical and behavioural against other HIV prevention,

  • Behavioural and structural against other HIV prevention,

  • Biomedical and structural against other HIV prevention, and

  • Biomedical, behavioural and structural against other HIV prevention

If possible, we will also conduct a multiple interventions meta-analysis in the Bayesian framework, grouping interventions in the evidence network based on the combination of components. Put otherwise, each node in the network of evidence represents a specific combination of prevention components (biomedical, behavioural, or structural), with nodes for both no/minimal HIV prevention and for single-component HIV prevention interventions against which combination interventions have been tested. This analysis will be implemented in WinBUGS using code developed by the Centre for Research Synthesis and Decision Analysis at the University of Bristol. Analyses will be run on each of the primary outcomes on two independent chains for at least 100,000 iterations, with the first 50,000 discarded. Convergence will be assessed using Brooks-Gelman-Rubin plots. Information presented will include pairwise ORs between each node in the network, the probability of rank for each node, and the surface under the cumulative ranking line (SUCRA) for each node. Before analysis, the network will be evaluated for transitivity; if applicable, the effect modifiers listed below will be used for network meta-regression.

Subgroup analysis and investigation of heterogeneity

To assess the impact of potential study-level moderators of effect size subgroup analyses will be performed. Previous reviews of behavioural HIV prevention interventions for MSM have found differential effects on the basis of various participant and study characteristics (Herbst 2005; Johnson 2008). Moreover, we recognise the inequitable geographic distribution of the HIV/AIDS intervention trials performed previously, which have historically been disproportionately conducted in high-income countries (especially the United States), neglecting the low and middle income countries in which the burden of HIV/AIDS tends to be greatest (Ahmad 2011). It is useful, therefore, to map the geographic distribution of included studies, both to identify cultural and geographic populations for whom, and contexts in which, further research is required, and to explore whether previous trials have tended to produce differential effects by geographic location.

We will perform subgroup analyses based on the following variables:

Study characteristics

  • Date of study commencement

  • Geographic location in which the study was conducted

  • Study design (i.e., randomised versus non-randomised)

  • Duration of follow-up

Participant characteristics

  • Mean age

  • HIV serostatus

  • Proportion who are gay-identified

  • Education level

If possible, the effect of potential study-level covariates will be assessed by conducting multivariable random-effects meta-regression analyses (Thompson 2002) using the ‘metareg’ macro for STATA 12.0. We will include the following potential effect modifiers in our analyses

Intervention characteristics

  • Types and combinations of intervention components present

  • Duration of intervention

Participant characteristics

  • Mean age

  • Proportion who are gay-identified

  • Proportion who are living with HIV

Study characteristics

  • Income stratum of country in which study is conducted, as defined by the World Bank

Sensitivity analysis

If possible we will perform sensitivity analyses to test the robustness of our meta-analyses by calculating summary effects for all studies and then progressively excluding studies of lowest methodological quality and assessing the impact on the summary effect size.

We will also assess differences in the summary effect estimates on the basis of changes in assumptions related to missing outcome data. Results of an ‘available cases’ analysis will be compared to those of imputed case analyses under a ‘best case’ scenario (i.e., intervention participants lost to follow-up are assumed to be treatment successes and comparison participants lost to follow-up are assumed to be treatment failures) and a ‘worst case’ scenario (i.e., intervention participants lost to follow-up are assumed to be treatment failures and comparison participants lost to follow-up are assumed to be treatment successes).

For data from cluster RCTs in which imputation of an ICC estimate is necessary we will we assess the robustness of our findings to changes in the assumed ICC value.

Declarations of interest

We declare that we have no conflicts of interest.