Circumcision for HIV prevention in men who have sex with men: an updated meta-analysis

Search strategy:

We identified studies in five ways:

1. We assessed all articles included in the 2019 meta-analyses^7,8 or a 2022 systematic review¹².

2. We conducted an updated literature search (details in online appendix 1) in Academic Search Complete, Pubmed, Embase, PsychInfo, Cochrane, OVID Global Health, WHO Index Medicus and Scopus for variations of the terms “circumcision”, “MSM” and “HIV” during January 1, 2018--October 16, 2024.

3. We manually searched International AIDS Society (IAS), Conference on Retroviruses and Opportunistic Infections, and HIV Research for Prevention conference abstract archives from 2018 through 2024 for abstracts with the same terms above.

4. We searched biobehavioral surveys (BBSs) among MSM in a CDC-maintained repository. These surveys typically use respondent-driven sampling to assess HIV prevalence and collect risk and behavioral data. We attempted to acquire additional BBS reports but were unsuccessful.

5. We included our own primary work shared at IAS 2025.¹³

Analysis strategy:

We planned a descriptive analysis of all observational studies reporting on the association between circumcision and HIV among MSM, followed by the two meta-analyses described above. This strategy was chosen because the former could illustrate but not address the challenges that have plagued previous analyses, including any confounding not addressed in the studies themselves; whereas the latter mitigated this by comparing findings by sex role within the same study.

Inclusion and exclusion criteria:

One author (SD) performed screening. Products (articles or abstracts) reporting primary data on any HIV outcome (prevalent or incident infection, including self-reported positive result history) with any observational study design using any metric (odds ratio, hazard ratio, etc.) were includable. Exclusion criteria are shown in table 1. Products were not excluded for being available only through a prior meta-analysis as abstracted data (e.g., dissertations not retrievable by us).

Data abstraction:

One author (SD) abstracted measures of association for all studies in the descriptive analysis. We standardized coefficients of association so that values <1 represent lower risk of HIV infection with circumcision, whereas values >1 represent higher risk. We used adjusted coefficient of association measures where available; otherwise, we calculated an odds ratio from cell counts. Coefficients of association were also categorized as “lower-risk, significant” (coefficient < 0.94, p<.05), “lower-risk, not significant” (p≥.05), “neutral” (coefficient of association 0.94–1.06), “higher-risk, not significant” (coefficient of association > 1.1), or “higher risk, significant”.

Other fields abstracted included number of participants, country and region, years of data collection, study design, sampling and recruitment approach, variables adjusted for in calculating coefficients of association, stratifications provided, criteria for designation as “primarily insertive” when applicable, and prevalence of versatile role (neither exclusively insertive nor receptive) among participants. Study design was based on circumcision as the exposure of interest (e.g., an RCT of another intervention was termed cross-sectional if only baseline data were used.)

For studies included in the meta-analyses, initial abstraction of stratum-specific coefficients of association and cell counts for PI-MSM and other MSM was repeated by two other authors (DP, MP) and discrepancies reconciled through discussion. Cell counts were back-calculated from proportions and denominators where necessary. For studies which collected data by sex role but did not report stratum-specific coefficients of association, we contacted authors to request the additional data needed for the confounder-adjusted meta-analysis.

Risk of bias assessment:

We did not assess risk of bias on studies already included in the prior meta-analyses, but reviewed their risk-of-bias assessments. We assessed newly-identified included studies using the Newcastle-Ottawa scale for prospective cohort studies¹⁴ and a modified version for cross-sectional studies.¹⁵

Data analysis:

We performed descriptive analyses on the full set of studies in Excel, including distributions of results by number of studies, number of participants, and other variables of interest. Where relevant, some comparisons were made between what we defined as regions: US/Canada/Europe/Australia vs. Asia/Africa/Latin America.

We then completed an unadjusted random-effects meta-analysis in R using the metafor package, estimating pooled log odds ratios (logORs) and sampling variances among PI-MSM. We converted effect measures to logORs by log-transforming published ORs with their variances derived from reported 95% confidence intervals; we likewise converted risk ratios, hazard ratios, and prevalence ratios to logORs using the generic inverse-variance method. We selected adjusted effect estimates when available and used cell-count estimates otherwise. Where an adjusted estimate was unavailable, 2 × 2 contingency tables were used to compute logORs, replacing zero cells with 0.01 to allow estimation. Finally, we assessed the robustness of findings via two sensitivity analyses: i) a cell-count-only analysis, which excluded studies without raw counts, and ii) a prefer-cell-count analysis, which used cell-count estimates when available and adjusted estimates otherwise.

All models were fitted with restricted maximum likelihood (REML) and a random intercept for study, because several studies contributed multiple strata. The forest plot was ordered by region and variance.

Finally, to explore and mitigate confounding, we conducted the primary analysis, which we termed a “confounder-adjusted” meta-analysis, with the following rationale: Circumcision status is influenced, in childhood and adulthood, by a range of factors that may also be associated with HIV risk. These factors are not necessarily all captured or adjusted for in the existing data, confounding the analysis and potentially driving the data inconsistencies seen, including regional variation. They may include:

1. Social or religious group: Certain groups commonly practice infant, child or adolescent circumcision (e.g., Jewish religion, some ethnic groups). Membership in these groups could be associated – positively or negatively - with adult risk behaviors, and/or with tendency to seek sexual partners from among the same group (assortive mixing), leading to differences in partner pool HIV prevalence between groups which also affect risk.

2. VMMC as risk mitigation: As described more fully in Table 1, individuals who circumcise to reduce their HIV risk are by definition HIV-negative at that time, likely leading to low HIV prevalence upon entry to the “circumcised” group as compared to the general population, introducing self-selection bias. This bias would, to our knowledge, be confined to sub-Saharan African countries implementing VMMC, in the absence of large-scale circumcision uptake done specifically for HIV prevention elsewhere.

To address these issues, the confounder-adjusted meta-analysis used the same rules described above but estimated the pooled difference of logORs between PI-MSM and other MSM, among studies providing the needed stratification. The reasoning was that while in purely non-insertive MSM, circumcision status would not directly affect HIV risk, they may have lower or higher HIV prevalence than non-circumcised MSM due to the confounding factors above. If those factors affect HIV risk in insertive and non-insertive individuals in the same way, then we could “subtract” them out by subtracting the log odds ratio among purely non-insertive MSM from the log odds ratio among insertive individuals, yielding the ratio of the odds ratios. Importantly, this approach does not depend on identifying, collecting or adjusting for individual confounding factors; it only requires that they affect both role groups equally.

If these confounding effects are not substantial, the confounder-adjusted analysis should have broadly similar findings to the unadjusted analysis, including with respect to regional heterogeneity and publication bias. Additionally, if there are confounders which do not apply equally to insertive and non-insertive individuals, they would not be subtracted out and would remain as residual confounding. The full mathematical framework is provided in online appendix 2a.

For both meta-analyses, publication bias was assessed with funnel plots with asymmetry testing (for all studies, and by region). Regional variation was assessed with estimate heterogeneity testing to determine whether our approach to confounder adjustment mitigated it. For the confounder-adjusted meta-analysis, findings were compared across study types (prospective vs. observational). Finally, we repeated the confounder-adjusted analysis dropping the de Leon¹³ et al dataset, which was our own work and the only one drawn from a VMMC country after substantial VMMC scaleup. Code to reproduce the analysis is available in online appendix 2b and the underlying data in appendix 2c.

Role of the funding source:

PEPFAR-supported staff designed and conducted all analyses.

Included data:

For the descriptive analysis, forty-nine observational products representing unique datasets were includable^{16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63} with nine additional products excluded due to criterion 2 only (inclusion flowchart in figure 1).^{64,65,66,67,68,69,70,71,72} Of includable products, five were not included in the prior meta-analyses (two BBSs^21,39 and three manuscripts^27,57,61); all products excluded by criterion 2 only were also new. Included products totaled over 150,000 participants (N’s for the four unretrievable products^33,24,25,34 were not reported); the five new includable products contributed 6,544 of these. Four of the five were cross-sectional; one was prospective.⁵⁷

There was minimal missing data for most variables, primarily occurring in studies available only as abstracted data,^24,33,25,34, none of which met inclusion criteria for the meta-analyses. Included studies and those excluded based on location (Table 1, criterion 2) are provided in appendix 3. Additional results are in appendix 4.

The majority of products (26/49) provided only unadjusted analyses (17/27 among studies recruiting only in Asia/Africa/Latin America, vs 8/19 among those recruiting only in the US/Canada/Europe Australia group). Only two completed data collected after 2014, despite later publication dates.

For the unadjusted meta-analysis, 12/49 observational products were includable, totaling 13 with the trial. (Vs. eight and 11 in the Yuan and Zheng meta-analyses respectively). Of these, all except three^29,46,53 prospective studies were cross-sectional. None of the five products newly identified in our search were includable for either meta-analysis.

Definitions for PI-MSM were heterogeneous. Of the 13 included studies, three used cutoffs for percent insertive or non-receptive among recent encounters, ranging from 50–90%^9,42,62; three used general preference^44,49,53; one²³ used “normally” taking only a “top” role; and the remaining six ^{29,36,37,41,46} used exclusive insertivity/non-receptivity over periods ranging from two to 12 months, or a floating period based on time of infection.¹⁸ These distributions reflect a pool of participants selected to be primarily but not exclusively insertive.

For the confounder-adjusted meta-analysis, eight of the twelve observational products included in the unadjusted meta-analysis were again includable, after obtaining additional data from three products’ authors. Also includable were the 2024 trial and our own work (de Leon et al¹³), for a total of 10. Of these, all except two^29,53 prospective studies were cross-sectional.

Using the same categories for PI-MSM definitions described above, of the 10 included studies, numbers of studies were: two using percent cutoffs of 50–90%^9,42,62, three using general preference^44,49,53; zero “top”-only normal role; and five^{13,18,29,36,41} using exclusive insertivity/non-receptivity over the 2–12 past months or a floating period as above. These distributions reflect a similar primarily, not exclusively, insertive pool.

No study included in either meta-analysis adjusted for either injection drug use or PrEP use.

Descriptive analysis:

The 49 observational products were from North America, South America and the Caribbean, Europe, Asia/Pacific, and sub-Saharan Africa. Sub-Saharan Africa had only six includable products; Latin America had five. With respect to taking a versatile sex role, medians and ranges of 57% (40–64%) self-reported this role in US/Canada/Europe/Australia, vs. 27% (5–50%) in Asia/Africa/Latin America.

Of the five new includable studies, two were in Asia^27,61, one in the UK,⁵⁷ and two in sub-Saharan African, falling either outside the VMMC-implementing subregion (Cote d’Ivoire)²¹ or prior to VMMC scaleup (Namibia).³⁹

Of the total includable pool, eight studies were “lower risk, significant”, 23 “lower risk, not significant”, 10 neutral, 7 “higher risk, not significant”, and 1 “higher risk, significant” (figure 2a). Among the five new products included in this total, one was “lower risk, significant”; two “lower risk, not significant”; and two neutral or mixed. This distribution supports a protective association, but when measured by total participant N instead, this pattern is greatly attenuated (figure 2b). The median coefficient of association was odds ratio (OR)=0.79; among studies with significant results, it was OR=0.46, both indicating lower odds of HIV with circumcision. Protective associations were commonly seen in studies conducted in Latin America and Asia (14/17); compared to the United States, Europe and Australia (11/23); and in smaller studies (13/16 among studies with N≤500) compared to larger ones (10/26 among those with N>500). Among the 12/49 studies reporting on PI-MSM – of which all were retrievable manuscripts or abstracts - nine assessed significance: 1 “lower risk, significant”, 4 “lower risk, not significant”, 2 neutral or mixed lower and higher risk not significant (depending on stratum, when stratified by race), 2 “higher risk, significant”. The median coefficient of association was OR=0.61, indicating lower odds of HIV with circumcision.

Risk of intra-study bias:

Our review of the prior meta-analyses’ assessments of risk of bias found that Yuan et al rated 52% of their included studies (comprising 33/49 of our included studies) as low-risk for bias, and the remainder high-risk. Zhang et al did not report assessments. Our own bias assessment in the five new included studies found that 4 (cross-sectional) scored below 7/10 (high-risk); one (prospective) scored 8/9 (low-risk). Details are in Appendix 5.

Unadjusted meta-analysis:

This analysis returned an OR of 0.57 (95% CI 0.33–0.98, p=0.043) for HIV infection in circumcised vs. uncircumcised MSM (figure 3). The test for heterogeneity returned p=0.02, supporting between-study effect heterogeneity, and the test for asymmetry returned p<0.0001, supporting publication bias. Secondary analyses 1) using cell counts only and 2) preferring cell counts where available returned ORs of 0.56 (0.32–0.98) and 0.62 (0.37–1.03) respectively. Effect differences by region are presented in the forest plot; tests for funnel plot asymmetry returned p<0.01 for both regions, demonstrating publication bias. The effect of region was not significant ( p=0.15), but the difference in point estimates was large (logOR-ratio −1.07 for Asia/Africa/Latin America vs. −0.21 for US/Canada/Europe/Australia).

Confounder-adjusted meta-analysis:

The geographic distribution of studies included in both meta-analyses is shown in supplemental figure 1 (appendix 6). Nine of the ten included studies had ratio of ratio point estimates indicating a protective effect. The meta-analysis returned a pooled ratio of ratios estimate of 0.53 (p = 0.006; 95% CI 0.34–0.83), indicating a significant protective association (forest plot, figure 4). There was no evidence for between-study heterogeneity of effect (p=0.32) indicating that the studies were relatively uniform in their estimation of the circumcision effect. Additionally, the test for funnel plot asymmetry returned p = 0.34 (funnel plot, figure 5) showing a lack of evidence for any publication bias.

Differences by region were also not significant (p=0.65), with a small difference in point estimates (logOR-ratio −0.77 (−1.52- −0.03) for Asia/Africa/Latin America vs −0.54 (−1.44– 0.36) for US/Canada/Europe/Australia). These findings indicate protective point estimates for circumcision in both regions, with no regional heterogeneity. Dropping de Leon et al did not impact the estimate (log ratio of ratios = −0.54) or test for heterogeneity; it did impact significance (p = 0.055). Both prospective studies (Templeton and Jameson) had point estimates in the direction of a protective effect (ratio of ratios of .1 and .78 respectively), with magnitudes consistent with the cross-sectional studies.