Receptive anal sex contributes substantially to heterosexually acquired HIV infections among at‐risk women in twenty US cities: Results from a modelling analysis

Receptive anal intercourse (RAI) is more efficient than receptive vaginal intercourse (RVI) at transmitting HIV, but its contribution to heterosexually acquired HIV infections among at‐risk women in the USA is unclear.


| INTRODUC TI ON
Current evidence suggests that penile-anal intercourse (receptive anal intercourse, RAI) increases the risk of HIV acquisition per sex act by up to 18-fold compared to one act of penile-vaginal intercourse (receptive vaginal intercourse, RVI). 1,2 Previous modelling studies suggested that even if 5%-10% of all heterosexual sex acts are RAI, this may reduce the effectiveness of HIV interventions that are only efficacious for RVI, such as daily or long-lasting vaginal microbicides, by up to 50%. 3,4 Recent systematic reviews suggest RAI is commonly practised by women across a variety of populations and contexts throughout their lifetime. 5,6 For example, 2%-36% of South African women and 20%-24% of sexually active women under 25 years old worldwide report ever engaging in RAI. 5,6 National surveys in the USA and UK indicate that around 12% and 11% of women aged 18-59 years, respectively, engaged in RAI over the previous year. 7-9 RAI prevalence among women may also have increased since the 1990s, 8,10,11 even doubling in places. 8,10 Women reporting RAI often also report other practices such as exchange sex, 12 high numbers of sexual partners and more frequent sex acts, 13,14 substance use, 12,13,15,16 coerced sex 17 and low condom use 14,18 associated with increased risk of acquiring sexually transmitted infections (STI), including HIV. Nevertheless, RAI has, until recently, been sidelined from receptive partner-controlled HIV prevention and product innovation both for men who have sex with men (MSM) and women. [19][20][21][22] Understanding the epidemiological context of RAI among heterosexual women and its contribution to HIV and other STI is necessary to tailor prevention messaging and product development, such as rectal microbicides. 23,24 In the USA, heterosexual transmission accounted for 24% of all adult and adolescent HIV infections diagnosed in 2017, while MSM and people who inject drugs (PWID) accounted for 70% and 6% of cases. 25 Black/African American (henceforth, Black) and Hispanic/ Latino (henceforth, Hispanic) populations are disproportionately affected by HIV. 26 In 2017, the HIV diagnosis rates (per 100 000) for adult and adolescent Black and Hispanic women were, respectively, 15 and 3 times the rate for White women. 25 The Centers for

Disease Control and Prevention (CDC) established the National HIV
Behavioral Surveillance (NHBS), a comprehensive system for conducting behavioural surveillance among people at high risk for HIV infection in the USA and identifying risk factors (including RAI) associated with infection. 27,28 Using NHBS data from low-income women at increased risk of HIV infection living in 20 US cities, we (a) describe RAI practices across key demographic and risk factor groups among women reporting heterosexual intercourse in the past year and (b) use these results to inform a mathematical model and estimate the annual fraction of new heterosexually acquired HIV infections that are due to RAI among at-risk women in the NHBS sample overall and in 20 cities.

| Data used
NHBS has conducted independent serial cross-sectional behavioural surveys among heterosexual women (NHBS-HET) living in high HIV prevalence metropolitan statistical areas (MSAs) every 3 years since 2007. 26,27 To derive average estimates for 2010 and 2013, our analysis combines the data from 20 MSAs included in the 2010 and 2013 study cycles. The 2007 cycle was excluded from this analysis due to differences in the sampling methods. Detailed data collection procedures have already been described. 26,29 In short, participants were recruited through respondent-driven sampling (RDS). 30 Recruitment prioritised women with household income below the federal poverty guidelines 31 or with no more than high school education. Individuals aged 18-60 years were eligible to participate if they lived in a participating Results: Receptive anal intercourse prevalence (overall: 32%, city range: 19%-60%) and RAI fraction (overall: 27%, city range: 18%-34%) were high overall and across cities, and positively associated with exchange sex. RAI accounted for an estimated 41% (uncertainty range: 18%-55%) of new infections overall (city range: 21%-57%).
Variability in PAF RAI estimates was most influenced by uncertainty in the estimate of the per-act increased risk of RAI relative to RVI and the number of sex acts.

| Statistical analyses
In this study, the level of RAI in the population was characterised by (a) RAI prevalence, the proportion of women reporting RAI with at least one partner in the past year, and (b) RAI fraction, the fraction of all acts (RVI and RAI) at last sexual episode that were RAI among those reporting RAI in the past year. We calculated RAI prevalence and RAI fraction estimates stratified by key demographic and risk factors (age, race/ethnicity, exchange sex and partner type), overall and by city for the two cycles combined. We also used two binary outcome variables measuring (a) whether or not women practised RAI in the past year (RAI and non-RAI women, respectively); (b) among sex acts of RAI women at last sexual episode, whether or not the sex act was RAI. Bivariate and stratified analyses, using chi-squared and Mantel-Haenszel tests, respectively, 32 compared both outcomes across the levels of demographic and risk factors except for condom use during the last sexual episode across which only RAI prevalence was compared; two-sample t test tested differences in the mean annual number of partners. Stratification controlled for city and additionally for each of the factors in turn. Among RAI women, bivariate and stratified analyses compared condom use at last vaginal sex with condom use at last anal sex. Partner's race/ethnicity was excluded from analyses as it was only collected in one cycle. We treated the data as a convenience sample; we report unadjusted and stratified odds ratios as well as 95% confidence intervals and P-values based on normal approximation. We report estimates and confidence intervals unadjusted for network size or clustering of RDS recruitment chains because these were used to derive prior ranges for parameters in the risk equation model and in model sensitivity analyses and not to make inferences about the wider population of low-income heterosexual women at risk of HIV. These intervals are therefore likely to be narrower than if recruitment-chain clustering was accounted for. This decision was taken because in sensitivity analyses, quadrupling the standard error did not affect our model predictions of the PAF. or city (J = 20 groups) and finally by exchange sex within city (details in Tables S1 and S2, parameter ranges for exchange sex within city available on request). CIR i,j depends on the annual number of sexual partners (m i,j,k ) of type "k" (ie main or casual) for women in groups ij, the probability that a male partner j′ is HIV-infected (p j′ ), the annual number of sex acts per partnership of type k (n i,j,k ), the fraction of sex acts which are RAI (f a i,1,k ), the probability of using a condom during RVI (f cv i,1,k ) and RAI (f ca i,1,k ) per partner type, condom efficacy in reducing HIV transmission during one RAI or RVI act (e c ),

| Risk equation model
HIV transmission probability per RVI (β) and the relative risk of HIV infection during RAI compared to RVI (RR RAI ). For each factor, the overall HIV risk is the average of the cumulative incidences over the J groups weighted by relative group size and RAI status (F i,j ) (Equation in Appendix S1 part A).

| Parameter assumptions and model calibration
Uniform ranges of plausible values were specified for each parameter (ie prior parameter range) based on the unadjusted 95% CI of the estimates from NHBS data (overall and by demographic or risk group) when available and sourced from the literature otherwise At the fitting stage, we simultaneously sampled prior parameter ranges using Latin-hypercube sampling 39 to generate 10 000 parameter sets that were used to produce model predictions of the annual cumulative HIV incidence risk (CIR i,j ). Predicted cumulative incidence risk estimates were converted to annual incidence rates to be comparable with data observed in 2010 from HPTN-064 (Appendix S1 part C). 40 We retained entire parameter sets if predicted rates and risk ratio fell within the 95% CI of HPTN-064 HIV incidence rate 38 and incidence risk ratio, 37 and if the total number of sex acts in a year across all partnerships and the ratio of sex acts reported by RAI women and non-RAI women generated by the model agreed with available data from similar US populations and other sources (prior parameter ranges in Table 1; details in Appendix S1 parts A, B). Given the lack of city-specific HIV incidence data, we estimated it by applying a scaling factor, to the HPTN-064 incidence rate estimates, based on 2013 HIV diagnosis rate among adults for each city (details in Appendix S1 part C). The

TA B L E 1 Model parameters and their ranges for the overall model
Parameter (for individual in RAI group i = 1,2 and overall model J = 1) Parameter ranges

<.001
Note: All P-values are derived from the chi-squared statistic, 95% confidence interval using the Wilson interval, and 95% confidence intervals for odds ratios are estimated using normal approximation (Wald). Abbreviation: NA, not available. c By definition, this refers to acts with main or casual sexual partners at last sexual episode.
resulting sets of fitting parameters define the baseline scenario in our modelling analysis.

| Modelling analysis
For each retained parameter set, we derived two population at- We report the median PAF and 10th-90th percentile uncertainty intervals (80% UI). We conducted an uncertainty analysis first to assess which parameter most influenced the variation across overall PAF estimates using Pearson's correlation coefficients. Second, we conducted a more general sensitivity analysis using wider parameter ranges and fitting only on the incidence rate and incidence risk ratio (as opposed to fitting additionally to the number of acts and act ratio). The first and second analyses help determine which additional and new data to prioritise for collection in the context of the NHBS study and in settings where little data are available,

respectively.
Although our analysis focused on heterosexual transmission, we also assessed the potential influence of HIV transmission by needle sharing on PAF RAI estimates (see detailed methods in Appendix S1 part D). We explored scenarios where we assumed that HIV incidence rate among PWID was the same, twice or five times larger than among women who do not inject (NIDU).
All analyses were conducted in R version 3.4.0 41 using R-studio version 1.0.143. 42

| Study sample
Women had a mean age of 37 years, 24% were under 25 years old, 73% were Black, 41% reported having had only main partners in the last year, and 23% reported exchanging sex. HIV prevalence was 3% overall and similar between women who did or did not practise RAI in the past year (OR = 0.99, 95% CI 0.77-1.29). Apart from race/ ethnicity, other city-level demographic and risk factor patterns were broadly similar to overall patterns. In all but four cities (Denver, Los Angeles, San Diego and San Juan), the percentage of Black women was at least 60%. In these four cities, the percentage of Hispanic women was >40% (Figure S1A-D). only main partners (19%) or no exchange partners (26%) but consistent across marital status ( Table 2).

| How common is RAI?
RAI prevalence was consistently high across cities, ranging from 18% in New Orleans (95% CI 16%-22%) to 60% in San Juan (95% CI 55%-65%; Figure 1 and Table S2). Among 18-to 19-year-olds, RAI prevalence varied substantially across cities (mean ranges: 6% in Miami to 61% in San Juan) and was lower than among 25to 29-year-olds in 5 cities; prevalence differed less among 25-to 29-year-olds and older age groups ( Figure S2A). Across cities with at least 10 participants, RAI prevalence ranged between 18% and 50% among Blacks, 15% and 60% among Hispanics, 14% and 69% among Whites and 20% and 64% among other races/ethnicities ( Figure S2B). In all cities, RAI prevalence was higher among women who had casual-only and main and casual partners over the past year than women who had main partners only ( Figure S2C) and was higher among women exchanging sex than women who did not ( Figure S2D).
Across cities, the RAI fraction varied between 18% (95% CI 12%-25%) in Denver and 34% (95% CI 29%-39%) in San Juan. Patterns of RAI fraction by age group, race/ethnicity and exchange sex within cities resembled the overall patterns (though comparisons across cities were limited by small numbers) (data not shown).

| Uncertainty and sensitivity analyses
First, in our uncertainty analysis, RR RAI (correlation = 0.89) and the total number of RVI acts among non-RAI women (correlation = 0.28) were the parameters that were most associated with the variation in PAF RAI estimates ( Figure 3A). PAF RAI 80% uncertainty interval estimates increased from 20%-30% to 43%-58% if RR RAI increased from 5 to more than 15, respectively ( Figure 3B), and decreased from 28%-55% to 15%-51% if number of vaginal acts increased from 50 to 150 ( Figure 3C). Second, in our more general sensitivity analysis with wider parameter ranges the correlation with RR RAI was substantially lower (correlation = 0.59) and RAI prevalence and RAI fraction became more influential (correlation = 0.35, −0.27, respectively) followed by HIV prevalence among male partners ( Figure 3D).
Finally, we assessed the influence of transmission due to needle sharing among PWID on our PAF RAI estimates. Assuming 100% of infections among PWID come from injecting behaviours, the PAF RAI due to RAI would range from 39% (UI: 17%-52%), 37% (UI: 16%-49%) and 32% (UI: 14%-43%) if the HIV incidence rate among PWID was the same, twice as large and five times larger than in NIDU.

| D ISCUSS I ON
Our results suggest that approximately one in three low-income women at increased risk for HIV infection in the US NHBS sample practised RAI at least once a year, overall. RAI was commonly practised across cities (~1 in 2 to ~1 in 5), including by young women (18-19 years old). Women who practised RAI in the past year did so frequently (overall approximately one RAI/unprotected RAI in 4 sex acts/unprotected sex acts), which means that in the whole sample of women (reporting and not reporting RAI), about 1 in 10 of all sex acts/ unprotected sex acts were RAI/unprotected RAI. Women practising RAI also reported riskier sexual behaviours than non-RAI women: more sexual partners annually and lower condom use with main and casual sexual partners, even though condom use was universally low.
Despite only 11% of all unprotected acts being unprotected RAI, RAI alone may contribute to 41% (80% UI: 18%-55%; city range: 21%-57%) of heterosexually acquired incident HIV infections annually among low-income women in the NHBS sample, due primarily to the high risk of HIV transmission during RAI. The higher sexual risk behaviour of women practising RAI had a negligible impact on the PAF. The PAF RAI was even higher among women exchanging sex (overall: ~60%, city range: 30%-64%) partly because RAI prevalence was the highest and about twice as large among women who exchanged sex (overall and across cities) than among women who did not. Even though younger and Black women tended to report lower RAI prevalence than older or White and Hispanic women, there were no major differences in PAF estimates by age or race/ethnicity given the relatively small differences in RAI prevalence and no difference in the fraction of last sex acts that were RAI across these groups.
RAI prevalence among low-income women in this study is higher overall and across all ages than general populations, 7 but comparable to other at-risk populations 13 in the USA. Our estimates of prevalence among 18-to 19-year-old women are consistent with recent review estimates among youth (20%-24%) 5 and add to growing evidence of a wide (and widening) sexual repertoire among adolescents in the USA 43,44 and elsewhere. 5,45,46 Similar to our findings, other studies among at-risk and general populations also found Black women were less likely than White women to report lifetime 14,47,48 or recent RAI 12 and yet others observe no differences. 13,49 While our results do not exclude RAI F I G U R E 2 A, Model estimates of the contribution of RAI to new annual HIV infections due to RAI alone (PAF RAI ) and due to RAI and riskier behaviours of RAI women (PAF RAI+Beh ) overall and among those who do (ES) and do not exchange sex (No ES). B, PAF RAI in 20 US cities. Boxplots (median, and 10th-90th percentiles) are shown in ascending order of city-specific median past-year RAI prevalence (Pearson's correlation coefficient between RAI prevalence and median PAF across cities). City abbreviations are as in Figure 1 F I G U R E 3 A, Tornado plot showing the correlation between uncertainty in PAF RAI overall model (J = 1) and key model parameters from NHBS-HET analysis. Input parameter ranges for each of the parameters are shown in Table 1. B, C, scatter plots of the most influential parameters (endpoints represent 80% UI): RR RAI (in B), total number of acts among non-RAI women (in C). D, Tornado plot showing correlation of PAF RAI variability with wider ranges of variables taken from the minimum and maximum values across cities as a HIV risk factor for Black women, other factors driving risk such as partner concurrency 49 and the sexual network may also be influential. 50 Our results are consistent with findings from other studies that women who practise RAI tend to report higher numbers of sexual partners, 49

| Strengths and limitations
Our analysis has several strengths and some limitations chiefly due to shortcomings in data. We report for the first time about 4 in 10 new heterosexually acquired HIV infections among a specific population of low-income women at increased risk of HIV infection in the USA may be due to RAI despite a minority of all unprotected sex acts being unprotected RAI. We benefited from detailed high-quality sexual behaviour data from multiple cities, which allowed us to account for parameter uncertainties and for detailed differences in the sexual behaviour of women reporting and not reporting RAI, across different demographic and risk groups and twenty different cities. We were able to draw on a comparable study (HPTN-064) to provide estimates of HIV incidence and the relative risk of RAI to calibrate the model. 37,38 While we cannot exclude the possibility of reporting biases of RAI from FTFI, which can produce lower estimates for sensitive behaviours than more confidential methods, 55 NHBS estimates were similar to those from HPTN-064, which used more confidential methods (RAI prevalence: 38%). Our analysis likely reflects the average behaviour prevailing over the 2010-2013 period rather than a specific year. Prior parameter ranges were derived from confidence intervals of estimates that were unadjusted for RDS design. However, this did not impact our modelling results as our PAF and UI estimates were very similar even after quadrupling standard errors (ie prior parameter ranges) in additional sensitivity analyses (results not shown). The lack of impact is because the main source of uncertainty in PAF estimates was due to uncertainty in estimates of the biological increased risk of RAI compared to RVI. Although the range for this parameter was informed by pooled estimates from systematic literature reviews, it remains uncertain because it is based on few studies. 1,2,33- 35 We did not have data on the number of RVI and RAI acts (used to estimate the total number of sex acts and the fraction of sex acts that are RAI).
Instead, we approximated the fraction of acts that were RAI using the fraction of RAI at the last sexual episode, which could be biased due to over-or under-reporting of certain practices at last sex. Our range for the RAI fraction is slightly higher but overlaps the confidence intervals of an estimate from a 1999 study among STI clinic attendees 13 and is comparable with estimates from general populations. 54  Despite the importance of RAI to heterosexual HIV transmission, key data to estimate its contribution to HIV epidemics such as the frequency of RAI sex acts are missing. The extent that RAI is underreported and imprecisely measured 5,67 in different contexts and subgroups affects the accuracy of our estimates of RAI practices across populations and our estimates of its contribution to HIV epidemics.

ACK N OWLED G M ENTS
Research reported in this publication was supported by the National