Estimating the effect of pre-exposure prophylaxis in Black men who have sex with men

Abstract

Background

Black men who have sex with men (MSM) are disproportionately burdened by the HIV epidemic in the USA. The effectiveness of pre-exposure prophylaxis (PrEP) in preventing HIV infection has been demonstrated through randomized placebo-controlled clinical trials in several populations. Importantly, no such trial has been conducted exclusively among Black MSM in the USA, and it would be unethical and infeasible to do so now.

Methods

To estimate the causal effects of PrEP access, initiation, and adherence on HIV risk, we utilized causal inference methods to combine data from two non-randomized studies that exclusively enrolled Black MSM.

Results

The estimated relative risks of HIV were: (i) 0.52 (95% confidence interval: 0.21, 1.22) for individuals with versus without PrEP access, (ii) 0.48 (0.12, 0.89) for individuals who initiated PrEP but were not adherent versus those who did not initiate, and (iii) 0.23 (0.02, 0.80) for individuals who were adherent to PrEP versus those who did not initiate.

Conclusion

Beyond addressing the knowledge gap around the effect of PrEP in Black MSM in the USA, which may have ramifications for public health, we have provided a framework to combine data from multiple non-randomized studies to estimate causal effects, which has broad utility.

Key Messages.

• Although pre-exposure prophylaxis (PrEP) has been shown to be highly effective in preventing HIV in various populations, randomized trials have had insufficient numbers of Black men who have sex with men (MSM) to estimate the effect of PrEP in this key population in the USA.

• We used two non-randomized trials conducted by the HIV Prevention Trials Network in addition to causal inference methods to estimate the causal effect of PrEP access, initiation, and adherence, and found large reductions in HIV risk, particularly for initiation and adherence.

• These findings are of public health importance as they may motivate Black MSM in the USA to initiate and adhere to PrEP and encourage clinicians to support these men in this endeavor. This work also illustrates use of a framework that can be used broadly to estimate causal effects based on data from multiple non-randomized studies.

Background

Men who have sex with men (MSM) in the USA have been disproportionately affected by the HIV epidemic since it began. MSM in the USA accounted for 71% of all new HIV diagnoses in 2020.¹ Black MSM are particularly at risk, accounting for 26% of all new HIV diagnoses and 39% of diagnoses among MSM in 2020,¹ despite Black people comprising only 13.6% of the US population.² A higher prevalence of sexually transmitted infections (STIs), higher levels of unrecognized HIV in sexual networks and the consequent delay in interventions to reduce the risk of transmission, sex-partner demographics, and a constellation of marginalizing factors, including lower income, underemployment, educational inequalities, inadequate access to healthcare, incarceration, stigma, and discrimination, all likely contribute to an elevated risk of HIV among Black MSM.^3–7

The high risk and disproportionate burden of HIV among Black MSM have persisted even after pre-exposure prophylaxis (PrEP) was approved by the Food and Drug Administration in 2012. PrEP is highly effective in preventing HIV infection in various populations, including MSM.⁸ However, although previous trials have included Black MSM in the USA, they have not been sufficiently represented in randomized, placebo-controlled trials to date. Given extensive evidence that PrEP is broadly effective, a randomized, placebo-controlled trial among Black MSM is unethical. Thus, a knowledge gap regarding the effectiveness of PrEP specifically within the unique context of this vulnerable population persists. Beyond providing insights into the generalizability of effectiveness estimates from other populations, a better understanding of the effectiveness of PrEP in Black MSM could motivate initiation of and adherence to PrEP in this population and encourage clinicians to support such behavior, reducing HIV incidence.

We estimated the effect of PrEP on HIV risk in Black MSM in the USA by combining data from two large studies that were conducted exclusively in Black MSM by the HIV Prevention Trials Network (HPTN), including a single-arm study and an observational study. We sought to quantify the effect of PrEP access, PrEP initiation without adherence, and PrEP adherence. To estimate causal effects from non-randomized data while leveraging data from a single-arm study, we utilized inverse probability weighting (IPW).^9,10 Beyond estimating the effectiveness of PrEP in reducing the risk of HIV in Black MSM in the USA, our work illustrates a general framework to estimate causal effects using data from multiple, non-randomized studies.

Methods

Data sources

HPTN 061 (n = 1134) was an observational study of Black MSM in the USA that was designed to evaluate the feasibility and acceptability of an HIV prevention intervention.¹¹ The study was conducted between 2009 and 2011; thus, we assumed functionally no access to PrEP for study participants. HPTN 073 (n = 226) was a non-randomized, open-label PrEP study of Black MSM in the USA¹² that was conducted between 2013 and 2015, i.e. immediately after FDA approval of PrEP, but close in time to HPTN 061. Participants in HPTN 073 had access to PrEP (daily oral co-formulated emtricitabine and tenofovir disoproxil fumarate) through the study and could decide whether to initiate and/or adhere to PrEP at any point. PrEP access in HPTN 073 was of an ‘active’ form: beyond PrEP being available, PrEP was offered at each study visit and staff were available to assist participants who were interested in initiating. HPTN 073 participants had access to the C4 program—a light-touch support program¹² that is unlikely to have a direct effect on HIV risk. Blood samples were taken to measure PrEP adherence (at least four doses per week)¹³ among those who initiated. Over 1 year of follow-up, measurements (including those of relevant covariates and HIV status) were obtained at the enrollment and at 6-month and 12-month visits in both HPTN 073 and HPTN 061; retention was high ($\ge$75%), particularly in HPTN 073. We restricted attention to covariates that were measured in both studies. See Supplementary Section 1 for additional information on the design of HPTN 061 and HPTN 073, and the harmonization of data from the two studies.

Our “target population,” i.e. that in which we would like to make inference, is Black MSM in the USA.¹⁴ Our “study population,” in which inference will be made, is the HPTN 073 population, i.e. the population from which the HPTN 073 study participants were (randomly) sampled. We chose the HPTN 073 population as the HPTN 073 study is more recent. Our analysis was based on the HPTN 073 sample and the HPTN 061 sample, which are representative of their respective populations. See Supplementary Section 1 for additional details.

Analysis

We sought to estimate the effects of PrEP access, initiation without adherence (‘initiation’ hereafter), and adherence on risk of HIV in Black MSM in the USA. Let $A$ denote PrEP access: $A=1$ indicates PrEP access and $A=0$ indicates no PrEP access. Consequently, $A$ is equivalent to a study indicator as $A=1$ for participants in HPTN 073 and $A=0$ for participants in HPTN 061. The variable $D$ denotes PrEP initiation and adherence: $D=0$ indicates initiation, $D=1$ indicates initiation without adherence, and $D=2$ indicates adherence. In HPTN 061, $D=0$. In HPTN 073, $D\in \{0, 1, 2\}$. $Y$ denotes HIV status ($Y=1$ if the participant has HIV and $Y=0$ otherwise) and $\mathit{X}$ is a vector of baseline and time-varying covariates.

The baseline, 6-month, and 12-month time points are denoted by $t=0, 1, \mathrm{and} 2$, respectively. We conceptualize the time between baseline and 6 months as the ‘first interval’ and the time between 6 and 12 months as the ‘second interval’. Access ($A$) is time-invariant. We denote the vector of covariate values at time $t$ (including both time-varying and time-invariant covariates) as ${\mathit{X}}_t$. Likewise, $D_t$ denotes PrEP initiation/adherence assessed at time $t, t=1, 2$; e.g. $D_t=1$ if the participant initiated PrEP prior to time $t$ but was not adherent at time $t$. Finally, $Y_t$ indicates HIV status at time $t$. The subscript $i$ refers to individual $i$, $i=1,\dots , n$, where $n=n_{061}+n_{073}$, the total number of participants in the HPTN 061 and HPTN 073 samples.

A directed acyclic graph that characterizes the relationships between the variables is given in Figure 1. Dashed lines between $A$ and $D_1$ and $D_2$ indicate a (partially) deterministic relationship: if $A=0$, then $D_1=D_2=0$. Likewise, dashed lines connect $Y_1$ to $D_2$ and $Y_2$: if $Y_1=1$, then $Y_2=1$ and $D_2$ is not relevant (a person with HIV would not keep taking PrEP). We suppose that the effect PrEP access on HIV risk is through PrEP initiation and adherence. It has been suggested that PrEP use influences the probability of engaging in high-risk behaviors (leading to an arrow from $D_1$ to ${\mathit{X}}_1$ if ${\mathit{X}}_1$ includes some of these behaviors); such influence, however, is likely to be minimal.^15,16

Figure 1.

Directed acyclic graph representing the relationships among the exposures ($A, D_1, D_2$), the outcomes ($Y_1, Y_2$), and the covariates (${\mathit{X}}_0, {\mathit{X}}_1$), where dashed lines indicate deterministic relationships.

$A$ analysis

We quantify the effect of access ($A$) on risk of HIV as:

$$\frac{E\left[Y_2\right(a=1\left)\right|A=1]}{E\left[Y_2\right(a=0\left)\right|A=1]},\mathrm{ }\left(1\right)$$

where $Y_2\left(a\right)$ is the potential outcome of $Y_2$ under exposure $A=a$, i.e. the outcome that would have been observed had $A=a.$¹⁷ This is the causal relative risk (RR) of HIV infection in 1 year ($Y_2$) comparing PrEP access versus no access in the HPTN 073 population, i.e. conditioned upon $A=1$.¹⁸ Thus, we compare the risk of HIV under access to PrEP (regardless of initiation or adherence) to the risk of HIV with no such access.

IPW, which uses weights to balance the distribution of covariates between HPTN 061 and HPTN 073, can be used to estimate Equation (1). IPW seeks to mimic a randomized trial in which the only (systematic) difference between the exposure groups is the exposure itself,¹⁹ though IPW can at most provide balance on the measured covariates only.²⁰ Our goal in weighting was to make the distribution of covariates in the HPTN 061 sample similar to that in the HPTN 073 sample as the HPTN 073 population was our target of inference. In other words, the weights yield a sample from a ‘pseudopopulation’ in which the distribution of the covariates in both the exposed ($A=1$) and the unexposed ($A=0$) reflects the distribution in the HPTN 073 population, avoiding confounding by these covariates.^10,21

We can estimate Equation (1) by calculating weights^18,22 as follows:

$$w_i^A=\left\{\begin{array}{cc}\frac{\mathrm{Pr}(A=1|{\mathit{X}}_0={\mathit{x}}_{i0})}{\mathrm{Pr}(A=0|{\mathit{X}}_0={\mathit{x}}_{i0})}\mathrm{ }& \mathrm{for}\mathrm{ }\mathrm{HPTN}\mathrm{ }061\\ \mathrm{ }1\mathrm{ }& \mathrm{for}\mathrm{ }\mathrm{HPTN}\mathrm{ }073\end{array}\right.,$$

$\mathrm{where}\mathrm{ }i=1,\dots , n$ and using these in a weighted log-linear model of the form:

$$\log \left[\mathrm{Pr}\left(Y_2=1|A\right)\right]={\beta }_0+{\beta }_{1}A,\mathrm{ }\left(2\right)$$

where $e^{{\beta }_1}$ is the causal RR (1). We must assume conditional exchangeability, i.e. no unmeasured confounding, and positivity, i.e. a positive probability of $A=0$.^23,24 A full description of all assumptions is provided in Supplementary Section 2 in addition to the derivation of $w_i^A$ (Supplementary Section 7). To obtain weights that make the HPTN 061 sample similar to the HPTN 073 sample in terms of the covariate distribution, we used the covariate-balancing propensity score (CBPS), which estimates $\mathrm{Pr}(A=1|{\mathit{X}}_0)$ by modeling treatment assignment while optimizing (mean) covariate balance.^25,26 We used the CBPS package in R.^27,28

$D$ analysis

The causal effect of PrEP initiation and adherence ($D$) on HIV risk can be written as:

$$\frac{E\left[Y_t\right(d_t\left)\right|A=1,Y_{t-1}=0]}{E\left[Y_t\left(d_t=0\right)\right|A=1,Y_{t-1}=0]},\mathrm{ }\left(3\right)$$

where $t=1, 2 \mathrm{and} d_t=1, 2$; this corresponds to the causal RR of HIV infection during the next 6 months for $d_t\in \{1, 2\}$ vs. $d_t=0$.²⁹ The potential outcome $Y_t\left(d_t\right)$ is the outcome (HIV infection by visit $t$) that an individual would have experienced had they received exposure $d_t$ during the prior time interval. We conditioned on $Y_{t-1}=0$ as individuals who have already tested positive for HIV are no longer at risk. Conditioning on $A=1$ makes the HPTN 073 population our target of inference.

To use data from HPTN 073 (a non-randomized study) and HPTN 061 (a single-arm study in which PrEP was not available), two sets of weights were required. The first made the distribution of the covariates in the HPTN 061 population similar to that in the HPTN 073 population:

$$w_{it}^{D,1}=\left\{\begin{array}{cc}\frac{\mathrm{Pr}(A=1|{\mathit{X}}_0={\mathit{x}}_{i0},Y_{t-1}=0)}{\mathrm{Pr}(A=0|{\mathit{X}}_0={\mathit{x}}_{i0},Y_{t-1}=0)}\mathrm{ }& \mathrm{for}\mathrm{ }\mathrm{HPTN}\mathrm{ }061\\ \mathrm{ }1\mathrm{ }& \mathrm{for}\mathrm{ }\mathrm{HPTN}\mathrm{ }073\end{array},\right.$$

$\mathrm{where}\mathrm{ }t=1, 2 \mathrm{and} i=1,\dots , n$ . The second provided balance in the distribution of measured covariates for $D_t=$ 0, 1, and 2 in the HPTN 073 sample is:

$$w_{it}^{D,2}=\left\{\begin{array}{cc}\mathrm{ }1\mathrm{ }& \mathrm{for}\mathrm{ }\mathrm{HPTN}\mathrm{ }061\\ \frac{1}{\mathrm{Pr}\left(D_t=d_{it}|A=1,\mathrm{ }{\mathit{X}}_{t-1}={\mathit{x}}_{i,t-1},Y_{t-1}=0\right)}\mathrm{ }& \mathrm{for}\mathrm{ }\mathrm{HPTN}\mathrm{ }073,\end{array}\right.$$

where $i=1,\dots , n, t=1, 2, \mathrm{and} d_{it}\in \{0, 1, 2\}$. The final weights were $w_{it}^D=w_{it}^{D,1}\times w_{it}^{D,2}$. We can then use $w_{it}^D$ in a weighted log-linear model³⁰ among individuals who have not yet tested positive for HIV, similar to a pooled logistic regression model,^21,31 as follows:

$$\log \left[\mathrm{Pr}\left(Y_t=1|D_t,\mathrm{ }{Y}_{t-1}=0\right)\right]={\beta }_{0t}+{\beta }_1{D}_t,\mathrm{ }\left(4\right)$$

where $e^{{\beta }_1{d}_t}, d_t\in \{1, 2\}$ is the causal RR (3). We must assume conditional exchangeability for both $A$ and $D$, i.e. no unmeasured confounding, and positivity, i.e. a positive probability of $A=0$ and $D\in \left\{\mathrm{0,1},2\right\}$.^23,32 A full description of all assumptions is given in Supplementary Section 2, in addition to the derivation and estimation of $w_{it}^D$ (Supplementary Section 7).

Covariate selection, missing data, and quantifying uncertainty

For inverse probability weights, covariates related to both the outcome and the exposure should be included in the weight model, as these are likely confounders.³³ Although the inclusion of covariates related only to the outcome can increase efficiency,³⁴ the inclusion of too many covariates risks positivity violations.³⁵ Thus, we focused on covariates that are thought to influence HIV risk,^36,37 of which the confounders will be a subset. Broadly, HIV risk among Black MSM in the USA is influenced by an individual’s sexual network and the occurrence of condomless sexual encounters, and may be augmented by cofactors such as concomitant STIs.^11,38 To capture these drivers of risk, we included region (categorical), age (continuous), number of partners (0–1, 2–3, 4 or more; ordinal), sexual orientation (gay vs. other), condomless receptive anal intercourse (URAI; binary), and the presence of an STI (binary), each of which captures a dimension of the features described. The number of partners, URAI, and STI were time-varying covariates.

Multiple imputation was generally used to address missing data and bootstrap 95% confidence intervals (CIs) were estimated to obtain intervals that reflected variability in both the estimated weights and the estimated effects.^39–41 See Supplementary Section 3 for additional details on missing data and interval estimation.

Results

Effect of PrEP initiation and adherence

More than three-quarters of HPTN 073 participants initiated PrEP by $t=1$ (Table 1). Nearly 44% of all participants (57.2% of PrEP initiators) were adherent to PrEP at $t=1$; this decreased to 36.3% of all participants (46.1% of PrEP initiators) at $t=2$. In HPTN 061, 18 individuals tested positive for HIV by $t=1$; by $t=2$, a total of 28 individuals had tested positive (i.e. there were 10 new seroconversions between the 6- and 12-month visits). In HPTN 073, five individuals tested positive by $t=1$ and a total of eight tested positive by $t=2$. Missing HIV status was more common in HPTN 061, particularly at $t=2$, and there were differences in covariate distributions and the extent of missingness. Consistently with best practices, HIV status was included in our imputation model and imputed where missing.^42,43 For both studies, covariate missingness was substantially higher at $t=1$ vs. $t=0$.

Table 1.

Cohort summaries for the HIV Prevention Trials Network 061 study sample and the HIV Prevention Trials Network 073 study sample

	HPTN 061 (n = 1134)			HPTN 073 (n = 226)
	$\mathit{t}=0$	$\mathit{t}=1$	$\mathit{t}=2$ a	$\mathit{t}=0$	$\mathit{t}=1$	$\mathit{t}=2$ a
Initiated PrEP, n (%)b	–	0	0	–	173 (76.5%)	178 (78.8%)
Missing		0	0		0	0
Adherent to PrEP, n (%)	–	0	0	–	99 (43.8%)	82 (36.3%)
Missing		0	0		4 (1.8%)	4 (1.8%)
Covariates
Age (years), mean (SD)	37.0 (12.2)	–	–	29.4 (9.9)	–	–
Missing	0			0
Region, n (%)
Northeast	565 (49.8%)	–	–	75 (33.2%)	–	—
South	217 (19.1%)			75 (33.2%)
West	352 (31.0%)			76 (33.6%)
Missing	0			0
Gay, n (%)	356 (31.4%)	–	–	158 (69.9%)	–	–
Missing	2 (0.2%)			0
Number of partners, n (%)
0–1	192 (16.9%)	353 (31.1%)	–	75 (33.2%)	88 (38.9%)	—
2–3	407 (35.9%)	297 (26.2%)		74 (32.7%)	51 (22.6%)
≥4	532 (46.9%)	270 (23.8%)		73 (32.3%)	46 (20.4%)
Missing	3 (0.3%)	214 (18.9%)		4 (1.8%)	41 (18.1%)
URAI, n (%)	459 (40.5%)	230 (20.3%)	–	95 (42.0%)	95 (42.0%)	–
Missing	35 (3.1%)	230 (20.3%)		5 (2.2%)	38 (16.8%)
STI, n (%)	126 (11.1%)	30 (2.6%)	–	32 (14.2%)	30 (13.3%)	–
Missing	80 (7.1%)	244 (21.5%)		8 (3.5%)	30 (13.3%)
Outcome: HIV, n (%)	–	18 (1.6%)	28 (2.5%)	–	5 (2.2%)	8 (3.5%)
Missing		150 (13.2%)	263 (23.2%)		14 (6.2%)	21 (9.3%)

HPTN, HIV Prevention Trials Network; PrEP, pre-exposure prophylaxis; STI, sexually transmitted infection; URAI, unprotected (condomless) receptive anal intercourse.

^aPeople with $Y_1=1$ were not removed from the $\mathit{t}=2$ column.

^bPrEP initiations: $n$ = number of initiations prior to $\mathit{t}$.

We estimated a 48% reduction in the risk of HIV in 1 year due to PrEP access (RR = 0.520, 95% CI: 0.213, 1.217) (Table 2). PrEP initiation during a 6-month interval resulted in an estimated 52% reduction in the risk of HIV at the end of the interval compared with individuals who did not initiate PrEP (95% CI for the RR: 0.122, 0.893) and PrEP adherence during a 6-month interval resulted in an estimated 77% reduction in the risk of HIV at the end of the interval compared with individuals who did not initiate PrEP (95% CI for the RR: 0.015, 0.798). Thus, we observed substantial reductions in the estimated risk of HIV due to PrEP access, initiation, and adherence, though the CI for PrEP access was wide. Importantly, these estimates pertain to the HPTN 073 population.

Table 2.

Effect estimates for pre-exposure prophylaxis access, initiation, and adherence

Comparison	Outcome	Estimated RR (95% CI)
PrEP access: $A=1$vs. $A=0$	Risk of HIV in 1 year ($Y_2$)	0.520 (0.213, 1.217)
PrEP initiation without adherence: $D=1$vs. $D=0$	Risk of HIV in 6 months ($Y_t)$	0.477 (0.122, 0.893)
PrEP adherence: $D=2$vs. $D=0$	Risk of HIV in 6 months ($Y_t$)	0.227 (0.015, 0.798)

PrEP, pre-exposure prophylaxis; CI, confidence interval; RR, relative risk.

Diagnostics and sensitivity analyses

We used established diagnostics to evaluate the balance of measured covariates and the presence of extreme probabilities (i.e. positivity violations or near-violations); full details and results are presented in Supplementary Section 4. (See online supplementary material for a color version of this section.) Briefly, good covariate balance was achieved (Supplementary Figure S1) and, although some extreme probabilities were found, these were generally in the bootstrap samples and so would primarily influence CI width (Supplementary Figure S2 and Table S4). (See online supplementary material for color versions of these figures and table.) Finally, we evaluated the impact of unmeasured confounders^9,30,44,45 and found that the causal effects of PrEP on HIV risk were attenuated in the presence of an unmeasured confounder with a strong negative association with HIV risk (Supplementary Figure S3). (See online supplementary material for a color version of this figure.)

Conclusions

We estimated the causal effects of PrEP access, initiation, and adherence in Black MSM in the USA based on a careful analysis by combining data from two non-randomized studies: HPTN 061 and HPTN 073. Using a weighting approach, we found that, in the HPTN 073 population, access to PrEP reduced the risk of HIV by an estimated 48% and, compared with Black MSM who did not initiate PrEP, initiation reduced HIV risk by an estimated 52% and adherence reduced HIV risk by an estimated 77%, though the CI for PrEP access was wide. Such large reductions in risk of HIV are not surprising, given the large protective effects of PrEP observed in other populations.^8,46,47 Our results are of considerable public health importance as they may motivate Black MSM—a population that bears a disproportionate burden of the HIV epidemic in the USA—to initiate and adhere to PrEP, reducing their risk of HIV. Although our analysis was well motivated and made judicious use of existing data, alternative approaches exist and are discussed in Supplementary Section 5.

Our analysis also illustrates a framework that could be broadly used to combine data from multiple non-randomized studies to estimate causal effects of various exposures. Such approaches are particularly valuable in the context of large networks such as HPTN and offer the opportunity to utilize data from multiple studies to answer causal questions without a randomized trial. When data sets like these exist, there is an ethical obligation to utilize the data as fully as possible, using robust and rigorous methods. Such stewardship is particularly necessary when the data come from minoritized communities, as these communities may be overburdened by requests to participate in research or wary of participation altogether due to past encounters with the medical community.

Strengths

Our analysis involved data from two carefully conducted studies, one of which was fairly large. Since both studies were conducted by HPTN and temporally close, data collection and recruitment and retention strategies were rigorous and similar between the two studies, making their combination highly compelling. Although some challenges related to combining the data from HPTN 061 and HPTN 073 persist (as discussed below), this analysis would not have been possible by using only one of these data sets given the size of HPTN 073 and the lack of PrEP in HPTN 061. Importantly, individuals who enroll in observational studies such as HPTN 061 and HPTN 073 may be more representative of the general population than those who would enroll in a randomized trial, enhancing generalizability.

We utilized a rigorous approach to effect estimation, quantification of uncertainty, and evaluation of key assumptions, providing a framework that can be used more generally to estimate causal effects based on data from multiple non-randomized studies. Even when a randomized trial is ethical and feasible, if data from non-randomized studies exist, then it may be advisable to employ a framework like ours to either motivate a randomized trial or, if the data are from well-conducted studies, obviate the need for a trial altogether, preserving financial resources, reducing burden, and allowing the expedient estimation of effects.

Limitations

The HPTN 073 sample may not be representative of the HPTN 073 population^14,48 and the HPTN 073 population may differ from our target population: Black MSM in the USA.^14,49 Generalizing our results to the population of Black MSM in the USA requires careful consideration of selection bias⁵⁰ and, relatedly, sampling schemes, enrollment processes, and eligibility criteria. Loss of generalizability will largely only be a concern insofar as there is effect modification and/or unmeasured confounding.⁵¹ Residual imbalances between HPTN 061 and HPTN 073 owing to imperfect overlap in variables measured and differences between the two studies in the distribution of effect modifiers that were not included in our weight models could lead to bias and loss of generalizability.¹⁴ The potential impact of imbalances between HPTN 061 and HPTN 073 can be quantified by considering that, among Black men in the USA, HIV incidence decreased by 10.8% between 2010 and 2014—the midpoints of HPTN 061 and HPTN 073, respectively.⁵² Much of this decrease is likely accounted for by the variables included in our weight models but differences in unmeasured factors, such as the rise in treatment as prevention during this period, may have played a role. If we suppose that the entirety of the 10.8% decrease in HIV incidence is due to unmeasured confounders and that the denominator of the estimand for the effect of $D$(3) is driven entirely by HPTN 061 (both extreme, unlikely scenarios), then our results would be attenuated but not qualitatively changed: the estimated RRs and 95% CIs would be 0.583 (0.239, 1.364) for PrEP access, 0.535 (0.137, 1.001) for PrEP initiation without adherence, and 0.254 (0.017, 0.895) for PrEP adherence.

Differences in the frequency of data collection between the two studies and use of self-report for several covariates may have resulted in measurement error (see Supplementary Section 1 for a discussion).^11,53 Additionally, the reliance on assessments of PrEP initiation, PrEP adherence, and HIV status every 6 months may have led to measurement error and some degree of reverse causation if individuals obtained off-study HIV tests. Future analyses using data with more frequent assessments would be informative.

Although the lack of cases of HIV among individuals with $D=2$ required modeling $D$ as an ordinal variable, a more flexible modeling approach would have been preferable. As most previous studies have found very large effects for PrEP adherence,^54,55 the constraints imposed by our model may have to some degree overestimated the effect of initiation and underestimated the effect of adherence (see Supplementary Section 5).

We estimated the causal effects of PrEP access, initiation, and adherence in a key population in the US HIV epidemic—Black MSM—finding large reductions in risk, particularly for initiation and adherence. More broadly, we illustrated a rigorous analytic framework for combining data from multiple studies to estimate causal effects. This approach could be applied generally to estimate causal effects using several data sources (see Supplementary Section 6), particularly when a randomized trial is unethical or infeasible.

Ethics approval

The studies (HPTN 061 and HPTN 073) were reviewed and approved by the Institutional Review Boards of each study site and all participants provided written informed consent, including for the data analysis presented in this work.

Author contributions

KM and DW led HPTN 061 and HPTN 073, respectively. AM designed and conducted the analysis. SZ and DD provided expertise on the statistical analysis, in particular the use of data from HPTN 061 and HPTN 073. FX and KCGC provided causal inference expertise. AM drafted the manuscript and all authors reviewed the manuscript.

Supplementary data

Supplementary data is available at IJE online.

Conflict of interest

None declared.

Funding

We acknowledge funding from the National Institutes of Health (NIH) under NIH U01 HL146242 and NIH P30 MH123248. Overall support for the HPTN is provided by the NIH under Award Numbers UM1AI068619 (HPTN Leadership and Operations Center), UM1AI068617 (HPTN Statistical and Data Management Center), and UM1AI068613 (HPTN Laboratory Center). Additional support for this study was provided by the NIH under UM1AI069412.

Data availability

Data and code to run this analysis are available at: https://github.com/meisnera/BMSM_PrEP.