Skip to main content
Download PDF

How likely is unmeasured confounding to explain meta-analysis-derived associations between alcohol, other substances, and mood-related conditions with HIV risk behaviors?

BMC Medical Research Methodology volume 25, Article number: 62 (2025) Cite this article

Abstract

Background

HIV transmission and disease progression may be driven by associations HIV risk behaviors have with a constellation of alcohol, other substance, and mood-related conditions (CASM). However, observational study-based measures of these associations are often prone to unmeasured confounding. While meta-analysis offers a systematic approach to summarize effect sizes across studies, the validity of these estimates can be compromised if similar biases exist across studies. Our analysis assesses the likelihood that unmeasured confounding explains meta-analysis-derived measures of association between CASM and HIV risk behaviors, and provides bias-adjusted estimates.

Methods

We first conducted systematic reviews and meta-analyses to assess associations between CASM conditions and four HIV risk behaviors (medication non-adherence, unprotected sex, transactional sex, and multiple sexual partners). We then adjusted for potential unmeasured confounders using two methods designed for meta-analyses - Point Estimate and Proportion of Meaningfully Strong Effects methods. We selected “risk propensity” as an illustrative and potentially important unmeasured confounder based on the extant literature and mechanistic plausibility.

Results

In analyses unadjusted for unmeasured confounding, 89% (24/27) of odds ratios (ORs) show strong evidence of a positive association, with alcohol use and stimulant use emerging as dominant risk factors for HIV risk behaviors. After adjusting for unmeasured confounding by risk propensity, 81% (22/27) of ORs still showed strong evidence of a positive association. Associations between mood-related conditions and HIV risk behaviors were more robust to unmeasured confounding than associations between alcohol use and other substance use and HIV risk behaviors.

Conclusion

Despite residual confounding present in constituent studies, there remains strong evidence of associations between CASM and HIV risk behaviors as well as the clustered nature of CASM conditions. Our analysis provides an example of how to assess unmeasured confounding in meta-analysis-derived measures of association.

Peer Review reports

Background

Human immunodeficiency virus (HIV) remains a globally persistent public health issue, affecting an estimated 39 million people living with HIV (PLHIV) at the end of 2022, and leading to 630,000 deaths in 2022 from HIV-related causes [1]. While progress has been made to reach the 95-95-95 goals, shortcomings remain, as there were an estimated 1.3 million additional cases in 2022, 660,000 of which were reported in the African region [1]. A constellation of alcohol (e.g., alcohol use disorder), other substance (e.g., tobacco, opioid, and stimulant use disorders), and mood-related (e.g. depressive and generalized anxiety disorders and chronic pain) conditions (CASM), have a high co-prevalence with HIV and appear to be important facilitators for HIV transmission and disease progression. Alcohol users in sub-Saharan Africa (SSA) were found to have 1.61 (95% CI 1.44–1.80) times the odds of having HIV compared to non-alcohol users [2], and PLHIV in high-income countries have an estimated 1.2–2.4-fold greater prevalence of alcohol use disorder, major depression, generalized anxiety, and drug use disorder than that among non-PLHIV [3]. An estimated 11% and 26% of PLHIV and male PLHIV respectively in SSA smoke cigarettes [4], and 22% of PLHIV in Nigeria have anxiety disorders [5]. An estimated 31% (95% CI 26–38%) of PLHIV on antiretroviral therapy (ART) in SSA have significant depressive symptoms [6], and 18–81% of various PLHIV populations have major depression [7]. An estimated 25–90% [8], 50–70% [9], 16.8% [10], and 17–64% of PLHIV in the United States have chronic pain, are current smokers, have substance use disorders, and used an illicit drug in their lifetime, respectively. These comorbidities remain present across age, sex, and gender differences [3, 7, 11,12,13,14], and studies have suggested biological pathways that CASM contributes to HIV disease and progression [15,16,17,18,19].

While associations between CASM and HIV risk behaviors have been studied in depth, including medication non-adherence [20,21,22,23,24,25,26,27,28,29,30,31,32] and risky sexual behaviors [33,34,35,36,37,38,39,40], observational study-based measures are inherently prone to unmeasured confounding. That is, the influence of factors not included in an analysis that may be the true drivers of results observed. Observational study results risk such scenarios since only a finite group of infinite potential risk factors are hypothesized and measured. Furthermore, composite exposure measures (e.g. substance use) prevent single exposure association estimates, and variation in sensitivity and specificity of CASM screening and diagnosis tools and HIV risk behavior measurement scales add to the ambiguity of conclusions drawn by these observational studies. Despite these shortcomings, observational studies remain essential ways to understand public health phenomena, for ethical and practical reasons. While meta-analysis offers a systematic approach to summarize effect sizes across studies by aggregating data, the validity of these estimates can be compromised by bias from residual unmeasured confounding in observational constituent studies. Furthermore, confirmation bias has led to systematic inclusion of more common potential confounders while excluding others from routine consideration [41].

Our objective was to assess the likelihood that unmeasured confounding explains meta-analysis-derived measures of association between CASM and HIV risk behaviors, and to provide corresponding bias-adjusted estimates. Understanding the extent to which presumed causal associations between CASM and HIV risk behaviors are confounded can help identify integrative preventive strategies that complement the traditional HIV care cascade by addressing CASM (Fig. 1). We hypothesized that risk propensity (e.g., an increased preference to act in ways having elevated potential for negative consequences [42,43,44,45]) is an important unmeasured confounder between CASM conditions and HIV risk behaviors and selected it as our primary potential confounder of interest. Risk propensity has been associated with CASM [42, 43] and HIV risk behaviors [44, 45] and was not adjusted for in any constituent studies across our unadjusted analyses. We also hypothesized that the strength of unmeasured confounding by risk propensity would serve as a useful benchmark for the strength of other potential unmeasured confounders. We hypothesized that distrust of medical institutions (DIM) may also be an important confounder and selected it as a secondary potential confounder of interest, as it has been associated with both our exposures and outcomes of interest [57,58,59,60].

Fig. 1
figure 1

Conceptual diagram of CASM and HIV risk behavior

Full size image

Methods

First, we describe our analyses unadjusted for potential unmeasured confounders, conducted to derive summary estimates for associations between CASM conditions and HIV risk behaviors. Second, we describe how we analyzed these summary estimates for potential unmeasured confounding [46, 47].

Analyses unadjusted for potential unmeasured confounding: systematic reviews and meta-analyses

We carried out 28 systematic reviews and performed corresponding meta-analyses to estimate the association between seven CASM (alcohol, depression, anxiety, pain, tobacco, opioids, stimulants) and four HIV risk behaviors (medication non-adherence, unprotected sex, transactional sex, and multiple sexual partners) (Supplementary Box 1; Figures S1-S26; Supplementary Box 2; Figures S27-S52). Our recently published meta-analysis characterizing the association between stimulants and HIV risk behaviors describes the methodology in detail and is reported elsewhere [48]. Briefly, we searched the PubMed database between July 2019-January 2021, filtering for English-language and human studies with no restrictions on setting, population, study design, or year. A single reviewer viewed abstracts from all search results and reviewed in full studies meeting inclusion criteria. Inclusion criteria required (1) the exposure of interest with a comparison group, (2) the outcome of interest, and (3) an odds ratio (OR) with corresponding 95% CI or other data from which ORs could be ascertained. To prevent overrepresentation of any study, one measure of association was used per constituent study. We used adjusted estimates when available. If multiple population groups or variations of exposure or outcome were presented, a pooled OR and 95% CI estimate was calculated [49]. If pooling was not possible, a series of decision rules was used: (1) we focused on the subgroup most commonly reported across all other constituent studies, the timeframe with the highest likelihood of exposure preceding the outcome, and the highest-level exposure and/or outcome stratum; (2) we prioritized an unadjusted OR including all subgroups over any particular subgroup-adjusted OR; (3) we harmonized directionalities (e.g., protected sex rather than unprotected sex) [49]; and (4) we excluded studies with composite exposure or outcome measures that might misclassify the exposure or dilute its effect.

Medication non-adherence refers to non-adherence to medication for HIV and/or hepatitis C virus (HCV), and no restrictions were placed on reporting method (e.g., self-reported adherence, visual analog scales, MEMs caps). Unprotected sex refers to condomless sex and includes all relevant measures (e.g., inconsistent condom use, condomless sex within specified recall times). Transactional sex refers to measures of sexual activity in which money or gifts were exchanged for sexual acts; we aggregated data on “buyers” and “sellers.” Multiple sexual partners refers to having multiple concurrent or subsequent sexual partners, and we considered different magnitudes and recall times. Exposure definitions were used as per individual study.

Our 28 systematic reviews yielded 27 measures of association (there were no eligible studies associating pain and transactional sex), 26 of which included ≥ 2 studies and therefore were amenable to meta-analyses (only one eligible study associated pain and multiple sexual partners). Random effects meta-analyses were conducted in RStudio, Version 1.3.1093, and Stata/IC, Version 15.1. We assessed heterogeneity in each meta-analysis by I-squared, tau-squared (τ²), and the Q-statistic.

Analyses adjusted for potential unmeasured confounding

To assess how robust our meta-analysis-derived measures of association were to unmeasured confounding, we used the Point Estimate method and the Proportion of Meaningfully Strong Effects method (PMSE) [46, 47, 50,51,52,53,54]. The Point Estimate method is most suitable for single studies or for meta-analyses with low or no heterogeneity (e.g., τ² = 0), whereas PMSE is most suitable for meta-analyses with greater heterogeneity, albeit with a requirement of ≥ 10 studies. Neither approach requires specifying the prevalence of the unmeasured confounder. Both approaches allow interactions between the exposure (X), unmeasured confounder (U), and outcome (Y) [46, 47, 52, 54], and require the assumption that unmeasured confounders of interest are not completely collinear with any variables controlled for in constituent studies. To employ these methods, ORs were converted to risk ratios (RR) [54] using the square root transformation [55]. Figure 2 illustrates the interdependence of methods and parameters, and Table 1 lists key inputs used.

Fig. 2
figure 2

Parameter map for sensitivity analyses accounting for potential unmeasured confounding. Literature-derived input parameters: RRCXY: unadjusted meta-analysis-derived risk ratio (RR), presumed to have unmeasured confounding; RRTXY: true RR estimate, corrected for presumed unmeasured confounding; RRXU: RR between the exposure and U; RRUY: RR between U and the outcome; τ²: heterogeneity between studies. User-specified input parameters: q: minimum threshold (RR scale) of a meaningfully strong effect; r: the minimum proportion of constituent studies with true effects above q required to indicate moderate to strong evidence of an effect; p: proportion of heterogeneity (τ²) due to variation in confounding bias. Output parameters: E-value: Minimum confounding strength (risk ratio (RR) scale) by an unmeasured confounder that could explain RRCXY); Bias factor: maximum amount the unmeasured confounder could impact RRCXY, derived from RRXU and RRUY, applied to RRCXY to calculate RRTXY; p̂(q): the proportion of studies that have scientifically meaningful strong effects (RR > q); Ĝ(r, q): the minimum confounding strength required to lower p̂ to < r; \({\rm{\hat T}}\)(r, q): the minimum bias factor required to lower p̂ to < r

Full size image
Table 1 Input parameters used in sensitivity analyses
Full size table
Table 2 Assessment of unmeasured confounding among meta-analyses assessed by Point Estimate sensitivity analysis only
Full size table
Table 3 Assessment of unmeasured confounding among meta-analyses assessed by Point Estimate and PMSE sensitivity analyses
Full size table

Point estimate method

The Point Estimate method entails calculating an E-value and an E-value 95% CI lower limit (LL) for associations of interest. Meta-analyses yield “g-values” rather than “E-values,” which pertain to single studies. Since they are mathematically equivalent to E-values, both terms are grouped together hereon. E-values quantify how strongly an unmeasured confounder would need to be associated with the exposure (minimum risk ratio RRXU) and outcome (minimum risk ratio RRUY) to fully explain the observed association (RRXY) [56], conditional on the measured covariates [54]. The E-value 95% CI lower limit (LL) quantifies the strength of the unmeasured confounder necessary for the 95% CI of the observed and likely confounded association to contain the null value. An E-value greater than both RRXU and RRUY indicates that unmeasured confounding cannot fully explain RRXY [54]. First, the E-value and 95% CI LL were calculated using the formula:

E-value =

$$\text{R}\text{R}_\text{XY}\:+\:\text{S}\text{Q}\text{R}\text{T}\left(\text{R}\text{R}_\text{XY}\:\times\:\right(\text{R}\text{R}_\text{XY}\:-\hspace{0.17em}1\left)\right)$$
(1)

We used the inverse of the RR when the point estimate or 95% CI LL was below 1 to yield defined values, and 95% CI LLs were set to 1 if the 95% CI LL RR values were less than 1 [54], since an uncertainty range including the null already indicates sufficient association with the unmeasured confounder to explain RRXY.

We performed a sensitivity analysis to assess the relative magnitude of our E-values by using distrust in medical institutions (DIM) as an alternative unmeasured confounder [56]. To gauge if an E-value is “small” or “large,” it is useful to place it in context of RRXU and RRUY values corresponding to other, similar, unmeasured confounders [56]. This could indicate, for example, that an association with a seemingly large E-value is rather weak to confounding if associations between the exposure and outcome with a comparable potential confounder are relatively high-magnitude [51]. In accord with recommendations, we describe E-values that exceed “reference RRs” (RRXU and RRUY where the unmeasured confounder is DIM rather than risk propensity) as “very strong,” those within the range of reference RRs as “moderately strong,” and those less than reference RRs as “likely not strong” to unmeasured confounding.

Second, we estimated the lower bound estimate of the true measure of association (RRTXY) by calculating the bias-adjusted RR [47, 54], using a bias factor to “correct” the confounded measure of association (RRcXY). The bias factor quantifies the maximum amount by which an unmeasured confounder could alter RRTXY [54] and is derived from RRXU and RRUY (Table 1). We assumed that RRXU differs between two subcategories of CASM conditions (1: alcohol, tobacco, opioids, and stimulants (ATOS); 2: depression, anxiety, and pain (DAP)) because they had distinct clustering patterns, potentially representing differential associations with risk propensity. The bias factor was calculated by [54]:

$${\rm{B = }}\left( {{\rm{R}}{{\rm{R}}_{{\rm{UY}}}}{\rm{\times R}}{{\rm{R}}_{{\rm{XU}}}}} \right){\rm{ / }}\left( {{\rm{R}}{{\rm{R}}_{{\rm{UY}}}}{\rm{ + }}\left( {{\rm{R}}{{\rm{R}}_{{\rm{XU}}}}{\rm{ - 1}}} \right)} \right){\rm{ }}$$
(2)

The bias-adjusted RRs (and 95% CIs) were calculated by:

$$\:\text{R}\text{R}^\text{T}\:_\text{XY}\:\ge\:\:\text{R}\text{R}^\text{C}\:_\text{XY}\:/\:\text{B }\:(\text{i}\text{f}\:\text{R}\text{R}_\text{XY}\:>1)$$
(3)
$$\:\text{R}\text{R}^\text{T}\:_\text{XY}\:\le\:\:\text{R}\text{R}^\text{C}\:_\text{XY}\:\times\:\text{B }\:(\text{i}\text{f}\:\text{R}\text{R}_\text{XY}\:<1)$$

Proportion of meaningfully strong effects method (PMSE)

We used PMSE for meta-analyses with heterogeneity (τ² > 0), sufficient numbers (≥ 10 studies, k), and robust parametric confidence intervals (as determined by 0.15 ≤ p̂(q) ≤ 0.85; p̂(q) defined below). The PMSE method assumes a normal distribution of population effects across studies to account for (1) dispersion of individual study point estimates from a hypothesized “true effect” size and (2) statistical error commensurate with the number of studies. This is done to prevent using “statistically significant” pooled estimates that comprise few studies with meaningfully strong effects and/or effects in the opposite direction of association (52).

The PMSE first assessed the strength of evidence for each meta-analysis by estimating p̂(q): the proportion of studies in a meta-analysis likely to meet or exceed a user-specified threshold of meaningfully strong effect (q). We set q = 1.10 (risk ratio (RR)) since all meta-analyses analyzed had positive directions of association, and additionally assessed q = 0.90 to consider the possibility of an association in the opposite direction. For example, p̂(q = 1.10) = 0.2 means that 20% of individual studies in a given meta-analysis are likely to yield risk ratios of 1.10 or higher. Larger p̂(q) values indicate stronger evidence and suggest the meta-analysis is robust to unmeasured confounding (47). We used the following recommended criteria to categorize strength of evidence by p̂(q): p̂(q) < r (weak), r ≤ p̂(q) ≤ 0.70 (moderate), p̂(q) > 0.70 (strong), referencing values used in existing literature (47). Here, “r” is a user-specified threshold for the minimum proportion of true effects above q needed to suggest evidence for causation. Given the observational nature of included studies in our analysis, we interpret evidence for causation as evidence for association suggesting causation. Per pre-established guidelines (47), we set r equal to 0.20 for meta-analyses with < 15 studies and equal to 0.10 for meta-analyses with ≥ 15 studies. Thus, for a meta-analysis of 30 studies, p̂(q = 1.10) > 0.10 implies sufficient evidence of causation in RRCXY since more than 10% of studies have true effect RRs above 1.10.

Second, for each meta-analysis, we used E-value analog parameters \({\rm{\hat T}}\)(r, q) and Ĝ(r, q) to determine the magnitude of unmeasured confounding necessary to reduce p̂(q) below r (suggesting RRCXY contains residual confounding). \({\rm{\hat T}}\)(r, q), analogous to the bias factor used in single studies, reflects the minimum bias capable of reducing p̂(q) to less than r. Ĝ(r, q), analogous to the E-value used in single studies, indicates the minimum values of RRUY and RRXU capable of reducing p̂(q) to less than r. Put another way, a meta-analysis can be considered robust to unmeasured confounding if \({\rm{\hat T}}\)(r, q) > the bias factor or Ĝ(r, q) > RRXU and RRUY. Given r = 0.10, q = 1.10, and k ≥ 15, these conditions would indicate that more than 10% of studies have true RRs > 1.10 and that RRCXY is a plausible estimate of RRTXY.

Analogous to the method used for the E-value, in order to gauge relative magnitude of Ĝ(r, q) values in the context of other potential unmeasured confounders, we describe Ĝ(r, q) that exceed the reference RRs (pertaining to DIM) as “very strong,” those within the range of reference RRs as “moderately strong,” and those less than reference RRs as “likely not strong” to unmeasured confounding. We apply the same categorizations for assessing the relative magnitude of \({\rm{\hat T}}\)(r, q) but use reference bias factors (pertaining to DIM) in lieu of reference RRs.

The p̂(q) value, \({\rm{\hat T}}\)(r, q), and Ĝ(r, q) values were calculated using the online tool accessible at: https://www.evalue-calculator.com/meta/ [47]. Results yielding p̂(q) < 0.15 or > 0.85 did not have robust parametric confidence intervals and are accordingly not reported.

Secondary analyses

Our secondary analysis considered a subset of meta-analyses in which constituent studies adjust for ≥ 1 other CASM to explore the impact of measured confounding by other CASM (e.g., constituent studies examine the association between tobacco and medication non-adherence, adjusting for alcohol use disorder). This comprised 26 CASM-adjusted ORs (one association yielded no eligible studies and one association yielded no CASM-adjusted estimates) and 25 “meta-analysis sets” comparing the primary meta-analysis OR and the CASM-adjusted-subgroup meta-analysis OR (one association yielded a CASM-adjusted OR without a reference unadjusted OR) (Figure S53; Supplementary Box 3). We define notable confounding by other CASM as a ≥ 0.20 change-in-effect between ORs of a given meta-analysis set (adjusted value - unadjusted value). We discuss these results in more detail in Supplementary Boxes 3 and 4.

Results

Analyses unadjusted for potential unmeasured confounding: systematic reviews and meta-analyses

Overall, 89% (24/27) of pooled ORs show strong evidence of a positive association with 95% CIs excluding the null, and varying levels of heterogeneity (Fig. 3). Pooled ORs of associations between CASM and medication non-adherence range from 1.45 (95% CI 1.17–1.79) to 2.52 (95% CI 2.07–3.07), with no confidence intervals spanning the null. Pooled ORs of associations between CASM and unprotected sex range from 1.22 (95% CI 0.80–1.87) to 2.08 (95% CI 1.88–2.31), with one of seven confidence intervals spanning the null. Pooled ORs for associations between CASM and transactional sex range from 1.48 (95% CI 1.16–1.89) to 3.05 (95% CI 1.75–5.31), with no confidence intervals spanning the null (there were no eligible studies for pain and transactional sex). Pooled ORs for associations between CASM and multiple sexual partners range from 0.94 (95% CI 0.66–1.35) to 2.69 (95% CI 2.04–3.55), with two of seven confidence intervals spanning the null. Alcohol, stimulants, depression, and anxiety have greater magnitudes of association with medication non-adherence compared to the other CASM, and alcohol and stimulants have greater magnitudes of association with unprotected sex compared to the other CASM. Alcohol, tobacco, and stimulants have greater magnitudes of association with transactional sex and multiple sexual partners compared to the other CASM (Fig. 3).

Fig. 3
figure 3

Meta-analysis pooled estimates of CASM-HIV risky behavior associations

Full size image

Analyses adjusted for potential unmeasured confounding

Point estimate method

Twenty-seven associations were assessed by the E-value, among which 81% (22/27) are robust to unmeasured confounding by risk propensity, as indicated by E-values and/or E-value 95% CI LLs exceeding RRXU and RRUY (Tables 2 and 3). The pain and multiple sexual partners systematic review yielded a single relevant study for which the E-value exceeded RRXU and RRUY. Among the non-robust associations, only the opioids and unprotected sex OR has an E-value (rather than the E-value 95% CI LL) exceeded by RRXU and RRUY. E-values range from 1.21 (95% CI LL 1.00) to 2.89 (95% CI LL 1.98). 80% (4/5) of non-robust associations involved associations with ATOS exposures rather than DAP exposures, and all non-robust associations comprise sexual behavior outcomes rather than the medication non-adherence outcome. In the context of reference RRs of DIM (Table S4; Supplementary Box 5), our E-values appear relatively robust (more in the discussion).

Bias factors ranged from 1.01 (smaller amount of bias for associations with DAP exposures) to 1.11 and 1.17 (greater bias for associations with ATOS exposures (Table 1). Bias-adjusted RRs suggest the CASM with more robust associations are stimulants, depression, and anxiety for medication non-adherence, pain, stimulants and depression for unprotected sex, stimulants and tobacco for transactional sex, and stimulants, tobacco, and pain for multiple sexual partners (Tables 2 and 3).

Proportion of meaningfully strong effects method

Twelve meta-analyses had sufficient heterogeneity, comprised ≥ 10 studies, and were therefore assessed by the PMSE method. The strength of evidence for causation varies from moderate (r ≤ p̂(q) ≤ 0.70) to strong (p̂(q) > 0.70)) with p̂(q = 1.10) ranging from 0.315 (for opioids and unprotected sex) to 0.831 (for stimulants and multiple sexual partners) (Table 3). That is, 31.5–83.1% of individual studies in corresponding meta-analyses are likely to yield RRs ≥ 1.10. Of the 12 associations, 50% (6/12) have moderate evidence for causation (alcohol, anxiety, and tobacco, and opioids for unprotected sex; opioids for transactional sex, and depression for multiple sexual partners), and 50% (6/12) have strong evidence for causation (opioids for medication non-adherence, depression, and stimulants for unprotected sex; tobacco for transactional sex; stimulants and alcohol for multiple sexual partners) (Table 3). All meta-analyses demonstrate robustness to bias from unmeasured confounding by risk propensity, as all Ĝ(r, q) values exceed their respective RRXU and RRUY values, ranging from 1.819 (opioids and medication non-adherence) to 3.805 (tobacco and transactional sex) (Table 3). Furthermore, all \({\rm{\hat T}}\)(r, q) values exceed their respective bias factor values, ranging from 1.254 (opioids and medication non-adherence) to 2.190 (tobacco and transactional sex). In the context of reference RRs and bias factors of DIM (Table S4; Supplementary Box 5; Table S5), our Ĝ(r, q) and \({\rm{\hat T}}\)(r, q) values are relatively robust (more in the discussion).

The true association is not likely to be in the opposite direction of the observed association, as analyses altering q to RR 0.90 yield very low p̂(q) values (between 0 and 0.091) among these meta-analyses, except for tobacco and unprotected sex (0.214), opioids and unprotected sex (0.447), opioids and transactional sex (0.211).

Secondary analyses

In secondary analyses of CASM-adjusted estimates (Figure S53; Supplementary Box 3; Tables S2-S3; Supplementary Box 4), we found that 73% (19/26) of pooled ORs show strong evidence of a positive association with 95% CIs excluding the null, and varying levels of heterogeneity. Among the 25 meta-analyses sets, 84% (21/25) of CASM-adjusted meta-analysis estimates have notable confounding by other CASM. Most of the notable shifts, 76% (16/21), occurred towards the null, suggesting positive confounding by other CASM. ATOS associations are more likely to have notable confounding by CASM than DAP associations (100% (15/15) versus 60% (6/10)), and among those with notable shifts, ATOS associations are also more likely to indicate positive confounding by other CASM (87% (13/15) versus 50% (3/6)). Among those with notable shifts, transactional sex, and multiple sexual partner associations are more likely than medication non-adherence and unprotected sex associations to indicate positive confounding by other CASM (100% (9/9) versus 57% (4/7) and 60% (3/5), respectively). In regard to E-values, 62% (16/26) are robust to unmeasured confounding by risk propensity, and 60% (6/10) of non-robust E-values comprised ATOS associations. E-values range from 1.00 (95% CI LL 1.00) to 2.97 (95% CI LL 2.24) (Tables S2-S3). Only two associations were assessed by PMSE, both showing strong evidence of causation (p̂(q) > 0.70) and robustness to unmeasured confounding by sufficiently large \({\rm{\hat T}}\)(r, q) Ĝ(r, q) values (Tables S2-S3; Supplementary Box 4).

Discussion

Our analyses unadjusted for potential unmeasured confounding demonstrate evidence for associations between CASM and HIV risk behaviors (Fig. 3), consistent with past findings [20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40]. Our findings add to the literature by increasing precision and generalizability and suggest that alcohol and stimulants are dominant risk factors for all investigated HIV risk behaviors. Our analyses accounting for potential unmeasured confounding by risk propensity indicate that most associations were robust to unmeasured confounding. Associations between ATOS exposures and HIV risk behaviors are less robust to unmeasured confounding than associations between DAP exposures and HIV risk behaviors. Bias-adjusted RRs reflect varying levels of bias by risk propensity, particular to exposure or outcome type. While some bias factors seem negligible in magnitude (e.g., for DAP and medication non-adherence), we present all bias-adjusted RRs to identify trends across CASM conditions and HIV risk behaviors.

Our sensitivity analyses designed to investigate unmeasured confounding in meta-analyses suggest that unmeasured confounding by risk propensity is unlikely to explain the meta-analyzed associations. Furthermore, no meta-analyses showed evidence of a “protective” effect (q = 0.90), suggesting it is unlikely that the true direction of association was negative. Though the Point Estimate method suggests ATOS associations are less robust to unmeasured confounding than DAP associations, PMSE results suggest all ATOS associations are robust to unmeasured confounding. In the PMSE analysis, all associations appear strong to unmeasured confounding, with 75% (9/12) of meta-analyzed studies comprising ATOS associations. This apparent “relative leniency” of PMSE, compared to the Point Estimate method, that deems any residual confounding insufficient to explain results is likely explained by its consideration for inter-study heterogeneous effects that are intrinsic to meta-analyses. Unlike the derivation of Point Estimate method parameters which solely use the central measure of association (RRs), the derivation of PMSE parameters include measures of variance and heterogeneity [47]. Accordingly, the presence of heterogeneity across our meta-analyses (Table S1) suggests more weight should be placed on the PMSE results, as they better characterize the heterogeneous nature of meta-analyzed data. Furthermore, use of meta-analysis-specific parameters facilitate comparison of our data with other literature.

Inferences from the sensitivity analyses hinge to an extent on the relative strength of the unmeasured confounder. To assess the robustness of our Point Estimate method findings, we used DIM as an alternative unmeasured confounder [56], comparing our E-values to corresponding exposure-covariable RRs (RRXUATOS-DIM or RRXUDAP-DIM) and covariable-outcome (RRXUDIM-medication non-adherence or RRXUDIM-risky sexual behavior) (Table S4; Supplementary Box 5) [57,58,59,60]. DIM appears to influence the associations less than risk propensity, as all (27/27) E-values appear robust relative to DIM reference RRs. This suggests that our assessment that 81% of associations are robust may be conservative. Some differences emerge by outcome. For risky sexual behavior outcomes, unlike with risk propensity, we see that ATOS associations are generally more robust to DIM confounding than DAP associations. For medication non-adherence, all exposures except for pain are “very strong” and/or “moderately strong,” and pain is “likely not strong.” These differences illustrate the importance of the selection of the unmeasured confounder proxy used to interpret E-values. We believe risk propensity remains a suitable proxy for unmeasured confounding, as it covers behavioral factors transcending the medical sector.

The inferences from PMSE also hinge on the relative strength of the unmeasured confounder. There is almost perfect alignment between the relative assessment of Ĝ(r, q) and the relative assessment of E-values (as defined in methods) across meta-analyses (both compared to DIM E-values, Table S4, Supplementary Box 5). The only differences were depression and transactional sex and depression and multiple sexual partners (assessment was “likely not strong” by E-values and “very strong” by Ĝ(r, q) values). There is perfect alignment between the relative assessment of \({\rm{\hat T}}\)(r, q) and the relative assessment of E-values (compared to DIM bias factors in Table S5 and DIM E-values in Table S4, respectively).

Our secondary analysis of CASM-adjusted estimates suggests that compared to DAP, ATOS associations are more sensitive to measured confounding (by CASM) and unmeasured confounding due to a greater prevalence of ATOS among meta-analysis sets suggesting notable, positive confounding. This suggests that ATOS has greater clustering with other CASM than DAP. Our results also indicate that associations between CASM and medication non-adherence associations are more sensitive to measured confounding (by other CASM factors) than associations between CASM and risky sexual behaviors are; however, they are less sensitive to unmeasured confounding than associations between CASM and risky sexual behaviors. This suggests that measures of association that adjust for other CASM may sufficiently account for residual confounding in CASM-medication non-adherence associations. Although the ATOS-DAP distinction was not seen by PMSE, this is likely reflective of its limited scope, as only two meta-analyses were included by this method in the secondary analysis. Our secondary analysis findings are slightly different when considered relative to DIM. First, CASM-adjusted meta-analyses have a lower prevalence of robust estimates (as determined by the E-value) at 73% (19/26), compared to 100% (27/27) in the primary analysis. While we hypothesized the opposite, that CASM adjustment increases robustness to unmeasured confounding, these results could indicate that non-CASM confounders influence our associations more than CASM confounders. Alternatively, or in combination, CASM-adjusted studies may account for non-CASM conditions less than studies that do not adjust for CASM. Second, we see a higher prevalence of robustness to unmeasured confounding relative to DIM across secondary analyses (73% (19/26) for DIM and 62% (16/26) for risk propensity), supporting previous findings that risk propensity appears to influence our associations more than DIM. Despite these differences, CASM-adjusted meta-analyses are still “very strong” to unmeasured confounding (Tables S2-S3; Supplementary Box 4; Table S4; Supplementary Box 5; Table S5) according to Ĝ(r, q) and \({\rm{\hat T}}\)(r, q) values.

Our analyses were not without limitations. Our systematic review searches were limited to PubMed and English articles. Most constituent studies were cross-sectional and have potential for reverse causality, measurement error, selection bias, and publication bias [56]. Limiting inclusion of studies to those reporting ORs or convertible data may have introduced bias by requiring a dichotomous outcome but is not likely to affect findings. We did not perform risk of bias qualitative assessments for each meta-analysis, although we did assess the impact of CASM adjustment via secondary analysis. We had a single reviewer per meta-analysis and do not have a kappa test value of internal validity. However, a standardized data collection document and protocol were used. Decision rules used to standardize data collection across the meta-analyses have the potential to introduce bias compounded by publication bias (e.g., by choosing the mode population group or higher exposure strata to avoid overrepresentation of a study). Our secondary analysis-based interpretations on the presence and direction of measured confounding by other CASM are non-definitive, as they imperfectly assess confounding due to (1) the mutual inclusivity of CASM-adjusted studies in both primary and secondary analyses, (2) comparing studies in lieu of individuals for change-in-effect, and (3) a small study count for some secondary meta-analyses. Furthermore, we did not explore the possibility of unmeasured negative confounding that may bias results towards rather than away from the null effect. In such a case, the literature-based input value RRUY would have the reverse direction of association with the outcome compared to the direction of association between the exposure and outcome. For example, if the RR between the exposure and outcome is positive, RRUY would be negative. In this case, the E-values and PMSE-derived parameters would reflect the strength of unmeasured negative confounders in skewing effect estimates towards the null - that is, the extent to which they underestimate the true effect. For our point estimate analyses, associations with relatively high E-values may still have residual unmeasured confounding arising from other sources, such as the cross-sectional design of many constituent studies [61]. Interpretation of the E-value depends on the magnitude of difference between the exposed and unexposed groups, but this nuance was not assessed due to variation across constituent studies. We used one or two studies to estimate alternative RRXU, RRUY, and bias factors. However, this may be an oversimplification if these studies do not align with the specific study populations across all constituent studies they were applied to. It may be argued that potential unmeasured confounders other than risk propensity and DIM could be suitable candidates in this analysis. While our selection was based on plausibility, literature, and consideration to avoid colliders, alternative candidates could be applied to our illustrative analysis. As with the Point Estimate method, we used one or two studies to estimate alternative RRXU, RRUY, and bias factors for PMSE which may not capture the full plausible range of values. Despite its utility for quantitative assessment of unmeasured confounding in heterogeneous contexts, p̂(q) benchmark values and user-specified thresholds were somewhat arbitrary, though guideline- and literature-based to procure results compatible with other literature. 95% CIs were not available for interpretation in PMSE since the variance of τ² and the pooled point estimate were not available for each meta-analysis. We also only assessed a single threshold for a meaningful effect, RR 1.10, which could be a relatively low threshold. Only 42% (11/26) of studies met the criteria for PMSE, so this analysis was limited in scope. However, we reported E-values for all 27 meta-analyses and single studies to provide a uniform parameter for comparison. Although some constituent studies adjust for various potential confounders, we do not believe they adequately consider all unmeasured confounding and accordingly do not presume that our bias-adjusted RRs over-adjust for confounders.

Conclusion

CASM affect a sizeable portion of PLHIV, with effects reverberating from the individual to society. The associations between CASM and HIV transmission and progression risk factors highlight prevention opportunities, especially in light of evidence for effective interventions that treat CASM among PLHIV [62]. In fact, our measures of association have been used to inform a model examining the potential health benefits of integrated screening strategies for CASM conditions among PLHIV, using upper bound estimates for residual confounding [63]. Our analyses accounting for potential unmeasured confounding by risk propensity suggest that alcohol and stimulants are dominant risk factors for all investigated HIV risk behaviors, highlighting potential high-impact targets for integrative intervention strategies to reduce HIV morbidity and mortality. Associations with ATOS exposures are more sensitive than associations with DAP exposures to unmeasured confounding and more likely to be positively confounded by other CASM, suggesting greater clustering of ATOS conditions with other CASM and underscoring the value of targeting alcohol and stimulants in integrative intervention strategies. Despite evidence of measured and unmeasured confounding, we demonstrate that most associations are sufficiently strong, with or without adjustment for bias.

Data availability

The supplementary file contains additional information on the systematic reviews and meta-analyses (search terms, PRISMA diagrams, constituent article reference list, meta-analysis forest plots) as well as the analyses adjusting for potential unmeasured confounding (meta-analysis heterogeneity values, full secondary analysis, and measures of association, E-values, and bias factors for the comparable potential unmeasured confounder).

Abbreviations

HIV:

Human immunodeficiency virus

PLHIV:

People living with HIV

CASM:

Constellation of alcohol (e.g., alcohol use disorder), other substance (e.g., tobacco, opioid, and stimulant use disorders), and mood-related (e.g. depressive and generalized anxiety disorders and chronic pain) conditions

ATOS:

Alcohol, tobacco, opioids, and stimulants

DAP:

Depression, anxiety, and pain

DIM:

Distrust of medical institutions

PMSE:

Proportion of Meaningfully Strong Effects method

References

  1. World Health Organization. HIV data and statistics [Internet]. 2023 [cited 2023 Jul 29]. Available from: https://www.who.int/teams/global-hiv-hepatitis-and-stis-programmes/hiv/strategic-information/hiv-data-and-statistics

  2. Woolf-King SE, Steinmaus CM, Reingold AL, Hahn JA. An update on alcohol use and risk of HIV infection in sub-saharan Africa: Meta- analysis and future research directions. Int J Alcohol Drug Res. 2013;2(1):99–110.

    Article  Google Scholar 

  3. Hinkin CH, Castellon SA, Atkinson JH, Goodkin K. Neuropsychiatric aspects of HIV infection among older adults. J Clin Epidemiol. 2001;54(12):S44–52.

    Article  PubMed  PubMed Central  Google Scholar 

  4. Murphy JD, Liu B, Parascandola M. Smoking and HIV in Sub-saharan Africa: a 25-Country analysis of the demographic health surveys. Nicotine Tob Res. 2018;21(8):1093–102.

    Article  PubMed Central  Google Scholar 

  5. Olagunju AT, Adeyemi JD, Ogbolu RE, Campbell EA. A study on epidemiological profile of anxiety disorders among people living with HIV/AIDS in a sub-saharan Africa HIV clinic. AIDS Behav. 2012;16(8):2192–7.

    Article  PubMed  Google Scholar 

  6. Nakimuli-Mpungu E, Bass JK, Alexandre P, Mills EJ, Musisi S, Ram M, et al. Depression, Alcohol Use and Adherence to Antiretroviral Therapy in Sub-saharan Africa: a systematic review. AIDS Behav. 2012;16(8):2101–18.

    Article  PubMed  Google Scholar 

  7. Arseniou S, Arvaniti A, Samakouri M. HIV infection and depression. Psychiatry Clin Neurosci. 2014;68(2):96–109.

    Article  PubMed  Google Scholar 

  8. Addis DR, DeBerry JJ, Aggarwal S. Chronic Pain in HIV. Mol Pain. 2020;16:174480692092727.

    Article  Google Scholar 

  9. Reynolds NR. Cigarette smoking and HIV: more evidence for action. AIDS Educ Prev. 2009;21(3supplement):106–21.

    Article  PubMed  PubMed Central  Google Scholar 

  10. Denis CM, Morales KH, Wu Q, Metzger DS, Cheatle MD. Association between diagnoses of Chronic Noncancer Pain, Substance Use Disorder, and HIV-Related outcomes in people living with HIV. J Acquir Immune Defic Syndr 1999. 2019;82(2):S142–7.

    Article  Google Scholar 

  11. Ciesla JA, Roberts JE. Meta-analysis of the Relationship between HIV infection and risk for Depressive disorders. Am J Psychiatry. 2001;158(5):725–30.

    Article  PubMed  CAS  Google Scholar 

  12. Mishkin K, Nugmanova Z, Urbaeva J, Nugumanova G, Abdumananova M, Kim E, et al. Anxiety and depression among women living with HIV in Kazakhstan. AIDS Care. 2021;33(2):172–9.

    Article  PubMed  Google Scholar 

  13. Ngocho JS, Watt MH, Minja L, Knettel BA, Mmbaga BT, Williams PP et al. Depression and anxiety among pregnant women living with HIV in Kilimanjaro region, Tanzania. Yotebieng M, editor. PLOS ONE. 2019;14(10):e0224515.

  14. Yousuf A, Musa R, Isa MLMd, Arifin SRM. Anxiety and depression among women living with HIV: prevalence and correlations. Clin Pract Epidemiol Ment Health. 2020;16(1):59–66.

    Article  PubMed  PubMed Central  Google Scholar 

  15. An Q, Song R, Hernandez A, Hall HI. Trends and differences among three New indicators of HIV infection progression. Public Health Rep. 2015;130(5):468–74.

    Article  PubMed  PubMed Central  Google Scholar 

  16. Braithwaite RS, Conigliaro J, Roberts MS, Shechter S, Schaefer A, McGinnis K, et al. Estimating the impact of alcohol consumption on survival for HIV + individuals. AIDS Care. 2007;19(4):459–66.

    Article  PubMed  PubMed Central  CAS  Google Scholar 

  17. Hileman CO, McComsey GA. The opioid epidemic: impact on inflammation and Cardiovascular Disease Risk in HIV. Curr HIV/AIDS Rep. 2019;16(5):381–8.

    Article  PubMed  PubMed Central  Google Scholar 

  18. Leserman J. HIV disease progression: depression, stress, and possible mechanisms. Biol Psychiatry. 2003;54(3):295–306.

    Article  PubMed  Google Scholar 

  19. Yousuf A, Mohd Arifin SR, Musa R, Md. Isa ML. Depression and HIV Disease Progression: a Mini-review. Clin Pract Epidemiol Ment Health. 2019;15(1):153–9.

    Article  PubMed  PubMed Central  Google Scholar 

  20. Braithwaite RS, McGinnis KA, Conigliaro J, Maisto SA, Crystal S, Day N, et al. A temporal and Dose-Response Association between Alcohol Consumption and Medication Adherence among veterans in Care. Alcohol Clin Exp Res. 2005;29(7):1190–7.

    Article  PubMed  Google Scholar 

  21. Chander G, Lau B, Moore RD. Hazardous Alcohol Use: A Risk Factor for Non-Adherence and Lack of Suppression in HIV Infection. J Acquir Immune Defic Syndr. 1999. 2006;43(4):411–7.

  22. Hicks PL, Mulvey KP, Chander G, Fleishman JA, Josephs JS, Korthuis PT, et al. The impact of illicit drug use and substance abuse treatment on adherence to HAART. AIDS Care. 2007;19(9):1134–40.

    Article  PubMed  PubMed Central  CAS  Google Scholar 

  23. Hinkin CH, Hardy DJ, Mason KI, Castellon SA, Durvasula RS, Lam MN, et al. Medication adherence in HIV-infected adults: effect of patient age, cognitive status, and substance abuse. AIDS Lond Engl. 2004;18(Suppl 1):S19–25.

    Article  Google Scholar 

  24. Hinkin CH, Barclay TR, Castellon SA, Levine AJ, Durvasula RS, Marion SD, et al. Drug use and medication adherence among HIV-1 infected individuals. AIDS Behav. 2007;11(2):185–94.

    Article  PubMed  PubMed Central  Google Scholar 

  25. Ingersoll K. The impact of psychiatric symptoms, drug use, and medication regimen on non-adherence to HIV treatment. AIDS Care. 2004;16(2):199–211.

    Article  PubMed  CAS  Google Scholar 

  26. Lazo M, Gange SJ, Wilson TE, Anastos K, Ostrow DG, Witt MD, et al. Patterns and predictors of changes in adherence to highly active antiretroviral therapy: longitudinal study of men and women. Clin Infect Dis off Publ Infect Dis Soc Am. 2007;45(10):1377–85.

    Article  Google Scholar 

  27. Maru DSR, Bruce RD, Walton M, Mezger JA, Springer SA, Shield D, et al. Initiation, adherence, and retention in a randomized controlled trial of directly administered antiretroviral therapy. AIDS Behav. 2008;12(2):284–93.

    Article  PubMed  Google Scholar 

  28. Mohammed H, Kieltyka L, Richardson-Alston G, Magnus M, Fawal H, Vermund SH, et al. Adherence to HAART among HIV-infected persons in rural Louisiana. AIDS Patient Care STDs. 2004;18(5):289–96.

    Article  PubMed  Google Scholar 

  29. Murphy DA, Marelich WD, Hoffman D, Steers WN. Predictors of antiretroviral adherence. AIDS Care. 2004;16(4):471–84.

    Article  PubMed  CAS  Google Scholar 

  30. Peretti-Watel P, Spire B, Lert F, Obadia Y, VESPA Group. Drug use patterns and adherence to treatment among HIV-positive patients: evidence from a large sample of French outpatients (ANRS-EN12-VESPA 2003). Drug Alcohol Depend. 2006;82(Suppl 1):S71–79.

    Article  PubMed  Google Scholar 

  31. Tucker JS, Burnam MA, Sherbourne CD, Kung FY, Gifford AL. Substance use and mental health correlates of nonadherence to antiretroviral medications in a sample of patients with human immunodeficiency virus infection. Am J Med. 2003;114(7):573–80.

    Article  PubMed  Google Scholar 

  32. Wang H, He G, Li X, Yang A, Chen X, Fennie KP, et al. Self-reported adherence to antiretroviral treatment among HIV-infected people in Central China. AIDS Patient Care STDs. 2008;22(1):71–80.

    Article  PubMed  Google Scholar 

  33. Baru A, Adeoye IA, Adekunle AO. Risky sexual behavior and associated factors among sexually-active unmarried young female internal migrants working in Burayu Town, Ethiopia. PLoS ONE. 2020;15(10):e0240695.

    Article  PubMed  PubMed Central  CAS  Google Scholar 

  34. Castillo Mezzich A, Tarter RE, Giancola PR, Lu S, Kirisci L, Parks S. Substance use and risky sexual behavior in female adolescents. Drug Alcohol Depend. 1997;44(2–3):157–66.

    Article  PubMed  CAS  Google Scholar 

  35. Dallelucci CC, Bragiato EC, Areco KCN, Fidalgo TM, da Silveira DX. Sexual risky behavior, cocaine and alcohol use among substance users in an outpatient facility: a cross section study. Subst Abuse Treat Prev Policy. 2019;14(1):46.

    Article  PubMed  PubMed Central  Google Scholar 

  36. Muche AA, Kassa GM, Berhe AK, Fekadu GA. Prevalence and determinants of risky sexual practice in Ethiopia: systematic review and Meta-analysis. Reprod Health. 2017;14(1):113.

    Article  PubMed  PubMed Central  Google Scholar 

  37. Pokhrel P, Bennett BL, Regmi S, Idrisov B, Galimov A, Akhmadeeva L, et al. Individualism-Collectivism, Social Self-Control and adolescent substance use and risky sexual behavior. Subst Use Misuse. 2018;53(7):1057–67.

    Article  PubMed  Google Scholar 

  38. Ritchwood TD, Ford H, DeCoster J, Sutton M, Lochman JE. Risky sexual behavior and substance use among adolescents: a Meta-analysis. Child Youth Serv Rev. 2015;52:74–88.

    Article  PubMed  PubMed Central  Google Scholar 

  39. Shorey RC, Fite PJ, Choi H, Cohen JR, Stuart GL, Temple JR. Dating violence and substance use as longitudinal predictors of adolescents’ risky sexual behavior. Prev Sci off J Soc Prev Res. 2015;16(6):853–61.

    Article  Google Scholar 

  40. Turpin RE, Salerno JP, Rosario AD, Boekeloo B, Victimization. Substance Use, Depression, and sexual risk in adolescent males who have sex with males: a Syndemic Latent Profile Analysis. Arch Sex Behav. 2021;50(3):961–71.

    Article  PubMed  Google Scholar 

  41. Braithwaite RS, Ban KF, Stevens ER, Caniglia EC. Rounding up the usual suspects: confirmation bias in epidemiological research. Int J Epidemiol. 2021;50(4):1053–7.

    Article  PubMed  PubMed Central  Google Scholar 

  42. Cobb-Clark DA, Dahmann SC, Kettlewell N, Depression. Depression, Risk Preferences and Risk-Taking Behavior. Journal of Human Resources 2022;57(5):1566–160

  43. Kaplan CP, NaÂpoles-Springer A, Stewart SL, PeÂrez-Stable EJ. Smoking acquisition among adolescents and young latinas the role of socioenvironmental and personal factors. Addict Behav. 2001;20.

  44. Pelletier-Fleury N, Le Vaillant M, Goupil F, Paris A, Pigeane T, Gagnadoux F et al. Risk-seeking attitude in health and safety domain is associated with continuous positive airway pressure discontinuation in patients with obstructive sleep apnea—a multicenter prospective cohort study. Sleep [Internet]. 2021 Feb 12 [cited 2021 Aug 27];44(2). Available from: https://academic-oup-com.ezproxy.med.nyu.edu/sleep/article/44/2/zsaa156/5896478

  45. Risser JMH, Timpson SC, McCurdy SA, Ross MW, Williams ML. Psychological correlates of Trading Sex for Money among African American Crack Cocaine smokers. Am J Drug Alcohol Abuse. 2006;32(4):645–53.

    Article  PubMed  Google Scholar 

  46. Mathur MB, VanderWeele TJ. New metrics for meta-analyses of heterogeneous effects. Stat Med. 2019;38(8):1336–42.

    Article  PubMed  Google Scholar 

  47. Mathur MB, VanderWeele TJ. Sensitivity analysis for unmeasured confounding in Meta-analyses. J Am Stat Assoc. 2020;115(529):163–72.

    Article  PubMed  CAS  Google Scholar 

  48. Feelemyer JP, Richard E, Khan MR, Scheidell JD, Caniglia EC, Manandhar-Sasaki P et al. Does the Association between stimulant use and high risk sexual behavior vary by Injection Drug Use, sexual minority status, or HIV infection status? A Meta-analysis. AIDS Behav. 2023 Sep;27(9):2883-2890.

  49. Bland JM, Altman DG. The odds ratio. BMJ. 2000;320(7247):1468.

    Article  PubMed  PubMed Central  CAS  Google Scholar 

  50. Ding P, VanderWeele TJ. Sensitivity analysis without assumptions. Epidemiol Camb Mass. 2016;27(3):368–77.

    Article  Google Scholar 

  51. Haneuse S, VanderWeele TJ, Arterburn D. Using the E-Value to assess the potential effect of unmeasured confounding in Observational studies. JAMA. 2019;321(6):602–3.

    Article  PubMed  Google Scholar 

  52. Mathur MB, VanderWeele TJ. Robust Metrics and sensitivity analyses for Meta-analyses of Heterogeneous effects. Epidemiol Camb Mass. 2020;31(3):356–8.

    Article  Google Scholar 

  53. Mathur MB, VanderWeele TJ. Methods to address confounding and other biases in Meta-analyses: review and recommendations. Annu Rev Public Health. 2022;43(1).

  54. VanderWeele TJ, Ding P. Sensitivity analysis in Observational Research: introducing the E-Value. Ann Intern Med. 2017;167(4):268–74.

    Article  PubMed  Google Scholar 

  55. VanderWeele TJ. On a Square-Root Transformation of the odds ratio for a common outcome. Epidemiol Camb Mass. 2017;28(6):e58–60.

    Article  Google Scholar 

  56. VanderWeele TJ, Mathur MB, Commentary. Developing best-practice guidelines for the reporting of E-values. Int J Epidemiol. 2020;49(5):1495–7.

    Article  PubMed  PubMed Central  Google Scholar 

  57. Ahnquist J, Lindström M, Wamala SP. Institutional trust and alcohol consumption in Sweden: the Swedish National Public Health Survey 2006. BMC Public Health. 2008;8(1):283.

    Article  PubMed  PubMed Central  Google Scholar 

  58. Eaton LA, Driffin DD, Kegler C, Smith H, Conway-Washington C, White D, et al. The role of Stigma and Medical Mistrust in the Routine Health Care Engagement of Black men who have sex with men. Am J Public Health. 2015;105(2):e75–82.

    Article  PubMed  PubMed Central  Google Scholar 

  59. Elder K, Ramamonjiarivelo Z, Wiltshire J, Piper C, Horn WS, Gilbert KL, et al. Trust, Medication Adherence, and Hypertension Control in Southern African American men. Am J Public Health. 2012;102(12):2242–5.

    Article  PubMed  PubMed Central  Google Scholar 

  60. Piette JD, Heisler M, Krein S, Kerr EA. The role of patient-Physician Trust in moderating medication nonadherence due to cost pressures. Arch Intern Med. 2005;165(15):1749–55.

    Article  PubMed  Google Scholar 

  61. VanderWeele TJ, Mathur MB, Ding P. Correcting misinterpretations of the E-Value. Ann Intern Med. 2019;170(2).

  62. Chibanda D, Benjamin L, Weiss HA, Abas M. Mental, neurological, and Substance Use disorders in people living with HIV/AIDS in Low- and Middle-Income Countries. JAIDS J Acquir Immune Defic Syndr. 2014;67:S54.

    Article  PubMed  Google Scholar 

  63. Bershteyn A, Richard E, Zhou Q, Khan MR, Scheidell JD, Manandhar-Sasaki P, et al. Potential health benefits of integrated screening strategies for alcohol, tobacco, other substance use, depression, anxiety, and chronic pain among people living with HIV in the USA: a mathematical modelling study. Lancet HIV. 2023;10(2):e118–25.

    Article  PubMed  CAS  Google Scholar 

Download references

Acknowledgements

We are grateful to Maya Mathur for feedback on methodology inquiries, Maria Khan for feedback on meta-analysis results, and Dyanna Charles for management support.

Funding

Funding was provided by the National Institutes of Health (NIH) and the National Institute on Alcohol Abuse and Alcoholism (NIAAA), on grants R01-AA024706 and U01-AA020799. The funding bodies played no role in the design of the study and collection, analysis, and interpretation of data and in writing the manuscript.

Author information

Authors and Affiliations

  1. Department of Population Health, New York University Grossman School of Medicine, New York, NY, USA

    Prima Manandhar-Sasaki, Kaoon Francois Ban, Emma Richard & R. Scott Braithwaite

  2. Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, USA

    Ellen C. Caniglia

Authors
  1. Prima Manandhar-Sasaki
    View author publications

    You can also search for this author inPubMed Google Scholar

  2. Kaoon Francois Ban
    View author publications

    You can also search for this author inPubMed Google Scholar

  3. Emma Richard
    View author publications

    You can also search for this author inPubMed Google Scholar

  4. R. Scott Braithwaite
    View author publications

    You can also search for this author inPubMed Google Scholar

  5. Ellen C. Caniglia
    View author publications

    You can also search for this author inPubMed Google Scholar

Contributions

PM wrote the manuscript text, created all tables and figures, and prepared the appendix; RSB and ECC conceptualized the analysis; PM conducted the analysis and interpretation; PM, KFB, and ER did the systematic reviews and meta-analyses; all co-authors assisted in edits to the manuscript and reviewed the final draft prior to submission to the journal.

Corresponding author

Correspondence to Prima Manandhar-Sasaki.

Ethics declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.

Additional information

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Supplementary Material 1

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Manandhar-Sasaki, P., Ban, K.F., Richard, E. et al. How likely is unmeasured confounding to explain meta-analysis-derived associations between alcohol, other substances, and mood-related conditions with HIV risk behaviors?. BMC Med Res Methodol 25, 62 (2025). https://doiorg.publicaciones.saludcastillayleon.es/10.1186/s12874-025-02490-9

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doiorg.publicaciones.saludcastillayleon.es/10.1186/s12874-025-02490-9

Keywords

BMC Medical Research Methodology

ISSN: 1471-2288

Contact us