Systematic review of hormonal contraception and risk of HIV transmission: when to resist meta-analysis

Introduction

HIV is most often transmitted by sexual contact, and women are probably at higher risk of acquiring HIV from an infected male partner than vice versa. The mechanisms involved in vaginal transmission of HIV are poorly understood, but oestrogen and progesterone both have effects on the female genital tract that could plausibly influence HIV transmission [^1,2]. Since hormonal contraceptives (including oral, injectable and implantable forms) are in widespread use, it is clearly important to know whether they have any significant role in the vaginal transmission of HIV. Numerous studies relating to contraception and HIV have been published [^3,4]. The main purpose of this article is to present a systematic review of the epidemiological evidence concerning hormonal contraception and the risk of HIV transmission to women. Other lines of evidence and issues facing further research are also considered, including the appropriate use of meta-analysis.

Review of epidemiological evidence

Search strategy

All published studies of HIV infection and contraception were sought through a combination of computerized and hand searches. On-line databases MEDLINE, EMBASE and the Science Citation Index were searched from 1983 to 1997, using exploded MeSH search terms in combination with Cochrane search histories for publications related to causality and risk. Hard copies of suitable studies and review articles, plus relevant citations from those papers were obtained. Abstract books of the International Conferences on AIDS were hand-searched from 1988 to 1996, and journals were hand-searched from January to June 1997 to identify relevant papers that might not have been available on-line. Prospective, case-control and cross-sectional designs were included if they reported crude or adjusted odds ratios (OR), or rate ratios, with 95% confidence intervals (CI), or contained sufficient numeric data from which that information could be calculated, or if the necessary data could be obtained from the authors. Studies were excluded if they appeared to predate a later publication of the same study population. Further details of the search strategy are available from the author.

Assessment of methodological quality

Several predefined criteria were used to assess the quality of individual studies. One was whether the period relating to the woman's contraceptive history clearly predated any HIV infection, because this sequence of events is necessary for a causal explanation of the association between contraception and infection. Another criterion was whether the study was specifically designed to examine the relationship between hormonal contraception and risk of HIV transmission. Prospective studies fulfilling the first criteria, by defining a risk period for transmission starting with an HIV-negative test and ending with a second HIV test, were then assessed for completeness of follow-up. All studies were judged on how they had dealt with potential bias or confounding in five important respects: (i) level of exposure to HIV, (ii) definition of hormonal contraceptive users, (iii) definition of comparison group, (iv) history or presence of other sexually transmitted diseases (STD), and (v) condom use. The first criterion strongly affects the risk of transmission, but is hard to estimate. Partner studies, in which all women have a single HIV-infected partner, were considered to provide a setting in which HIV exposure would be more uniform or quantifiable than studies of women with one or more sexual partners of unknown HIV status. The second and third criteria relate to misclassification of the two groups [⁵]. Other STD (fourth criterion) are a major cofactor for HIV transmission [⁶] and therefore a potentially strong confounder, whereas consistent condom use (fifth criterion) is highly protective [⁷] against HIV transmission and therefore another potentially strong confounder.

A total of 32 studies relating to hormonal contraception and risk of HIV transmission were included in the review. Ten studies were prospective (Table 1) and 22 were cross-sectional or retrospective (Table 2). A few of the studies included both prospective and cross-sectional analyses relating to the same study population, and therefore appear separately in Tables 1 and 2, respectively. Two of the prospective studies [^7,8] and five of the others [^9-12] were partner studies in which each woman had a regular sexual partner who was known to be HIV-infected. Most of the prospective studies had considerable (25-37%) or unspecified losses to follow-up. Follow-up was more complete in the two partner studies.

Table 1

Table 2

Table 2

Only five of the 32 studies (two case-control and three cross-sectional) [^9,13-16] were specifically designed to examine the relationship between contraception and HIV transmission. Taneepanichskul et al. [¹⁶] appear to have taken care to reduce confounding in their case-control study by matching each case (HIV-positive women) to two controls (HIV-negative women) on a number of risk factors. However, the reported OR were incorrect because the analysis did not take the matched design into account. When a matched case-control study is analysed as if there has been no matching, the OR is generally biased towards 1.0, and the true relationship between putative risk factor (oral contraceptive use) and disease (HIV infection) is underestimated [¹⁷]. The other case-control study [¹³] was a pilot for a larger study, from which the authors chose not to draw firm conclusions. The partner study by Nagachinta et al. [⁹] estimated the probability of transmission per intercourse act during a risk period defined by male partners who tested HIV-positive before marriage or had known seroconversion dates (i.e., the start of the risk period was not defined by women having a negative HIV test). The remaining two studies [^14,15], which stated that examination of the relationship between contraceptive use and HIV infection was a study objective, were conducted in African women attending family planning clinics. Both examined duration of hormonal contraceptive use with extensive adjustment for potential confounding factors.

In four (40%) of the prospective studies and 10 (46%) of the other studies, no multivariate analyses of the relationship between hormonal contraception and HIV infection were presented. Methods of controlling for potential confounding by level of HIV exposure varied from none to multiple variables relating to sexual practices (particularly in studies of prostitutes). Three of the partner studies [^9-11] also adjusted for factors related to the infectiousness of the male partners. Apart from the studies that were specifically set up to examine this research question, contraceptive use was generally poorly defined, and the choice of comparison group varied from never users of any contraceptive method to combinations of users of non-hormonal methods of unspecified duration. Treatment of condom users was also highly variable. In some studies [^14,18-20] condom use was very low and therefore ignored, and in three studies [^9,12,21] condom users were excluded. Most of the studies that reported data on other STD took account of this information in multivariate analyses.

Findings of the studies

No clear or consistent pattern emerge from these studies. If the findings are considered according to methodological quality, then the two strongest studies are the prospective partner studies [^7,8]. Pooling the published data from one cohort [⁸] with unpublished raw data supplied by the other cohort [⁷] gave a crude serocon-version risk ratio of 0.30 (95% CI, 0.05-2.1) for women who used the oral contraceptive pill during a defined exposure period, compared with those who did not. However, the combined number of seroconversions is relatively small (n = 26) and the wide CI are compatible with both a 95% reduction and a 100% increase in the risk of seroconversion associated with oral contraceptive use.

Taking the 10 prospective studies together (Table 1), three reported a three-to fourfold increased risk in association with oral contraceptive use [^13,22,23], and two found a similar increase associated with injected medroxy progesterone acetate [^18,24]. However, two of these studies [^9,23] had substantial (34-37%) losses to follow-up, and two studies [^23,24] gave no information on this. In one study [²³] the association was dependent on cervical ectopy, and in another [²⁴] there was no information about confounding factors. One was a pilot study [¹³] where control of confounding was probably incomplete and the authors chose not to draw conclusions. Of the five other prospective studies, three found a protective effect (OR, 0.2-0.9) in association with hormonal contraceptives, while two found modestly increased risks (OR, 1.3-2.0). There was a tendency for higher OR/rate ratios to be reported in higher risk women (prostitutes) and lower ratios to be reported in women attending family planning, antenatal or other health services.

Only three of the 22 cross-sectional and retrospective studies (Table 2) [^19,25,26] reported significantly increased risks after adjusting for confounding factors. The majority reported OR around 1.4 with CI including 1.0 (a ratio of 1.0 indicating no effect of hormonal contraception on HIV infection). As with the prospective studies, higher OR tend to be found in higher risk women (prostitutes and STD clinic attenders) and studies with more extensive adjustment for confounders [^10,14,15,18] tend to report lower OR.

Strengths and weaknesses of the available data

The methodological problems facing studies of this question are formidable. They include the temporal sequence between contraceptive use and HIV infection, confounding between sexual behaviour and choice of contraceptive and the related problem of choosing an appropriate control group, and lack of precise measures of level of HIV exposure or other risk factors. The first of these, that is, lack of information about the temporal sequence of contraceptive use and HIV infection, is a fundamental drawback in retrospective studies. Clearly, hypotheses about contraceptives affecting susceptibility to HIV infection can only be examined when contraceptive use is known to precede HIV infection. The problem of confounding applies to all observational studies in which women's choice of contraception is likely to be related independently to sexual behaviour and thereby risk of HIV infection. For example, if sexual risk-taking among women taking injectable contraceptives (such as injectable medroxy progesterone acetate) were higher than in women using other forms of contraception, any association between injectable medroxy progesterone acetate and HIV infection might simply reflect differences in sexual risk-taking and not a hormonal effect. Multivariate analyses attempt to control or adjust for such confounding factors, but adjustment is rarely complete when the confounders themselves are hard to measure. The degree to which a women is exposed to HIV is directly related to her risk of infection, but the concentration of HIV in her partner's ejaculate is hardly accessible to measurement. Some studies use proxy measures of the male partner's infectiousness such as low CD4 cell count (indicating advanced disease and presumably higher viral load), but in most studies, only remote measures such as number of sexual partners or frequency of sexual intercourse are available. Inability to measure strong risk factors precisely leads to residual confounding [²⁷] of the 'adjusted' estimate of the relationship between hormonal contraception and HIV infection. There is indirect evidence of residual confounding in this review. Studies in prostitutes, where levels of exposure to HIV and other STD are high but imprecisely measured, tended to produce higher OR. Conversely, studies in which adjustment for confounding was extensive (e.g., multiple variables used as proxy for HIV exposure) tended to give lower OR. The most reliable way to eliminate confounding between choice of contraceptive and sexual behaviour is by random allocation of women to one or other form of contraceptive, but this approach is unlikely to be ethically acceptable.

Role of meta-analysis

Meta-analysis has become an increasingly popular [^28-30] method of summarizing results from different studies of a specified problem or research question. The purpose is to produce a single, comprehensive 'best' estimate, usually with statistical confidence limits, that is meant to summarize all that is known about the problem. Meta-analysis has undoubtedly been useful in detecting relatively modest treatment effects from randomized trials with a precision that simply could not have been achieved in any one trial [³⁰]. However, the method has also attracted considerable criticism [^28,29,31,32], particularly when applied to observational studies rather than randomized trials [^29,31,32]. Declaring meta-analysis to be 'statistical alchemy for the 21st century', Feinstein [³²] argues that the way forward with observational research is to concentrate not on criteria for meta-analysis, but on criteria for scientific inference and the scientific acceptability of individual observational studies. His sentiment is clearly relevant to the systematic review presented here. Very few of the studies actually set out to investigate the relation between hormonal contraceptives and HIV infection. This is a far cry from having several randomized trials specifically designed to estimate the effect of an intervention. Whatever the type of study, the quality of the meta-analysis clearly depends on the quality of the component studies. It may be a statistically powerful technique, but it cannot overcome the problems of bias and confounding described above.

Summary of the epidemiological evidence

The clearest conclusion from this review is that the quality of the 32 studies, in terms of their ability to investigate the relationship between hormonal contraception and risk of HIV transmission, is mostly poor and inappropriate for statistical meta-analysis. Quantitative findings are inconsistent across studies. Sparse data from the methodologically strongest studies are compatible with both a 95% reduction and a 100% increase in the risk of seroconversion associated with oral contraceptive use. In studies reporting increased HIV risk with hormonal contraception, residual confounding may account for much of the association, because studies with more extensive statistical adjustment for other risk factors tend to report lower OR than studies with little or no adjustment.

Evidence from shedding studies

A separate but related question is whether hormonal contraception affects the infectiousness of HIV-infected women. Several studies in Africa have examined whether there is an association between hormonal contraceptive use and the degree of HIV shedding in the genital tract [^33-36]. The results of the different studies are surprisingly conflicting. Early studies in Nairobi reported a large positive association (OR, 11.6; 95% CI, 1.7-77.6) with detection of cervical, but not vaginal, HIV in women who had previously been enrolled in a partner study [³³], but a weak negative association in a similarly sized sample (n = 97) of local prostitutes [³⁴]. In the positive study [³³], only eight women were using oral contraceptives, and six of those had cervical ectopy, whereas the negative study included 65 oral contraceptive users, but only two women with cervical ectopy. The conflicting findings of these two studies could be reconciled if an association between hormonal contraception and HIV shedding were dependent on the presence of cervical ectopy, and there is some epidemiological evidence to support such a mechanism [³⁷]. Another study in Abidjan found no significant association (OR, 1.2; 95% CI, 0.6-2.7) amongst 32 oral contraceptive users of whom only five had cervical ectopy [³⁵]. However, more recent data from an STD clinic in Mombasa [³⁶] indicated a marked dose-dependent relationship between hormonal contraception and HIV shedding from both the cervix and vagina. The conflicting results of these studies are not readily explained, but they extend to other findings: for example, HIV detection was related to the presence of other STD in most, but not all studies [³⁴]. Many of the reported CI are wide, and much of the variation between studies may reflect relatively small study samples.

Evidence from animal models

In animal studies by Marx et al. [³⁸], 14 out of 18 macaques pretreated with subcutaneous progesterone pellets and then inoculated vaginally with simian immunodeficiency virus (SIV) during the follicular phase of the menstrual cycle became infected, compared with one out of 10 macaques pretreated with placebo pellets before inoculation. Progesterone treatment produced levels within or slightly above the range expected for macaques during the normal menstrual cycle. The eightfold increase in infection in the progesterone-treated animals was associated with marked thinning of the vaginal epithelium. In a separate experiment, progesterone treatment was associated with both thinning of the vaginal epithelium and a sixfold increase in SIV DNA-positive cells in the vaginal lamina propria. One clear hypothesis arising from these experiments is that progesterone reduces the efficacy of the vaginal barrier.

In considering the relevance of the findings to women, the authors suggested that the dose of SIV used was not excessive compared with the likely concentration of HIV in several millilitres of semen. Probably the most important criticism [³⁹] of the original macaque experiments is that the relative risk of infection associated with progesterone treatment was much lower (relative risk, 2.1; 95% CI, 0.9-4.9) if the macaques were exposed to SIV throughout the menstrual cycle rather than just during the follicular phase. These findings may be more relevant to women who risk exposure to HIV during most of the menstrual cycle.

The main concern following the animal studies has been about women taking contraceptives only containing progestin (synthetic progesterone). However, both oestrogen and progesterone have effects on the vaginal tract that could plausibly influence HIV transmission [^1,2], and many more women take the combined oral contraceptive pill. Besides thinning the vaginal epithelium and increasing vaginal pH, which might enhance HIV transmission, progesterone also increases the viscosity of cervical mucus, which may inhibit transmission. Oestrogen has opposite effects. Injectable progestogens induce an anovulatory state through their anti-oestrogenic actions. In a hypo-oestrogenic state, the vaginal mucosa may become more atrophic (as at menopause) and susceptible to mucosal tears during intercourse, creating a route for HIV infection [²]. Changes in bleeding pattern induced by progesterone-based or combined hormonal contraceptives could theoretically play a role. Hormonal influences on cervical ectropion or the mucosal immunology of the vagina and cervix might also be relevant.

Future research

The question of whether women taking oral or injectable hormonal contraceptives are at increased risk of acquiring HIV infection will be hard to answer. Statistical meta-analysis of existing observational studies cannot overcome their formidable methodological problems. Randomized controlled trials are not an option for ethical and practical reasons. Further retrospective studies are unlikely to resolve the problem of confounding between sexual risk-taking and choice of contraceptive method, or to estimate the level of HIV exposure before infection accurately enough.

What other avenues of research should be pursued to try to answer this question? The association between hormonal contraceptives and HIV infection could be examined by time [⁴⁰] and place [⁴¹], but ecological and time trend analyses based on aggregate, not individual, data are highly susceptible to confounding and may therefore be unable to settle the question. The strongest epidemiological study design, after randomized trials, would be prospective partner studies with more precise information about contraceptive use and risk factors for HIV transmission. Cohort studies of high-risk uninfected women have the important advantage of being able to record contraceptive use before HIV infection. But they would need to be large to accrue sufficient seroconversion end-points while advising consistent condom use, and will still have to tackle the problem of confounding. Moreover, it is important to consider these obstacles to further epidemiological studies in the light of other research priorities. Other potentially protective interventions (e.g., vaginal virucidal agents) are more amenable to evaluation and might therefore claim a higher research priority.

Further research is needed to improve understanding of the mechanisms of vaginal transmission of HIV [⁴²]. It is clear that infection can occur without breaching of the epithelial barrier [^2,43], but little is known about hormonal control of events. Future studies should investigate local defence mechanisms in the female genital tract and physiological changes during the menstrual cycle. Biologically plausible hypotheses about the effect of hormonal contraceptives on these mechanisms need further investigation. For example, the effect of progesterone-based contraceptives on vaginal wall thickness could be examined in women. Animal models could be used to investigate mechanisms by which viruses cross the vaginal epithelium, and hormonal effects on the mucosal immunology of the vagina and cervix.

Current recommendations

The oral contraceptive pill is a highly efficacious and safe form of contraception that has transformed the lives of many women world-wide. Its relationship to HIV transmission remains uncertain. Current approaches to promote hormonal contraception when it is appropriate to family planning needs should therefore continue. If further animal experiments were to present a more compelling case to indicate that hormonal contraception increases the risk of HIV transmission, women could be advised that such contraceptives do not protect against HIV and might increase susceptibility. But such advice would need to be weighed against the risks of rejecting safe and reliable contraception (including induced abortion). Clearly, the importance and effectiveness of consistent condom use against HIV transmission should be stressed whether other contraceptive methods are being used or not.

Acknowledgements

The author thanks R. French for help with the search strategy and B. Leynaert for providing data from the European Study of Heterosexual Transmission of HIV.