Bacterial vaginosis (BV) is a common vaginal condition with an estimated prevalence of 29% among U.S. women during 2001–2004 [1]. BV has been linked to a range of adverse reproductive outcomes, including infertility, spontaneous abortion, preterm premature rupture of the membranes, amniotic fluid infection, low birth weight, and preterm delivery [2–8]. BV also might increase women's risk for pelvic inflammatory disease although evidence on this possible association is inconsistent [9–11]. In addition, evidence suggests that BV increases women's risk for sexually transmitted infections (STIs), including gonorrhea, chlamydial infection, trichomoniasis, human papillomavirus, herpes simplex virus, and HIV [12–19].

Although the etiology of BV remains unknown, two competing hypotheses currently prevail [20–22]. In the first, BV is viewed as an imbalance of the vaginal microbiota caused by the colonization of endogenous organisms from the intestinal tract [23]. This imbalance could be precipitated by a variety of events, including coitus and vaginal cleansing or douching. The second hypothesis holds that BV is caused by the sexual transmission of a specific pathogen (e.g., Gardnerella vaginalis or unknown bacteria). The similarity between the epidemiology of BV and that of STIs supports the hypothesis that BV is sexually transmitted. For example, BV has been associated with risky sexual behaviors, including having new or a relatively high number of sexual partners, having sex frequently, not using condoms, using drugs during sex, and having sex with uncircumcised partners [1, 16, 22, 24, 25]. However, because these associations often have been found in observational studies, they could be the result of uncontrolled confounding. Use of a case-crossover analysis (in which each woman serves as her own control) would minimize the effects of time-independent confounders [25]. We conducted both cohort and case-crossover analyses to identify time-variant correlates of BV among a cohort of high-risk women in the U.S., who participated in a longitudinal study of the effects of HIV infection on women's health [26].

We analyzed data from the HIV Epidemiology Research Study (HERS), which was conducted at 4 U.S. sites (Bronx, NY; Detroit, MI; Baltimore, MD; and Providence, RI, USA) in 1993–2000 [26]. Participants consisted of 871 HIV-infected women and 439 uninfected women who, at the time they enrolled in the study, were 16–55 years of age, did not have an AIDS-defining clinical diagnosis, and either injected drugs or engaged in high-risk sexual behaviors (i.e., had >5 sexual partners in the previous 5 years, traded sex for money or drugs, or had sex with a male who injected drugs or who was suspected of being or known to be infected with HIV). After enrollment, participants completed follow-up visits scheduled at 6-month intervals. During these visits, HERS staff conducted interviews to collect demographic, health, and behavioral information, conducted physical examinations, and collected specimens to be tested for infections, including BV, HIV, human papillomavirus (HPV), and trichomoniasis. Study visits were not used to diagnose or treat symptoms, and less than 1% of participants reported using metronidazole or topical clindamycin [27]. Ethical review boards at the study sites and the Centers for Disease Control and Prevention approved the study, and only women who gave informed consent were enrolled.

Gram-stained slides prepared from swabs of posterior vaginal fornix specimens were air dried, fixed in methanol, and shipped to a central laboratory where a single technician used oil immersion with ×1000 magnification to quantify and score the specimens. Specimens with a Nugent score of 7–10 were considered positive for BV [28]. Gram stains also were evaluated for morphological identification of spermatozoa, which is specific for recent exposure to semen [29]. Spermatozoa usually clear from vaginal secretions by 12–36 hours after exposure to semen although they have been detected microscopically up to 10 days after exposure [30, 31]. Wet mount was used for diagnosing trichomoniasis, and vaginal specimens were cultured for Candida organisms. Aliquots of cervicovaginal lavage fluid were frozen for later testing for HPV by polymerase chain reaction.

We limited our analyses to data collected during participants' first 10 follow-up visits and excluded the 12 women who HIV seroconverted during the study. Participants' incident BV status was assessed at follow-up visits only if Nugent scores of samples collected at their preceding visit indicated that they were BV negative. If they tested positive for BV or their Nugent scores were missing, their incident BV status was coded as missing. We used unconditional (using generalized estimating equations to account for intrasubject correlation from multiple visits) and conditional logistic regression to analyze the data as if they were derived from a cohort and case-crossover study, respectively. For both analyses, we constructed individual models to evaluate the correlates of incident BV for HIV-infected and -uninfected women separately. While the analytic population for the cohort analysis included all follow-up visits with nonmissing data on incident BV, the case-crossover analysis was limited to follow-up visits from women who had ≥1 follow-up visit with and ≥1 follow-up visit without incident BV.

For both the cohort and case-crossover analyses, we fitted individual models to assess the bivariable relationship between each potential correlate and incident BV. For the multivariable analyses, we fitted full models with all potential correlates and used manual, backward elimination to exclude factors that were not significantly associated (based on an alpha of 0.05) with incident BV. Potential correlates were selected because of their prior identification in the literature. The cohort analyses included both time-independent and -dependent variables. However, because individual participants in the case-crossover analyses served both as case subjects and matching control subjects, the variables evaluated in these analyses were limited to time-dependent factors, which had the potential to vary between the participant's visits.

Because of the differences in findings by HIV status in both the cohort and case-crossover analyses, we present results separately for HIV-infected and -uninfected participants. The cohort analyses were based on data collected during 3,050 visits by 799 HIV-infected women and 1,564 visits by 375 uninfected women. The case-crossover analyses were based on data collected during 1,543 visits by 332 HIV-infected women and 753 visits by 159 uninfected women. The incidence of BV during the study follow-up period was 21% among HIV-infected women and 19% among uninfected women.

The sole correlate of incident BV that emerged in both the cohort and case-crossover analyses among HIV-infected and -uninfected women was the detection of spermatozoa on Gram stain, which is a biological marker of recent exposure to semen. The cohort analyses among HIV-infected and -uninfected women also found that incident BV was more common among those reporting frequent coitus (regardless of the consistency of condom use); however, this association was not found in the case-crossover analyses. A protective effect of condoms against BV has been demonstrated in some prior studies (including a case-crossover analysis), but not in all studies [22, 32–34]. The failure to find a relationship between condom use and decreased BV risk in earlier studies could have been the result of misclassification in participant reporting of coitus and condom use. This misclassification could have occurred if studies collected inaccurate reports of condom use, as a result of social desirability or recall bias, or did not collect comprehensive data on condom use, including possible malfunctions or misuse. A protective effect of condoms against BV also could have been obscured in previous studies because of the role of recurring cases of BV. That is, if unprotected coitus can cause incident BV but is not a necessary component for its recurrence, establishing the link between unprotected coitus, and incident BV could be difficult.

We found fewer correlates of incident BV in our case-crossover analyses than in our cohort analyses. Results of the adjusted case-crossover analyses of incident BV among HIV-infected women showed only spermatozoa detection, current injection drug use, and crack use within the previous 6 months to be associated with incident BV, whereas results of the cohort analyses among HIV-infected women also showed trichomoniasis at the previous visit, HPV at the previous visit and coitus ≥4 times per month during the previous 6 months to be associated with incident BV. Similarly, results of the case-crossover analyses of risk among HIV-uninfected women only found spermatozoa detection and coitus ≥4 times per month during the previous 6 months to be associated with incident BV, whereas results of the cohort analyses among HIV-uninfected women also showed trichomoniasis at the previous visit, current hormonal contraception use, and cigarette use within the previous 6 months to be associated with incident BV. The case-crossover analyses might have identified fewer correlates of incident BV as a result of reduced confounding from each woman serving as both a case subject (visits with incident BV) and a matching control subject (visits without incident BV) [25]. Alternatively, reduced power in the case-crossover analyses might have prevented the detection of correlates of BV.

A major study limitation was that BV was only assessed at six-month intervals. Studies with frequent sampling have suggested that women may have rapid fluctuations in vaginal microbiota, including short episodes of BV that resolve spontaneously [35, 36]. Thus, our study might have missed cases of BV. We were also unable to determine the temporal relationship between exposure to semen and the development of BV; as a result, we cannot rule out the possibility that the association between the two factors is the result of BV causing longer persistence of spermatozoa in vaginal fluid rather than semen exposure actually causing BV. Previous case-crossover analyses also suggest that recent menses, use of vaginal lubricants, rectal sex, douching for cleansing after menstruation, and psychosocial stress could be risk factors for incident BV [36–38]. None of these factors, though, were evaluated in the present analysis. Finally, although the detection of spermatozoa is specific for recent exposure to semen, it is not a sensitive marker and cases of exposure might have been missed [29, 39]. Strengths of our study included our use of data from a large, prospective study in which semen exposure was assessed by an objective measure and our use of case-crossover analyses, which allowed us to reduce possible effects of unmeasured time-independent confounding by comparing women to their own control visits.

The detection of BV among women who have reported being sexually abstinent has been an argument against the role of sexual activity as a necessary component in causing BV [40, 41]. However, results of a recent study among young adults with a laboratory-diagnosed case of chlamydial infection, gonorrhea, or trichomoniasis showed that 10% reported having abstained from penile-vaginal intercourse in the previous year and that 6% reported never having had intercourse [42]. Thus, imperfect validity of self reports could explain the occurrence of BV among women reporting abstinence in prior studies. While study findings implicate the role of sexual exposure in the development of incident BV, this does not necessarily mean that BV is caused by the transmission of specific organism(s) during intercourse. Semen exposure could also increase women's risk for incident BV by increasing vaginal pH levels, changing the growth patterns in bacteria populations, or exposing women to an unidentified component of semen. The present study found biological evidence of an association between semen exposure and incident BV, which provides new support for the sexual transmission of BV; however, the mechanism remains unknown.