We recorded the method and timing of BV diagnosis. In studies with multiple methods of determining BV, we collected the results for each method. The gram stain/wet mount result was preferentially used in the final analysis, as this is the most relevant technique for clinical assessment of BV. Vaginal culture data could only be correlated to outcome by individual microbes, so we limited data recording to Gardnerella vaginalis only. We recorded the timing of the diagnosis of BV as presented (ie, weeks of gestational age, trimester, during labor), then grouped the data by trimester. If more than one sample was collected in an individual study, we recorded all results but used the earliest in the analysis.
We recorded outcome data dichotomously in 2 x 2 tables for all 4 study outcomes. Extracted information on the handling of confounding included the method used, the adjusted odds ratio with confidence interval, and the variables included in the final model.
In studies that reported preterm delivery results at multiple gestational age cutoffs, we used only the 37-week cutoff. If studies used a definition other than 37 weeks, their data were noted for subanalysis. Some case control studies defined cases as women in preterm labor, but reported data separately from those cases that delivered prematurely from those that delivered after 37 weeks. For the preterm delivery outcome, the preterm labor cases who delivered at term were included in the control group.
Validity Assessment
We used validity assessment worksheets that were developed specifically for this project using a summary of previously published criteria.12-14 Each included study was critically appraised independently by the 2 investigators, and their assessments were compared. A third investigator (LM) settled disagreements. We did not use validity criteria to exclude any study from analysis.
Data Synthesis
We calculated summary estimates of risk as odds ratios using both the fixed and random effects models.15-17 Additionally, we combined cohort studies to generate summary relative risk estimates using both models. Precision is reported as 95% confidence intervals for each statistic. These calculations were generated using Review Manager 3.0 software.18 We evaluated homogeneity using the chi-square statistic:15 the greater the P value, the more homogeneous the studies.
We conducted subanalyses by study design, baseline population risk of prematurity, method and timing of BV diagnosis, and country of study population. When the subgroupings resulted in any category having fewer than 3 studies, we did not calculate a summary statistic for that group. We excluded those studies and generated a new pooled risk assessment for the alternate groups only.
To address the issue of confounders, we used the general variance-based model15,19,20 to combine the adjusted odds ratios of individual studies into a summary odds ratio with its 95% confidence interval.
Results
Data Sources and Study Selection
Our literature review identified 233 studies; no unpublished data were discovered. Reviewing the abstracts identified 39 studies for possible inclusion (27 observational studies and 12 trials). We excluded 11 observational studies because they had an inadequate or no control group (2),21,22 no vaginal assessment of BV (2),23,24 repeated data (1),25 or the inability to link BV with pregnancy outcomes (6).9,26-30 Nine of the 12 trials did not present their control group cohort data in a way that distinguished the outcomes by the presence or absence of BV and were therefore excluded.31-39 We included 19 studies10,11,40-56 in the final analysis: 8 case control trials and 11 cohort studies. Three from this latter group consisted of the placebo group of randomized controlled trials.
We estimated the likelihood of publication bias by generating a funnel plot.15 The graph of study size versus the logarithm of the ratio results, although funnel-shaped, is not completely symmetric.* Data from small studies demonstrating a protective effect of BV in pregnancy are missing.
Data Extraction
Twelve of the 19 included studies drew patients from a university or tertiary care hospital setting; the 7 others were clinic based Table 1. Two studies were conducted in nonindustrialized countries (Nairobi and Indonesia). The remaining studies were performed in the United States (11), Australia (3), the United Kingdom (1), Sweden (1), and Finland (1). The baseline prevalence of preterm delivery ranged from 1.1% to 64.9%, with a mean of 20.0% and a median of 13.8%.
To diagnose bacterial vaginosis, 13 studies used gram stain alone or in combination with a wet mount, 2 used gas-liquid chromatography, and 4 used vaginal swab cultures. The timing of risk factor detection varied among the studies, ranging from the first prenatal visit to the time of labor. We grouped the studies by trimester as precisely as possible, resulting in 10 studies diagnosing BV in the first or second trimester, 4 with second or third trimester assessment, and 4 studies discovering BV in the third trimester only (which in most cases was at the time of labor).