Cheryl A. Flynn, MD, MS Amy L. Helwig, MD Linda N. Meurer, MD, MPH Syracuse, New York, and Milwaukee, Wisconsin From the Center for Evidence Based Practice, Department of Medicine, State University of New York, Health Science Center at Syracuse (C.A.F.), and the Department of Family and Community Medicine, Medical College of Wisconsin, Milwaukee (A.L.H., L.N.M.). Reprint requests should be addressed to Cheryl Flynn, MD, MS, CEBP, SUNY HSC at Syracuse, 475 Irving Avenue, Suite 200, Syracuse, NY 13210. E-mail: flynnc@mailbox.hscsyr.edu.
References
Finally, BV remained a significant risk for preterm delivery regardless of the subanalysis groupings. Pooling data from different populations with variable baseline risk or in different settings may lend confidence to the generalizability of these estimates. Caution is warranted, however, when drawing conclusions from any specific subanalysis.
All studies included in this analysis were observational in design, which raises 2 particular concerns: causality and confounding. Because neither the Mantel-Haenszel or DerSimonian and Laird methods for pooling risk estimates can incorporate confounding, we calculated separately the adjusted odds ratio pooled from risk estimates generated by regression analyses in individual studies. This summary of adjusted odds ratios still demonstrated a statistically significant elevated risk of 1.6 for preterm delivery in women with BV. This value may be overestimated, because 4 studies did not report regression results for the variable of BV. Presumably, these values were near the null but were likely not statistically significant, or they would have been reported. A lower summary adjusted odds ratio may have resulted were inclusion of these results possible.
Although causality cannot be proven by observational studies or by meta-analytic combination of such studies, several of the criteria suggesting causality are met.57,58 The strength of the association is relatively small, ranging from 1.4 to 2.4. Although it is possible that confounders account for all this association, we think that is unlikely given that the controlled summary odds ratio remained significant. This meta-analysis did not address the dose response question directly, although some individual studies reported a stronger association with prematurity outcomes for those who had heavier colonization or higher BV scores. In all but one study this risk factor preceded the outcome, although in some cases vaginal assessment was done at the time of labor. Results are consistent, as can be seen from the summary figures: No study found a protective effect, and only one had a null value. The presence of BV makes biological sense as a contributor to preterm labor and thus to preterm premature rupture of membranes and preterm delivery. One proposed mechanism for this association is the production of phospholipase by the bacteria associated with BV. These enzymes can initiate prostaglandin synthesis, which is one step in the physiology of normal labor activation.53 Infections in pregnancy are also associated with fetuses that were small for the gestational age,59 which is another mechanism for low birth weight. Taken together, these factors lend support to a causal association for BV and prematurity.
Identifying BV in pregnancy as a modifiable risk factor for prematurity raises the obvious question of intervention. Accumulating evidence demonstrates that treating pregnant women who have BV with certain oral antibiotics can decrease the risk of prematurity. Clindamycin taken orally by pregnant women with BV decreased preterm deliveries and low birth weight infants by approximately 50%.38,55 Hauth42 combined oral erythromycin and metronidazole and found a decreased rate in preterm births among a group of high-risk women with BV. Neither oral amoxicillin31 or intravaginal clindamycin33 have been shown to affect pregnancy outcomes.
Limitations
As with any meta-analysis, one major limitation of this work is the appropriateness of combining results from different studies. Statistical homogeneity was met in 1 of the 4 analyses for summary odds ratios (pooling all studies possible), but in 3 of the 4 analyses when only cohort studies were pooled. This discrepancy suggests that study design was likely a key source of heterogeneity in this review. In addition to study design differences, we expected heterogeneity given the range of risks reported in individual studies, the varying population risks, the disparate methods, the timing of BV diagnosis, and the different definitions for the measured outcomes. One method for addressing heterogeneity is to use the random effects model, which accounts for variability between studies when estimating the precision of the risk. When we analyzed the data using this method, none of our conclusions changed.
A second limitation of this project is the possibility of publication bias. Our funnel plot reflects the absence of studies finding that BV protects pregnant women from delivering preterm infants. By chance alone, some studies may find this result, but it is unlikely that such a study would be published. Using Orwin’s method60 to calculate a fail-safe N of 60 and our weighted summary effect size of 0.052 (number needed to harm [NNH] = 19), more than 75 studies showing no effect would be needed to drop the risk difference to 0.01 (NNH = 100). Given our systematic and complete search, we think it unlikely that publication bias accounts for our findings, despite the asymmetry of the funnel plot.
