Human papillomavirus infection and bladder cancer: an alternate perspective from a modified meta-analysis.

TO THE EDITOR—We read with interest the recent meta-analysis by Li et al [1] regarding the association between human papillomavirus (HPV) infection and bladder cancer incidence. The authors reported a summary odds ratio (OR) of 2.8 (95% confidence interval [CI], 1.4–5.8) for the relation between any type of HPV infection and bladder cancer on the basis of data from 17 casecontrol studies and interpreted the finding as a “clear link between bladder cancer and HPV infection” [1]. We are skeptical of basing such a definitive conclusion on the reported summary estimate because of the potential synergistic consequences of small samples and sparse-data bias among the individual case-control studies used in the metaanalysis. The consistently large variances (indicated by large CI ratios [2]) among the individual case-control studies used in the meta-analysis are evident of the small sample sizes. Additionally, the prevalence of HPV infections among cases and controls was low in most studies. Low exposure prevalence likely resulted in small numbers within strata (ie, sparse-data bias), which is well documented to result in an upward bias of the ORs [3, 4], and synergism between small samples and sparse data could result in implausibly high estimates [5]. Consequently, the summary estimate reported by Li et al [1] may be influenced by the summarized bias among individual studies rather than an effect of HPV infection on bladder cancer incidence. Given these concerns, we reanalyzed the meta-analysis by Li et al [1] using a semi-Bayesian approach to evaluate the sensitivity of the individual and metaanalytic estimates to sparse-data bias. In brief, we abstracted ORs and CIs for the 17 case-controls studies reported by Li et al in Figure 1. of their article [1]. We adjusted the study-specific estimates by using a semi-Bayesian method [6] that combined the study-specific OR and standard error with a null-centered prior data stratum, which assumed a variance of 0.5 for ln(ORprior) and corresponds to a 95% CI of 0.25–4.0 for ORprior. Prior limits are intended to reduce the magnitude of sparse-data bias by placing statistical constraints on the expected study-specific estimates [4, 6]. The prior limits we used were informed by evidence that ORs for the well-documented relation between HPV infection and nongenital cancers, such as oral cancers, rarely exceed 4.0 [7]. The resulting posterior ORs and 95% posterior intervals (PIs) from each study were subsequently summarized in fixed-effect and random-effects models, using standard inverse-variance weighting [8] instead of sample-size weighting used in the analysis by Li et al [1]. Finally, we