Response to Smith and Brogly Et Al. Commentaries on Zedler Et Al.

Important considerations should be taken into account when assessing the scientific literature related to harms, particularly in an especially challenging population such as pregnant women with opioid use disorder. However, simulation analyses to adjust for confounding bias should be conducted and interpreted with caution, especially when key parameter values are difficult to estimate or otherwise substantiate empirically with deserved rigor. We thank Dr Smith as well as Dr Brogly and colleagues for their thoughtful commentaries 1, 2 on our paper 3. Both experts highlight important considerations when assessing the scientific literature related to harms, particularly in an especially challenging population such as pregnant women with opioid use disorder. We agree with Dr Smith 1 that the newest extension to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines 4 contains several essential elements to improve harms reporting in systematic reviews. This extension was published after submission of our manuscript for peer review and should be considered in future similar work focused on assessing harms of health interventions, including but not limited to pregnant women receiving opioid agonist medication-assisted treatment. We also agree with Dr Brogly and colleagues 2 that accounting for unmeasured confounding is an important potential source of bias that should be taken into account in systematic reviews and meta-analyses. Further, we applaud the attempt of these commentators to adjust for confounding by indication via bias analysis simulation 5. When we applied the method outlined in their 2014 paper based on the work of VanderWeele & Arah 6, we observed similar results to theirs for the three outcomes examined in both papers. In short, adjusted treatment effect estimates for preterm birth, birth weight and head circumference were less marked than the unadjusted meta-analysis estimates, but generally continued to support better outcomes for buprenorphine-treated versus methadone-treated pregnancies. However, the bias simulation analysis adopted by Brogly et al. 5 is based on admittedly subjective and potentially severe assumptions for key parameter values. Perhaps chief among them is the estimate for the proportion of patients subject to bias by indication. Brogly et al. assume that 40% of buprenorphine-treated pregnant women are subject to bias by indication such that they are likely to experience better birth outcomes than pregnant women receiving methadone. The authors also assume that 40% of the methadone-treated women are similarly subject to bias by indication such that they are likely to have worse birth outcomes than those treated with buprenorphine. Brogly et al. note transparently that empirical evidence for the degree of bias is not available, and that the researchers believed that their proposed parameter values were plausible. When we conducted additional simulations using the other, similarly arbitrarily selected, bias estimates of 35 and 30%, the results followed analogous trends to the analysis in which the estimate of 40% was used. However, taking birth weight as an example (Fig. 1), if 30% were selected as a parameter estimate for the percentage of patients subject to bias by indication in each respective group, one might be less convinced that higher birth weights in neonates exposed to buprenorphine during pregnancy can be explained substantially by confounding by indication. Moreover, while Brogly et al. 5 claimed to use estimates for the effect of confounding based on published results, it is not clear specifically how the chosen ranges of values were actually derived. Again using birth weight as an example, Brogly et al. assumed that neonates subject to bias by indication will be 100–300 g heavier than all other neonates exposed to buprenorphine. They assumed further that neonates subject to bias by indication will be 100–300 g less heavy than all other neonates exposed to methadone. When the lower bounds of these assumptions are relaxed modestly to 50 g, respectively, again one might be less convinced that higher birth weights in neonates exposed to buprenorphine during pregnancy can be explained substantially by confounding by indication (Fig. 2). In summary, we endorse the valuable notion that unmeasured confounding is a considerable challenge in summarizing and interpreting the available data in this difficult population. However, we are concerned that the methodology of Brogly et al. represents little more than a guess, although perhaps reasonable, as to what impact confounding by indication may have in this context. Simulation analyses to assess bias such as that adopted by Brogly et al. are conceptually appealing and should continue to be pursued and developed. However, they should be conducted and interpreted with caution, especially when key parameter values are difficult or perhaps impossible to estimate or otherwise empirically substantiate with deserved rigor. At the time the work to which this commentary refers was conducted, B.K.Z., H.R.A., A.R.J., and E.L.M. were paid consultants of Venebio Group LLC, which has had research and consulting agreements with Indivior PLC. H.E.J. has no financial ties to either Indivior PLC or Venebio Group LLC, and did not receive any form of remuneration in the preparation or writing of the paper to which this commentary refers. Partial funding for the review referred to in this commentary was provided by Indivior PLC (formerly Reckitt Benckiser Pharmaceuticals) to Venebio Group LLC. The review was conceived, designed, executed and reported by the authors, who had sole control over the literature selected, data analysis, interpretation and manuscript preparation. Indivior PLC was asked to review the final manuscript for proprietary information. The opinions and conclusions of the authors are their own and do not necessarily reflect the position of Indivior.