Some data suggest that babies born to women in Australian public hospitals, largely cared for by midwives either alone or jointly with the woman's general practitioner, tend to die more often than those born in private hospitals where intrapartum care is provided by obstetricians. To test this finding with refined statistical methods, a study by Adams et al. (BJOG 2018;125:149–58) contrasts perinatal mortality rates among women that delivered in public versus private obstetrician-led hospitals in Brisbane, Australia. Their study included 69 037 public and 62 399 private hospital births. The authors applied a propensity score (PS) matched method to minimise bias due to measured confounders, and reported that the risk of perinatal mortality was substantially higher among births in public than in private hospitals. This increased risk, although attenuated, was still evident following adjustments for major congenital anomalies, birth method, and gestational age. Propensity score methodology enables an observational study to be converted to a ‘pseudo-randomised’ trial that is free of measured confounding (Rosenbaum & Rubin. Biometrika 1983;70:41–55; Austin. Multivariate Behavior Res 2011;46: 399–424). However, PS analysis does not account for unmeasured confounding, an important source of bias in observational epidemiology that results from variables that may be associated with both the exposure and the outcome, but which were not measured in the study. In randomised trials, when treatment allocation is done properly (i.e. via randomisation), the treatment groups are virtually balanced with respect to both measured and unmeasured confounders. In comparison, PS analysis guarantees the former, but not the latter, and unmeasured confounding still remains a serious concern. An elegant and simple-to-use method to assess the extent to which unmeasured confounding may play a role in the association between hospital type and risk of perinatal mortality, is the ‘E-value’ (VanderWeele & Ding. Ann Intern Med 2017;167:268–74). This measure provides an estimate of how large an unmeasured confounder (of both the exposure-unmeasured confounder, and the unmeasured confounder-outcome scenarios of the exposure-outcome association) must be to render the observed odds ratio (OR) a null estimate, as well as on the lower limit of 95% confidence interval (CI) estimate to include the null value of the observed OR. In the study by Adams et al., the observed OR for the association between deliveries in public hospitals (versus those in private hospitals) and perinatal mortality was 1.53 (95% CI 1.29–1.80). The corresponding E-values for the OR and the lower limit of the 95% CI were 2.43 and 1.90, respectively (see VanderWeele & Ding. Ann Intern Med 2017;167:268–74 for details). This suggests that in order for the observed OR to be reduced to null, and for the lower 95% CI estimate to cross null, the corresponding ORs for unmeasured confounding must be at least 2.43 and 1.90, respectively. Given the magnitude of these ORs, the association of increased perinatal mortality in Australian public hospitals for deliveries cared for by midwives either alone or jointly with the woman's general practitioner by Adams and colleagues is less likely to be the result of unmeasured confounding. None declared. Completed disclosure of interests form available to view online as supporting information. Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.
Read full abstract