Sensitivity analysis: Probabilistic bias analysis
Chronic hypertension is subject to misclassification in these data files, with sensitivity of 44% (and agreement of 98%) in comparison to data abstracted from hospital medical charts.(28) To address potential exposure misclassification and unmeasured confounding biases, we undertook a sensitivity analysis through a probabilistic bias analysis.(37, 38) Exposure (chronic hypertension) misclassification was assumed to be differential with respect to the outcome (perinatal mortality). Based on a uniform distribution, we assumed the priors for sensitivity for chronic hypertension to range between 0.30 and 0.95 among those with perinatal deaths and 0.20 to 0.90 among live births; the priors for specificity for chronic hypertension was assumed to range between 0.98 and 1.00 both for deaths and live births, respectively.
Corrections for unmeasured confounding bias was based on the following assumptions: (i) we assumed that the prevalence of the unmeasured confounder among those with and without chronic hypertension, under a log-normal distribution, ranged between 5% and 15%, and 3% and 10%, respectively; and (ii) the RR for the confounder-outcome association was varied between 1.25 and 3.00. Under these assumptions, we drew the bias parameters 100,000 times from the prior distributions to address exposure misclassification and unmeasured confounding (computational strategy are provided in the R package “episenr”.(39)) From these analyses, we report the median bias-corrected RR (RRbc) and 95% CIbc.
Log-linear regression models and the mediation analysis were fit in SAS (version 9.4; SAS Institute, Cary, NC) using the GENMOD and the CAUSALMED procedures, respectively. The probabilistic bias analysis was implemented in R (R Foundation for Statistical Computing, Vienna, Austria) using the episensr package.(39)