Analyzing Missing Data in Perinatal Pharmacoepidemiology Research: Methodological Considerations to Limit the Risk of Bias

Abstract

Pharmacoepidemiological studies on the safety of medication during pregnancy are all susceptible to missing data (ie, data that should have been recorded but for some reason were not). Missing data are ubiquitous, irrespective of the data source used. Bias can arise when incomplete data on confounders, outcome measures, pregnancy duration, or even cohort selection criteria are used to estimate prenatal exposure effects that would be obtained from the fully observed data, if these were available for each mother–child dyad. This commentary describes general missing data mechanisms and methods, and illustrates how missing data were handled in recent medication in pregnancy research, according to the utilized data source. We further present one applied example on missing data analysis within MoBa (the Norwegian Mother, Father and Child Cohort Study), and finally illustrate how the causal diagram framework can be helpful in assessing risk of bias due to missing data in perinatal pharmacoepidemiology research. We recommend that applied researchers limit missing data during data collection, carefully diagnose missingness, apply strategies for missing data mitigation under different assumptions, and finally include evaluations of robustness results under these assumptions. Following this set of recommendations can aid future perinatal pharmacoepidemiology research in avoiding the problems that result from failure to consider this important source of bias.