Propensity Score Approaches In The Presence Of Missing Data: Comparison Of Balance And Treatment Effect Estimates

The use of propensity score methods can potentially improve balance between groups in observational study data, thus minimising confounding. However, a frequent problem with such studies is the presence of missing data. We compare three approaches to generating the propensity score accounting for missing data within the context of a clinical case study examining the effect of telemonitoring on hypertension. Overall, 4,642 patients diagnosed with hypertension receiving online health support in My Health Guardian – a telephone chronic disease support program offered by private health insurer HCF – were offered a telemonitoring intervention. Of these, 2,729 accepted and started treatment between July 2014 to April 2015 (designated “cases”), and 1,913 declined (designated “controls”). Data were available from cases and controls on several baseline variables including demographic, lifestyle and clinical characteristics. Outcomes were the number of hospitalisations, total length of stay and total cost of hospitalisation between 1 January to 31 December 2016. Propensity score methods were used to balance baseline variables between groups. Three approaches were used to generate propensity scores from a logistic regression model accounting for missing data: 1) categorisation of all variables with “Missing” as a category, 2) multiple imputation of treatment effect (MIte) where treatment effect estimates are combined over 20 imputed datasets, and 3) multiple imputation of propensity score (MIps) where propensity scores are first averaged over imputed datasets prior to estimation of treatment effects. The propensity score from each approach was then used in two ways: a) matching cases to controls, and b) inverse probability of treatment weighting (IPTW). The balance achieved by each approach was compared using standardised differences of means and proportions in baseline characteristics between groups. The treatment effect estimates from each approach were also compared. The discussion will canvas our findings and recommendations for handling missing data when using propensity score approaches.