Evaluation of different missing data strategies in propensity score analyses

Johanna H.M. Driessen, Elizabeth J. Williamson, James R. Carpenter, Frank De Vries

Research output: Contribution to journalMeeting AbstractAcademic

Abstract

Background: Propensity scores are used to adjust for confounding. However, data on confounders is often not fully available. Different techniques are available to handle missing data, but there is little data available about the relative performance of missing data methods within the context of propensity score analyses. Observational studies have shown decreased risks of fracture with use of statins while large randomized clinical trials (RCTs) found a relative risk of 1.0. BMI is an important confounder with often incomplete data and this might explain the differences between the results of the RCTs and the observational studies. Objectives: To explore the sensitivity of estimated treatments effects to the missing data approach used. Methods: A retrospective cohort study using data from the UK Clinical Practice Research Datalink (CPRD) (1992-2014) was conducted. Statin users, aged 50 years or older, having at least one statin prescription since 1992 were selected and matched 1:1 by year of birth, sex and practice to non-users. Cox regression models were used to estimate the hazard ratios (HR's) of hip fracture in statin users versus nonstatin users. Missing data were handled by complete case (CC) analysis, adding an indicator (IND) and multiple imputation (MI). Adjustments by propensity scores (inverse probability weighting) were compared to adjustments including confounders in the regression model. Results: The confounder adjusted methods showed all a decreased risk of hip fracture for statin users as compared to non-users (CC: HR 0.92 95% CI (0.85 - 0.99); IND: HR 0.92 (0.86 - 0.99); MI: HR 0.92 (0.86 - 0.99)). Propensity score adjusted models showed no association in the CC analysis (HR 0.98 95% CI: 0.93 - 1.02), whereas the IND and MI analysis showed a decreased risk of hip fracture with statin use (IND: HR 0.94 (0.91 - 0.98); MI: HR 0.95 (0.91 - 0.99)). Conclusions: The point estimates of the three different missing data techniques did not differ much, suggesting that the different techniques used in the present study did not greatly influence the estimated treatment effect.
Original languageEnglish
Pages (from-to)154-155
Number of pages2
JournalPharmacoepidemiology and Drug Safety
Volume25
DOIs
Publication statusPublished - 1 Aug 2016
Event32nd International conference on Pharmacoepidemiology & Therapeutic Risk Management - The convention centre Dublin, Dublin, Ireland
Duration: 25 Aug 201628 Aug 2016

Keywords

  • hydroxymethylglutaryl coenzyme A reductase inhibitor
  • adult
  • adverse drug reaction
  • body mass
  • case study
  • clinical practice
  • clinical trial
  • cohort analysis
  • controlled clinical trial
  • controlled study
  • disease model
  • hazard ratio
  • hip fracture
  • human
  • middle aged
  • observational study
  • prescription
  • probability
  • propensity score
  • proportional hazards model
  • randomized controlled trial
  • risk factor
  • side effect
  • statistical model

Fingerprint

Dive into the research topics of 'Evaluation of different missing data strategies in propensity score analyses'. Together they form a unique fingerprint.

Cite this