Reinvestigation of causal effects of Right Heart catheterization: a matching approach

Open Access
Yang, Qiong
Graduate Program:
Master of Science
Document Type:
Master Thesis
Date of Defense:
Committee Members:
  • Debashis Ghosh, Thesis Advisor
  • Aleksandra B Slavkovic, Thesis Advisor
  • Spiro E Stefanou, Thesis Advisor
  • Right Heart Catheterization
  • Causal Inference
  • Matching
  • Sensitivity Analysis
Right Heart Catheterization (RHC) is a common procedure applied in critically ill patients. In US, over 1 million cases of RHC procedures are conducted annually at present. In particular, patients with low blood pressure, lung water, hearts abnormalities, kidney abnormalities are usually the targeted group to receive the procedure. Although there is no absolute clinical contradiction with the use of RHC, its effect has not been statistically validated using randomized controlled trials due to the lack of randomized data. Instead, studies on RHC have been using observational data to quantify its causal effect, which should be taken with caution because of the issues in estimating true causal effect in observational studies. In my study, I reinvestigate the causal effect of RHC on subsequent survival using observational data and several multivariate matching schemes including nearest neighbor matching, optimal matching, full matching and genetic matching. Data for the study is from the Study to Understand Prognoses and Preferences for Outcomes and Risks of Treatments (SUPPORT), with a total of 5735 subjects. Our study has three main findings. First, RHC increased the risk of dth30, after controlling for the observed confounders under each of the four matching schemes, which coincides with earlier studies using SUPPORT data. Second, the Monte Carlo simulation experiment conducted in our analysis suggests that our conclusion on the causal effect of RHC is robust because: 1) The bias, defined by the difference between true and estimated causal effect, is small with acceptable variance; 2) Our simulation results are robust to a variety of specifications on additivity and/or linearity in the relationship between confounders and the treatment indicator. However, for the case where strong nonlinearity is present, the model performance is not as good as in other scenarios. Third, this negative causal effect is sensitive to the possible hidden bias identified by sensitivity analysis. However, the performances of these four matching schemes are different confronting hidden bias. Full matching gives the most robust result compared to other three matching schemes.