ABSTRACT
Objective@#To introduce the methods for sensitivity analysis, discuss and compare the advantages and disadvantages of different methods.@*Methods@#The difference between confounding function method and bounding factor method in accuracy of identifying unmeasured confounding factors in observational studies through simulation trials and actual clinical data was compared.@*Results@#The results of simulation trials and actual clinical data showed that when there was unmeasured confounding between exposure (X) and outcome (Y), the results of confounding function and the bounding factor analysis were similar in terms of the effect of unmeasured confounding factor to lead to the complete change of the magnitude and direction of the observed effect value. However, the confounding function method needed smaller confounding effect to fully interpret the observed effect value than the bounding factor needed. In addition, the bounding factor method needed to analyze two confounding parameters, while only one parameter was needed in the confounding function method. The confounding function method was simpler and more sensitive than the bounding factor method.@*Conclusion@#For real-world observational data, the sensitivity analysis process is essential in analyzing the causal effects between exposure (X) and outcome (Y). In terms of the calculation process and result interpretation the sensitivity analysis method of confounding function is worth to recommend.
ABSTRACT
Objective@#This project aimed to explore the effectiveness of estimating individual treatment effect on real data, among the heterogeneous population, with Causal Forests (CF) method, to find out the characteristics of heterogeneous population.@*Methods@#We designed and conducted four computer simulation schemes to verify the effect of estimating on individual treatment, using the CF under four different environments of the treatment effects. Real data was then analyzed for the catheterization on right heart.@*Results@#Results from the simulation process showed that the values on individual treatment effect that were estimated by causal forests were consistent with the population effect as well as in line with the expected distribution under the setting of four different effect values. Results of real data analysis showed that values of individual treatment effect among most patients appeared positive, so the use of RHC could cause an increase of the '180-day mortality rate’ in the sampled population. Patients with lower predicted probability of 2-mo survival and albumin were more likely to have a lower risk of death after using the RHC.@*Conclusion@#CF method could be effectively used to estimate the individual treatment effect and helping the individuals to make decision on the receipt of treatment.