Abstract:
|
We compared commonly used propensity score methods that consider covariate balance: generalized boosting models (GBM) in which the number of trees optimizes covariate balance and covariate balancing propensity score (CBPS), where covariate balance is directly optimized by solving estimating equations. We considered four scenarios differing by the complexity of a propensity score model and a range of exposure prevalence. Propensity score weights were estimated using four methods: the maximum likelihood (ML), CBPS of logistic regression, GBM and generalized additive logistic regression models (GAM). We used these propensity weights to obtain the estimates for the average treatment effect on a binary outcome. We evaluated these four propensity score methods in terms of power, bias-adjusted power, bias, root mean squared error (RMSE), 95% confidence interval coverage and average standardized absolute mean difference. Our results suggest that in the scenarios of complex treatment assignment models the CBPS generally is more powerful and precise than the other methods and that with large sample sizes GAM and GBM potentially can be more accurate in terms of RMSE than both ML and CBPS.
|