|
Abstract:
|
The G-computation algorithm is a general method to estimate causal effects. For a fixed-time binary treatment, it can be derived from the Law of Total Expectation. The G-computation algorithm requires specification of only the conditional distribution of the outcome given the treatment and confounders, also known as the Q-model. Simple parametric Q-models for dichotomous outcomes have been used in papers demonstrating G-computation that include the treatment variable as a covariate, such as logistic regression and even linear regression. However, Q-models should not be so simplistic in practice. We propose a flexible and robust method to perform G-computation that imputes potential outcomes for each treatment group separately. Our novel method specifies separate Q-models for each treatment group that combine discretization with local logistic regression adjustment. Using extensive simulations of data generated from non-logistic probability distributions, we perform G-computation using our novel method and with ordinary logistic regression when estimating causal marginal risk differences and risk ratios and compare their performances in terms of coverage, bias, and interval width.
|
