Covariate adjustment for linear models in estimating treatment effects in randomised clinical trials. Some useful theory to guide simulation

TL;DR

This paper develops a theoretical framework to evaluate covariate adjustment in randomized clinical trials, focusing on residual error, variance inflation, and precision, supported by formulas and simulations.

Contribution

It introduces formulas for the expected value and variance of the variance inflation factor (VIF) for covariates, including categorical variables, and demonstrates their accuracy through simulations.

Findings

Formulas accurately predict VIF mean and variance across models.

VIF relates to chi-square contingency tables for categorical covariates.

The three-aspect system guides covariate adjustment analysis.

Abstract

Building on key papers that were published in special issues of Biometrics in 1957 and 1982 we propose and develop a three-aspect system for evaluating the effect of fitting covariates in the analysis of designed experiments, in particular randomised clinical trials. The three aspects are: first the effect on residual mean square error, second the effect on the variance inflation factor (VIF) and third the effect on second order precision. We concentrate, in particular, on the VIF and highlight not only an existing formula for its expected value based on assuming covariates have a Normal distribution but also develop a formula for its variance. We show how VIFs for categorical variable are related to the chi-square contingency table with rows as treatment and columns as categories. We illustrate the value of these formulae using a randomised clinical trial with five covariates, one of which is binary, and show that both mean and variance formulae predict results well for all $2^5=32$ possible models for each of three forms of simulation, random permutation, sampling from a Normal distribution and bootstrap resampling. Finally, we illustrate how the three-aspect system may be used to address various questions of interest when considering covariate adjustment.