A comparison of covariate adjustment approaches under model misspecification in individually randomized trials

Table 1 Summary of analytic methods and their properties

Method	Properties
Unadjusted	Unbiased in all settings.
Unadjusted	Typically reduced power compared to adjusted approaches.
ANCOVA/Adjusted	Typically leads to increases in power.
	Retains good properties if many covariates adjusted for.
	No issues with estimation of standard errors in small samples.
	Bias in non-linear interaction setting.
	Marginal odds ratio cannot be targeted.
	Convergence issues if risk ratio is of interest.
G-computation	Undercoverage and high type I error in small sample sizes.
G-computation	Bias in non-linear interaction setting; alleviated by allowing for interaction.
IPTW	Covariate–outcome relationship need not be specified.
	Undercoverage and high type I error in small sample sizes and adjusting for a few covariates.
	Overcoverage if adjusting for many covariates.
	Convergence issues if there are many covariates.
	Slight bias in non-linear interaction setting.
AIPTW	Either covariate–treatment or covariate–outcome relationship needs to be correct.
	Undercoverage and high type I error in small sample sizes.
	Convergence issues if there are many covariates.
	Slight bias in non-linear interaction setting.
TMLE	Either covariate–treatment or covariate–outcome relationship needs to be correct.
	Standard errors can be underestimated if efficient influence function based estimators are used.
	Slight bias in non-linear interaction setting.

ISSN: 1745-6215