Accounting for covariate measurement error in a Cox model analysis of recurrence of depression

When a covariate measured with error is used as a predictor in a survival a nalysis using the Cox model, the parameter estimate is usually biased. In c linical research, covariates measured without error such as treatment proce dure or sex are often used in conjunction with a covariate measured with er ror. In a randomized clinical trial of two types of treatments, we account for the measurement error in the covariate, log-transformed total rapid eye movement (REM) activity counts, in a Cox model analysis of the time to rec urrence of major depression in an elderly population. Regression calibratio n and two variants of a likelihood-based approach are used to account for m easurement error. The likelihood-based approach is extended to account far the correlation between replicate measures of the covariate. Using the repl icate data decreases the standard error of the parameter estimate for log(t otal REM) counts while maintaining the bias reduction of the estimate. We c onclude that covariate measurement error and the correlation between replic ates can affect results in a Cox model analysis and should be accounted for . In the depression data, these methods render comparable results that have less bias than the results when measurement error is ignored. (C) 2001 Els evier Science Ltd. All rights reserved.