Issues in Adjusting for Covariates Arising Postrandomization in Clinical Trials
- PDF / 587,773 Bytes
- 10 Pages / 504 x 720 pts Page_size
- 25 Downloads / 208 Views
Drug Informarion J o u m l . Vol. 33, pp. 1219-1228. 1999 Printed in h e USA. All rights reserved.
ISSUES IN ADJUSTING FOR COVARIATES ARISING POSTRANDOMIZATION IN CLINICAL TRIALS JAMES ROCHON,PHD The Biostatistics Center, The George Washington University, Rockville, Maryland
While there is broad consensus on analytic techniques for adjusting for covariates at baseline, the situation for covariates arising postrandomization is considerably more dificult. Examples include the level of patient “compliance” measured through pill counts and other biochemical markers, the occurrence of missing dnta over patient follow-up, and early withdrawal from medication. The “intention-to-treat” (ITT)principle requires that all randomized patients be included in all analyses irrespective of their confounder experience. This approach, however; seems at odds with good scientific method and is a considerable source offriction with medical investigators. In this paper; we review the interpretation of this analysis strategy and suggest that the statistical community has been careless in its interpretation of these results. We outline a conservative strategy that is consistent with IlTprinciples. Nevertheless, any analysis that adjusts for these covariates must be considered speculative in nature and followed by a properly designed confirmutory study. For this reason, we argue that these analyses are of greater relevance early in a drug development program. Key Words: Clinical trials; Confounding; Time-dependent covariates; Intention-to-treat analysis; Statistical adjustment
INTRODUCTION THERE ARE TWO important reasons for adjusting for covariates in randomized clinical trials (RCTs). The first addresses the issue of bias, that is, despite randomization, imbalances may exist across the study arms with respect to important prognostic factors. Any comparison of the treatment groups at the end of the study may be influenced by these prevailing differences. Moreover, even if there is no imbalance across the study arms, the covariate may account for variability in the primary outcome. This allows the treat-
Reprint address: Dr. James Rochon, The Biostatistics Center, The George Washington University, 61 10 Executive Blvd., Suite 750. Rockville, MD 20852. E-mail: [email protected].
ment difference to be estimated with greater precision. While there is broad consensus on analytic techniques for adjusting for covariates at baseline, the situation for covariates arising postrandomization is considerably more difficult. Examples include the level of patient “compliance” measured through pill counts and other biochemical markers, the amount of concomitant medications and “rescue” therapies, the Occurrence of missing data over patient follow-up, loss to follow-up, cross-over to the alternative therapy and premature withdrawal from medication, and the incidence of side effects and adverse reactions. In general, one must contend with the same bias and precision issues described above. Because these covariates are observed postrandomizatio
Data Loading...
