| To Register SMDM Homepage | |
|
|
|
||||
Purpose: To explore the relative merits of using a Local Control (LC) approach based upon patient clustering rather than traditional smooth regression models for Covariate Adjustment (CA) in assessing treatment effects.
Methods: We used patient data (n=1658) from a randomized placebo-controlled trial of drotrecogin alfa (activated) (DAA) for the treatment of severe sepsis. For simplicity, we selected as covariates three patient characteristics found to be the important independent clinical predictors of mortality: age, APACHE II acute physiology score, and platelet count. In the LC approach, patients’ baseline values for these three continuous variables were used to determine their location in 3-dimensional X-space, and dissimilarities between patients were determined by calculating Mahalanobis distances, which were then used in clustering patients. A range of estimates of treatment effect were generated by varying the size (and number) of clusters and measuring the weighted average within-cluster difference in survival across all clusters containing both treated and untreated patients. These estimates were then compared with those obtained for CA using logistic regression.
Results: Despite randomization, small treatment group differences on important baseline predictors of mortality were noted. CA using logistic regression reduced the estimated survival advantage of DAA from 5.2% (unadjusted p=0.024) to 4.6% (p=0.038.) In contrast, the LC approach supported a range of DAA survival advantages from 4.4% (60 clusters) to 7.3% (500 clusters). Higher estimates (with lower uncertainty) emerged as cluster size decreased even though more patients were forced into clusters non-informative about local treatment differences (i.e., containing only treated or only untreated patients). Whereas CA requires pre-specification of presumed relationships between predictors and outcome, LC allows for arbitrary, even discontinuous, relationships between predictors and outcome. As such, LC is an attractive choice for exploratory analyses and for assessing the robustness of intent-to-treat analyses. Conversely, taking advantage of available clinical knowledge in forming clusters and interpreting results can be problematic. Additional challenges include uncertainty around the best approach to weighting across clusters and in determining an “optimal” number of clusters. Finally, the LC approach requires analysis and graphical display capability not currently available in many statistical packages.
Conclusions: LC can be used to explore treatment effects and may be useful in assessing the robustness of baseline efficacy estimates from randomized clinical trials.
See more of Oral Concurrent Session B - Methodological Advances
See more of The 26th Annual Meeting of the Society for Medical Decision Making (October 17-20, 2004)