An Approach to Obtaining Initial Values for the Covariance Parameters for Repeated Measures Analysis with Unstructured Covariance Model

• Activity Number: 379
• Type: Contributed
• Date/Time: Wednesday, August 1, 2007 : 8:30 AM to 10:20 AM
• Sponsor: Biopharmaceutical Section
• Author(s): Kaifeng Lu*+
• Companies: Merck & Co., Inc.
• Keywords: repeated measures ; covariance structure ; missing at random

Abstract

In randomized clinical trials, patient data are usually collected over time. Dropouts and the resulting missing data often pose a challenge in data analysis. Repeated measures model can provide valid inference when the missing data are missing at random. An unstructured covariance model is often used to avoid bias due to misspecification of the covariance structure. Convergence problems may arise for various reasons. In some situations, convergence may be achieved via the use of reasonable initial values for the covariance parameters. One approach to obtaining initial values is to fit a series of univariate analyses of covariance models for the measurement collected at the current time point conditional on baseline covariates and measurements collected at previous time points. The covariance matrix for the repeated measures can then be reconstructed from the sequence of ANCOVA models.