Abstract:
|
Data for chemical ingredients measured in several manufacturing batches of tobacco products often include only two or three batches (clusters) in one product and three to eight samples in each cluster. In addition, the intraclass correlation can be different between products, which reflects different between- and within-cluster variances. For comparisons between products, limited statistical methods are available when both the numbers of clusters and samples in each cluster are small and often assume the same intraclass correlation and between- and within-variances among products. Simulations reveal large inflation of type I error in these methods when variances are different. To address the problems, we have developed an approach to adjusting the Cochran-Cox t-test to improve upon the available methods such as the Kish and Hedges methods. In this presentation, we present simulation results of the adjusted t-test in comparison to the Kish and Hedges methods, provide insights into the issue of inflation in type I error, and highlight future directions for combining data science approaches in this research effort.
|