Power Estimation Using the Exemplary Dataset for Ordinal Data in the Presence of Nonproportional Odds

Abstract

Ordinal categorical data are extremely common in clinical trials where subjective outcomes are being measured. The most common analyses for ordinal categorical data are based on the proportional odds model. However, the assumption of proportional odds is an assumption of the model and not necessarily a property of the data. In many situations, clinical researchers may suspect that an effect on an ordinal scale will not satisfy the proportional odds assumption. In these situations, statisticians need to be able to investigate alternative models and estimate the power for these alternative models. We examine the trend odds model and the saturated model and compare these models to the proportional odds model under a wide class of alternative hypotheses. We also examine the exemplary dataset (ED) approach for power estimation for these models and determine how well this approach approximates the actual power. A recently developed ordinal scale for assessing nausea in the pediatric population illustrates the issues. The ED method was found to be accurate down to small sample sizes and the proportional odds test statistic was generally competitive even in the presence of nonproportional odds.