Analyzing Bounded Outcome Score Data

Clinical trial endpoints are often bounded outcome scores (BOS), which are variables having limited values within finite intervals. BOS are conceptually ordered categorical variables, often with a large number of possible values (> 10). Although they are commonly analyzed as continuous data, their distributions are often skewed, posing challenges for analysis. Three major approaches, namely data transformation, zero-inflated, and latent-variable have been used in practice with varying degrees of success. Additionally, predicting derived endpoints, such as achieving specific levels of improvement from baseline, is often crucial. Choosing an appropriate analysis method is a complex and often counter-intuitive task, which is the primary focus of this review.

One key factor that distinguishes BOS data analysis methods is the treatment of the data as either continuous or categorical. Formal likelihood-based method comparisons should only be conducted when the data are treated as the same type. Categorical analysis methods have the advantage of naturally aligning with BOS data. Nevertheless, the specific methods may differ in their effectiveness in handling skewed data. The latent variable categorical analysis methods, particularly those utilizing beta distributions, have demonstrated the ability to describe both the BOS data and the derived endpoints.

When selecting analysis methods, it is crucial to consider the analysis objective, which may involve predicting data within its natural range, addressing skewed data, and predicting derived endpoints. The methods using the latent variable approach, especially the latent-beta method, may hold the most promise.