Statistical analysis of clinical prediction rules for rehabilitation interventions: current state of the literature

Deriving clinical prediction rules (CPRs) to identify specific characteristics of patients who would likely respond to certain interventions has become a research priority in physical rehabilitation. Understanding the appropriate statistical principles and methods of analyses underlying the derivation of CPRs is important for future rehabilitation research and clinical applications. In this article, we aimed to provide an overview of statistical techniques used for the derivation of CPRs to predict success following physical therapy interventions and to generate recommendations for improvements in CPR derivation research and statistical analysis in rehabilitation. We have summarized the current state of CPR intervention-related research by reviewing 26 studies. A common technique was found in most studies and included univariate association of factors with treatment success, stepwise logistic regression to determine the most parsimonious set of predictors for success, and calculation of accuracy statistics (focusing on positive likelihood ratios). We identified several shortcomings related to inadequate ratio of events by number of predictors, lack of standardization regarding acceptable interobserver reliability of predictors, questionable handling of predictors including reliance on univariate analysis and early categorization, and not accounting for dependence and collinearity of predictors in multivariable model construction. Interpretation of the derived CPRs was found to be difficult due to lack of precision of estimates and paradoxical findings when a subset of the predictors yielded a larger positive likelihood ratio than did the full set of predictors. Finally, we make recommendations regarding how to strengthen the use of statistical principles and methods to create consistency across rehabilitation research for CPR derivations.