Two-part Statistical Model for Identifying Baseline Predictors of Chronic Postsurgical Pain.

backgroundA substantial proportion of patients report no pain after surgery, resulting in an excess of zero values that pose challenges for analysis using traditional statistical models. The current study was designed to test the hypothesis that a two-part model, commonly used in healthcare expenditures research, would demonstrate superior performance in predicting postsurgical pain when compared with traditional models and would secondarily better identify predictors of this clinically important outcome.

methodsThis study analyzed a prospectively collected single-center data set (n = 3,925) of chronic postsurgical pain to compare a novel two-part modeling framework with logistic and linear regression. The two-part model first estimated the probability of experiencing postsurgical pain at 3 months (binary outcome), followed by the severity of pain among those affected (on a 1 to 10 numeric rating scale). To obtain an unbiased assessment of model performance, the data were randomly split into training (n = 3,000) and testing (n = 925) data sets. The models were trained on the training data set and evaluated on the testing set. This process was repeated 400 times to compute average performance estimates. As a secondary aim, the study assessed the associations of 15 baseline factors, including validated measures of patient-reported pain, functional and psychological measures, comorbidities, and surgical details, across the different modeling approaches.

resultsThe two-part model demonstrated superior predictive performance compared with linear regression alone, with a higher mean R 2 value (0.075 vs . 0.050; P < 0.0001), lower root mean square error (1.466 vs . 1.485; P < 0.0001), and lower mean absolute error (1.020 vs . 1.030; P < 0.0001). Statistical comparison with logistic regression alone was not possible due to the shared binary component. The two-part model identified seven baseline preoperative covariates as independently associated with postsurgical pain that were missed by either logistic or linear models or both: self-reported race, education level, American Society of Anesthesiologists classification, overall body pain, widespread pain, symptom severity, and anxiety level. Significant baseline factors identified in all three models were patient sex, surgical type, and surgical site pain.

conclusionsA two-part model may offer a superior statistical approach to linear and logistic regression, better identifying patient- and clinical care-related risk factors of chronic postsurgical pain, primarily by distinguishing factors that drive the occurrence of chronic pain from those that are associated with pain severity.