Methodologic Challenges in the Analysis of Count Data in Radiology Health Services Research

Abstract

The authors explain that negative binomial (NB) and zero-inflated NB (ZINB) distributions are probably the most commonly seen distributions of outcomes in radiology health services research. Using simulation data, the authors demonstrate the potential errors in adopting an inappropriate model in the analysis of count outcomes in this field of research. A hypothetical database with 5,000 records was generated to evaluate the associations between the number of head CT studies (with Poisson, NB, and ZINB distributions) and age, gender, mechanism of injury, and injury severity. Linear, Poisson, NB, and ZINB regression models were used to analyze these hypothetical data. For analysis of the number of head CT studies with an NB distribution, using linear regression resulted in biased estimates. Poisson regression resulted in artificially narrow confidence intervals. For the analyses of the number of head CT studies with a ZINB distribution, Poisson and NB regression models overestimated the association between the number of head CT studies and the predictors, while linear regression resulted in incorrect point estimates. With substantial increases in health care costs and the upcoming health care overhaul, pressure on radiology health services research will increase. To provide valid estimates of the predictors of utilization pattern, researchers should adopt models that appropriately deal with the skewed count outcomes, or the results might be incorrect.