Examining radiographic outcomes over time.

Statistical analysis plays a critical role in data interpretation in all fields and particularly so for clinical data where important treatment decisions are made. We provide here an in-depth and illustrative analysis to examine patterns and radiographic scores in an early disease rheumatoid arthritis cohort over a 3-year follow-up period. The total Sharp radiographic scores were interpolated from the rates at 6 months, 1, 2, and 3 years and were transformed to count data after rounding. The generalized estimating equations approach and two-part models were applied to analyze the longitudinal radiographic scores using the clinical, demographic, and therapeutic characteristics of the patients after adjusting for the pattern outcomes. Total Sharp scores were modeled, assuming that they were Poisson distributed or had a negative binomial distribution with either an AR(1) working correlation matrix or an exchangeable working correlation matrix. To account for the excessive zero counts, we used two-part models that include the zero-inflated Poisson and the zero-inflated negative binomial to fit the data. This is an innovation because two-part models have not been used in rheumatology even though they are highly appropriate for analyzing data from rheumatic studies. In addition, we analyzed data using generalized estimating equations and compared results from different models using formal statistical goodness-of-fit criteria and arrive at the best model for predicting purposes.