Robust analyses for radiographic progression in rheumatoid arthritis

Abstract

Demonstrating inhibition of the structural damage to joints as a statistically significant difference in radiographic progression as measured by the van der Heijde modified Total Sharp Score (mTSS) is a common objective in trials for rheumatoid arthritis treatments. The frequently used analysis of the covariance model with missing data imputed using linear extrapolation (analyses of covariance, ANCOVA+LE) may not be ideal for long-term extension studies or for paediatric studies. The random coefficient (RC) model may represent a better alternative. A two-arm (active treatment and placebo) setting with a week 44 study period was considered. RC model, ANCOVA+LE and ANCOVA with last observation carried forward imputation were compared under different scenarios in bias, root mean square error (RMSE), power and type I error rate. The RC model outperformed ANCOVA+LE in metrics measuring bias, RMSE, power and type I error rate under the evaluated scenarios. ANCOVA and RC provide similar performance when there are no missing data. With missing data, RC+observed (OBS) provides similar or better results than ANCOVA+LE in power and bias. Our simulations support that RC is both a more sensitive and a more precise alternative to the commonly used ANCOVA+LE as a primary method for analysing mTSS in long-term extension and paediatric studies with a higher likelihood of missing data. The RC model can provide a reference at time points with missing data by estimating a slope; mTSS change by one unit change in time. ANCOVA+LE is recommended as a sensitivity analysis.