Objectives: Progression of joint destruction is an important phenotypic feature in rheumatoid arthritis (RA). When factors have small effect sizes, both the avoidance of phenotypic misclassification and discerning true effects from noise are challenging. Assembling radiological measurements repeatedly in time harbours a smaller risk of misclassification than single measurements. Given serial measurements, different methods of analysis can be applied. This study evaluates different statistical methods of analysing longitudinal data.
Methods: Three statistical methods were studied: linear regression (LR), generalized estimating equations (GEE), and multivariate normal regression analysis (MRA). All were applied longitudinally, testing for differences in radiological progression rates. As genetic variants are known to have small effect sizes, two genetic variants were studied as examples: rs675520 (located in the TNFAIP3-OLIG3 region) and the presence of the human leucocyte antigen (HLA) shared epitope (SE) alleles. Radiological data for 602 early RA patients with yearly radiographs and 7-years of follow-up were used. The powers obtained with the methods and the robustness against missingness were evaluated as outcome measures.
Results: The presence of the rs675520 polymorphism and the HLA-SE risk genotype was associated with a 0.65-0.77 and 1.17-1.51 fold increased rate of joint destruction, respectively. The analyses performed with MRA resulted in smaller 95% confidence intervals (CIs) than the analyses using LR or GEE. In addition, the 95% CIs increased with the number of radiographs per patient. The power of MRA was higher than that of GEE. MRA was more robust against selective missingness than GEE or LR with a two-step approach (LR(ts)).
Conclusions: A multivariate normal regression model on subsequent radiographs is a powerful and robust method for analysing longitudinal joint destruction data.