Background Cohort studies often use simple models, for example, completers-only or aggregated scores, to assess changes in imaging outcomes. With more data available, this introduces selection bias. Integrated analyses allow scores from multiple reading rounds; therefore, we assessed their effect on the change estimates’ precision compared with the completers-only analysis. Second, we compared change estimates from models using aggregated scores and scores analysed per reader. Methods Patients with axial spondyloarthritis (axSpA) from the SPondyloArthritis Caught Early cohort were followed for 2 years. Imaging of the sacroiliac joints and spine was scored by ≥2 central readers in two rounds. Change in imaging outcomes was analysed using generalised estimating equations, with different patient and data selections for integrated and completers-only analyses. The integrated and multilevel completers-only analyses assessed outcomes per reader. Change estimates of the completers-only analysis of aggregated scores of readers were compared with estimates of the multilevel completers-only analysis. Results The integrated analysis included 332 patients (vs 279 completers-only). Compared with the multilevel completers-only analysis, the integrated analysis included 12%–16% more patients with a similar precision. Within the completers-only population, change estimates from models using reader and aggregated scores were similar for continuous outcomes. However, similar comparisons in dichotomous outcomes showed that models with aggregated scores underestimated spinal changes (maximum 82% less) and overestimated sacroiliac joint changes (maximum 37% more). Conclusion Multilevel models retain level variability by integrating information across patients, time points, readers and rounds to reduce selection bias and improve precision of change, which would otherwise be lost in simpler models.
Bruin et al. (Thu,) studied this question.