2.5. Statistical methods

We used two types of models for statistical analysis. In model 1, a mixed-effects logistic regression, all persons with non-missing data were included. Apart from the knee-specific covariates (ipsilateral meniscus damage, tibial width, and knee alignment), we adjusted for the following person-specific covariates: age, body weight, body height, and physical activity. As sensitivity analysis we repeated this model with adjustment for knee injury during the follow-up period.

In model 2, a fixed-effects logistic regression, only persons with discordant knee outcome were included (i.e. BML progression in one knee but not in the other) [14]. In this model only knee-specific covariates were included, because all potential confounding at the person level (both measured and unmeasured) was taken care of through the model specification. The power in this model is however lower (because no data on between subject variation are used and only persons with discordant knees are included). We consider this model as a sensitivity analysis to evaluate if the results of the mixed-effects model could be confounded by unmeasured person-specific factors.

The presented estimates are odds ratios (ORs) with 95% confidence intervals (95% CIs). As the occurrence of incident/enlarging BMLs was low (less than 10%) these ORs can be interpreted as risk ratios.