Statistical Principles

Due to selection effects, MS patients receiving different treatments are likely to have different underlying characteristics and to experience outcomes at different rates even before allowing for the effects of treatment. Thus, confounding is expected between treatment selection and outcomes, leading to biased treatment effect estimates. Confounding variables may include demographics, disease and treatment history, and time variables representing period and cohort effects (5).

Controlling for the effects of confounders can be particularly difficult in longitudinal studies where past treatment exposures and covariates may influence future exposures and/or covariates as well as future outcomes. This is known as time-varying causal pathway confounding, and the bias it introduces may not be adequately controlled by the standard multivariate covariate adjustment approach (6–8). This type of confounding is expected to occur in the Optimise:MS cohort, given the nature of MS as a chronic progressive disease and the factors that are suspected to influence treatment decisions. Methods for controlling confounders have been chosen to mitigate this problem (further details below).

The statistical analyses fall into three classes: cohort analyses, signal detection analyses, and pregnancy analyses. These are described under the headings below.