In the context of observational cohort studies, it is often difficult to distinguish potential adverse effects of GCs from the effects directly related to the activity of the underlying disease, prompting prescription of GCs. This channelling bias which means that more severely ill patients are treated with more effective, but also potentially harmful agents is present in all real-world studies of drug exposure.
In rheumatoid arthritis, as an example, adverse outcomes that are linked to highly active disease can be prevented by drugs that efficiently reduce disease activity, such as GCs. However, the same adverse outcome can also be a direct result of the immunosuppressive effect of GCs. Therefore, the challenge is to disentangle treatment and disease-associated effects.
Due to some practice variation among physicians, large observational cohorts can be used to analyse the effects of GCs (e.g. on the risk for infection or mortality) separately from the effects of high disease activity. Multivariate statistical models are applied to estimate the individual effects of disease and treatment. Since GCs are not given at fixed dosages, statistical modelling has further to consider time-varying risks resulting from changing dosages of GCs.
With examples from the German biologics register RABBIT on serious infections and on mortality I will demonstrate how confounding by indication and real effect can be analysed separately and what this means for the assessment of the adverse potential of GCs. The data show that, after control for baseline confounding and time-varying risks, there is a considerable remaining risk for higher dosages of GCs if given over longer periods of time.
Disclosure of Interest None declared