When studying the causal effect of drug use in observational data, marginal structural modeling (MSM) can be used to adjust for time-dependent confounders that are affected by previous treatment. The objective of this study was to compare traditional Cox proportional hazard models (with and without time-dependent covariates) with MSM to study causal effects of time-dependent drug use. The example of primary prevention of cardiovascular disease (CVD) with statins was examined using up to 17.7 years of follow-up from 4,654 participants of the observational prospective population-based Rotterdam Study. In the MSM model, the weight was based on measurements of established cardiovascular risk factors and co-morbidity. In general, we could not demonstrate important differences in results from the Cox models and MSM. Results from analysis on duration of statin use suggested that substantial residual confounding by indication was not accounted for during the period shortly after statin initiation. In conclusion, although on theoretical grounds MSM is an elegant technique, lack of data on the precise time-dependent confounders, such as indication of treatment or other considerations of the prescribing physician jeopardizes the calculation of valid weights. Confounding remains a hurdle in observational effectiveness research on preventive drugs with a multitude of prescription determinants.

Additional Metadata
Keywords Cardiovascular disease, Confounding, Marginal structural modeling, Methods, Observational studies, Time-dependent exposure
Persistent URL dx.doi.org/10.1007/s10654-014-9951-y, hdl.handle.net/1765/92084
Journal European Journal of Epidemiology
de Keyser, C.E, Leening, M.J.G, Romio, S.A, Jukema, J.W, Hofman, A, Ikram, M.A, … Stricker, B.H.Ch. (2014). Comparing a marginal structural model with a Cox proportional hazard model to estimate the effect of time-dependent drug use in observational studies: statin use for primary prevention of cardiovascular disease as an example from the Rotterdam Study. European Journal of Epidemiology, 29(11), 841–850. doi:10.1007/s10654-014-9951-y