| Description |
Background: Time-varying confounding/mediating can often be encountered when using observation studies. Standard regression techniques for time-to-event analyses such as Cox regression may provide biased estimates of the exposure-outcome relationship due to improper control of time-dependent confounding/mediating. Method: A cohort of women from the Framingham Heart Study was used to assess the effect of a history of preeclampsia on subsequent stroke in life. The study involves over 40 years of follow-up for most individuals and is subject to time-varying confounding/mediating due to the number of follow-ups. There is also a portion of the cohort that is censored from the study. To account for both time- dependent confounding/mediating and censoring history of the cohort, inverse-probability-weights were calculated and assigned to fit a marginal structural model to assess the causal effect of preeclampsia on stroke. A complete method to fit a marginal structural model in Stata is shown. Results: After controlling for confounding/mediating due to baseline differences and time-dependent changes via inverse-probability-weighting, the average causal effect of preeclampsia on subsequent stroke is a relative risk estimate of 3.8 (95% CI [1.2, 11.6]) Conclusion: Proper balance and control of possible confounders/mediators is crucial when assessing an exposure-outcome relationship. Through the methods of marginal structural modelling, balance was achieved between the exposure groups to provide a least biased estimate of the causal effect of preeclampsia on stroke. Introduction Observational studies include following individuals over time without any input or influence by the researcher. Characteristics of interest are discussed and recorded from their entrance into the study (prospective study) or the first time point available in a previously recorded dataset (retrospective study); this value is considered the subjects' baseline measurement. The subjects of the study are followed up through the duration of the study until the subject experiences the outcome of interest or they are censored from the study. Censoring most commonly occurs when subjects fail to be present at a follow up visit or remove themselves from the study for a variety of reason. Finally, the time-to-event is defined as the time from the first visit to the event occurrence. In the span from first visit to the event, a set of characteristics are measured during each visit. Some of these variables can be considered confounders and/or mediators and need to be accounted for during analysis of the exposure-outcome relationship. Confounding is present when there is a covariate that is associated with both the exposure and outcome but does not lay between the two in the direct pathway of the exposure-outcome relationship. Can introduce correlation where isn't any or can introduce bias. For example, abnormal blood pressure can affect the chance of preeclampsia as well as stroke. However, abnormal blood pressure does not lie in the direct pathway of a possible relationship between preeclampsia and subsequent chance of stroke[4]. Mediation is present if there exists a relationship between the exposure and outcome where the exposure variable influences a mediator variable which in turn influences the outcome. This divides the possible effect of the main exposure onto the outcome into a direct effect from the exposure to the outcome and the indirect effect which contains the mediator variable in the exposure-outcome pathway. Mediation analysis is used to attempt to understand the underlying mechanisms by which the exposure is associated with the outcome through a mediator variable[5,6]. |
