Article Text

Download PDFPDF
Marginal structural models to control for time-varying confounding in occupational and environmental epidemiology
  1. Kyle Steenland
  1. Correspondence to Dr Kyle Steenland, Environmental and Occupational Health, University of Emory, Rollins School of Public Health, Health, Atlanta, GA 30322, USA; nsteenl{at}

Statistics from

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.

Dumas et al1 have provided a nice example of the use of marginal structural models (MSMs) in occupational epidemiology. The clear presentation will help occupational/environmental epidemiologists become more familiar with these models, which have not been used much to date in this field. Although sometimes difficult to understand and to ‘intuit’, MSMs are worth understanding, if only because they help clarify in general the nature of confounding and selection bias.2

These models in (non-randomised) longitudinal studies are needed when time-varying confounders are themselves predicted by previous exposure, at which point they become intermediate variables on the pathway between exposure and disease. In this circumstance, the epidemiologist is left with the unenviable choice of adjusting for the cofounder (necessary to avoid confounding) or not adjusting for it (necessary to avoid adjusting for an intermediate variable). MSMs were developed to solve this conundrum,2 ,3 which has been increasing common as more detailed longitudinal data have been collected in epidemiological studies. These models apply inverse-probability weights based on an ‘exposure’ model which assesses the probability for each subject that they have received their own exposure and confounder history up to each time t, with the follow-up period divided into T (t=1 to T) categories.

These weights are then used in standard regression models (eg, pooled logistic regression models across the categories of follow-up time) assessing treatment/exposure effects on disease, and their use amounts to analysing a …

View Full Text


  • Competing interests None.

  • Provenance and peer review Commissioned; internally peer reviewed.

Linked Articles