West Virginia University biostatistics faculty, Michael Regier, PhD and his colleagues, E. Moodie, PhD and  R. Platt, PhD, both of McGill University, recently published an empirical study investigating the effects of confounder measurement error on the estimation of the causal parameter (effect) when using marginal structural models and inverse probability-of-treatment weighting to adjust for confounding. Marginal structural models are important family of longitudinal models for causal inference when exposures (treatments), confounders, and outcomes are repeatedly measured over time and causal inference about the effects of the exposure on the outcome are desired.  Frequently, they are used in situations where a randomized controlled trial is either difficult or impossible; more recently, they are being used for observational studies to move model interpretation from purely associative to causal.

An area of understanding that has not received much attention concerns the problem of measurement error in the time-varying confounders.  The researchers aimed to fill this gap and investigated the effects of confounder measurement error.  They observed the well-understood effects of attenuation and augmentation, and identified two new, undocumented effects of measurement error: null effects and sign reversals.  Through simulation studies, it was determined that the type of measurement error observed is a function of the correlation amongst confounders and the magnitude of the noise of the measurement error process.  Based on these results and the associated theory, findings show there is no a priori determination of the effect of confounder measurement error.  Caution should be exercised with both the methods used to correct for this problem and with the emphasis placed on the unbiased estimation of the causal parameter.  The full article is found in the International Journal of Biostatistics (PMID: 24445244).