Negative control exposure studies in the presence of measurement error: implications for attempted effect estimate calibration

Abstract

Negative control exposure studies are increasingly being used in epidemiological studies to strengthen causal inference regarding an exposure-outcome association when unobserved confounding is thought to be present. Negative control exposure studies contrast the magnitude of association of the negative control, which has no causal effect on the outcome but is associated with the unmeasured confounders in the same way as the exposure, with the magnitude of the association of the exposure with the outcome. A markedly larger effect of the exposure on the outcome than the negative control on the outcome strengthens inference that the exposure has a causal effect on the outcome. We investigate the effect of measurement error in the exposure and negative control variables on the results obtained from a negative control exposure study. We do this in models with continuous and binary exposure and negative control variables using analysis of the bias of the estimated coefficients and Monte Carlo simulations. Our results show that measurement error in either the exposure or negative control variables can bias the estimated results from the negative control exposure study. Measurement error is common in the variables used in epidemiological studies; these results show that negative control exposure studies cannot be used to precisely determine the size of the effect of the exposure variable, or adequately adjust for unobserved confounding; however, they can be used as part of a body of evidence to aid inference as to whether a causal effect of the exposure on the outcome is present.