Causal Inference
Confounding or bias amplification? Clues for the researcher seeking causal inference. Krista Christensen* Krista Christensen Michael Leung Elizabeth Radke Michael Wright Tom Bateson Marc Weisskopf
In epidemiology studies, associations between exposure and outcome could be affected by confounding, but also by amplification of bias due to the presence of unknown or unmeasured confounders and correlation between the exposure of interest and co-occurring exposures. However, determining which of these types of bias – or both! – may exist is not straightforward. An example where both ‘traditional’ confounding and bias amplification may occur, is examination of health effects due to perfluoroalkyl and polyfluoroalkyl substances (PFAS), where correlation between PFAS is usually present, and sources (and therefore potential confounders) are not always well understood. As an example, several epidemiological studies have reported associations between PFAS biomarker levels and immune endpoints including vaccine response. In one of these studies, correlations between PFAS were moderate to high (range: 0.22 to 0.78), and results were presented for both single PFAS and multi-PFAS models. In some cases, adjusting for other PFAS changed point estimates (e.g., a 37% attenuation of the point estimate of PFDA in multi-PFAS models, compared with the single-PFAS model) while in other cases (e.g., PFOA), results were similar. We performed a simulation exercise with a dataset constructed to mimic the data from this study and examined how point estimates were changed with different causal assumptions. We found that when amplification bias was present the ‘better’ choice is the single PFAS model. However, if the direction of the amplification bias, strength of associations and/or degree of correlation were changed, the balance could shift such that the multi-PFAS model yielded lower bias. Furthermore, we show that it may be possible to detect bias amplification if there exists a PFAS in the mixture that is not causally related to the outcome. *Disclaimer: The views expressed in this abstract are those of the authors and do not necessarily reflect the views or policies of the U.S. EPA.