Overview of New Approaches to Causal Models (Causal Pathways DAGs, Mathematical Models, FPDR Bayesian Approaches)
• Miguel Hernan
Associate Professor of Epidemiology
Department of Epidemiology, Harvard School of Public Health, Boston, MA, United States

Causal Diagrams in Epidemiology: an Application to the Direct Effect of SNPs

After finding an association between a SNP at certain genetic locus and a disease, investigators often want to identify the physiologic pathways through which the SNP, or another genetic locus in its proximity, causes the disease. A common strategy is to fit a regression model in which the disease is the outcome and both an indicator for the SNP and a measurement of some physiologic parameter are included as covariates. If an association between the SNP and the disease persists in such model, it is often concluded that the SNP (or its causal correlate) has a direct effect on disease risk that is not mediated by the physiologic parameter. For example, to estimate the direct effect of a SNP at the low density lipoprotein (LDL) receptor gene loci on the risk of coronary heart disease (CHD) that is not mediated by LDL-cholesterol levels, one would fit a logistic model for CHD risk and would include an indicator for the SNP and a measurement of LDL-cholesterol as covariates in the model. If an association between the SNP and CHD risk (odds ratio not equal to 1) is found conditional on LDL-cholesterol, one might argue that the effect of the LDL receptor on CHD is not entirely mediated through LDL-cholesterol levels. However, this simple strategy is only valid under strong assumptions that are unlikely to hold in many realistic situations.

This talk uses causal directed acyclic graphs to describe several settings in which a conditional association between the SNP and CHD risk is expected to exist even in the absence of a direct effect of the SNP (or its causal correlate). These situations include 1) the presence of a common effect (other than the SNP) of the physiologic parameter and the disease, 2) measurement error for the physiologic parameter, and 3) a time-varying dose-response relation between the physiologic parameter and the disease. We use causal diagrams to show that this problem is just another example of selection bias, and discuss the methods and conditions for estimation of the direct effect of a SNP on disease risk.

Miguel Hernán is Associate Professor of Epidemiology at the Harvard School of Public Health, Associated Director of the HSPH Program on Causal Inference, and Editor of the journal EPIDEMIOLOGY. His research is focused on methods for causal inference from longitudinal data. He combines observational data, mostly untestable assumptions, and statistical methods to emulate hypothetical randomized experiments.