Although there have been considerable strides in epidemiological methods for causal inference with observational data, many challenges remain. In observational studies, different confounding mechanisms may operate across ranges of a risk factor, and the direction of causality may reverse for some subsets of the population. Failure to account for these different mechanisms can create the impression that low to moderate exposure is beneficial.
The purpose of this study is to use UK Biobank data to explore practical methods to mitigate this source of bias. Our goal is to better understand how to distinguish when the often-seen J-shaped curve reflects a real change in the direction of the causal effect of a risk factor and when it reflects possible bias from different confounding mechanisms over various ranges of the risk factor.
The causal effect that moderate wine intake improves cardiovascular health has never been fully established and continues to be a point of discussion in the medical literature. The American Heart Association states: “Some studies have shown that those who drink moderate amounts of alcohol have lower rates of heart disease than nondrinkers. However, evidence is still insufficient to generalize the results to everyone.” Furthermore, the AHA recommends limiting alcohol to no more than two drinks per day for men and one for women.
Our study aims to assess cardiovascular risk related to alcohol consumption. In particular, we want to assess (a) whether the observed benefits of low consumption are real or an artifact of analysis and bias in data collection, and (b) identifying a change point value of alcohol consumption where cardiovascular risk begins to rise more sharply. We will use Targeted Maximum Likelihood Estimation (TMLE), Augmented Inverse Probability Weighting (AIPW), and Double/Debiased Machine Learning (DML) applied to data that possibly exhibit reversals in causal direction. The project is expected to last 18 months.