|
Abstract:
|
Many of a person's health outcomes form a time series (e.g., wearable-device data, chronic conditions). Statistical analyses of such n-of-1 observational data may reveal cause-effect relationships, but causality is often concluded without justification, leading to subsequent testing of spurious causes. Counterfactual-based causal inference (CI) provides a framework for determining when associations might be causal effects. We hope to demonstrate how to estimate individual-focused effects from time series data using CI methods. We review how careful application of causal concepts permits statistical assessment of causal hypotheses, and introduce the average unit treatment effect (AUTE) for a subject under two treatments. The AUTE is loosely defined as a contrast of two mean outcomes, each corresponding to the subject always under one treatment. We illustrate empirical properties of two AUTE estimators by simulation, for time series of lengths 200, 1000, and 5000. The estimators are shown to be consistent, with fair coverage. Analysis of the author's own health data reveals that increases in his proportion of exercise days may have caused decreases in weight change 10-12 weeks later.
|
