November 23, 2017
Exponent health scientists Dr. Suresh Moolgavkar, Dr. Ellen Chang, and Dr. Heather Watson have co-authored a paper that evaluates the consequences when assumptions that underlie a statistical method commonly used in epidemiologic research are found not to be true.
The Cox proportional hazards regression model is often used in statistical analyses of risk factors in long-term cohort studies where individuals are followed over time for disease occurrence. However, this model often relies on commonly-made assumptions that relative risks of disease are constant over time and that risk depends solely on cumulative exposures rather than the pattern of exposure — assumptions that are unlikely to hold in many real-life scenarios.
In their new article, entitled "An Assessment of the Cox Proportional Hazards Regression Model for Epidemiologic Studies," Dr. Moolgavkar, Dr. Chang, Dr. Watson, and Mr. Lau constructed a simulated cohort of 500,000 adults in which they investigated the impact of inadequate control for a strong time-dependent risk factor, namely, cigarette smoking, on relative risk estimates for another, weaker risk factor. They found that the effect of smoking was strongly modified by age, and that residual confounding from inadequate control for the time-varying effects of smoking using the Cox model resulted in spurious modest relative risk estimates for a correlated variable. Thus, this paper shows that reliance on the Cox model can result in biased results when common assumptions of the proportional hazards model are violated, and that other statistical approaches are needed to better understand the time-varying effects of risk factors.
The article is published in the journal Risk Analysis and is available here