Healing with Ceiling Analysis
Genetics, behavior, social context, health care, environment and stress all contribute to our health. So, “How can I be healthier?” is a complex question.
Many studies investigate the contribution of a single determinant to health outcomes, in an effort to identify a root cause for disease. For example, do genetics play a role in a person developing obesity?
It is equally important to investigate relative contribution of determinants to our health. For example, how much do genetics contribute to obesity, versus diet & exercise behaviors? Performing ceiling analysis, commonly used in software development and machine learning modeling, may help us better understand the relative contribution of individual determinants to our health.
In data science, a predictive model is made up of several functions stitched together to forecast an outcome. The predictive quality of each of these individual functions is imperfect. When combined together, the model has an overall error rate. To know where to spend energy in improving the model’s forecast, we must know how much each individual function contributes to the overall error rate. Stanford professor Andrew Ng encourages his students to perform a ceiling analysis to ascertain relative contribution. The analysis addresses the question — if each function in the model was perfectly predictive, how much would it improve the overall forecast outcome? By looking at the relative contribution of each function to performance of the overall model, it helps us spend time improving models in the most impactful way.
Apply this concept to health outcomes. How much does behavior affect health outcomes, versus genetics? Does the relative contribution of each determinant change from person to person? How does this contribution ratio shift as we investigate different diseases?
If we better knew the relative contribution of determinants to health outcomes, and how these contribution ratios varied across populations and diseases, we could identify levers to more effectively allocate resources to improve health. Work has begun in this direction — for example, McGovern et al published an insightful piece in Health Affairs on this topic(1) — but there is much to be done. Armed with the knowledge of relative contribution, there is a powerful opportunity to take a step towards rationalization of investment at the individual, institutional and policy level.
1. McGovern L, Miller G, Hughes-Cromwick P. The Relative Contribution of Multiple Determinants to Health Outcomes. Health Aff. 2014;August 21:1–9. doi:10.1377/hpb2014.17