OBSSR Methodology Seminar:
Predictive Modeling for Behavioral and Social Sciences Health Research
October 12, 2018
National Institutes of Health
Porter Neuroscience Research Center, Bldg. 35A, Room 610
9:00 AM – 4:00 PM
Link to full agenda: https://drive.google.com/file/d/1krHjFeBCdNXNuQ7VukqRpJzkDyHlOyyq/view
Objectives: This one-day methodology seminar was sponsored by the National Institutes of Health (NIH) Office of Behavioral and Social Sciences Research (OBSSR) will showcase principles and techniques for prediction modeling from machine learning via specific case examples presented by scientists who are applying predictive algorithms to health-related behavioral and social sciences data. This seminar was intended for scientific program and review officers and other interested NIH staff, fellows, or intramural scientists engaged in evaluating research proposals or conducting research featuring predictive algorithms. Attendees gained a broad understanding and appreciation for the capabilities of prediction modeling to advance research in health by complementing the more traditional and exclusive focus on explanation. This workshop included a public access video archive which is found here:
https://videocast.nih.gov/summary.asp?live=27987
Background: Behavioral and social science research in health funded by the NIH historically has been concerned with explaining the causal mechanisms that give rise to behavior. Under this tradition the primary focus is investigating mediating and moderating variables that explain attitudes, beliefs, behaviors and health outcomes in observational studies, field experiments, and randomized clinical trials. This near-total emphasis on explaining the causes of health behavior has led to research programs that offer intricate theories of mechanisms but that have little ability to predict future behaviors in the general population with any appreciable accuracy.
Principles and techniques from machine learning may help health behaviors research become more predictive, and in so doing, may generate theories about social and behavioral aspects of health to inform our causal experiments. In social science disciplines outside of health research, where machine learning has been more readily applied, a field of “computational social science” is growing and beginning to reverse the traditional bias against predictive modeling. The addition of prediction models as analytic tools to complement pure explanatory statistical models may lead to better, more replicable behavioral and social science health research.