Abstract:
|
The key strategy in statistical learning to ensure the quality of prediction models is rigorous validation using the data at hand and also using fresh test data sets not used in model development. In causal inference, clarifying identifying assumptions and conducting sensitivity analysis has been the core strategy to establish the internal validity, but less emphasis has been given to external validity. However, we see growing attention to external validity in causal inference such as in Stuart et al. (2011) and Cole and Stuart (2010). In their work, generalizability is assessed as similarity between observed outcome in a particular trial and predicted outcome when the trial sample is made look like a target population based on propensity scores. Validation methods in statistical learning and propensity score approaches have different emphases and strengths, although little is known about how they are related and how their strengths can be combined to better assess internal and external validity. In this work, we compare these two validation approaches and show that hybrid methods that jointly utilize the two can provide richer information on reproducibility and generalizability.
|
ASA Meetings Department
732 North Washington Street, Alexandria, VA 22314
(703) 684-1221 • meetings@amstat.org
Copyright © American Statistical Association.