Open policyglot opened 5 years ago
As per the presentation slides the Dehejia and Wahba 1999 paper (https://ssl.uh.edu/~adkugler/Dehejia&Wahba_JASA.pdf) which uses the LaLonde dataset is a classical example in which propensity score matching produces unreliable results. You may want to read the Smith & Todd rejoinder and related papers (http://www-personal.umich.edu/~econjeff/Papers/nsw_rejoinder_092203.pdf).
Why is this question here?
Great innovation here, Ari! I wouldn't have thought of connecting machine learning with these more classical methods.
Propensity-score based matching has been used extensively in policy assessments. On the repository of the International Initiative for Impact Evaluation, the number stands at 499. http://www.3ieimpact.org/evidence-hub/impact-evaluation-repository
Could you mention a few of those studies where the use of SVMs may alleviate some of the concerns you outlined? Or, in which contexts you would expect to see this result?