“…If there are match-specific unobservables, the likelihood function involves a numerical integral of 10,000 dimensions and an integrand that involves computing whether 1 out of 10 158 assignments occur. Of course, simulation estimators (sometimes based on moment conditions chosen for tractability) can be used for matching when required at least when markets are small and computational resources are large, as in Boyd, Lankford, Loeb and Wyckoff (2013), Sørensen (2007), Agarwal and Diamond (2013) and Fox, Hsu and Yang (2015). The computational advantages of matching maximum score are also present in single agent, multinomial choice maximum score (Manski, 1975;Matzkin, 1993;Briesch, Chintagunta and Matzkin, 2002;Fox, 2007).…”