Old and new

The paper presentation I found most interesting at the PolMeth meetings is this one by Gary King and co-authors. One way to think about it, which is not how it was presented, is that it seeks to codify the cell matching plus weighted regression schemes used in some of the ill-fated evaluations of CETA (= the Comprehensive Employment and Training Act, the predecessor to JTPA) back in the 1970s. Those evaluations are ably surveyed in Burt Barnow's (1987) Journal of Human Resources piece.

The basic idea is that you create cells using discrete variables plus discretized (or coarsened, in the language of the paper) continuous variables as a first step in a two step process, the second step of which is some parametric model that aims to catch any remaining bias. The method is marketed as simplifying the imposition of the common support - it is just those cells with both treated and untreated units - and also as avoiding the necessity of balancing tests as the (coarsened) covariates are necessarily balanced within cells.

The trick of course comes with choosing just how to coarsen. In the langauge of non-parametrics, this presents a k-dimensional bandwidth choice where the optimal choice for each bandwidth (which corresponds to the amount of coarsening in each of the k individual variables) depends on the choice of all the other bandwidths. In any given finite sample, institutional knowledge may suggest particular cuts - think of how you would group years of schooling into categories like < 10, 10-11, 12, 13-15 and so on) - but it is less clear how to think about the asymptotics. Also, once you rely on institutional knowledge and intution to guide the coarsening, rather than a more or less automated procedure like cross-validation, the opportunities for disagreement across researchers, and for misbehavior by researchers looking for particular answers, increase.

Nonetheless, the presentation was lively and got me thinking. I look forward to reading the (apparently already revised!) paper more closely.