Matching
Matching is a pre-processing technique.
Contents
Description
Matching is generically a method for identifying similar observations, based on covariates.
Exact matching means taking a treated case and identifying an untreated case that has equivalent values for all covariates. Analysis is then performed on just the matched sample; all unmatched cases are deleted. A necessary follow-up step is ensuring that the matched sample is balanced.
Coarsened Exact Matching (CEM) is a modification of that approach; covariates are 'coarsened' by binning, so as to encourage more 'exact' matches. Analysis should be performed with the matched sample and the original, un-coarsened variables.
Propensity score matching (PSM) models the propensity of being 'assigned' the treatment, giving a 'propensity score'. Cases are then matched based on similarity of the scores.
Usage
Weighting
When redistributing weights from non-respondents to respondents, a uniform adjustment is often used. A more nuanced method is to design weighting classes using propensity scores (wherein 'treatment' is response), and then calculating a uniform adjustment per class. This ensures weight transfers between similar cases.