Hello David,
I've included below the reply from one of the team experts.
(1) Matching based on Euclidian/Mahalanobis: you can probably get away with not bothering to incorporate any survey information here because you’re effectively looking for a match based on covariates X and ideally the distance will be close to zero. Mahalanobis requires calculating the variance-covariance matrix of the matching variables but one suspects the difference in matches obtained using the unweighted and weighted estimates of this matrix won’t be huge.
(2) Matching based on propensity scores: the key results that (i) residual e is independent D | p(X) and (ii) D independent X | p(X) are both based on the population propensity score p(X), so you need to use the survey weights when estimating p(X).
(3) Final-stage regression estimation is for making population inference so, with the matches obtained above substituted in, requires using the survey weights. Moreover, the clustering and stratification must also be set to ensure the standard errors are suitably adjusted. (If inverse probability weighting is used then both the 1/p(X) and 1 / (1 – p(X)) weights for treated and untreated should be such that the overall weight is w/p(X) and w/(1-p(X))).
I hope this information is helpful.
Best wishes,
Roberto Cavazos
Understanding Society User Support Team
Updated by