Propensity score matching questions

[Nikos Korompos]

I am trying to conduct a propensity score matching analysis, and I am quite confused on some issues.


(1) A first think relates to "noreplacement" option. In particular, I find quite different results in terms of ATT when conducting an analysis with and without the option. Is it reasonable? And if this is case, then which of them should I prefer?

psmatch2 `matching_var', outcome(birth_weight) noreplacement common

psmatch2 `matching_var', outcome(birth_weight) common

(2) Is there any criterion for specifying the number in "neighbor" option? I guess, if I specify neighbor(1), this corresponds to one-to-one matching, right?

(3) I am trying to use the kernel matching algorithm using the following command. However, stata does not perform the analysis (it's loading), possibly because my file is quite large. Any ideas for this?

psmatch2 `matching_var', kernel outcome(birth_weight) common

Thank you in advance!

Kind regards,

Nikos

[David Radwin]

I can answer 1 and 2.

1. See p. 164 of Stuart, E. A., & Rubin, D. B. (2008). Best practices in quasi-experimental designs. Best practices in quantitative methods, 155-176. http://www.corwin.com/sites/default/...Chapter_11.pdf

2. Yes, but see http://stephenporter.org/understandi...atas-psmatch2/ .

[Nikos Korompos]

Many thanks David. I would also like your opinion on (1)

For (1), in another paper ('Matching Methods for Causal Inference: A Review and a Look Forward'), Stuart (2010) argues the following:

Additionally, when matching with replacement, the order in which the treated individuals are matched does not matter. However, inference becomes more complex when matching with replacement, because the matched controls are no longer independent—some are in the matched sample more than once and this needs to be accounted for in the outcome analysis, for example, by using frequency weights. When matching with replacement, it is also possible that the treatment effect estimate will be based on just a small number of controls; the number of times each control is matched should be monitored.

Technically, does it mean that the treatment effect after psmatch2 with replacement may not reflect the true effect?

[David Radwin]

If we could know the true effect, then we wouldn't have to use matching or any other statistical estimation method!

The basic issue is that to date, there is no consensus on exactly how to implement matching in practice to get the least biased result (not to mention minimizing variance). Much of the evidence for and against matching in the last 20 years has been comparing matching estimates to values from the "gold standard" of randomized controlled trials. The results are mixed and still unresolved. See, e.g.,

Dehejia, R.H., and Wahba, S. (1999). Causal Effects in Nonexperimental Studies: Reevaluating the Evaluation of Training Programs. Journal of the American Statistical Association, 94(448): 1053–1062.

Sekhon, J.S. (2009). Opiates for the Matches: Matching Methods for Causal Inference. Annual Review of Political Science, 12: 487–508.

Smith, J.A., and Todd, P.E. (2005). Does Matching Overcome LaLonde’s Critique of Nonexperimental Estimators? Journal of Econometrics, 125(1): 305–353.