• 5 years - 25 posts - Latest - RSS
• OP: 7 Goods vs 4 No Goods
• Thread: 104 Goods vs 73 No Goods

1. Sociologist
   ee55
   

   ^ beakman, your point is valid if you have the option of running an experiment. But there are plenty of social science "treatments" that can't actually be randomized. In those situations, matching techniques + OLS are definitely better than just OLS.
   Not sure what you mean by better here. Matching techniques and OLS draw on the same information (measured variables) and suffer from the same limitation (unmeasured variables). You might mean that matching techniques make some of the assumptions more transparent than a standard regression model, but it's not as if a matching estimator somehow makes more of the available information than is possible with a standard regression model.
   Have you read Gary King's work? He has a different take on matching than you.

   Yes ... I've read quite a bit of King's work on matching (and other statistical topics). I'm pretty certain he does not have a different take. Can you point to some statement of his where he claims that a matching estimator gets more out of the data than a standard regression? I recall that he makes the case, as do many other statisticians and econometricians, that matching estimators make assumptions more transparent (e.g., whether covariate distributions are balanced, regions of common support, ...), but I would be quite surprised if he argues for more than this. Alternatively, perhaps you can try to make the case yourself.

   5 years ago # QUOTE 3 GOOD 3 NO GOOD !

2. Sociologist
   fafa
   

   ee55. If you read his paper, you can clearly see that he does not think matching technique can solve the problem of unobserved heterogeneity, and in that regard, matching might be almost the same with regression if you only think of omitted variable bias (or post-treatment bias). But, he think matching is a "non-parametric" tool to increase the balance and to reduce the number of cases without common support, as you noted. So, he emphasizes his coarsened exact matching technique is effective to reduce the "model dependence" (and therefore he does not like the propensity score matching because of its assumption on specific functional forms). He argues that matching should be used with other regression techniques to reduce the iterpolation bias and extrapolation bias. So, it is all about model dependence.

   5 years ago # QUOTE 3 GOOD 2 NO GOOD !

3. Sociologist
   0350
   

   ee55. If you read his paper, you can clearly see that he does not think matching technique can solve the problem of unobserved heterogeneity, and in that regard, matching might be almost the same with regression if you only think of omitted variable bias (or post-treatment bias). But, he think matching is a "non-parametric" tool to increase the balance and to reduce the number of cases without common support, as you noted. So, he emphasizes his coarsened exact matching technique is effective to reduce the "model dependence" (and therefore he does not like the propensity score matching because of its assumption on specific functional forms). He argues that matching should be used with other regression techniques to reduce the iterpolation bias and extrapolation bias. So, it is all about model dependence.

   This is the most idiotic thing I've ever read on SJMR. Seriously please stop pretending to know anything about statistics.

   5 years ago # QUOTE 2 GOOD 8 NO GOOD !

4. Sociologist
   e86e
   

   No one is arguing that matching solves omitted variable bias. My original comment to beakman was just intended as saying, "Matching has utility even if it is inferior to randomization." Gary King clearly states that matching reduces model dependence and that matching + regression reduces estimation bias. If by "getting more out of the data," ee55 just means OVB, then no, matching does not get "more out of the data." If ee55 means reducing bias, then yes, Gary King and others say matching does.

   PS - On rereading beakman more closely, I see that he was talking about people who think that matching = randomization and therefore matching solves OVB. I have never run into anyone who makes this claim, so I misinterpreted him as dismissing matching wholesale. Sorry beakman.

   5 years ago # QUOTE 2 GOOD 2 NO GOOD !

5. Sociologist
   3e61
   

   I find this thread encouraging. I am pleased to see all you youngsters thinking this through. A few corrections are in order, and I encourage others to read the interior of the following pieces that have been mentioned above, going beyond the basic conclusions that are asserted. On weighting, the Berk-Freedman pieces go after weights in regression equations where all coefficients are of interest as interpretable effects. They are far less critical of the usage of weights for the estimation of single causal effects, where all other variables are confounders. They almost endorse using weights for the calculation of differences, citing Imbens. On King's recent article, the entire argument is premised on nature having created a fully blocked experiment, after which matching is used in an attempt to recover it. This is quite unfair to propensity scores, which are not attempting to recover a blocked experiment but rather being used to solve the curse of dimensionality, unconditionally without respect to whether the observational study is based on an underlying reality that is blocked or not. Rubin has nothing but opprobrium for the King piece, which is not surprising, and he'll surely publish and demolish in due time. Most importantly for this thread, the King article is completely irrelevant to OP's question. King does not assert that regression is or is not superior to matching, although the rest of his work makes it clear how very much he dislikes simple-minded regression. In general, the literature is clear that omitted variable bias cannot be solved by matching or by regression with observables, and yet there is value in using all techniques to condition the data and represent evidence that is not driven by parametric assumptions of any particular technique.

   5 years ago # QUOTE 5 GOOD 0 NO GOOD !

Reply