IO is a subfield of applied micro btw.
Applied micro is a joke field
-
I don't understand why you are stuck on the falsehood that structural guys don't know the source of their identifying variation. Must I remind you that structural economists are the ones who developed IV based estimation? Plenty of structural papers use policy discontinuities, exogenous cost shifters, etc for identification. But these aren't the only ways to identify parameters. You can use theoretically derived exclusion restrictions as in arellano bond under appropriate assumptions for example. The funny truth is that reduced form people know less about identification than structural people.
Identification in nonlinear models is much more difficult to obtain even in much simpler cases than in the average structural model - see Honore or Matzkin, or Bonhomme's work. If the top econometricians in nonparametric identification and identification of nonlinear panel data models think we know very little about identification in these classes of models, why should I be confident that you know so much about the identification of your much more complicated model?
The more honest way to do this would be to stick to the assumptions that are justified by economic theory, i.e. shape restrictions. But partial identification in structural models is in its infancy (which isn't to discount the work being done in this area, which seems highly valuable).^ This.
^ This.
-
I don't understand why you are stuck on the falsehood that structural guys don't know the source of their identifying variation. Must I remind you that structural economists are the ones who developed IV based estimation? Plenty of structural papers use policy discontinuities, exogenous cost shifters, etc for identification. But these aren't the only ways to identify parameters. You can use theoretically derived exclusion restrictions as in arellano bond under appropriate assumptions for example. The funny truth is that reduced form people know less about identification than structural people.
Identification in nonlinear models is much more difficult to obtain even in much simpler cases than in the average structural model - see Honore or Matzkin, or Bonhomme's work. If the top econometricians in nonparametric identification and identification of nonlinear panel data models think we know very little about identification in these classes of models, why should I be confident that you know so much about the identification of your much more complicated model?
Applied reg monkeys don't do non-parametric estimation. They run diff-in-diff, RDD, or maybe, if they're lucky, have some rainfall IV to run 2SLS.
1. Including RDD in that group is technically ignorant - all RDDs start from the standpoint of some nonparametric estimation with some bandwidth choice, with polynomial approximations and local linear regressions having come to dominate the literature (perhaps partly because it is easier to calculate optimal bandwidths in those cases).
2. You are right that the majority of diff-in-diff and IV papers use parametric linear models. You seem to ignore why that is - identification in nonlinear diff-in-diff (Athey & Imbens 2006) and nonparametric IV (Horowitz, Newey, Shennach, Chesher) is much more difficult. So most "reduced-form" researchers content themselves with estimating just the mean of the distribution of the parameter of interest, since otherwise they'd have to invoke significantly stronger assumptions. Unfortunately, in much structural work, people jump straight to the estimation of a nonlinear IV via exclusion restrictions imposed in SMM/GMM estimation without discussing the economic content of those exclusion restrictions and how they are more demanding than the exclusion restrictions for linear models.
I respect structural work and think it is valuable and important, but we have to realize there are trade-offs: to say more with the same data you need stronger identifying assumptions. There are no free lunches in empirical work.
-
Lol, your assumptions are so transparent that you can’t characterize how the counterfactual is modeled. You can tell me what variation identifies your parameters.
I’d be open to some structural approaches, but your attitude is the biggest problem."Applied economists", let me give you an economist's answer as to why we don't find your work as valuable as you think. It is, as always, about incentives.
What did the credibility revolution promise? Well, it's in the name: credibility. Give up on your structural assumptions and models; give up on correlations; and just try to get a "credible" estimate of the relationship you are studying. The disadvantages are legion, but the advantage is, in an ideal world, a credibly-estimated magnitude.
This sounds all nice on paper. In an ideal world, there is a trade-off that can be worthwhile. In practice, though, the whole idea crumbles before the altar of incentives.
What are economists' incentives? To publish or perish. That's the institutional arrangement we are in. To publish, you need to crank out papers with interesting results. That means, if you have put months into a project to construct the data, you won't simply tolerate the pesky data not agreeing with you. P-hacking exists. For every single "credible identification" method, there are load of tricks to obtain the coveted three stars.
When I see an "applied economist" paper, can I trust the results? The whole point of the "credibility revolution" was to provide credible estimates. That was the upside that made us stomach giving up on structural economic models. But when I read an "applied economics" paper, how do I know the results are not p-hacked? I know the authors have *all* the incentives to torture the data to make it spit out three stars while using "credible" methods.
Compare and contrast this with structural models:
-To rig a structural result, you need to rig the assumptions. But the assumptions are in plain sight; you cannot just hide them.
-The data moments are less numerous and explicitly reported. If the authors are using a dubious target moment to rig the results, I can do a quick literature check and shoot them down.
-Structural models do not only produce a single regression coefficient, but complex objects with infinite dimension such as policy functions, transition paths, surface plots, etc. These properties usually lay bare any underlying problems with a structural model, which then lead to rejection. The method simply forces the authors to divulge too much information, which can then be used to evaluate whether there is something fishy in the paper.
Long story short, given the incentives of economists, applied econ papers are anything but "credible", whereas rigging in structural papers is much easier for the referees or readers to detect. Tell me: once the credibility part is out of the window, what is the advantage of applied econ papers over structural papers?
None.Honest question, is this a serious post? Who is confused about where identification comes from in a structural model?
An earlier post said structural work lacks internal validity. Every one of my papers are empirical, and even to me that is insane. Of course structural models are internally valid. The internal part is the model and the assumptions, no different from applied work, just very different assumptions !Your response encapsulates the problem in IO. Internal validity means are your assumptions valid, not whether you made them. Are the variable exogenous that you assume are exogenous? The fact that you didn’t understand this is why this whole thread of complaining, by bitter grad students who don’t understand what they are doing, exists.
-
I recently saw a thread on twitter complaining that applied micro acceptance rates are lowest at the top 5. Any respectable economist should understand why the field with lowest entry barriers has the lowest acceptance rates. Of course people take it as evidence as that field being harder... which is evidence that they are not respectable economists.
-
I don't understand why you are stuck on the falsehood that structural guys don't know the source of their identifying variation. Must I remind you that structural economists are the ones who developed IV based estimation? Plenty of structural papers use policy discontinuities, exogenous cost shifters, etc for identification. But these aren't the only ways to identify parameters. You can use theoretically derived exclusion restrictions as in arellano bond under appropriate assumptions for example. The funny truth is that reduced form people know less about identification than structural people.
Identification in nonlinear models is much more difficult to obtain even in much simpler cases than in the average structural model - see Honore or Matzkin, or Bonhomme's work. If the top econometricians in nonparametric identification and identification of nonlinear panel data models think we know very little about identification in these classes of models, why should I be confident that you know so much about the identification of your much more complicated model?
Applied reg monkeys don't do non-parametric estimation. They run diff-in-diff, RDD, or maybe, if they're lucky, have some rainfall IV to run 2SLS.
DD is non-parametric, buddy. Again, that you guys don’t understand this is the problem.
The weaknesses of DD is that it’s model for the counterfactual may not hold. But the model is just a difference. In two periods this is a general functional form.
-
If you become a prof and got to see with your own eyes the students who opt into structural, you would not believe in these higher barriers to entry you keep obsessing about.
I recently saw a thread on twitter complaining that applied micro acceptance rates are lowest at the top 5. Any respectable economist should understand why the field with lowest entry barriers has the lowest acceptance rates. Of course people take it as evidence as that field being harder... which is evidence that they are not respectable economists.
-
I don't understand why you are stuck on the falsehood that structural guys don't know the source of their identifying variation. Must I remind you that structural economists are the ones who developed IV based estimation? Plenty of structural papers use policy discontinuities, exogenous cost shifters, etc for identification. But these aren't the only ways to identify parameters. You can use theoretically derived exclusion restrictions as in arellano bond under appropriate assumptions for example. The funny truth is that reduced form people know less about identification than structural people.
Identification in nonlinear models is much more difficult to obtain even in much simpler cases than in the average structural model - see Honore or Matzkin, or Bonhomme's work. If the top econometricians in nonparametric identification and identification of nonlinear panel data models think we know very little about identification in these classes of models, why should I be confident that you know so much about the identification of your much more complicated model?
Applied reg monkeys don't do non-parametric estimation. They run diff-in-diff, RDD, or maybe, if they're lucky, have some rainfall IV to run 2SLS.
1. Including RDD in that group is technically ignorant - all RDDs start from the standpoint of some nonparametric estimation with some bandwidth choice, with polynomial approximations and local linear regressions having come to dominate the literature (perhaps partly because it is easier to calculate optimal bandwidths in those cases).
2. You are right that the majority of diff-in-diff and IV papers use parametric linear models. You seem to ignore why that is - identification in nonlinear diff-in-diff (Athey & Imbens 2006) and nonparametric IV (Horowitz, Newey, Shennach, Chesher) is much more difficult. So most "reduced-form" researchers content themselves with estimating just the mean of the distribution of the parameter of interest, since otherwise they'd have to invoke significantly stronger assumptions. Unfortunately, in much structural work, people jump straight to the estimation of a nonlinear IV via exclusion restrictions imposed in SMM/GMM estimation without discussing the economic content of those exclusion restrictions and how they are more demanding than the exclusion restrictions for linear models.
I respect structural work and think it is valuable and important, but we have to realize there are trade-offs: to say more with the same data you need stronger identifying assumptions. There are no free lunches in empirical work.I agree with you in sentiment but non linear DD is parametric while standard “linear” is not. You are right it allows you to say more about the distribution of the counterfactual, but by invoking more assumptions, as you say. But the mean impact in linear DD is given generally by linear DD with two periods.
-
What a lol thread. As Imbens said, better LATE than nothing! And exactly you Macro / theory people have nothing, give me one paper that could considered internally valid in Macro. Give me one structural model that does not start from such absurd assumptions that you just want to stop following the mathurbation? I think the flood in empirical papers is a reaction to the failure of so-called „theory“. Young economists like me are just sick of it.
THIS
young economists like me are sick of reg monkeys like you. people who don't care about economics should go to other departments. And don't give me the BS about structural models being worthless because they make strong assumptions. If you have an issue with the assumptions, you ought to estimate a more general version of the model. If existing methods don't allow for such generalization, you ought to develop new techniques. Instead, you decide to forgo economics altogether.
Define `economics.'
-
What a lol thread. As Imbens said, better LATE than nothing! And exactly you Macro / theory people have nothing, give me one paper that could considered internally valid in Macro. Give me one structural model that does not start from such absurd assumptions that you just want to stop following the mathurbation? I think the flood in empirical papers is a reaction to the failure of so-called „theory“. Young economists like me are just sick of it.
THIS
young economists like me are sick of reg monkeys like you. people who don't care about economics should go to other departments. And don't give me the BS about structural models being worthless because they make strong assumptions. If you have an issue with the assumptions, you ought to estimate a more general version of the model. If existing methods don't allow for such generalization, you ought to develop new techniques. Instead, you decide to forgo economics altogether.
Define `economics.'
About 2/3rds of what in the QJE these days.
Production and allocation of resources, value and prices, decision-making, incentives - and the associated tools that help us make sense of these problems.
Many empirical papers these days (particularly labor) are policy evaluations with an outcome that is relevant for economic actors. But everyone is an economic actor, and everything is relevant to someone. So, apparently, everything is economics.
-
Applied reg monkeys don't do non-parametric estimation. They run diff-in-diff, RDD, or maybe, if they're lucky, have some rainfall IV to run 2SLS.
DD is non-parametric, buddy. Again, that you guys don’t understand this is the problem.
The weaknesses of DD is that it’s model for the counterfactual may not hold. But the model is just a difference. In two periods this is a general functional form.Diff-in-Diff isn't non-parametric, at least not in the vanilla version popularized by Card & Krueger, and followed by millions of reg monkeys ever since. I mean, you're literally fitting a linear regression with a bunch of dummy variables. Who told you this was non-parametric?
-
Applied reg monkeys don't do non-parametric estimation. They run diff-in-diff, RDD, or maybe, if they're lucky, have some rainfall IV to run 2SLS.
DD is non-parametric, buddy. Again, that you guys don’t understand this is the problem.
The weaknesses of DD is that it’s model for the counterfactual may not hold. But the model is just a difference. In two periods this is a general functional form.Diff-in-Diff isn't non-parametric, at least not in the vanilla version popularized by Card & Krueger, and followed by millions of reg monkeys ever since. I mean, you're literally fitting a linear regression with a bunch of dummy variables. Who told you this was non-parametric?
Reg monkeys don't understand the distinction between parametric vs non-parametric. Why would they?
-
And did you understand Sonnenschein-Mantel-Debreu as well?
Anyone who complains about representative agents without understanding Gorman Aggregation and Negishi Weights deserves a special place in hell.
If you do understand that, go ahead and try to work on het agent models. But the answer is more math, not less.