A few parts of this half-hour interview on “The Current and Future of Econometrics” with Jeffrey Wooldridge, conducted by several questions for a nonprofit called SciEcon-AMA (of which I have no knowledge), may be more accessible to researchers in of that blend of economics and statistics known as “econometrics.” But most of it should be accessible to the advanced undergraduate who has taken the basic ‘metrics class (March 7, 2022, podcast and transcript available).

On Considering the Model and the Estimation Method Separately

You did mention the difference-in-difference work, so let me focus on what I’ve actually written about. I think this is generally a good lesson that I showed you can use traditional regression methods. In particular, you can expand the usual two-way fixed effects estimator. Actually, it’s not expanding the estimator, but expanding the model. My interpretation of the recent criticism of two-way fixed effects is that it’s not a criticism of the estimator but a criticism of the model. The model assumes a single treatment effect, regardless of how many cohorts there are and how long the time period is for the intervention. And I simply noted that if you set things up properly, you can apply regression methods to estimate a much more flexible model, and this largely overcomes the criticisms of the simple two way fixed effects analysis.

So, what I tried to emphasize with my students is that it’s very important to keep separate the notion of a model and an estimation method. And I sometimes forget myself. I will say things like OLS [ordinary least squares] model, but OLS is not a model. It’s an estimation method which we can apply to various kinds of models. It’s up to us to be creative and use the tools that we have so that we apply those methods to models that don’t make strong assumptions. I hope that this idea bridges again a lot of my research, which is pretty simple. It’s trying to find simpler ways to do more flexible analysis, at the point that it gets really hard.

On the Temptations of Simulations and Machine Learning

I was in the middle of doing some simulations for some recent nonlinear difference-in-differences methods that I’ve been working on. But then I was thinking, as I was doing the simulations and changing the parameters of the simulations, am I doing this to learn about how this estimator compares with other estimators, or am I trying to rig it so that my estimator looks the best? So, I was really just making a statement. Like you know, it’s human nature to want yours to be the best, right? One uses the machine to learn about that, and I’m partly making a statement. I’m trying to be as objective as I can by showing cases where the methods I’m proposing work better but also being upfront about cases where other methods will work better. …

When we publish papers, the best way to get your work published is to show that it works better than existing methods. Since the people writing the theory and deriving the methods are the same ones doing the simulations, it will probably be better if there’s some disconnection there. … I’ve always thought that we should have more competitions, such as blind competitions where people who participate don’t know what the truth is. They apply their favorite method across a bunch of different scenarios, so we can evaluate how the different methods do. I’m guessing that machine learning will come out pretty well with that, but that’s an impression. I’m not convinced that somebody using basic methods who has good intuition and is creative can’t do as well. …

I think the work on applying machine learning methods to causal inference has guaranteed that it will have a long history in econometrics and other fields that use data analysis. When I took visits to campuses, Amazon, Google, they’re using machine learning methods quite a bit. That’s no secret. These companies are in the business of earning profits, and they’re not going to employ methods that somehow aren’t working for them. So, I think the market is certainly speaking on that. For prediction purposes, they seem to work very well.

On Simplicity and Credibility of Methods

It’s interesting that if you look at the literature on intervention analysis and difference-in-difference, in some ways we’re trying to go back to simpler things. So, if you were to compare today with twenty years ago and see what econometrics people are doing, it seems to me that structural methods may be more out of favor now than they were fifteen years ago with this re-emergence of difference-in-difference. It seems that we are always looking for natural experiments and interventions to learn things about policy. … So, I wonder if our reaction to these complications in the real world is leading us to simplify the econometrics. Or, at least we are going to only believe analyses that have some clear way to identify the causal effect of an intervention rather than our relatively simple economic models.