Saturday Links

Economics, Econometrics, AI, Pop Culture, Self Help and News

It’s now a week since I got here to Torino Italy. Today I worked on some admin stuff (referee reports are now called “admin stuff” in my mind I guess). I’ve been collecting a lot of links all week, particularly in my browser, and I thought I’d share them. I’ve grouped them, though, into economics, econometrics, AI, pop culture, food, and other stuff I can’t remember below. I’ll put the economics and econometrics up top for all, and the rest behind the paywall. Wish me luck I can get the rest of my admin stuff done. I have been worn down this week like never before.

Economics

Last year, I was invited to Stirling Scotland to teach a four day workshop on difference-in-differences for Scottish PhD students, and this year I have been invited to do so again. As we covered difference-in-differences last year, this year it is going to follow material on unconfoundedness, instrumental variables and regression discontinuity. So if you’re in the United Kingdom, and would like to spend a week in summer school at the Scottish Graduate Program in Economics with me, then click on this link. Here’s a screen shot too of the page on the link, too, if you’re wanting informational out it like the dates (Monday July 1st to Thursday July 4th).

I have one economics article I wanted to share though before we move into econometrics. It’s by Kathryn R. Johnson and Claudia Persico and is entitled “Broadband Internet Access, Economic Growth and Wellbeing”. It paints an optimistic picture of high speed internet access’s effect on overall well being. Basically, they study what happened between 2000 and 2008 when access to high-speed broadband internet in the U.S. significantly increased. They find that it led to a 1 percent reduction in county suicide rates for every ten percent increase in broadband access, or an elasticity of suicide with respect to broadband access of 0.1 (inelastic). They attribute this to economic benefits like reductions in poverty and reductions in unemployment rates among the working age population.

On the same day I read that, though, I found this article. A remote tribe in the Amazon suddenly got access to high speed internet access thanks to Elon Musk’s starlink system. The effect was that the community’s men now spends all their time online on social media, playing video games, and becoming addicted to pornography. Not saying the two articles contradict one another at all — just kind of interesting to compare them.

Econometrics

A lot this week on papers related to identification of causal effects in the context of SUTVA violations (e.g., interference), platforms, and general equilibrium related issues. The train keeps going. I’ll start with the network and spillover stuff, then move into other things.

First, I found this review article by Áureo de Paula on the econometrics of network. It’s from 2016, and I’m not quite sure where it was published. But I also found this edited volume, of the same name, which came out in 2020, is available in pdf form online which was edited by Áureo de Paula, Elie Tamer and Marcel-Cristian Voia, and which was published in 2020. You can also buy it on Amazon for around $142. I’m assuming that it’s online because the authors wanted to share the pdf. But, to the point, if you’re interested in the econometrics of networks, there’s two sources that I suspect are great starting points.

Edward Kao graduated from Harvard’s doctoral program in statistics a few years ago. He studied under Don Rubin. And he is now a research staff member in the AI Software Architecture and Algorithms group at MIT Lincoln Laboratory. He works on a variety of topics related to causal inference and networks and someone sent me his dissertation entitled, “Causal Inference Under Network Interference: A Framework for Experiments on Social Networks”. Here’s a little bit from the abstract:

“This thesis develops a causal framework and inference methodology for experiments where interference takes place on a network of influence (i.e. network interference). In this framework, the network potential outcomes serve as the key quantity and flexible building blocks for causal estimands that represent a variety of primary, peer, and total treatment effects. These causal estimands are estimated via principled Bayesian imputation of missing outcomes. The theory on the unconfoundedness assumptions leading to simplified imputation highlights the importance of including relevant network covariates in the potential outcome model.”

Guido Imbens has a lot of new material in 2024. Mind blowing productivity. Here are a few:

• “Multiple Randomization Designs: Estimation and Inference with Interference” (with coauthors)

• “Causal Clustering: Design of Cluster Experiments Under Network Interference” (with coauthors)

• “A Design-Based Perspective on Synthetic Control Methods” (with coauthors, published Journal of Business, Economics and Statistics).

The synthetic control one is on my reading list.

Evan Munro, Stefan Wager, and Kuang Xu also have a 2023 working paper related to experimentation in market equilibrium here.

Ever wonder what a standard error is? Jeffrey Wooldridge takes a stab that question here. Andrew Gelman here. And James Powell here. All published in the Journal of Econometrics.

David Card, Stefano DellaVigna, Chenxi Jiang & Dmitry Taubinsky have a new working paper at NBER also on decision times at journals, continuing a long run Card and DellaVigna have been doing economics journals stretching back over a decade. I haven’t read the paper, and from the abstract I’m not sure if this is a theory paper, a lab experimental paper, or if it’s using data from journals (as they’ve done in the past). But that’s why I post it — this is stuff I’ve been meaning to read.

This week I taught on unconfoundedness at Torino in the Collegio Carlo Alberto (as well as regression discontinuity design towards the end of the week). Because Guido Imbens and Yiqing Xu had a new working paper about “what we’ve learned since LaLonde 1986” (I think that may be coming out soon in Journal of Economics Perspectives actually), I was thinking a lot about Guido and Yiqing’s discussion about the LaLonde data and the Dehejia and Wahba subsample suffering from severe overlap problems and that since unconfoundedness methods like propensity scores require strong overlap conditions, one explanation for failing to replicate the experimental result was because the overlap problems were so bad. A lot of the subsequent literature focused, therefore, on a variety of ways to address overlap problems in data, one being through trimming. But trimming can create its own biases (called trimming bias) depending on the distribution of potential outcomes in the trimmed out sample (and how they different from that in the remaining sample). I really hadn’t seen a paper before now discussing that bias, but Ma, Sant’Anna, Sasaki and Ura have a new working paper on it entitled, “Doubly Robust Estimators with Weak Overlap”. I copied and pasted the entire article into ChatGPT-4o and told him my audience is above average interested in causal inference and could he write five sentences summarizing the article without using flowery adjectives (restrained, composed, simple, straightforward) and to make reading the article seem compelling. Here’s what he wrote:

ChatGPT4o: This paper introduces a new class of doubly robust (DR) estimators for estimating treatment effects, particularly when there is weak overlap in covariates between treatment and control groups. The authors' method involves trimming observations with extreme propensity scores and applying a bias correction to handle the resulting biases. This approach can be applied to various research designs, such as unconfoundedness, local treatment effects, and difference-in-differences (DiD). Simulation results show that these estimators perform well in finite samples, aligning with their theoretical properties. The proposed DR estimators offer a robust tool for causal inference, especially in settings with weak covariate overlap, making them valuable for researchers in many fields.

Artificial intelligence

Now on to the artificial intelligence articles. This is behind the paywall for the subscribers, but consider becoming one!