I have been more disciplined in this history of economic thought class and managed to get through Mill and Bastiat for the first time. It wasn't easy teaching them compared to some of the earlier people. There aren't clear models to teach with them. It's more like you're seeing the emergence of certain approaches to economics being worked out to their logical conclusions, mostly related to utilitarianism and exchange, though they come to different conclusions.

Today marks the 60th U.S. presidential election! Over the past 100 days, my students and I ran an experiment using GPT-4o mini to predict each state’s winner daily based on 100 news articles. Here’s how we did it (pics spread throughout):


1.GPT-4o mini read the articles, then crafted “future narratives” as if a journalist was reporting post-election.
2.Each narrative included state winners, which we multiplied by electoral college votes to tally Harris vs. Trump totals.

One of the things I love about teaching the history of economic thought is that it's all the economics without any of the economics profession. It's like most of the time, you really can't even figure out who paid these people's paychecks. You aren't thinking about where the stuff was published, or all the myriad of little details that are make the economics profession, and it's various careerism things, completely fused with the economics itself.

This is an idea that I have been thinking about for a very long time, but which I was only able to figure out how to visually convey recently. I think maybe that this is sort of the core argument of my reasoning about the spread of causal inference in economics. It is not really accurate I think to talk in terms of "causal inference" as though it is a monolithic thing. It has many roots and many distinct meanings to many different people and I think it is alienating to insist on any one version of it.

That substack I did yesterday trying to figure which author/article was first at introducing explicit Neyman-Rubin potential outcomes notation into economics put me down a rabbit hole. Heckman is where I came down (Heckman and Robb 1985) but the cites he tends to give for his potential outcomes references tends to conflate multiple things.

I'm really very different when I teach causal inference because I make us use spreadsheets to calculate individual treatment effects off of potential outcomes, make us write down treatment assignments based on a hypothesized treatment assignment mechanism, use the switching… …equation to select a potential outcome based on that treatment assignment, and then use all of that to emphasize the difference between a potential outcome and a realized outcome.

I think Goldin and Katz idea of skill biased tech change is the right paradigm to use when understanding ChatGPT as a disruptive force. We don’t have a good understanding of its longrun effects bc all studies are short run so far as it’s so new. But consider this. 1/ @erikbryn et al have a study looking at a staggered rollout of generative AI at a firm of customer service agents; workers had access iow to a ChatGPT bot to assist them with customer complaints. As with many studies I’ve seen, big increased productivity gains. 2/

We are now t-1 days away from the first Mixtape Sessions workshop of 2023 - my flagship intro to causal inference aptly named "Causal Inference 1". Here's the link:

mixtapesessions.io/session/ci_i_f… Same pricing as always:

$1 for current residents of low income countries
$50 for students, postdocs, predocs and current residents of middle income countries
$595 for everyone else, which includes 30 min office hours if you want them.

I'm teaching a class this semester called "Issues in Economics" for undergrads, non-majors, from outside the business school. I had such an amazing experience teaching my history of econ thought class as a readings and discussion class, I decided to do it again. 1/ I googled "how many pages should I assign each week for college" and found that on average college students were reading between 30-40 pages a week in their classes. I wasn't sure if that was total pages, or per class. 2/

Just wanted to pop in and share abt some upcoming workshops at Mixtape Sessions. There's two coming in February that I wanted to first announce, then another one later. The first is my causal inference I workshops. You can see the description here:

mixtapesessions.io/session/ci_i_f… I'm currently rewriting the Mixtape. In my mind it's called "New Edition", and it's really taking a lot of willpower not to mine 1980s R&B boy band New Edition lyrics to find lines for each design. Knock on wood at least one is out there. Anyway, the workshops reflect updates.

No one probably cares about this, but while trying to learn more about inverse probability weighting with the propensity score, I wondered if you could bootstrap. But then I remembered an econometrica by Abadie and Imbens about the bootstrap failing for propensity scores 1/n So I read the paper, and it’s only inappropriate for nearest neighbor matching on the propensity score. It didn’t say anything about weighting and since all I am focused on is weighting, I wrote Abadie and Imbens and asked them about it. 2/n