James Bland
@jamesbland.bsky.social
Economist at UToledo. 🇦🇺 Bayesian Econometrics for economic experiments and Behavioral Economics
Free online book on this stuff here: https://jamesblandecon.github.io/StructuralBayesianTechniques/section.html
https://sites.google.com/site/jamesbland/
He/his
Free online book on this stuff here: https://jamesblandecon.github.io/StructuralBayesianTechniques/section.html
https://sites.google.com/site/jamesbland/
He/his
My whole research thing is making econometrics for economic experiments more accessible. If I can't tell people when not to go down the rabbit hole, what's the point?
November 28, 2025 at 11:55 PM
My whole research thing is making econometrics for economic experiments more accessible. If I can't tell people when not to go down the rabbit hole, what's the point?
I laugh too, but this is how we science. I try something. Irrespective of whether or not it works, I wrote about it.
November 28, 2025 at 11:34 PM
I laugh too, but this is how we science. I try something. Irrespective of whether or not it works, I wrote about it.
Thanks! I really hope this ends up getting published (to be honest I don't really care where). This is how science progresses, and how we make sure the PhD students go down *new* rabbit holes.
November 28, 2025 at 11:28 PM
Thanks! I really hope this ends up getting published (to be honest I don't really care where). This is how science progresses, and how we make sure the PhD students go down *new* rabbit holes.
Thanks! It helps that I am actually enjoying writing this up. I really like writing "how to" papers, and this is almost that, perhaps better phrased as "how not to".
November 28, 2025 at 6:42 PM
Thanks! It helps that I am actually enjoying writing this up. I really like writing "how to" papers, and this is almost that, perhaps better phrased as "how not to".
Yeah, I'm well aware of that, and don't want to contribute to it, especially since I have all of the empirical results now, so it will be an easy write-up. I'd hate for some PhD student to spend six months trying to get this to work only to find out that it doesn't.
November 28, 2025 at 1:59 PM
Yeah, I'm well aware of that, and don't want to contribute to it, especially since I have all of the empirical results now, so it will be an easy write-up. I'd hate for some PhD student to spend six months trying to get this to work only to find out that it doesn't.
Thanks! I really thought it was a really cool idea, but it turned out to perform really badly. I've sunk enough into this now that I'm pretty sure I will be able to write up at least a working paper on it without much more effort.
November 27, 2025 at 11:19 PM
Thanks! I really thought it was a really cool idea, but it turned out to perform really badly. I've sunk enough into this now that I'm pretty sure I will be able to write up at least a working paper on it without much more effort.
Also, if you think you can reasonably estimate prospect theory preferences from Harrison & Ng (2016), you've either got much more informative priors than I have, or you don't know what you're doing (this is what ChatGPT offered to do when I originally fed it the data).
November 13, 2025 at 11:25 PM
Also, if you think you can reasonably estimate prospect theory preferences from Harrison & Ng (2016), you've either got much more informative priors than I have, or you don't know what you're doing (this is what ChatGPT offered to do when I originally fed it the data).
In the meantime, you can read my book, which I hope makes things more accessible both in terms of computational time, and more importantly, understanding.
jamesblandecon.github.io/StructuralBa...
jamesblandecon.github.io/StructuralBa...
Structural Bayesian Techniques for Experimental and Behavioral Economics
Structural Bayesian Techniques for Experimental and Behavioral Economics
jamesblandecon.github.io
November 13, 2025 at 11:25 PM
In the meantime, you can read my book, which I hope makes things more accessible both in terms of computational time, and more importantly, understanding.
jamesblandecon.github.io/StructuralBa...
jamesblandecon.github.io/StructuralBa...
So use ChatGPT at your own peril. Will it give you the right answer? Maybe. But will it give you the right answer in a reasonable amount of time if you don't know what a reasonable amount of time is? Probably not.
November 13, 2025 at 11:25 PM
So use ChatGPT at your own peril. Will it give you the right answer? Maybe. But will it give you the right answer in a reasonable amount of time if you don't know what a reasonable amount of time is? Probably not.
The Stan program that ChatGPT first gave me would have run in orders of magnitude longer than the one I coded up myself. And for Bayesian computation, that's still a substantial amount of extra time.
November 13, 2025 at 11:25 PM
The Stan program that ChatGPT first gave me would have run in orders of magnitude longer than the one I coded up myself. And for Bayesian computation, that's still a substantial amount of extra time.
But you've got to know that these problems exist. And my feeling is that I am on the cutting edge of this. Why can I *practically* estimate hierarchical models of risk preferences? I can do it because I *know* that vectorization changes your computational time from a week to an over-nighter,
November 13, 2025 at 11:25 PM
But you've got to know that these problems exist. And my feeling is that I am on the cutting edge of this. Why can I *practically* estimate hierarchical models of risk preferences? I can do it because I *know* that vectorization changes your computational time from a week to an over-nighter,
So where does this leave me? ChatGPT can write a good-enough hierarchical model for estimating risk preferences. With a bit more understanding of the problem, it can write a really good one.
November 13, 2025 at 11:25 PM
So where does this leave me? ChatGPT can write a good-enough hierarchical model for estimating risk preferences. With a bit more understanding of the problem, it can write a really good one.
It came back and adjusted the Stan program to fit in with my suggested workflow.
November 13, 2025 at 11:25 PM
It came back and adjusted the Stan program to fit in with my suggested workflow.
I then directed it to where I (selfishly) think the cutting edge of the literature is.
November 13, 2025 at 11:25 PM
I then directed it to where I (selfishly) think the cutting edge of the literature is.
I then probed it on why it chose the priors that ended up in the code. This is something I feel like I am on the cutting edge of (priors for structural models), but it gave me a reasonable answer.
November 13, 2025 at 11:25 PM
I then probed it on why it chose the priors that ended up in the code. This is something I feel like I am on the cutting edge of (priors for structural models), but it gave me a reasonable answer.
I then asked it for a hierarchical model, because at this point it was doing really well. It even worked out that the non-centered parameterization was a good idea.
November 13, 2025 at 11:25 PM
I then asked it for a hierarchical model, because at this point it was doing really well. It even worked out that the non-centered parameterization was a good idea.
I then gave it very specific instructions where I recognized there were speed-ups to be taken advantage of. But to be honest, all you need to know is that vectorization speeds things up in Stan. Once you know to ask it that, it will do it for you.
November 13, 2025 at 11:25 PM
I then gave it very specific instructions where I recognized there were speed-ups to be taken advantage of. But to be honest, all you need to know is that vectorization speeds things up in Stan. Once you know to ask it that, it will do it for you.
I have a whole chapter in my book about this, and ChatGPT just did it when I asked. jamesblandecon.github.io/StructuralBa...
November 13, 2025 at 11:25 PM
I have a whole chapter in my book about this, and ChatGPT just did it when I asked. jamesblandecon.github.io/StructuralBa...
What I got was something that actually did the job! Like, it would run and everything. That said, it wasn't as efficient as I knew it could get, so I probed a bit more.
November 13, 2025 at 11:25 PM
What I got was something that actually did the job! Like, it would run and everything. That said, it wasn't as efficient as I knew it could get, so I probed a bit more.