Sunday links

I thought I’d do it weekend links today since I skipped yesterday. I’d been saving them up. This’ll be behind the paywall as I need to start making more exclusive content for the paying subscribers, and I think these weekend posts will likely be that. But there’s all the regular stuff I do on the weekend — economics, causal inference, mental health work, artificial intelligence, self help, and then a bunch of stuff about my cats, but that’s at the end. Hope everyone has a great day and is getting geared up for the week.

Fall playlist

But here are a few things. This is my evolving fall playlist which I’m really digging. It’s coming together I think. Expect it to keep growing. I’ll try to cap it at around 2.5 hours though. I think the optimal playlist is around that length or less. It really needs to be an album, not some epic scroll. It should be a mixtape. Anyway, this becoming my new one and I love this first song “Ex” by Kiana Ledé. I’m going to send it to the girls as soon as I hit publish on this.

Mental Health

I mentioned this the other day in a substack about principal stratification and partial monotonicity violations, but medical noncompliance for people with schizophrenia is very common as this article notes. I think these small sample studies are useful (n=110 with schizophrenia or schizoaffective disorder) but it’s time to go beyond them. Some countries now have administrative data available to researchers. And I think it’s time to generate risk scores using machine learning modeling and predict the likelihood that a person will not comply. Where we have administrative data, just having propensity score measurements of baseline risk — Prob(Non-Compliance|X) — would go a very long way. Throw a neural net at it and build a dashboard that can be dynamically updated throughout a person’s time. I suspect that we aren’t taking advantage of enough of the data we have. We are living in a very different period where there is a massive amount of data for these kinds of things.

And then if there are interventions aimed at compliance (or maybe even you’re interested in the effectiveness of the antipsychotics themselves), we should consider estimating individual treatment effects using causal forests. But keep in mind the unconfoundedness assumption is underlying both double debiased machine learning and causal forests. If you’re not comfortable using propensity scores to estimate causal effects, then I don’t think you’re going to like double debiased ML or causal forests since they also depend on unconfoundedness or “conditional independence”.

But still, maybe you’re one of those who thinks if you throw 1500 covariates at something, you’re willing to defend unconfoundedness, or you have some deep understanding of the selection mechanism and think selection on observables is good enough. Or maybe you’re just okay defending partial unconfoundedness with respect to Y(0). Whichever, you may want to consider combining predictive risk scores for predicting noncompliance with something like DML or causal forests to then estimate the individual treatment effects which you might could then use to predict who is likely to benefit from a given intervention to reduce noncompliance (assuming it works in the first place — which that may itself have deep heterogeneous treatment effects too, and perhaps the reason someone with schizophrenia is quitting the medication voluntarily is because they know it isn’t working). I want to see more on this done, so I’m just thinking out loud.

Self help and relationships

According to this article, the five stages of uncoupling are differentiating, circumscribing, stagnating, avoiding and terminating. Here’s the TL;DR.