Some background

My Ph.D. thesis contained some attempts to “do causal inference” with Bayesian modeling. It was interesting stuff, but the experience was disorienting. At the time, my field (political science) had very few cases where somebody did causal estimation with a Bayesian model. It had even fewer examples of anybody discussing “what it meant” to combine these things, if it meant anything at all.

I had neither the brains nor the self-sacrificial dedication to launch an academic career with this work. But I was writing a dissertation, and you try to push on a few things in a dissertation. Causal inference and Bayesian modeling are both big things that occupy a lot of brain space as you try to grok them. So I was kicking some ideas around.

This post describes the process for scraping “S.C.U.M.” score data from the Kill James Bond podcast website.

To perform the scraping, we design a functional interface that lets us easily generate sequences of data and map functions over those sequences.

Along the way, we employ a functional optimization concept called memoization. Memoization, crudely, is when when a function “caches its own return value” to make repeated calls more efficient. Think of it like a function that remembers every argument it ever saw and every value it ever computed. If that sounds fuzzy, we explain it with more detail below.

When many researchers first encounter Bayesian methods, they are nervous that their choice of prior distribution will ruin their posterior inferences. Estimation under non-Bayesian methods feels “automatic”, but Bayesian methods bear this additional risk of bad priors.¹ In this new uncertain territory, it initially feels safer to prefer flatter, more diffuse priors in your model. How bad can flat priors be if, in the extreme, you merely get the same answer that you would have gotten without Bayes?