Skip to content

Statistical Rethinking (McElreath) discussion group: proposed outline

David Reinstein edited this page Mar 7, 2023 · 16 revisions

(See this overview discussion, which I should probably move that into the current wiki.)

Discuss organization, installing relevant software packages, goals and interests

Chapter 1: The Golem of Prague (1-2 weeks)

Supplement with McElreath video on 'what standard statistical tests actually do' vs what people think they are doing.

2a: Small worlds, part 1

Garden of forking data (I suggest we skip the 'coding of building that diagram')

  • we really discuss our intuitive understanding of Bayesian updating and the examples, and cool stuff about it

Building a model (the globe-tossing story)

  • Do the estimating and plotting with the Kurz implementation, incorporate a succinct version of this into our own methods book/notes

Components of the model

  • Watching/coding the cool updating
  • Consider implications for 'what can we learn from small amounts of data' (and the VOI stuff)

2b: Small worlds, part 2, 'making the model go'

Let's dig into Grid approximation, Quadratic approximation, and (a bit) MCMC, and how these work and how to code them

3: Sampling the Imaginary

A. From a grid-approximate posterior)

B. To simulate prediction

This might be 1 or 2 weeks, I'm not sure

4a: Geocentric Models

Normal distributions

A language for describing models, the globe-tossing model

Gaussian model of height (first part)

4b (Geocentric Models): Linear prediction

Linear prediction:

  • This adds a predictor variable to the above model of height
  • Implements quadratic approximation and interpretation of results

4c: (Geocentric) Curves from lines

Curves from lines: Polynomial models and Splines

5: Causal inference ... "The Many Variables & The Spurious Waffles"

DR: This material is interesting, but I think there are better resources on causal inference we might want to supplement this with. Still, given the importance of the issue and how he returns, to it, I think we probably should devote 1-2 weeks to causal inference.

Or could this material be skipped and deferred to a separate causal inference reading group?

6: The Haunted DAG & The Causal Terror

7a: Complexity, regularization, prediction: "Ulysses' Compass"

7b...

Lots of material here, not sure how we should split it up but I think it needs at least 2 sessions

8a: Interaction/moderation: "Conditional Manatees"

1-2 sessions?

(DR: perhaps not to be confused with mediation, which is very hard to measure.)

9: Markov Chain Monte Carlo

0-1 sessions

DR: Let's just skim this and mainly treat it as black box (as Nik suggested)

10: Big Entropy and the Generalized Linear Model

This gets rather mathematically technical but let's try not to be scared.

1 sessions/weeks for this one?

_Nik: idk how useful the maxent stuff would be. (You can use it to "justify" things like priors, but I think a better response to detractors is just to ask what values they think are plausible a priori and assess sensitivity to prior choice that way, or do an “expert elicitation” type thing])

11: GLMs for non-continuous data (Binomial, Poisson, etc.) "God Spiked the Integers"

Also seems mathy but worth it. 1 week, not as comprehensive

12: Extensions of the above Glms for overdispersion, ordering, etc... "Monsters and Mixtures"

Also seems mathy but worth it. 1 week?

13: Multilevel (aka 'mixed') models: "Models With Memory"

DR: This is ~the chapter I'm most interested in. Any way we can get here earlier?

This seems very important to our work, so maybe give it 2 weeks.

14: "Adventures in Covariance"

This seems to go in a bunch of fancy directions. Probably needs at least 2 weeks.

  • Advanced varying slopes
  • Instruments and causal designs (what's the connection to the varying slopes thing?)

# 15: Missing Data and Other Opportunities

A useful application. Hopefully if we get this far we'll be so smart we can do it in 1 week?

16: Bespoke examples ("Generalized Linear Madness")

17: Some editorializing by the author ("Horoscopes")

This doesn't really need a lot of work but we could use this as an opportunity for a recap discussion