If you want to use the larger files with posterior draws to fully incorporate measurement error in your analyses, see this explainer of the "method of composition" (essentially you bootstrap your model over the posterior draws). See www.cambridge.org/core/journal... for a review of this technique - ThreadSky

rmkubinec.bsky.social • 90 days ago

If you want to use the larger files with posterior draws to fully incorporate measurement error in your analyses, see this explainer of the "method of composition" (essentially you bootstrap your model over the posterior draws).

See https://www.cambridge.org/core/journals/american-political-science-review/article/democracy-public-support-and-measurement-uncertainty/621616CF93D45E659B758F8C3F3B0C3D for a review of this technique

Comments

rmkubinec.bsky.social•90 days ago

https://rdrr.io/github/devincaughey/CaugheyTools/man/propagate.html

rmkubinec.bsky.social•90 days ago

There are only 15 legislators that aren't included, and these people are legislators who voted on less than 25 non-anonymous distinct votes. Mostly people who were appointed to fill a seat for a short time. Specifically, those in the screenshot below.

rmkubinec.bsky.social•90 days ago

To avoid any confusion and to maximize comparability with existing estimates like DW-NOMINATE, I've released the simplest model, a 2-degree spline /w out missing data adjustment. I am happy to add other models as well (any in the plot above) if there is any interest.

devincaughey.bsky.social•90 days ago

It’s true! Though I have come to the conclusion that MOC is more assumption-dependent than is commonly realized, and it is often better to use @mattblackwell.bsky.social et al.’s “multiple overimputation” or to use Bayesian measurement-error model, which brms makes pretty easy.

rmkubinec.bsky.social•89 days ago

I haven’t tried the Blackwell method. Of course I think brms is great but the me implementation is very top heavy and demanding with big data. Something that needs more work. Also not a lot of theory on what prior to use for measurement error (ie normal or something more flexible).

devincaughey.bsky.social•89 days ago

Yeah I hear you on me(). MO is faster. The advantage of both over MOC is allowing other variables in the analysis model to inform estimates of the missing values rather than assuming them to be independent as MOC does (see the appendix to Treier and Jackman).

rmkubinec.bsky.social•89 days ago

Yep definitely limited. Me() can be made much faster—I actually re wrote a brms stan program once to do it. This area needs some methods research!

In particular it would be interesting to discuss how in a Bayesian model the posterior distribution of the scores is different than the prior estimates.

devincaughey.bsky.social•87 days ago

Re methods research, Stephen Jessee and @jerzakconnor.bsky.social have a working paper on dealing with measurement error in latent variables. It too deprecates MOC and proposes an IV-based approach instead.

rmkubinec.bsky.social•87 days ago

On point! I’ll follow up with them, thanks much.

Comments

Posting Rules

Reply