Thread by @MattHindman

I am seeing a lot of skepticism about #cambridgeanalaytica’s statistical methods that, while correct, mostly misses the point.

Getting good data on 50 million FB profiles would be a big deal _even_if_ CA’s psychometrics were bunk.

This is not about building an ultra-clever statistical model, it’s about being able to AVOID having to build a fancy model in the first place.

Modern campaign strategy depends on knowing who supports you. For this you could use the partisanship models from voter file vendors. You could conduct a poll.

Or you could just knock on every. single. door. and ask who they are voting for.

In the (few) campaigns with lots of door IDs, a crude model using that data can outperform state-of-the-art models from the best vendors.

It’s way more accurate to ask Jane Voter directly instead of guessing her views from her sex, race, age, etc. (yes, even with lying, etc.)

This why the Cambridge Analytica story is a big deal. Super-simple models -- say, guessing your partisanship is just the average of your friends’ partisanship -- work great, IF you have everyone’s data.

.@PatrickRuffini confirms that partisanship models based on Facebook data are just as reliable --"if not more reliable" (!!!) -- than voter file data.

And FB has data on millions of people that the voter file can’t help with.

External Tweet loading...
If nothing shows, it may have been deleted
by @PatrickRuffini view original on Twitter

LOTS of caveats apply. We don’t yet know exactly what data CA took and how they used it. Microtargeting can backfire (read @eitanhersh). I’m officially agnostic on psychometrics, and of course other campaigns have used Facebook data too.

Still: if -- IF -- the Facebook data let the campaign directly measure millions of political views they'd otherwise have to guess about, it is a big deal and a big jump in accuracy.

That's true even if you think psychometrics is just hocus-pocus.