Thread by @BiggerDataSci

Eric Grenier does not have a degree in math or statistics. Is that why he misinterprets data for personal promotion? I'm not sure. But let's read this together.
cbc.ca/news/politics/…

We're going to talk about this table, and what Grenier thinks it means.

I openly invite @VoiceOfFranky to criticize my take.

There are two obvious facts that pop out, and Grenier catches both of them. But he misinterprets their cause, and fails to discuss their significance.
1. conservative GOTV protest vote
2. strategic vote switching to Libs

This table presents aggregate vote, so it misses demographic effects, such as the massive rural/urban split. As such, it misses the fact that the conservative protest vote - a rural phenomenon - was largely offset by an urban pro-Lib counter-protest involving strategic voting.

These two biases resulting from voter behaviour on voting day is missed in all 14 polls. All 14. Moreover, agregating the voting patterns across the country hides the depth of bias, which has its roots in population density & relative importance of Con talk radio and Facebook.

Grenier argues that pollsters got it right, mostly. Well, the thing they got wrong is the only thing that matters, and it's the thing they ALWAYS get wrong. Voting day polling biases resulting from contagious voter behaviour on (1) turnout and (2) strategic day-of-vote switching.

Famous results pollsters got very wrong, from similar effects:
Clark over Dix
Trump over Clinton
Ford over Horvath

So rural/urban contagious behaviour is a massive source of consistently biased polling error that decides elections.

And this election exhibited the same biases.

If you are trying to predict who will form government, then the aggregated popular vote is irrelevant. You must predict riding-level turnout & that includes estimating protest contagion in turnout & switching. If you don't get rural/urban sources of turnout bias, you will fail.

Because of all this, the appropriate metric of polling performance is seat count, predicted vs actual. Aggregate vote is not only irrelevant, it leads to bogus interpretations, such as "we correctly predicted feet-of-clay Trudeau losing his majority".

No. You failed to predict rural con turnout contagion and you failed to predict the urban liberal counter reaction. And covering up that detail allows you to pretend (a) your models are good and (b) Scheer won the popular vote. Both propositions are fake news.

Moreover, promoting Abacus over Ekos based on an irrelevant bulk metric shows an interpretive bias resulting from a conflict of self-interest. It's also fake news.

Show us seat count, or sit down, because that's the metric that matters.

CBC's entire take on this election is garbage. "Trudeau lost his majority due to feet of clay, just as we predicted" is fake news that misinforms Canadians of the facts of what happened ... to our own detriment.

"Urban Canada rescued Trudeau from rural conservative wrath" is the headline.

"Central & coastal Canada ward off fascist northern prairie propaganda" is the headline.

"Pollsters fail to predict urban/rural turnout contagion, yet again" is the headline.

If we don't look at the data honestly, we can't have a productive discussion about what the election results mean to the future of this country.

Speaking of self-promotion, here's another guy without a statistics degree busting into my TL with fake news:

A massive rural blue wave jacked on Facebook fake news crashed into an urban red wave fearful of rising American Republicanism in Canada. Foreign-owned & con-controlled #cdnmedia is denying its role in the former & refuses to discuss the latter.

That's the true story of #elxn43