Say Trump has 55% of men saying they'll vote for him, and 45% of women. That's sentiment, but what will the final Election Day ratio of men/women voters be? If it's a 50:50 split, the vote is dead even. But for every 1% more women that vote, Trump's vote share falls by 0.10%; for every 1% more men, it goes up by 0.10%
Here, have a spreadsheet. In fact, have a primitive voting model, because I just kept going once I got started.
It's not very exciting because gender is the only variable, and the scaling is linear. Each of the rows 5-14 is a single model. What Nate Quartz does is take one of them and match it to historical data; the highlighted row 8 is his likely choice since it matches the 2016
turnout with 52% women. His probability that Trump wins would be in column D.
What he
should be doing is running probability across multiple models, and weighting the outcome on their likelihood. This is where the "art" portion of polling comes into play: predicting what the likely electorate will be. My "Probability this happens" column has arbitrary numbers, just to keep things nice and linear, but if you play with the weightings you can see how reality diverges from polls if you guess the electorate wrong. In this case, weighting multiple projections gives me a flat 50% for Trump on Election Day.
Do the same for race, education, age, party affiliation, etc and you get a full model; all you have to do is plug in new numbers in row 3 from fresh polls and let them update. If you want to get fancy, you can plug in the real gender gap from
here. If you want to be professional, you replicate that 50 times with the unique characteristics of each individual state.
My original 0.20 divergence looks trivial, but that's the difference on
one variable, gender. When you're compounding those bits of errors across every possible demographic variable, they start to add up.
Nate thinks he's clever because he puts the average of multiple polls up in row 3 instead of a single number, but that doesn't erase the projection error. In fact, it introduces a
completely new error, when polls have different methodologies and slants but you average them together as if they were homogeneous. But that's a different rant for some other time.
