Polling: Unskewed Polls and comments on polling (moved from Harris Thread)

A place to talk about polling, not the latest polls, but poll taking and balance.

The narratives from Trump and his tribe are that any poll showing Harris-Walz in front are fake polls and they’ve changed the methodology etc etc.

I know polls are not indicative of final results, but I remember the rise of the right-wing “Unspun Polls” (probably either 2008 or 2012 - not sure). That exercise in flat-out denial didn’t turn out so well.

It’s key, though, to their “the election was stolen from us!” narrative. Seriously, they’re greasing the skids for the skullduggery they have planned. “How could we have lost when we were up—bigly!—in every honest poll?”

Wasn’t it in 2012, with all the chatter about how if you “unskewed” the polls, Romney was leading? I think @adaher was a big proponent of that argument.

It boiled down to : “That can’t be right! Obama can’t be that popular!”

Well, he was.

That was it - Unskewed Polls. Thanks!

Aged like milk, that did.

There’s a germ of truth in there, though.

Sitting here quite a few years later, many modern polls are “unskewed” to some extent. Really, they “weight” by certain criteria, which is basically unskewing by a more flattering name.

The difference is you ‘weight’ by pre-determined criteria (education, or age, or whatever), not to get the result you want.

Getting representative samples is very difficult these days. People don’t answer their phones. And those who do are generally not representative of the voting population, so some pollsters will ‘weight’ certain responses more than others so that they can get something representative.

Tricky business, though. And to be avoided as much as possible, which it may well not be given the difficulty these days in getting representative polling samples.

And definitely don’t make up numbers to fit the answer you want to hear. That’s just right out.

If you mean weighting responses is to be avoided as much as possible, that’s absolutely not true. For the reasons you mention, your collected results will not match the voting population, and if you want an accurate estimate of the actual vote, you must weight it.

Some of what the “unskewed” folks did was un-weight the results, so if 60 of 100 respondents favored Trump, that was 60% - never mind that 95% of the respondents were white rural voters. That’s just as wrong as using incorrect weights.

You’re right that determining the proper weights is very tricky business. That, among other things, is what sets trusted pollsters apart from hacks.

Yep. Properly determining weighting parameters is why real pollsters get paid so much money. Not very many people can do this properly.

I thought the polls were screwed up when Biden was losing and I still think the same now that they show Harris ahead. Not that she is losing (I did think Biden was going to win), just that everything is off. If you looked at the crosstabs on some of those polls every single one shows ridiculous things like Trump winning the young vote and Biden winning the boomer vote. Once you dig down into the details none of the polls have made any sense for the whole year.

Like the recent Siena (highly rated) polls out of PA, MI and WI that show Harris up 50/46 in all three states:

  • Only winning age 18-29 by +15. Biden 2020 was in the +30s.
  • Only winning age 30-44 by +6. Biden 2020 was in the +25s.
  • Winning age 65 and over by +13. Biden 2020 lost them by 11 points.
  • Losing white voters by only 1. Biden 2020 lost them by 13 points.
  • Only winning Black voters by +68. Biden 2020 was +82. Also she’s only getting 81% of the Black vote. Biden 2020 was in the over 90% range.
  • Other races (Latino, Asian, Mixed, Middle Eastern, etc) she only wins by +5. Biden 2020 was in the +30s.

Make that make sense…

Here’s a really good article from Pew Research (who know a lot about the topic). A few interesting points:

  • Different polling organizations conduct their surveys in quite different ways.
  • The real margin of error is often about double the one reported. The margin of error addresses only one source of potential error: the fact that random samples are likely to differ a little from the population just by chance. But there are three other, equally important sources of error in polling: nonresponse, coverage error (where not all the target population has a chance of being sampled) and mismeasurement. Not only does the margin of error fail to account for those other sources of potential error, it implies to the public that they do not exist, which is not true.
  • All good polling relies on statistical adjustment called “weighting” to make sure that samples align with the broader population on key characteristics.

That’s saying the same thing from the other way round.

In an ideal world, the polling sample will be representative.

But in the real world, this is effectively impossible (and has become more so in recent years), so weighting becomes necessary. But it also increases margins of error.

Those are still incorrect weights.

Just incorrect in a different way than somebody legitimately trying to get good weights and failing to do so.

You’re right, but I think missing the point of the @Great_Antibob: it’s preferred to have a polled sample that better matches the actual population.

(Edited to add: I should’ve previewed, but looks like I was reading correctly.)

Hey … what are we in the “Polling is broken?” thread? Chopped liver? :wink:

Crosstabs are really tiny sample sizes with very large margins of error.

This set of numbers has a plausible hypothesis at least - as Biden has said and said, the economy is doing “great”. But that’s “great” for people who own houses and investments (the older brackets) while the younger folks are really hurting from inflation.

This is a great thread for my hypothesis that the explanation for both Democratic over performance in the last few years of special elections, and Trump/extremist under performance in primaries, is that young, technologically savvy voters, whom I dub “Swifties”, are almost impossible to contact by pollsters, and thus these voters’ overwhelmingly anti Trump views are not being well captured by polls.

If this hypothesis is correct, then Harris is on the way to a landslide victory.

If the young folk can be bothered to vote.

Silver, G. Elliot Morris of 538, and several other forecasters have a system built into their models where “no state is an island”. Essentially, trends in polled neighboring states have been found to bleed through the borders and indicate something about what’s happening in an unpolled state.

Virginia’s lean in Silver’s analysis today is certainly being influenced by recent state-level polls in Pennsylvania, North Carolina and likely Ohio. And it’s not necessarily going by raw red/blue status. Ohio might today still poll as a comfortable Republican win, but if Trump is today getting (say) ~55% of Ohioans instead of (say) ~59% in recent past polling, that pro-Harris trend in Ohio will count some in scoring Virginia “bluer”.

This may be true, but it accuses the professional statisticians of not doing statistics.

They have all kinds of systems to weight the statistics. They proudly claim that they can accurately break down every conceivable specific population. (men with less than 2 years college, white woman in suburbia, bi-lingual, single parents, black, white, mixed race, income levels, registered voter, likely voters, and every combination of categories, etc)
So why not factor in one more category–the Swifties?

Sure, the swifties don’t answer their phones. But they do answer texts.
And it’s easy to find them: just go to college campuses!

And not just the ivory tower campuses, but also commuter community colleges , junior colleges, beautician training and trade-schools, christian bible schools, etc. Also, stand outside an iphone store ,with a clipboard and a questionaire, and collect statistics the old fashioned way.

Some of the work would require spending money for personal interviews and focus groups, but a lot could then be followed up with text messages. And as for expenses… surely a hundred field workers, each working 6 hours to interview 10 or 20 Swifties , would cost less than a single 30-second TV ad.

Statisticians want accurate numbers, and there are huge party organizations who desperately need them, and will pay for it.