Thread by @anthonyjwells

A longish thread about trying to explain the differences in party support - primarily Labour - in the polls.

It is often very difficult to get to the bottom of the differences between polls. There is rarely one big, easy simple answer, it's normally a bit of lots of little, different to explain things. Hence stupid & simple explanations prosper.

One current cause seems to be different approaches to past vote weighting. Most pollsters* weight by how people voted at the last election (ensuring that the sample has 44% people who voted Conservative in 2017, 41% Labour in 2017 and so on)

Problem is that people don't necessarily report this info very accurately to pollsters two years on - people forget, people lie, people regret their vote and so on.

This isn't a new issue. The original academic work on false recall was done in the 1970s and it was a topic of huge debate among pollsters in the 1990s. How much impact it has wavers over time though, as depending on levels of false recall it doesn't always make a difference.

I wrote about this earlier this week. YouGov asked panellists how they voted back in 2017, stored the data and uses that to weight. When we ask the same people afresh, the proportion of people reporting they voted Labour drops dramatically.
yougov.co.uk/topics/politic…

If lots of people who actually voted Lab in 2017 aren't admitting it to pollsters, and pollsters are weighting Lab to their actual 2017 %, you'd end up overstating them (because your sample would have 41% people who admit having voted Lab, *plus* those who did but don't admit it)

In the experiment we ran, weighting by how people voted in 2017 using data **collected in 2017** seemed to produce Labour support about three points lower than weighting by how people voted in 2017 using data **collected now**.

YouGov use 2017 recall data collected in 2017 to weight their data, most other companies who weight using 2017 recalled vote data collect it now. This could explain several points of difference.

A 2nd cause is likely to be the question prompting - that is, which options do pollsters list when they ask the question. One pollster might say "Would you vote Con, Lab, Lib Dem or another party?" another company might say "Would you vote Con, Lab, Lib Dem, Green or Brexit?"

The impact of this is most obvious for the Brexit party. Ipsos MORI and BMG do not currently include the Brexit party in their prompt and consequently show significantly lower support for the Brexit party.

However, there is also a difference between companies in whether the Green party are included in the main prompt. YouGov do include the Green party & given they draw voters from similar groups I expect this may have a knock on effect on Labour's share.

Those are two causes that I can put my finger on, but it is likely there are others that are not so easy to identify. These two are unlikely to explain it all.

For example - polling companies use different weights. YouGov use political attention when other companies do not. Ipsos MORI use newspaper readership; Survation use income. Most use education these days, but ComRes do not. Hard to tell what the impact is

One thing the differences are probably NOT down to is turnout models. You may remember that in 2017 pollsters adapted complicated turnout weights that weighted down younger voters and understated Labour support. Companies who did do that in 2017 have now dropped those methods.

Most people's turnout models now have pretty similar impacts, and are overwhelmingly based on how likely respondents themselves say they are to vote, rather than making assumptions about age.

So short answer is that the most likely explanation for the differences in VI numbers are fairly technical differences in how polling companies do their weighting or ask their voting intention questions.

...but that took 17 tweets to explain, when "Pollster A is biased and run by lizards, I only listen to Pollster B who is the most accurate" takes just one.