Post

Chris Prosser
@caprosser.com

Using ~8.5k election-poll-party observations from 372 elections in 32 countries, we find a persistent partisan asymmetry: polls tend to overestimate the left relative to the right. On average across elections, polls the month before an election overstated the left by about 1.4 points

Jack Bailey
@jack-bailey.co.uk

Can you say whether this is bias in favour of the left or bias against the right/others? I mentioned this to Stuart at EPSS, but, given that the data are compositional and must sum to 100%, either could be true but the implications are different!

Chris Prosser
@caprosser.com

I don't think we can say - we're agnostic to the cause of the errors in the paper because I don't think we can test it from polling data alone, and I think you need a causal story to say whether its driven by one side or the other, with the other bit driven by compositional processes.

0 reposts1 like

Jack Bailey
@jack-bailey.co.uk

Makes sense. Another thought I had was whether you could use log ratios to get rid of the compositional element in the data and see what happens.

Chris Prosser
@caprosser.com

Yeah I had similar thoughts at one stage because I've gone down some log ratio rabbit holes for other purposes, but the data is a nightmare from a compositional perspective because there are inconsistent choice sets between elections and lots of zeros (particularly poll shares for 'other' parties).