bob baxley



538 Polling Data

There is a scuffle going on between 538 and Huffington Post over polling data adjustments. See this tweet and this article. Reading the debate made me curious enough to check the data for myself so I downloaded 538’s data, which includes the adjustments and made the following viz to see what the trends look like.

I have defined the net adjustment as

Even after reading 538’s nicely written modeling methodology, It isn’t clear to me what independent source of data 538 has used to model/estimate the adjustment to poll data. I can imagine using past elections to model bias of pollsters, but in that case the adjustments would be fixed per election cycle–clearly the adjustments aren’t fixed.

Anyway, in looking at the data, I was surprised to see that the adjustments are applied almost uniformly across all pollsters. That is, almost all Georgia raw poll data has been adjusted towards Trump. This is also the case for almost all polls in all states (unselect Georgia to see all data).

If we only plot the adjustment per cadidate , then we see a trend that is more intuitive where pollsters have varying adjustment depending on their partisanship. What’s curious is that 538’s model always gives Trump the relative bump for all polls–whether they tend to be biased left or right.

Maybe there is an obvious reason for this that I am missing. Perhaps it has something to do with correcting for third party candidates?