Reply #29: A few thoughts

[Febble]

First is that the within-precinct error calcs are an analytical tool (to find the source of the error) not, AFAIK, part of the projection process. So the state-level projects will be subject to various forms of error, and also to various forms of correction of apparent error in the inputs (age-sex-race adjustment for non-responders, for example; and of course the notorious re-weighting in light of incoming vote-returns).

I've spent more time looking at what Freeman calls the Precinct Level Discrepancy (PLD) than elsewhere, as it would appear that it was at this level the discrepancy was greatest. Precinct selection does not appear to have been a problem and the effect size of the discrepancy appears to be smaller at state level (i.e. the discrepancy between the close-of-poll projections and the "final" projections) than at precinct level, suggesting that whatever weighting was applied to the raw data was effective at reducing the precinct level discrepancy.

So the precinct level analysis is what we want to look at - but we also need to see it nested within state level factors, for example gross state margins. It does appear that the precinct level discrepancies were greater where the state was bluer. A plot of mean WPEs, given in the E-M report against the counted margin shows quite a marked relationship.



I found a similar relationship in 1988, but the opposite relationship, oddly, in 1992. Note that 1992 was the big Perot year, so patterns of "reluctance" might have been different. Or other patterns of bias. Or indeed patterns of fraud.

Which is not exactly addressing your question, except that the factors you cite (and others) are sometimes state wide (weather; swing state status; state margin; distance of pollster from polling place) and others more precinct-specific (interviewer characteristics; interviewing rate; voting technology). And others are more generally demographic (rural/urban) but collinear with certain state characteristics (vote margin).

Which is still not addressing your question, which I think can only begin to be answered by a multiple regression model that includes both state-level and precinct level variables predictor variables and a precinct level measure of PLD as dependent variable (preferably NOT the WPE!)- which is what my original paper actually called for.