Reply #47: OK

[Febble]

I am assuming that De and Sv are not correlated (although they could be, but I won't get into that right now*) and that Df and Sf are. You are saying that Df and Sf could still still be correlated in the presence of a non-significant correlation between D and S.

Well, yes, they could. But the straightforward interpretation says that the portion of the variance in D that is UNshared with S is much larger than the portion of the variance in D that IS shared with S, and for now we are assuming that all the variance in D shared with S is fraud. So there might be a real effect, but it is swamped by the error (error in the technical sense of unexplained variance).

The non-straightforward interpretation is that there could be suppressor effects. What follows is entirely hypothetical, but if fraud was more common in marginal states, but Bush voters were also keener to participate (make their voice heard?) in marginal states, you might get no significant zero-order correlation between swing and bias, but if you added "marginal state" (sorry I'm stuck here having used up the term "swing" for swing....) to the regression model, you might find the correlation between swing and bias became apparent. In other words, having controlled for the (hypothetical) tendency for Bush voters to be keener to participate in marginal states, the underlying relationship between swing and bias is revealed.

However - this hypothesis depends on the assumption that something other than fraud contributes to the bias. In this hypothetical case, it depends on the assumption that Bush voters were less likely to respond to pollsters in non-swing states. Which is of course perfectly plausible if you believe in rBr. But if you believe in rBr, then aren't you a heretic?

And I'm not suggesting this happened - just that that's how suppressor effects can work. But it means, AFAICT, that the exit poll discrepancy is not a proxy for the magnitude of fraud. Or, if you will, that fraud is neither necessary nor sufficient to account for the exit poll discrepancy.


*For example, you might imagine that Bush GOTV efforts might encourage shy Bush voters to vote - but that they might still baulk at responding to a poll. In that case the better the GOTV in a precinct the bigger the swing - and the smaller the response rate from Bush voters. So a correlation De and Sv might be correlated, and result in a correlation between D and S - except that it doesn't seem to have done.