revisiting Ohio exit poll: NW Ohio off by 9%

[chorti]

I was reviewing the Ohio exit poll detail and noticed a couple interesting things there. <http://www.cnn.com/ELECTION/2004/pages/results/states/OH/P/00/epolls.0.html> The first is about the late deciding vote. There seems to be an impression that the late vote must have broken for Bush because otherwise the projections of Sam Wang and others would have been true. But that is not the case, of the 9 percent who decided in "the last 3 days or today" (Nov. 2), it broke 62 to 38 for Kerry. Almost at the hypothetical two-thirds incumbent rule, which Wang discussed. Of the ones (91%) who decided more than 3 days before the election, Bush won 53 to 47. (Remember this is the "final" CNN exit poll that some say was changed just before the last report.)

In terms of size of community, I think the samples must be goofed up but the one that is big enough for there to be a good sample is the Suburban vote. Bush won but only 51 to 49. This constitutes 49 percent of the sample. I'm not quite sure what the poll considers suburbs but it seems like the "official" results would put Bush at a much higher percentage than that.

OK, and then the regional poll. It divides the state up into 5 regions. I tried to match the regions as best I could and I think these are pretty close. The Central region, including Columbus and the southeast, broke much more towards Kerry (6%) than what the exit poll suggested. This can probably be explained by the massive turnout in Columbus which the exit pollsters did not plan for in their design. The southwest broke about 3.5% more towards Bush, but some of this discrepancy can probably be explained by a larger than expected turnout for Bush in the Miami valley. This is probably within the margin of error anyway.

But then there is the northwest. The exit poll says that the NW broke for Kerry 52 to 47. The final results show that the NW broke for Bush 56 to 43! What's up with that?! Maybe the pollsters had a small sample or a bad sample - but as a rule, you don't put results out there that are that unreliable. There are a bunch of these that appear suspicious to me, Sandusky, Auglaize, Mercer, Richland, Van Wert, Allen, Putnam, Wood, but I will follow up later on specific counties. Here's the data

EXIT POLL
VOTE BY REGION (% of turnout) BUSH KERRY

Cuyahoga County (12%) 35% 65%

Northeastern Ohio (27%) 50% 50%

Central Ohio (20%) 58% 40%

Northwestern Ohio (15%) 47% 52%

Southwestern Ohio (26%) 55% 45%

This is the data that I came up with, using official updated results

ACTUAL RESULTS
VOTE BY REGION (% of turnout) BUSH KERRY


Cuyahoga County (12%) 33% 67%

Northeastern Ohio (27%) 47% 52%

Central Ohio (20%) 52% 47%

Northwestern Ohio (15%) 56% 43%

Southwestern Ohio (26%) 58% 41%


List of Counties in NW Ohio (Kerry%, Bush%)
Allen 33% 66%
Auglaize 26% 74%
Crawford 36% 64%
Defiance 38% 62%
Erie 53% 46%
Fulton 37% 62%
Hancock 29% 70%
Hardin 37% 63%
Henry 34% 66%
Huron 41% 58%
Logan 32% 68%
Lucas 60% 40%
Marion 41% 59%
Mercer 24% 75%
Ottawa 48% 52%
Paulding 37% 63%
Putnam 23% 76%
Richland 40% 60%
Sandusky 44% 56%
Seneca 41% 59%
Union 29% 70%
Van Wert 28% 72%
Williams 35% 65%
Wood 46% 53%
Wyandot 34% 66%

[meganmonkey]

1. I think by 'late vote' they mean people who voted later in the day

not people who were undecided until the last minute.

[chorti]

3. oops

OK, so the people who decided at the last minute broke for Kerry 62 to 38. This makes is harder to believe that Bush won Ohio by over 2 percent.

[Straight Shooter]

5. I thought the late voters in Warren County

supposedly voted overwhelmingly for the shrub.

:shrug:

Nothing at all makes sense in this election, statistically speaking. Next thing you know, bush will overcome the laws of gravity, heal the sick, raise the dead, make the little girls talk outta their heads.

He's the one :eyes:

[myschkin]

2. Maybe you send this to Arnebeck...

[e-mail]

[ClintCooper2003]

4. That's not a 9% shift, it's a full-on 18% shift!

[chorti]

6. minor correction

I had a minor error in my formula. I'll add decimal points too.

ACTUAL RESULTS
VOTE BY REGION (% of turnout) BUSH KERRY

Cuyahoga County (11.9%) 33.1% 66.4%

Northeastern Ohio (26.8%) 47.5% 52.0%

Central Ohio (20.1%) 52.6% 46.9%

Northwestern Ohio (15.3%) 56.3% 43.3%

Southwestern Ohio (26.0%) 58.5% 41.0%

[Lerkfish]

7. the magical 51%

as prescribed by programmers.

[ewulf]

8. It hard to know what kinds of conclusions you can draw

from the adjusted exit poll numbers. There are a number of uncertainties as to exactly what was adjusted and where. If the adjustments were made to the overall state numbers, and not based on the returns by region, county, or precinct (as MIGHT be evidenced by the discrepancies between the adjusted poll numbers and the tallied votes), it looks like you have a very interesting case. Any discrepancy in favor of bush in the tallied numbers as compared to the exit polls would correspond to an even greater discrepancy between them and the real exit polls: this would imply fraud (although, granted, the sample sizes are smaller so the the errors go up, they only go up by the square root of the change in sample size: a 25% sample of the state would just have twice the error of the entire state sample).

The 7% discrepancy in favor of Kerry when one compares the actual results to the adjusted exit polls in Central Ohio could partly be the result of the overall readjustment: there is also a 2% benefit for Kerry in the Northeast and in Cuyahoga County (perhaps Cuyahoga wasn't the locus of major tabulational vote fraud ) which could be more readily explained by the adjustment process. It does look like the Northwest, and to a lesser extent the Southwest, deserve our attention.

Assuming that the discrepancy wasn't just because you swapped some precincts between Central and Northwest (this should be checked), there were 9% fewer votes counted for kerry than were shown by the adjusted exit polls in the Northwest. If we assume, very conservatively, that the overall "adjustment" in the exit polls was 2%, then the overall discrepancy is 11%. Since 15% of the full 2,020 respondents were in the northwest, that gives us a sample size of 300. 1/SQRT<300>=0.057, so the 2Sigma margin of error is 11%, there is a roughly 2% chance that this could happen due to sampling error alone.


Boy, sure would be nice to have the original Edison and Mitofsky data... sooner than three months from now.

[ewulf]

9. On second though:

Is it the change in the margin, and not the numbers for an individual candidate, that one compares with the margin of error?

If it is, then the shift was ~22% in the Northwest (twice what Kerry lost). With a margin of error of 2Sigma= 11%, this is really, really, unlikely.

[chorti]

10. one tail or two tail

I am trying to remember. I think that you use the 9% difference, not the overall swing of 18%.

All of the detailed exit poll data was also adjusted to fit the final results. I checked about ten of them, including this one, and they all are consistent. So it's only a 9% difference not 11%.

[ewulf]

12. Your probably right about the first part,

its the difference, not the swing that is important as it is actually what you are sampling. I guess you could go and say that the probability of the adjusted numbers varying on the bush side of the tail of the distribution would be ~1%, then.

As far as your second point, I'm not sure what you mean. My "11%" number was from taking the rough difference between Kerry's exit poll numbers and the tabulated numbers, 9%, and adding a certain amount to that (2%) to account for how the exit polls were adjusted in favor of *. If all the data in the exit polls were adjusted to make the total state data jell with the final, tabulated results, then this would be the right approach, unless I'm missing something.

[RaulVB]

11. Don't tell me...

Statistically impossible.