« Freeman's Data | Main | The Raw Exit Poll Data »

December 01, 2004

NEP Revises Texas Hispanic Estimates

I am a bit behind on this item, but the Associated Press issued a "correction" on Monday based on a revision of the exit poll estimate of President Bush's support among Hispanics in Texas. The correction reduces Bush's support among Hispanics by ten percentage points (down from the original estimate of 59% to 49%):

The Associated Press overstated President Bush's support among Texas Hispanics. Under a post-election adjustment by exit poll providers Edison Media Research and Mitofsky International, 49 percent of Hispanics in the state voted for Bush, not a majority. The revised result does not differ to a statistically significant degree from Bush's 43 percent support among Texas Hispanics in a 2000 exit poll.

The revised BC-TX-Exit-Poll Excerpts showed that 20 percent, not 23 percent, of all Texas voters were Hispanic. They voted 50 percent for Kerry and 49 percent for Bush, not 41-59 Kerry-Bush.

I notice that the Texas exit poll results on CNN.com now reflect this correction. At 49%, Bush's support among Texas Latino voters was still 6 9 percentage points higher than in 2000 (when it was 43%), but his gains are less than the original exit poll suggested.

Several weeks ago, a source with access to NEP numbers passed along a rumor that NEP would soon revise the national estimate of Bush's support downwards. On Monday I had also noticed that a recent Washington Post article put Bush's Hispanic support at 42% rather than the 44% cited both elswhere in the Post and on CNN.com (still 44% as of this writing). Blogger Steve Sailor estimates that the Texas correction alone should reduce Bush's share of the Hispanic vote to 39% (hat tip: Ruy). So perhaps a national correction is coming as well.

You can get more background on this issue from a recent piece by the Washington Post's Fears and from bloggers Ruy Teixeira and Steve Soto. I cannot say much about how or why NEP made this correction, though I will certainly pass along any information that comes my way. If anyone at NEP is reading, I am sure my readers would appreciate an explanation of how you arrived at this revision.

One thing I can explain is the special challenge exit polls face when it comes to small subgroups like Latinos and Jewish voters. The reason is the whole issue of "cluster sampling." Exit polls must sample voters in clusters:  They randomly sample precincts first, then voters at precincts. In a cluster sample, characteristics or opinions that tend to "cluster" geographically tend to have higher rates of sampling error.

The reason is not all that mysterious. Consider the example of Jewish voters in Ohio (a demographic that once included the Mystery Pollster and still includes all of his family, so he speaks from some experience). Most Jews in Ohio live in a few suburbs east of Cleveland and few neighborhoods near Columbus and Cincinnati that cumulatively represent (at most) 3-5% of the state. If the Ohio exit poll sampled only 100 precincts, then most of the Jewish subsample would have come from, at most, 3-5 precincts. Now throw in a kicker: Orthodox Jews tend to be more politically conservative and tend to live among other Orthodox Jews in even more concentrated geographic areas within the Jewish Community. Thus, the odds are good that the exit poll sample will either over or underestimate the share of Orthodox Jews depending which 3-5 precincts get randomly selected. The same problems occurs with Cuban and non-Cuban Latinos in Florida.

So the bottom-line: The potential for error is much greater for a highly clustered demographic group. The fewer clusters in the sample relative, the greater the chance for this sort of error.

Related Entries - Exit Polls

Posted by Mark Blumenthal on December 1, 2004 at 12:10 PM in Exit Polls | Permalink


"Blogger Steve Sailer estimates that the Texas correction alone should reduce Bush's share of the Hispanic vote to 39%..."

Actually, that's an oversimplification. Overall, I think Bush's national share of Hispanics was around 39%, but the Texas correction alone would knock 2 points off Bush's national share of Hispanics just by itself. The arithmetic would be that Texas was supposed to account for 18% of the national Hispanic vote total according to the pre-correction exit poll, so reducing the Texas share by 10 points would lower Bush's national share by 1.8 points. However, the other correction, reducing the fraction of the the vote in Texas accounted for by Hispanics from 23% to 20% would marginally lower the national share as well, so it's probably right around 2 points. (Three weeks ago I reviewed why the 59% share for Hispanics in Texas looked phony here: http://www.vdare.com/sailer/041107_election.htm

However, there is another big problem with Bush's Hispanic share besides Texas, which is that the national and regional numbers for Bush's Hispanic share are inflated relative to the broken-out state figures. For example, in the West region, Bush's share of Hispanics was supposedly 39% but if you add up the results for all the states in the West with enough Hispanics for the exit poll to report their Bush-Kerry breakdown (states which account for 97% of Hispanics in the West, Bush only won 34% in those states, meaning he'd have to win 170% of the Hispanics in the other states.

The same problem is found in the Midwest (the regional figure for Bush's share of Hispanics in the Midwest is 32%, but the broken out states only come out to 29%, and in the South, where Bush's regional share of Hispanics is an absurd 64%, even though the four broken out states (Texas, Florida, Georgia, and Oklahoma) only summed up to 58% even before the correction to Texas (now 52%). Only in the East did the broken-out states agree with regional total.

The individual state figures come from about 76,000 respondents, among whom about 62,000 filled in a short questionairre with about 20 questions. The other 13,660 respondents filled in the long questionnaire with about 60 questions. The national and regional figures comes from only those who filled in the long form. However, Hispanics who filled in the short form that was only counted in the state totals voted for Bush about 4 points fewer than did those who filled in the long form. Perhaps well-educated office-working Hispanics, who tend to vote Republican, were more likely to fill in the long form than poorly educated working class Hispanics, who tend to vote Democratic. The latter may have been intimidated by the long form, while white collar Hispanics breezed through it. Or, there could be other causes for the bias. You can read all about this problem here:

Taken together, I estimate they knock Bush's share of the Hispanic vote down from 44% to about 39%. That would be up four points over 2000, just as the exit poll says Bush's share of whites was up four points. Republicans' share of the Hispanic vote tends to go up and down in the same cycles as their share of the white vote, just well to the left of the white vote, but nobody pays attention to the white vote, so they get over-excited by shifts in the Hispanic vote.

Posted by: Steve Sailer | Dec 1, 2004 6:54:38 PM

This TX adjustment must have come all from Latino Bush support. That would indicate that an earlier adjustment came from Latino Kerry respondents. This would be how the Bush Hispanic percentage was made to rise to 44%; by a highly unbalanced deletion of minority Kerry respondents, to compensate for an oversampling of minorities which was large enough to cause the erroneous prediction of a Kerry win. How could the TX Hispanic percentage of the vote have gone from 10 to 20%; would it then be also able to double to 40% by '08, and so on? Similarly, if the non-hispanic white percentage of the vote countrywide can drop one point a year, would that mean that no minorities at all voted prior to 1980? If they used a more accurate model of 81% plus 9% black, 6% hispanic, 2% asian and 2% other, they would not get these contradictions in the data which require readjustment, and in an obviously unbalanced manner.

Posted by: John S Bolton | Dec 2, 2004 2:06:10 AM

Steve: Why should I read (or believe) you and your pals at Vdare? I can get the same stuff from David Duke...


Posted by: David T | Dec 7, 2004 11:52:15 PM

The comments to this entry are closed.