Let's start with the bottom line: The final value of our trend estimate for Texas (at least as of this writing) shows Hillary Clinton running slightly ahead of Barack Obama (47.6% to 45.9%), but I would advise readers against treating that as a solid prediction of the outcome. It may turn out that way, of course, but variation among individual polls and more importantly -- uncertainty at this hour about the racial composition of the Texas electorate -- means that the ultimate result is unknowable.
First, let's take a look at the latest version of my table comparing the demographic composition of most of the polls out over the last few weeks, updated with the surveys released since my post on Friday (and lets say a thank you to all the pollsters who have released this data -- what a change since Super Tuesday):
If you read through the lines a bit, you can see the different approaches that various pollsters take to "modeling" the likely electorate. Some arrive at a set of arbitrary weighting quotas for gender, age and race and apply these consistently on each survey. Notice the way the percentages for both Zogby and InsiderAdvantage are identical on all of their surveys, except one. The exception is the Latino percentage on the most recent InsiderAdvantage, which plunges from 37% to 27% (while all other demographics remain spot on identical). Perhaps someone there had a change of heart about their model?
Some -- such as ABC/Washington Post and SurveyUSA -- take a very different approach. The begin by interviewing a sample of all adults in Texas, weight the demographics of the adult sample to Census estimates for Texas, choose "likely voters" based on their answers to screen questions and allow the demographics of the likely voters "fall where they may." See this post for more details on SurveyUSA's approach.
An important underlying point here is that some pollsters have more confidence than others in the ability of their measurements to "predict" the likely electorate and its demographics. My own sense (and be advised that other pollsters may not agree) is that pre-elections polls are much better at measuring the opinions and preference of respondents than at precisely predicting who will vote and who will not. I will spare the detail this morning, but the bottom line is that screen questions are at best a crude measure of who will turn out. We can select "likely voters" with a greater probability of voting than those we screen out, but that's as good as it gets.
The best approach in situations like these, when voters demonstrate huge differences by racial and demographic subgroups, is to watch those differences and understand the potential range of outcomes. So let's do just that.
The good news, again, is that most of the pollsters have released both racial composition data and cross-tabulations of the vote by race, which allows for the following table. A few observations: Most surveys have been reasonably consistent (within the vagaries of sampling error) in their results for Latinos and African Americans. Most have shown Clinton with a roughly two-to-one lead among Latinos and most have shown Obama winning 75% to 85% of African Americans. Those results are generally consistent with exit poll results from other states (although Obama has typically done a few points better on the exit polls than in final pre-election polls).
However, pollsters have been less consistent in their measurements of white voters.For example, on poll released in the last 24 hours, SurveyUSA and Rasmussen show Clinton leading by just four points among white voters, while Mason Dixon, PPP and InsiderAdvantage show Clinton with margins of 15 or more points among white voters.
I have included three sets of averages at the bottom of the table: One for all of the polls listed, one for the final poll by each organization and one for final polls released over the last three days. Keep in mind that cross-tabs by race were not available for all surveys.
Let's use these results to put together a "what if" analysis of the turnout (similar to that found in the Belo/Public Strategies analysis). I have created a spreadsheet in Excel (download here) or Google Documents (edit here) that you can use to try and test your own assumptions (details on how to edit the Google docs version at the end of this post).
I set up the spreadsheet using the following assumptions:
(1) A racial composition of 51% Anglo, 29% Latino and 20% African American -- the average of the assumptions and findings of the pollsters in the first table; (2) A 56% to 44% Clinton margin among white voters -- which takes the average above and proportionately allocates unecideds; (3) A 67% to 33% Clinton margin among Latinos and (4) an 85% to 15% Obama margin among African Americans (which assumes that Obama overperrforms in this constituency by about as much as he has elsewhere this year). I am not making predictions here, just grabbing for reasonably defensible assumptions based on the available data as a starting point -- your "mileage" may vary.
As it turns out, these assumption produce a roughly two-point Clinton lead, the same as our current trend estimate. That's not a great surprise, since they are based on mash-ups of the demographics and results of most of the the polls used to generate the overall estimate. But now, play "what if" and see how making very small changes in any of the assumptions can easily alter the outcome.
For example, apply racial composition findings from the ABC/Washington Post survey (39% Latino, 17% African-American), leave all other assumptions as is, and you get a 6 point Clinton victory. On the other hand, apply the racial composition used on the Belo/Public Strategies polls (25% Latino, 22% African American), and you get a half-point Obama win. Leave the racial composition as-is, but assume that SurveyUSA or Rasmussen has the white vote right (Clinton leading by just four points) and Obama wins by two. But if InsiderAdvantage, PPP and Mason Dixon has Clinton's margin among white voters right (roughly 15 points), and all other assumptions remain constant, and Clinton wins by six. I could go on.
Again, your assumptions may be different on any of the above, so open up the spreadsheet and have at it. But now hopefully, you have a sense for why races like these give pollsters heartburn.
To edit the Google Docs spreadsheet: You will need a free Google account. Click here to display the published version, then click the "edit this page" link at the bottom right of the screen and you will see a "view only version of the spreadsheet." If you use the "File" pull-down at upper right to copy and rename the spreadsheet, you can edit, change and even publish as you see fit.
Hat-tip to David Ianelli and Elise Hu for providing the final Belo/Public Strategies numbers by race. Typos in the tables have been corrected.
Update: In the comments, Steve P asks some good questions:
Extrordinarily high turn out overall but I would think that [African-American] turnout is probably outpacing their regular share. But that Hispanic share is probably down.
I understand that this is difficult to guage since we have never had a primary go this deep into the schedule before on a competitive basis. Wouldn't you have to toss election models out because we have never seen anything like this before?
How well has final polls lined up with exit poll data demographically? Are we seeing the same demographic portions despite the much larger pie?
Should we throw out old demographic models? Certainly. But which model makes the most sense? To answer Steve's question briefly: While we have seen consistently greater proportions of younger and higher income voters participating in Democratic primaries as measured by exit polls, the change in racial composition has been inconsistent. We will know the answer in a few hours, of course, but a critical question is whether Texas will be more like California this year (where the Latino share surged from 16% to 30%, while the African American share dropped a point from 8% to 7%) or more like Arizona (where the Latino share increased one point, from 17% to 18%, but the African American contribution quadrupled, from 2% to 8%).