Lies, damn lies, and TV election polls

So Atlanta's Channel 11 came out with a new poll in the Georgia gubernatorial and Senate races with a big surprise. According to the detailed "cross-tabs," Georgia Latinos are going for our ultra-corrupt Republican governor and his "pants on fire" lying running mate in the race for U.S. Senator by a 4-to-3 margin.

I've gotta ask: dudes, what have you been smoking?

Because even more surprising than our supposed support for these two anti-Latino bigots is the poll's projection that Latinos, who are less than 1.8% of the registered voters in the state, will be 7% of those that go to the polls in November. Assuming every last Latino voter actually showed up, that would mean an overall turnout of 1.3 million, less than one in five adult residents of the state, impossibly low, even for Georgia.

And that brings us to the poll's surprising Latino results. The total sample was said to be 550, which gives us a theoretical "margin of error" of 4.3%, which in reality means 4.8%, since you have to add a rounding error of up to 0.5%.

But --assuming that the projection of a 7% Latino turnout was based on the number of Latinos the poll encountered-- that would be around 38 persons. So the question isn't what is the margin of error for a poll of 550 but what is the margin of error for a poll of those 38 Latinos. Using the standard formula, that would be around 15%, meaning the supposed Republican advantage is not statistically significant. It might be real. It might not.

Except that statistical science says, when samples get that small, the "margin of error" formula doesn't apply. The results are meaningless -- overwhelmed by statistical noise and the practical difficulties involved in public opinion surveys.

That's why at CNN, where my work included reporting precisely these sorts of numbers through 11 successive election cycles (1990-2010), we had a hard rule not to report cross-tab (subgroup) results of a poll where we had less than 100 interviews for that group. And even then, results had to be very dramatic, with big differences in the numbers, for us to rely on small samples.

Think of it this way. If the WXIA pollsters talked to 38 Latinos, Governor Deal's 40% would be about 15 people, Democrat challenger Jason Carter's 29% would have been 11, and 12 were undecided (reported as 31%). If you were calling 1,000 people, who were the first 20 who answered when you called, and agreed to be interviewed, isn't than significant. But if that is the majority of who you talk to, you have skewed the poll tremendously.

That's especially true when dealing with Latinos. Who was available to answer the phone? What languages are they comfortable with? The overwhelming majority of adult Georgia Latinos are immigrants, whose first language is not English.Was the survey conducted bilingually? Almost certainly not, because otherwise WXIA would have bragged about it. Pollsters rely on mechanisms like asking for the oldest or youngest adult to "randomize" a sample. In Latino households, that won't work.

And the Latino community in Georgia has been the most heavily hit of any state in the nation by Obama's "secure communities" deportation dragnet. It is a community that has been profoundly traumatized. Based on the information the government was forced to hand over thanks to a lawsuit by the Georgia Latino Alliance for Human Rights --with the support of the Georgia ACLU, NDLON and the Immigrant Rights Clinic at New York University School of Law-- I estimate that ICE, the immigration cops, have made some 100,000 arrests in Georgia since Obama was elected, overwhelmingly of Latinos, in a state that officially has less than one million Latinos.

 I do not think it is possible under such conditions to do an accurate public opinion survey of this community. There is simply too much fear.

But leaving that aside, I am sure WXIA's figures are, in essence, bogus. The Latino vote would need to be ten times its current size to be even measurable by a poll of the size done on behalf of WXIA. As a news organization, WXIA does not have the competence in polling nor the knowledge of our community to be making claims about how we're likely to vote.