by David F. Damore
The challenges of polling Latino voters have received less attention in 2014, because there are fewer competitive states and districts this cycles where Latinos are positioned to be influential. A notable exception is Colorado where a recent Colorado Public Radio story highlighting variation in polls conducted in that state suggested that much of the inconsistency across polls was a consequence of the difficulties associated with polling Latino voters.
The table below summarizes the results--courtesy of HuffPost Pollster, as of October 16--of recent Colorado gubernatorial and U.S. Senate polls. Across both contests, the polls suggest very different states of play. In some instances, the candidates' projected vote shares are outside of the polls' margins of error. The margin of error, of course, captures the random variation resulting from sampling assuming that the sample is representative of the electorate. However, when a sample contains systematic error, and is therefore unrepresentative of the electorate it seeks to describe, then the margin of error tells us very little.
As we at Latino Decisions have repeatedly argued, in districts or states like Colorado where Latino voters compose a large share of the electorate, the inability to accurately estimate the size of the Latino electorate and these voters' preferences create systematic error. In what follows, I use results from a U.S. Senate poll of Colorado Latino voters recently conducted by Latino Decisions for the NCLR Action Fund to highlight some of these issues.
First, there remains a great deal of uncertainty about the size of the Latino electorate both nationally and in specific states. Among the pollsters working in Colorado who released this information, the expected share of the Colorado electorate that is Latino varies from a low of five percent to a high of 16 percent. In comparison, the 2012 exit polls estimated that Latinos were 14 percent of the Colorado electorate and 12 percent in 2010.
To be sure, most pollsters and handicappers expect a decrease in Latino political participation in the 2014 election relative to 2012. Specifically, it is commonly assumed that younger, less educated, and lower-income voters and those with shorter voting histories--all characteristics of the Latino electorate--are less likely to vote in midterm elections. However, in our sample of 400 Latino registered voters, 80 percent reported that they were almost certain to vote in 2014 and another nine percent indicated that they probably would vote. These numbers closely match the 91 percent turnout among Colorado Latino registered voters in 2012. Among subgroups of Latino voters presumed to have a lower likelihood of voting in 2014, 82 percent or respondents with a high school degree or less indicated that they almost certain or probable to vote. Similarly, 93 percent of Latinos under the age of 30 indicated they were almost certain or probable to turnout this election.
Also note that by assuming a decrease in Latino turnout, pollsters are implicitly assuming that the size of the Latino electorate is constant. This assumption is unlikely to hold given the relative youth of the Latino population relative to other racial or ethnic groups. In short, through the process of generational replacement, younger non-white voters are replacing older white voters. Thus, regardless of any campaign effects (i.e., efforts to mobilize or depress the vote of certain blocs of voters), demographic change is reshaping the contours of the electorate in real time, all of which makes it more difficult to estimate what the electorate will look like in future elections based upon what happened two or four years prior.
Second, in terms of sampling, because of the aforementioned characteristics of Latino voters, many Latinos are unlikely to satisfy the "likely voter" screening questions to be included in the actual sample. Yet, as we know from the 2010 U.S. Senate elections in Colorado and Nevada these voters can be mobilized in midterm elections and when they are, any polls or prediction models assuming a decrease in Latino turnout will be wrong. Indeed, 43 percent of our sample reported that they are more enthusiastic about voting in 2014 as compared to 2012. Among voters under the age of 30, twice as many reported that they are more enthusiastic about voting in 2014 as compared to 2012.
Third, the relatively small number of Latinos included in most polls means that very few respondents are used to estimate the preferences of Latino voters. For instance, assuming an electorate that is 12 percent Latino and a sample size of 500 means that variation in the preferences of Latino voters is determined by 60 interviews. Thus, among those 60 respondents a poll must capture in proportion to the Latino electorate in the population all of the factors that may affect the distribution of the Latino vote such as partisanship, age, education, and income. Among the recent Colorado polls, only the FOX poll included an oversample of Latino voters. However, this oversample totaled 38 respondents, or 5 percent of the sample.
Fourth, ensuring a representative cross-section of Latinos necessitates dual-language interviews and a blend of cell-phone, landline, and web based interviews. While each of these factors adds to the costs and time of conducting a poll, failing to do systematically biases the Latinos who are sampled. In particular, cell-phone only, Spanish-speaking, lower socio-economic status Latinos, who are the most Democratic of all Latino voters, are the most difficult (and costly) voters to include in a poll. Thus, even if polls include both Spanish language and non-landline sampling frames, they are unlikely to capture a representative cross-section of Latino voters. Instead, polls are likely to over-sample more English speaking, higher income, and Republican leaning Latinos.
Some of these dynamics can be seen in our Colorado analysis. For instance, 35 percent of Colorado Latinos reported having a high school education or less and among the 7 percent who did not attend high school, 58 percent of those interviews were conducted in Spanish. Among respondents who were interviewed in Spanish, 76 percent indicated that they will vote for or are likely to vote for the Democratic Senate candidate Mark Udall; a level of support that is 13 percentage points greater than for respondents who were interviewed in English.
Fifth, pollsters often use weighting to adjust samples after the fact to bring them in line with expectations about what the electorate is supposed to look like. When a poll's results are weighted, some respondents are counted more or less in the aggregate results to adjust for any over or under sampling relative to what the pollster expected the sample to look like. For example, the CBS/NYT/YouGov Internet survey, weighted up Latinos from five to 13 percent meaning that each Latino respondent counted for the equivalent of 2.6 respondents in the weighted total. However, if a sample suffers from systematic error, such as failing to draw a representative sub-sample of certain groups, then weighting respondents from those groups exacerbates rather than alleviates the level of systematic error that a poll might contain.
The bottom line is that unfortunately what appears to be occurring in Colorado in 2014 is nothing new. What is surprising though is that pollsters are failing to learn the lessons from prior election cycles.