Search This Blog


Bad sampling boosts Edwards poll support

It’s a relatively bad sample.

That explains why results from a mid-April poll by Market Research Insight, directed by Verne Kennedy, varied significantly from a late April survey conducted by JMC Analytics’ John Couvillion. The former, among other things, called the reelection without a runoff by Democrat Gov. John Bel Edwards likely, while the latter concluded Edwards would be unlikely to avoid a runoff with one of two major Republican candidates.

Since other earlier polling generally had backed the JMC interpretation, observers questioned the MRI conclusion. That poll continued a series stretching back years, commissioned by business and political elites of whom now one is the owner of the Baton Rouge Advocate, John Georges, where the results appeared.

That story revealed almost nothing about the specifics surrounding the poll. Perhaps feeling the heat, The Advocate days later put out another story about it with an explanation of the analysis. Here lay the key to understanding the differences and allows for comparing the relative validity of the two polls.

The JMC effort (which released its analysis along with its results initially) polled 650 registered voters, 50 more than MRI’s “active” voters. An “active” voter was someone who had voted in at least two of four “recent” elections, presumably the 2017 treasurer special general election, the ensuing runoff election, the 2018 secretary of state special general election, and ensuing runoff election.

If these were there four, this is problematic for identifying a representative sample in the 2019 governor’s race. Only the third of these, which occurred along with federal elections – none of which were competitive – and also local contests in many areas, had a turnout level remotely comparable to what would happen in a governor’s race. That one hit 50 percent, while the others languished in the 13-17 percent range.

This means the MRI sample drew disproportionately from more activist citizens. Then it exacerbated that effect by oversampling with voters using cellphones.

The JMC poll used a ratio of 72 percent landline/28 cell phone for contact. But, apparently drawing from the federal government’s National Health Interview Survey guidelines (which many pollsters use as a reference on this question about what proportions of each contact method to use), MRI used a 47 percent landline/53 percent cell phone ratio.

Cell phone-only users as whole differ from the population in several ways. They are younger and more likely to live with unrelated people, to rent, be lower-income, and to be non-white. Directly related to voting behavior, these characteristics also are found disproportionately among two groups of individuals: Democrats and nonvoters. Compounding this, differences in contact, cooperation, and completion rates also suggest that voters more likely fitting a Democrat’s profile more likely accept and complete a cell phone interview (the completion rate of which – meaning successful contact and cooperation – is a little lower than for landlines although the combined rate has plunged to six percent). Finally, and specific to these two polls, completion rates increase the shorter the survey; JMC’s at seven questions appears less than half the length of the MRI effort.

Thus, two problems arise with the MRI approach of doubling up on the cell phone users. First, even if the national proportion comes in around 55 percent, that’s probably significantly higher than it is for the electorate that will turn out in October and November (JMC lowered the weighing by including more landlines that more accurately reflects the actual electorate composition likely to coalesce). MRI tries to avoid this by its active voter use strategy, and it can claim some success insofar as the sex, racial, and partisan registration statistics of its sample come close to JMC’s and overall registration numbers in these categories.

But it can’t avoid at all the second problem: within that sample, it likely picked an unrepresentative sample on characteristics such as likelihood of voting and ideological intensity. More specifically, it includes too many people who won’t vote and, of those who will, it oversamples liberals.

In the first instance, the same aspects about cell phone users that suggest disproportionate support for Democrats also make for greater likelihood of completing a call even if they don’t intend to vote. This is reflected two ways in the MRI sample: it’s oversampling of people who call themselves independent (as opposed to registered that way) and in the oversampling of younger voters. Both categories are disproportionately unlikely to vote in elections compared to the remainder of the population. Particularly in the former instance, many black independents who call themselves that because of extreme disinterest and lack of knowledge about politics are the least likely people to vote, while specifically in the latter instance younger registrants are the least likely to vote (and register, for that matter).

The former instance also finds reflection in the strikingly high proportion of MRI blacks calling themselves undecided, about half, a proportion highly unusual to find at this stage of such a contest. JMC found about half of that number undecided. The reason why is that MRI pulled a sample disproportionately of blacks who don’t intend to vote, leading Kennedy mistakenly to assign Edwards much more support than he actually has.

At the same time, the sample also disproportionately contains interested activists from across the political spectrum and undersamples the more causal voter. This creates a bias in favor of the political left, and thus in support for Edwards, because the casual voter more likely falls back on their roots when they cast their ballot. And in today’s Louisiana, when the casual voter – who knows little about the candidates but knows he should vote for one – sees the names on the touchscreen with a “D” and “R” next to them, they disproportionately vote for the Republican because partisanship is the most prominent cue for such people.

In short, MRI overestimates support for Edwards because its sample too heavily weighed towards cell phone users, which introduced biases of disproportionately more of those who expressed Democrat support but who won’t vote and disproportionately fewer casual voters who disproportionately will support a Republican. By contrast, JMC seems to have struck a better balance. That’s why JMC says its poll “almost guarantees that the race will go to a runoff” and the “race is likely to be closely contested” while MRI claims with its that “If Edwards has strong and viable opposition, he still has at least a 50% probability of re-election. However, as seen in the upcoming slides the two Republicans in the race do not provide the popularity or strength to seriously challenge Edwards at this time.”

Both can’t be right. Based on what pollsters and political scientists know about voting behavior and polling, JMC more likely is.

No comments: