Support for Republican presidential candidate Donald Trump is declining. How much it's declining depends on which polls you read and, maybe, on how those polls were conducted.
In theory, the ways that surveys are run can produce systematic differences in the results they get. In practice, Pew Research has shown this to be true. Last month, Democratic analyst Jonathan Robinson identified such a difference in Trump's numbers, which raised speculation about how polling methods might be influencing the primary contest.
Since Trump's numbers showed a clear pattern, HuffPost Pollster decided to take a closer look at the national horserace polling for all of the 2016 presidential candidates. We found that the differences by type of poll were far from systematic. They were actually all over the place.
The charts below compare the candidates' average support as measured by two different polling methods from March to October. The purple lines show the average for polls that used "traditional" telephone methods, which rely on live interviewers calling people on their cell or home phones. The green lines show the averages for polls that didn't involve live interviewers -- primarily, surveys done with automated phone technology or the Internet.
Among the Republicans, as Robinson found, the most obvious pattern is that Trump's support is higher on average in Internet and automated phone polls than in interviewer-conducted phone polls. Support for New Jersey Gov. Chris Christie shows a similar split. But the numbers for retired neurosurgeon Ben Carson and former Hewlett-Packard CEO Carly Fiorina go the opposite way: Their support is now higher in polls using the traditional methodology and lower in the online and automated polls.
Support for the remaining candidates doesn't vary much these days based on the types of polls. Former Florida Gov. Jeb Bush, Texas Sen. Ted Cruz and Kentucky Sen. Rand Paul showed differences earlier in 2015, but those occurred when there was less polling and a lot of uncertainty about who would run.
Even for the candidates with differing results, the patterns aren't consistent. There was no real difference in Trump's numbers between polling methods until the middle of August. Carson, Christie and Fiorina have only begun to show differences in the last month.
Among the candidates for the Democratic nomination, the contest leaders -- former Secretary of State Hillary Clinton, Vermont Sen. Bernie Sanders and Vice President Joe Biden -- show absolutely no relevant difference in their support by poll type. The results for former Maryland Gov. Martin O'Malley and former Rhode Island Gov. Lincoln Chafee show varying differences, with greater gaps in the spring when there was less polling available. And former Virginia Sen. Jim Webb's numbers have stayed between 1 and 2 percent regardless of how the poll was done.
It's not clear that how the polls were conducted has been the source of the particular differences that did arise this year. Certainly, survey methods can factor into a complicated primary polling scene. But polls also use various sampling, recruitment and weighting procedures that could affect the results. Some surveys try to identify the views of "likely voters," while others report results for all registered voters -- another source of possible variation.
The effect of polling methods on Trump's numbers right now is striking. But HuffPost Pollster's examination suggests that the difference between traditional telephone polls and online or automated polls may not matter as much in assessing the current state of the race overall. We won't ever know how right or wrong these estimates are, since they purport to measure a national primary election that never actually happens. Polling averages that incorporate all the polls are probably the best indicator of what's happening in the primary season.
R code for these analyses can be found here.