Erroneous Margins of Error

After the shocking Trump victory in 2016, some pundits wondered whether Spirited Reasoners could ever trust opinion polls again. What those pundits forgot is that the polls that year were largely accurate. Most predicted that Hillary Clinton would win a small majority of the popular votes. And, in fact, she did.

Now, many of the same polls are showing Joe Biden with a national lead averaging around 8 percentage points. What’s interesting to Spirited Reasoners is that most of these polls are now showing Biden victories greater than that; however, a few polls (notably Rassmussen and IDB) show a much tighter race, and thus drag the average down the 8 percentage point level mentioned above.

Given the previous election results, Spirited Reasoners are not content simply to read these polls and accept their predictions. We dig deeper, wondering about each poll’s methodology and bias. This week’s post will prove beyond a reasonable doubt that at least one major poll is seriously skewed, even if we can’t be sure which one it is.

Here’s the evidence:

My favorite listing of current polls can be found at the FiveThirtyEight website:

Over the past week, Joe Biden’s lead over Donald Trump has ranged from a low of 3 percentage points ((Rassmussen Reports/Pulse Opinion Research) to a high of 12 percentage points (IPSOS and USC/Dornsife polls).

According to Rassmussen, their margin of sampling error is +/- 2 percentage points with a 95% level of confidence. I could not find margin of error or confidence levels for the IPSOS and USC/Dornsife polls; however, given that the sampling size of the latter poll is much larger than most other polls, its margin of error should be tighter, i. e., smaller. Fortunately for Spirited Reasoners, we only need the margin of error from the Rassmussen poll to make our case.

Let’s assume, for a moment, that the Rassmussen claim is accurate. If so, Joe Biden’s lead must range between 1 percentage and 5 percentage points over Donald Trump. How is it, then, that almost all other polls show Joe Biden with leads not only exceeding that 3 percentage point prediction but also the 5 percentage point maximum we obtain by adding the margin of error?  Or, to put the question a bit more bluntly, how could we ascribe any meaning to the terms “margin of error” and “95% level of confidence” if Joe Biden were to end up winning by the 12 percentage point margin predicted by IPSOS and USC/Dornsife?

If Joe Biden were to win the popular vote by only 6 percentage points, the folks at Rassmussen could rightly claim that they were only slightly off the mark; as in, “we told you we had a 95% level of confidence, not a 100% level of confidence in our margin of error.” But if Joe Biden were to win by the 12-percentage-point margin predicted by IPSOS and USC/Dornsife, there would be no statistically valid way to justify the discrepancy among these polls. Such a margin would not just be outside the Rassmussen margin of error. It would be far enough outside the margin to indicate serious flaws and/or biases in polling methodology.

Of course, the reverse is also true. If Joe Biden were to win the popular vote by a margin of only 3 percentage points or less, then IPSOS and USC/Dornsife would have some explaining to do. As would a host of other polls, all of which would be experiencing margins of error greatly in excess of the Rassmussen claim.

Could it be true that only one or two polls are using valid methodology while all the others are seriously flawed?  

We’ll see.