If you read my last post, you know I’ve been looking at some very fishy survey data from Strategic Visions, LLC. The data seems to stink no matter how anyone looks at it, and mathematicians, statisticians, and programmers have been looking hard and every which way. Instead of throwing yet another heavy mathematical brick at the poor numbers, Let’s see how it stands up to a feather.
You don’t have to read many poll or survey results to be familiar with the phrase “totals may not equal 100% because of rounding.”
So guess what? The numbers in Strategic Visions’ results all add to exactly 100. Oops. That’s a huge red flag. Huge, like big enough to wrap a planet in.
Ok, I didn’t check all their polls. Just the most recent 73, which is how many I checked before stopping. For the record, I didn’t stop because poll #74 added to 99 or 101. That would have been statistical misconduct on my part. I didn’t look at poll #74 or any others, because I wanted to write this up. (If anyone checks further, let me know and I’ll post an update.)
If Strategic Visions were not lying (or being extremely sloppy in a systematic way, which is their only hope of explaining this—see below), the chance the 73 most recent polls would add to 100 is less than 1 in 2,000. (That assumes intentional rounding to minimize 99s and 101s. For any pre-determined rounding rule, however, we’re talking 1 in 10,000,000 or worse. Maybe a real pollster can fill me in on what’s industry practice among those who aren’t lying.)
One-in-two-thousand-ish stuff happens all the time, but believe me, I didn’t find this on a data dredging excursion. I noticed that the first few poll results I saw added to exactly 100, I formulated a plausible hypothesis based on all the evidence at hand, I carried out an experiment, and I calculated the p-value.* Small enough to be incriminating in my home court, and I tend to be a benefit-of-the-doubt guy.
Just two more things. First, the possibility of systematic sloppiness:
The consistent adding-to-100 could be the result of a systematic error, as opposed to cheating. Of the possible excuses, this is the one I suggest SV choose if they decide not to come clean. Logically it can’t be distinguished from lying, and they can attribute it to a whipping boy like the web site designer. (This excuse doesn’t defend against the mountain of mathematical bricks I mentioned earlier, however.) They can say that they made a regrettable decision that to avoid the appearance of error, they calculated one percentage in each survey from the other percentages, not from the survey results. I won’t believe it for a minute (though if they show me their programs or confirm that some commercial product makes this error, I’ll reconsider), but it might get them out of hot water.
And second and last, an example and a bit of the math behind my calculations:
Most survey results are short lists of whole number percentages that express fractions to the nearest whole percentage point. Suppose 600 likely voters were polled in a tight race between Tintin and Dora the Explorer. Tintin had the support of 287 people, Dora was close behind with 286, and the rest of the 600 people surveyed (that would be 27 of them) said they weren’t sure. To the nearest whole percentages, that’s 48% for Tintin, 48% for Dora, and 5% undecided. The sum of the rounded percentages is 101%, and that’s due to honest mathematics, not fraud.
Let me skip some really fun mathematics and tell you that for survey questions that have three answers, the percentages add to 100 most, but by no means all of the time. Exactly how often they don’t add to 100 depends on several factors, two of which matter much at all in this case: the number of people surveyed, and how numbers ending in .5 are rounded to whole numbers. Strategic Visions, LLC’s usual survey sample size is 800, and even if ending-in-.5 numbers are rounded differently in each poll to avoid a sum of 99 or 101 when possible, three-choice result percentages should still add to 99 or 101 at least one time out of ten.
* Not to be taken as evidence of a non-Bayesian persuasion on my part. The frequentist approach seemed to me pretty straightforward and justifiable here, that’s all.
November 27th, 2009 at 11:53 am
Here’s a nice article about localization: “Does Your Code Pass the Turkey Test?” http://www.moserware.com/2008/02/does-your-code-pass-turkey-test.html