When Standard ‘Pundit Talks To Pals’ Behavior Is Gussied Up With Numbers

Et tu, 538? One of the long-standing problems in political punditry is when pundits consider the group of people they know, a group which is almost always completely unrepresentative of the phenomenon they’re discussed, as a representative sample. Seth Masket, in a recent post, did this, except he gussied it up with numbers. In a poll of 35 activists (yes, n = 35), Masket found this:


Seems bad. Except that, with a sample of 35 people, dislike for Sanders isn’t significantly different from dislike for many candidates–you have to reach Gillibrand before the result is significant (2×2 contingency tests, how do they work?). And that doesn’t take into account multiple corrections, so this graph becomes even more ridiculous. Adding percentages makes the pattern look far more serious than it really is. Also, it’s interesting that the dislike for Biden is also high, but that goes unmentioned (time to add the disclaimer: were I to vote today, I would vote Warren; I just don’t like statistical abuse that is inevitably turned into bullshit arguments).

Had Masket taken a more sociological approach in methodology–treating this sample of 35 people as interview subjects–it would have been far more fruitful. It’s an interesting demographic and understanding some of the changes as well as the reasoning would have been useful (we do need Democratic Cletus safaris). But instead, we just get received conventional wisdom gussied up with (mostly meaningless) numbers.

Time for a blogger ethics panel.

This entry was posted in Democrats, Polling, Statistics. Bookmark the permalink.