19 February 2008

A Virtual Tie?

The Press's presentation of polling data along with the popular understanding of what poll results actually mean are both so awful that I place them into that category of things about which I consistently resist ranting. But this morning I seem to have lost my self-control, when I read this:
50 percent of likely Democratic primary voters support Clinton as their choice for the party's nominee, with 48 percent backing Obama.But taking into account the poll's sampling error of plus or minus 4½ percentage points for Democratic respondents, the race is a virtual tie.

First of all, I'm going to put aside the fact that if the sampling and polling techniques are not done properly than any poll tells us absolutely nothing (since it conveys information only about the sample and not about the entire population). Anyone who knows anything about polls knows this, of course. But the press never mentions it at all, and over the years I have discovered through my students that many readers of press coverage on polls therefore have no idea that a poll could ever be more wrong that the sampling error. (OK, I didn't put aside that fact.)

Let's assume, however, that this poll was properly conducted. Does it tell us that Obama and Clnton are tied? What, for that matter, is a virtual tie? Read straightforwardly the poll tells us that Clinton's support lies somewhere in the range between 45.5% and 54.5%; Obama's support lies somewhere in the range between 43.5% and 52.5%. Therefore, it is true that they could be tied. It is also true that either one could be ahead of the other slightly. Moreover:
  • Clinton might be ahead by 11 points
  • Obama might be ahead by 7 points

Is that a tie, or a virtual tie? No.

I think what the reporter means to say is that the poll cannot indicate clearly whether either candidate is ahead. Saying that they are in a virtual tie is like... (this analogy took me a while to come up with, and may not work, so bear with me)...the following:
The score in the football game is 7 to 10.
I hear a news report that one team has just scored, but I don't know which team.
So, assuming no safeties or missed extra points, the score could be
  • 14 to 10
  • 10 to 10
  • 7 to 13
  • 7 to 17
Whatever we want to call this situation, 'virtual tie' is the wrong choice.

Let me just close in noting that this is a ridiculous subject for me to blog about. I'm guessing there are about 5 readers of SecondAmericano: 3 of them know far more about politics and polling than I do, and one of them knows far more about statistics than I do. Still, if you can't preach to the choir, then what good is preaching (and besides, if I screwed something up here, I'm sure to be quickly informed). 

1 comment:

Transient Gadfly said...

The only thing I'd add is that the +/- is within a confidence interval (which is usually something like 95% for polling), so to use your metaphor and my preferred interpretation of statistics, you are following a football game, you just heard on the radio that there are 20 football games going on of which your game is one, and in every one the score is 7 to 10. In addition, in 19 of them, one team just scored--ergo, the game you are interested in is virtually tied.

What are the odds that the game you are interested in is actually tied? Using your four possible outcomes and assuming they're all equally likely given that one team scored, it's more or less (.25 * .95) = 24% chance (similar calculations tell you that there's a 52% chance they're behind and a 24% chance they're ahead).

So we find that "virtual tie" is actually code for something like, "this poll tells us that Hilary's more likely to be ahead of Obama, but we're only 50% sure of that. It could be the other way around, or they could be tied, which is, electorally speaking, not a valid result, so we pretty much have to throw that out, too." Unfortunately, that's just a convoluted way of saying, "Dude. This race is, like, really close." Which we already, like, totally knew.