Lying with statistics: Ur doin it wrong

Now that we're inside four weeks to the election, it's probably worth reviewing some of the Magic Phrases of political punditry. Here are three that should raise immediate red flags:

• "Real Clear Politics average"
• "Within the margin of error"
• "Statistical dead heat"

If you see any of those in a campaign thumbsucker, you should assume that the writer is either (a) amusingly ill-informed or (b) attempting to blow smoke up your butt on the assumption that you'll fall for anything that sounds entertaining. Which will it be there, National Review Online?

Senator Jeanne Shaheen (D., N.H.) and Republican challenger Scott Brown continue to find themselves neck-and-neck in the New Hampshire Senate race less than four weeks away from the election.

The latest NH1/New England College poll shows Shaheen leading Brown 48.5 percent to 46 percent, falling within the roughly three-point margin of error. Just 3 percent of voters remain undecided at this point.

Well, not quite. The lead is 48.5 to 45.9, according to the polling agency, and the margin of sampling error -- which you really should leave to one decimal place* -- is 2.7 points, meaning (at 95% confidence) that the result is right at the so-called "margin." Whether you're a tenth of a point above or a tenth of a point below the arbitrary number that's misapplied here, the best bet is that Shaheen is leading. All you have to do is turn the confidence level down to about two-thirds. If that's your idea of a dead heat, you shouldn't go to the track by yourself.

There's no such bogus statisculating in the second example:

Senator Pat Roberts (R., Kan.) has closed the gap with independent candidate Greg Orman, according to a new poll in what has become one of the increasingly competitive races this cycle.

After polls showing Roberts trailing by as many as ten percentage points last week, a new CNN/ORC poll of likely voters in the Sunflower State finds Roberts leading Orman 49 percent to 48 percent. The survey finds both candidates above water with their favorable-unfavorable ratings.

In case you don't feel like calculating it,** the standard error for that survey is about .019, for a margin of sampling error (95% confidence) of 3.7 percentage points. The "slight lead" isn't as good a bet as the "statistical tie."

Well, that's not unusual. The party press has often been the sort of place where you hear what you want, regardless of those pesky confidence intervals. And this isn't the National Review's first county fair. If it's only trying to lighten the rubes' wallets a little, that's between National Review and the rubes; it might get to the county line ahead of them, and it might not. But if it's in the habit of deceiving its masters, it might want to watch out. Those bills can come due in a hurry,  

* This poll has a sample of 1,286; you could get a margin of 3 points -- OK, 3.04 -- at 95% confidence with 250 fewer respondents.
** Real Clear Politics is a useful archive of current survey results and sample sizes, but averaging a bunch of estimates -- particularly estimates with different CIs -- is never a good idea. The commentary files can be pretty hilarious, though.