### Vacation

#### Mark Blumenthal | August 22, 2006

And finally . . .

As readers may have guessed from the lack of posts the last few days, I am trying to take a much needed vacation break this week. I interrupted it today to do finish up a few items I did not quite get to last week or -- in the case of the push polling post -- that seemed important enough to justify the interruption. My family probably disagrees.

And speaking of family, I am going to try hard to stay away from the blog for the rest of the week. I'll be back on Monday, and you will definitely want to stay tuned next week. Things are about to get a lot more interesting around here. Sorry to be a tease (again), but stay tuned . . .

## Comments

Excellent work,

greetings from Chile

Posted on August 25, 2006 11:37 PM

Rosenthal is wrong when he claims that poll percentages should be rounded off to the neares ONE percent..

He writes:

----------------------

False Precision

Beware of decimal places. When a polling story presents data down to tenths of a percentage point, what the pollster almost always demonstrates is not precision but pretension. A recent Zogby Interactive poll, for instance, showed that the candidates for the Senate in Missouri were separated by 3.8 percentage points. Yet the stated margin of sampling error meant the difference between the candidates could be seven points. The survey would have to interview unimaginably many thousands for that zero point eight to be useful.

Experienced researchers offer a rule of thumb: rather than trust improbably precise numbers, round them off. Even better, look for whole fractions.

------------------------------------------

Since when is Rosenthal an expert?

Zogby knows presicely what he is doing by rounding to 0.1%. He wants to provide the actual polling results, so that the numbers can be reconciled. There is nothing wrong with that.

There is a vast difference between a probability of 97.51% and 98.49%, although they both round to 98%.

There is a vast difference between a .51% MoE and a 1.49% MoE even though they both round to ONE percent.

But apparently its OK when exit poll naysayers add a 50% increment to the MoE to account for a "design/cluster" effect. Such hypocrisy.

Mitofsky knew exactly what he was doing when he rounded the National Exit Poll weights and vote shares to ONE percent, even though the National Vote share Margin of Error was ONE percent. He wanted to make it difficult for analysts who use weights and vote shares to compute probabilties.

Roundoff is a fudge. True analysts need all the accuracy they can get.

Which is a more meaningful result: a 50.49%-49.51% vote split or a 50%-50% roundoff? A +/-0.5% difference in a +/-3.0% MoE can be very significant, especially when replicated across multiple demographics or states.

Posted on September 6, 2006 9:14 AM

"Mitofsky knew exactly what he was doing when he rounded the National Exit Poll weights and vote shares to ONE percent...."

When he what?!?

Here, let me reproduce a weight from the national exit poll for your convenience:

.983492195606232 (as displayed by SPSS)

Goodness, man, do you know anything? do you care? where do you come up with this stuff?

If you want to dance on the man's grave, you should learn better steps.

Posted on September 6, 2006 11:45 AM

TruthIsAll writes:

"But apparently its OK when exit poll naysayers add a 50% increment to the MoE to account for a "design/cluster" effect."

Actually, you'd calculate the design effect rather carefully.

SPSS has a package that will do it for you here:

http://www.spss.com/complex_samples/data_analysis.htm

I am not aware that it limits the number of decimal places to one.

Posted on September 6, 2006 11:59 AM

Mark,

I suggest that you refer to the published National Exit Poll, as displayed on CNN.

All numbers are rounded to the nearest percent.

How come you are using SPSS?

What does that 10 digit number represent?

Why don't you just go to the SOURCE?

The final exit poll as published on CNN:

http://www.cnn.com/ELECTION/2004/pages/results/states/US/P/00/epolls.0.html

Posted on September 6, 2006 7:06 PM

The link you give is not the source. It is simply CNN's set of crosstabs made using the National NEP dataset. Did you not download that dataset while it was available? It gives all the weights (to a large number of decimal places). It was available in two formats, one of which was SPSS. Datasets for each state were available too.

Posted on September 7, 2006 4:01 AM

So, Elizabeth, are you saying that it was the decision of CNN to round to the nearest one percent?

CNN Final NEP: 1:25pm 11/3; 13660 respondents,

Wash Post/Edison-Mitofsky: 12:22am 11/3; 13047;

CNN: 4pm, 11/2; 8349;

CNN: 7:30pm, 11/2; 11027;

If so, why did they do it?

To save the printing of 2 characters (.X) ?

Posted on September 7, 2006 10:05 PM

I'm saying that if you had gone to the publicly available source data you could have had any degree of precision you wanted.

So it is a libel to allege that Mitofsky "wanted to make it difficult for analysts who use weights and vote shares to compute probabilties", and you should retract it. By insisting on making sure that the data, with weights, was placed in a publicly accessible archive each year, Mitofsky made it possible for any interested analyst to go to the data and run their own crosstabs, with or without the weights.

Some time back, I personally gave you the fast-track link to the archive so that you could download the datasets. They are available in ASCII format, so you could have opened them in Excel - they did not require SPSS. You do not appear to have bothered to do so. The fast-track link is now closed. If you want them now you will have to pay for the CD.

Posted on September 8, 2006 5:04 AM

## Post a comment