Friday, March 03, 2006

Why Polls Stink

DovBear today posted a recent LATimes/Bloomberg poll that came out, claiming Bush's approval rate was at 38% overall. After the CBS poll earlier this week was found to be skewed, with a too-high sampling of Democrats, I figured I would check the numbers on this one as well. Here's what I figured out, using the first example on Page 6 as an example: (sorry for the formatting)
Q2. Do you approve or disapprove of the way George W. Bush is handling his job as president? (IF APPROVE OR
DISAPPROVE) Do you (approve/disapprove) strongly or do you (approve/disapprove) somewhat?
Approve (Net) 38 12 43 77 15 35 60 44 32
Approve strongly 19 1 18 51 7 11 39 21 18
Approve somewhat 18 11 25 26 7 24 21 23 14
Disapprove (Net) 58 86 52 19 83 61 38 52 64
Disapprove somewhat 16 16 14 6 17 17 14 14 17
Disapprove strongly 43 70 38 13 66 44 24 38 47
Don’t know 4 2 5 4 2 4 2 4 4
Let's use the first example, net approval of 38%. Assuming an equal proportion of Democrats, Independents, and Republicans, the net approval should be 44%:

Dem: 1/3 (12%) = 4%
Ind: 1/3 (43%) = 14.33%
Rep: 1/3 (77%) = 25.67%

4 + 14.33 + 25.67 = 44

A factor that would work, however, is the following:

Dem: 9/20 (12%) = 5.40%
Ind: 6/20 (43%) = 12.90%
Rep: 5/20 (77%) = 19.75%

5.4 + 12.9 + 19.75 = 38.05

In other words, 45% of the respondents would have to be Democrats, while just 25% were Republicans. Is that a fair poll? Well, yes. It's simply a skewed sample... which is why polls such as this are meaningless.


  1. Don't most polls correct for party affiliation? Zogby's been showing approval ratings around 40% at least since January.

  2. From my blog:

    Ezzie is correct if you think the mathematics in polling never advanced beyong the level of 4th graders.

    The link to that poll contains methodology at the bottom. They randomly select phone numbers. A very common method is to weight responses based on demographic details. For example, if your sample contains 47% men, but the US census says the population is 49% male, then the responses of the men are weighted slightly more. This is a quick way to correct for some random sampling biases without needing to quadruple the sample size. They don't weight by political affilation because that changes over time and no Census quality documents contain that information.

    Each poll contains only about 1000 people in the entire nation so they say there is +/-3% error for the total responses (I'm assuming 95% confidence intervals) and more error in subpopulations. For a 38% approval rates that means there a 95% chance that the true approval rating is within 35%-41%.

    As more polls with more respondants all come out with the same results, it increases the likelihood that a result is accurate:

    39% approval ...poll_030206.pdf

    43% approval (dropped from 52% in January) pd...bush_022706.pdf

    There is absolutely no question that his approval rating is dropping and definitely below 46%.

    Anon Y. Mouse | | Email | 03.03.06 - 1:56 pm | #

  3. As I responded at DovBear:

    What did he say that contradicted anything I said? The LATimes poll does NOT show how they did the weights at all, and as opposed to the CBS poll, which did adjust the numbers, it does not show any evidence of doing so. Even in the CBS poll, you can see that such corrections do NOT fully make up for the sampling. There's not much that can be done about sampling; by sampling 1,000 or so people, you run the risk that you'll get a heavy percentage of one side or the other.

    Do I think Bush would poll well now? No. Do I think it's 38%? No.

    Does it make any difference? No.

  4. JA - They often do, though it can only be done to a certain extent; and furthermore, the LATimes poll shows no evidence of having done so.

  5. And CharlieHall noted that it's been around 40% on Zogby as well - I'm not doubting that, nor that it's somewhat close to accurate. But these polls that are coming in so low and getting such heavy press are off, and it shouldn't be ignored. It's clear Bush won't poll well now, but to have 60% of the country think the economy is doing poorly either shows that polls are dumb or people are dumb - if not both.

  6. Anon Y. Mouse is ripping you into little tiny pieces over at my outstanding, reality-based, blog.

  7. DovBear is wrong, Ezzie is right.

    Bush's approval rating is higher than most polls show and for a very simple reason:

    They don't poll likely voters. If they did, they would encounter higher Republican numbers. The general populace's opinion is wonderful to know, but the people who care enough to actually go to the polls select the representatives in this country.

  8. Thanks for the post, Ezzie. I often think about how skewed these posts are, but had never thought about how to calculate it specifically. Much appreciated.

  9. Croaky - Thanks, another good point.

    Scottage - My pleasure.