Skip to main content

View Diary: Do states have 'house effects' when it comes to polling? (68 comments)

Comment Preferences

  •  Not that different (3+ / 0-)
    Recommended by:
    radarlady, llywrch, ontheleftcoast

    Actually usually worse than getting a true 10k sample, since the ten firms will likely poll some of the same people.

    But a truly random distribution of 10k is MUCH MUCH more significant than 1K in terms of margins of error.

    The worry with modern polling is the knowledge that the sample is NOT random, and the various models used to estimate the fraction that doesn't answer polls for whatever reason seem to have issues (this article is saying that in 2012, all polls seemed to come in low for the actual winner.  That's a fairly strange bias and thus interesting to study)

    •  Non random samples - why ten 1K polls are better (1+ / 0-)
      Recommended by:

      If you had ONE 10K sample - whatever bias was in the selection would go unbalanced.  By having TEN 1K polls, sample bias is balanced out (as long as you don't have lots of intentional bias, a la Razzy and Gravis)

      •  Only if the bias isn't in the same direction (0+ / 0-)

        You can't assume that.

        Most polling firms try pretty hard on the sampling (the likely voter screen is more likely to have partisan bias).  The systematic errors are far more likely to be tied to methodology than to an attempt to skew the outcome.

        Because of that, because the sampling methodology is similar between polling firms (the primary differences are robocalling vs live interviewer and whether the interviewer can handle languages other than english) there is a concern that even a larger sample won't be free of bias (multiple poll averaging won't eliminate errors, they'll just reinforce them)

    •  Oh lord no. This is dead wrong. (0+ / 0-)

      Sample size has strongly diminishing returns. Standard deviation of a poll, where p is how many people chose candidate A and n the sample size, is sqrt(p*(1-p)/n). That should make fairly obvious that one large poll performs worse than several small ones-- the error doesn't shrink proportionally, it's proportional to the square root.

Subscribe or Donate to support Daily Kos.

Click here for the mobile view of the site