Skip to main content

View Diary: Can the Democrats Take Back the House? (75 comments)

Comment Preferences

  •  Your wrong (0+ / 0-)

    take a look at the PPP polling. Remember many of the young haven't registered yet. Marginal voters often aren't registered.  The PPP polling is some states is showing a turnout of half what of what it was in 2008

    More to the point, you don't understand the larger issue.  Likely voter screens are picking up intention at the moment.  But we know from past experience that  this will change.  

    The likely voter argument falls apart when you look at races involving real candidates - eg for President.  The analysis presented with the exception of 2010 cannot reject the null hypothesis - that there is not statitical difference - given the margin of error in these polls.

    The bitter truth of deep inequality has been disguised by an era of cheap imported goods and the anyone-can-make-it celebrity myth - Polly Toynbee

    by fladem on Tue May 17, 2011 at 06:40:20 AM PDT

    [ Parent ]

    •  Trust me, I do understand the larger issue. (0+ / 0-)

      We haven't pulled the stuff in the diary out of thin air. That's literally thousands of hours of work on elections and statistics. There is a gap between likely and registered voter polling in favor of the Republicans.
      You can't even tell, because you need a complete statistical model that separates variance, House effects and movement in opinion to be able to quantify that gap. And given the fact that there are dozens to hundreds to sometimes upwards of thousand polls in a cycle a difference of 1 point is highly, highly statistically different from zero.

      •  And so have I (0+ / 0-)

        And you are still wrong.

        The bitter truth of deep inequality has been disguised by an era of cheap imported goods and the anyone-can-make-it celebrity myth - Polly Toynbee

        by fladem on Tue May 17, 2011 at 07:12:34 AM PDT

        [ Parent ]

      •  Too extend (0+ / 0-)

        on that comment, among the mistakes you are making is one of the most common I see: you assume that you have thousands of polls in your sample.

        But in a way you only have 4.  You have only 4 elections in you sample.  This is the sort of stuff I once spoke to Nassim Taleb about with respect to elections.

        He  shook his head in anger.

        The bitter truth of deep inequality has been disguised by an era of cheap imported goods and the anyone-can-make-it celebrity myth - Polly Toynbee

        by fladem on Tue May 17, 2011 at 07:15:50 AM PDT

        [ Parent ]

        •  Well, the issue with a low number of elections (0+ / 0-)

          is an issue when we do stuff like correlating GDP growth or Unemployment rate with election results, or trying to establish rules like "Incumbents win big or lose big".

          It is not, however, a problem when dealing with opinion polling, since we don't actually need an election to happen to look at the gap between likely and registered voter polling. We can estimate the gap even before the election.

          •  NO (0+ / 0-)

            you don't undertstand.

            You are assuming the same relationship exists between these two: but you have only 4 sets of data.

            If you go further back you find the dynamic changes.

            Because elections are not frequent, they are severe limits on the lessons you can learn from statistical anaysis.

            Ask the guys who created the models of the chance of mortgage defaults.

            Your model may spit out high R squareds, but it is very flawed.

            Again, read Taleb.

            The bitter truth of deep inequality has been disguised by an era of cheap imported goods and the anyone-can-make-it celebrity myth - Polly Toynbee

            by fladem on Tue May 17, 2011 at 07:43:40 AM PDT

            [ Parent ]

            •  Your mention of R-squared makes me suspect (0+ / 0-)

              that we're talking of different kinds of models. These aren't frequentist OLS regressions, but highly complex Bayesian MCMC-filters of opinion polling.

              Sure, with just some regressions it's impossible to get anything significant out of the data, but with something more advanced you can.

              Also, "if you go further back"- datasets from before 2004 are crap. Both the quantity and more importantly the quality of polling back then was terrible to the point that I suspect that many pollsters just made shit up back then.

            •  We're still in Kansas (0+ / 0-)

              All of the distributions involved are normal or Laplace, and we're only dealing with simple objective functions. Mediocrestan and all that jazz.

        •  Could you be clear (0+ / 0-)

          When you say that there are only four elections in my sample, what exactly do you mean?

          Let me be more clear as to what model we're working with:

          Public opinion follows a boring random walk

          alpha[t] = alpha[t-1]+[randomly distributed noise]

          The measurement model is a little more complicated

          Poll~norm(mu, PIE^2*p(1-p)/n)

          mu = alpha[day of poll]+houseeffect[organization of the poll]

          So then, the idea is that you set priors on the house effect(And constrain an overidentified model!):

          Houseeffect[j]~norm(0, sigma) if it's a Registered voter poll,
          Houseeffect[j]~Norm(LVV,sigma) if it's a likely voter poll.

          The hyperprior LVV then will be the "Enthusiasm gap"

Subscribe or Donate to support Daily Kos.

Click here for the mobile view of the site