Thanks for the response. I’ve followed you since you were at the RCP and am familiar with and admire your work. You are right to point out that you shouldn’t be grouped with the traditional pundits. If the world is still standing after tonight, I’d like to discuss in greater detail why you have soured on the the statistical analyses that many of my colleagues are increasingly turning to. Nate Silver and I had a good discussion on this topic earlier, and it would be great to get your thoughts.

Meanwhile – good luck tonight!

- Matt

]]>Thank you for mentioning my piece in your post! A few points of clarification:

I am not a “traditional pundit” insofar as I am quite familiar with the methodologies of the “quants,” and have written at length about why I think their approaches are lacking. Most recently, I wrote a post explaining how the polls of Ohio are non-normally distributed, thus creating potentially intractable difficulties in putting together a probability estimate from the polls (which appear to be bi-modal).

In fact, I started blogging in 2004 by doing a very version of what Silver, Linzer, etc. are doing — confidence intervals off state poll averages. It is quite easy to do, in fact. I do not do it anymore because I no longer think the approach is a good one.

In particular: one could very easily put together a statistically powerful predictive model built around the final Gallup poll of likely voters. It shows a Romney victory. Similarly, various predictive models built on fundamentals — e.g. Hibbs’s Bread & Peace — show a comfortable Romney win.

This gets to my fundamental methodological objection with the “quants,” which as I stated in the column you site, it creates a false sense of precision. One can use different variables to build a model that predicts the past with great accuracy AND points to dramatically different results today.

Regards,

Jay Cost

More details here.

]]>