The Fox News Effect, revisited

Earlier today I reported on a "new" study of how Fox News influences voting patterns; the authors concluded:

We find a significant effect of the introduction of Fox News on the vote share in Presidential elections between 1996 and 2000.

On Wikipedia, however, you can find this link, to a May 2005 version of the paper, by the same authors.  (An alert reader, "MN," pointed me to this.)  Then the authors concluded:

We find no significant effect of the introduction of Fox News on the vote share in Presidential elections between 1996 and 2000.

Hmm…there is much to be said for changing your mind.  Given my motivated dentist and my forthcoming trip to Chicago, I don’t have the time to get to the bottom of this discrepancy, but comments are open in case you can explain how and why the two papers differ.  I am glad I titled that earlier post "Fox News Seems to Matter".

Addendum: Mark Thoma had noted the same a few days ago.


I wonder if taking John Lott on as a research associate made a difference?

just kidding

Either way, it's always very difficult to show causation. Correlation between 2 events is usually quite trivial to show but it does not mean than one event is the cause for the other. In that case it could be that Fox decided to roll out in those places where the audience would be the most receptive. In other words, it's because the audience had shifted that Fox was successful.

There is also a larger sample in the later version - 8,634 in the May 2005 version and 9,256 in the April 2006 version.

In theory, hetero should only change the standard errors not the coefficients. To change the coefficients you have to correct for hetero by weighting which can often be arbitrary and not robust.

Alex; Even when the weights arent known, they can be estimated from the data, isnt that the whole point of Feasible Generalized Least Squares ? Or are my econometrics too rusty ?

Yes... good questions. The change from non-signficance to signficance of our results come from both increases in our point estimates as well as large reductions in our standard errors.

Table 2 in the appendix of the paper outlines how to get from the old to the new estimates. The main differences are (1.) weighting by town size (actually, turnout but weighting by population gives the same results), (2.) we found 4 new states - by far, most of the obserevations came from the state of NJ which, being a highly democratic state, had a large impact of fox news.

In terms of (1.), we noticed strong patterns of heteroskedasticity by town size. Small towns had much larger fluctations in vote shares. Intuitively, random turnout effects play much larger roles in determining vote share in small towns. So, we decided to weight by turnout.

Comments for this post are closed