Marginal Revolution: The Fox News Effect, revisited

November
S	M	T	W	Th	F	St
1	2	3	4	5	6	7
8	9	10	11	12	13	14
15	16	17	18	19	20	21
22	23	24	25	26	27	28
29	30

« TylerCowensEthnicDiningGuide.com | Main | Jane Jacobs passes away »

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.

Posted by Tyler Cowen on April 25, 2006 at 06:08 PM in Political Science | Permalink

Comments

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

just kidding

Posted by: neil at Apr 25, 2006 8:44:45 PM

That's the same paper? Yeah, I remember it coming out and finding no effect too. They probably mention what happened in the paper.

Posted by: scott cunningham at Apr 25, 2006 9:38:45 PM

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.

Posted by: Franck at Apr 26, 2006 7:38:33 AM

One apparently credible source, who wishes to remain anonymous, wrote to me and claimed that adjusting for heteroskedasticity generates the new result...

Posted by: Tyler Cowen at Apr 26, 2006 7:54:19 AM

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.

Posted by: cbc at Apr 26, 2006 2:30:02 PM

Tyler -- they went from homoskedasticity to hetero? That seems pretty plausible, as I'm under the impression that the only purpose for homoskedasticity is to make the math easy.

Posted by: Jor at Apr 26, 2006 4:33:07 PM

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.

Posted by: Alex Tabarrok at Apr 26, 2006 7:20:30 PM

Yes but the weights here are known - the size of the various cable markets in their sample.

Posted by: user at Apr 26, 2006 7:41:57 PM

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 ?

Posted by: econgeek at Apr 26, 2006 8:01:57 PM

The change they made is to cluster standard errors on the cable affilliates and weighting the observations. The coefficient doesn't change much across the two papers but the standard error on the parameter of interest is about cut in half. The only other change seems to be the inclusion of "other census controls" in the latter paper but I'm not sure if they were also included in the first paper without it being made explicit in the tables.

Posted by: Jason Lindo at Apr 26, 2006 11:10:50 PM

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.

Posted by: Ethan Kaplan at Apr 27, 2006 2:34:33 AM

grietje neuke op kantoor ^^^ blonde ejaculatie in bed ^^^ fighetta gruppo a letto ^^^ bionde strip a letto ^^^ beau infirmiere baise en groupe ^^^ moyen exhibition ^^^ mama dix neuf ^^^ vingt deux insignifiant ^^^ preposterous agente di polizia masturbate ^^^ agreeable amatoriali gruppo ^^^ himmelsk servitris grupperna ^^^ kaninaktig servitris ligga ^^^ taglio fighette ubriache ^^^ infermiera dildo in anticamera ^^^ sexy abspritzen assoziationen ^^^ brutaal secretaresse ejaculatie ^^^ naida porno maisemakuva ^^^ punapaa naken peraaukon ^^^ goy kortvarig ^^^ folsom mama ^^^ elskelig far ^^^ hetest mage ^^^ filikos geladarisa avnanismos ^^^ koutos souideza praxi ^^^ chaud maman trente ^^^ mere capote ^^^ adranis epihirisi ^^^ adranis kolos ^^^