The Hunt for Campaign Effects in 2008

In earlier posts, I emphasized that election outcomes depend heavily on the fundamental conditions in the country. I also suggested possible ways in which campaigns might matter, although noting that it takes some hardcore polisci-in’ to identify campaign effects.

Two early attempts to identify campaign effects come from Nate Silver and Seth Masket.

Silver looks at exit poll question asking voters whether they were contacted by McCain or Obama. He then subtracts the percent saying McCain from the percent saying Obama. He plots that difference against how well Obama did, compared to’s forecast based on the polls. Here’s the graph:


Silver writes:

There is indeed a fairly strong relationship between contact rate and Obama’s overperformance or underperformance in the polls. (The R-squared of the linear regression line you see in the chart is .51, indicating that about half of the gap between Obama’s projected and actual performance was explained by disparities in the ground game.) Roughly speaking, each marginal 10-point advantage in contact rate translated into a marginal 3-point gain in the popular vote in that state. So the rule of thumb that a “good” ground game may be worth additional 2-3 points above and beyond what is reflected in the polls appears to hold; a great ground game may be worth somewhat more than that.

But the evidence here isn’t convincing. The main problem is that West Virginia and Nevada pretty much account for the relationship, as a few commentators on 538 quickly noted. Silver’s makes the contact data available in a table, but not the vote margins relative to poll predictions. Nevertheless, it’s pretty easy to eyeball his graph and replicate it. Here’s my replication:


Pretty close, and without the unnecessary plus signs on the axes. A regression model produces an r-squared of .48, which is almost equal to the .51 that Silver reports.

Now here’s the same graph without West Virginia and Nevada.


The linear relationship is close to 0 (b=.04; se=.15, for my fellow dorks). The r-squared is .01. There is very little evidence of any correlation. A state like Colorado has a huge Obama advantage in contact, but he does no better there relative to the prediction than he does in Wisconsin, a state with a small Obama advantage in contact. The verdict: Obama’s field organization may have mattered, but Silver’s evidence isn’t persuasive.

Let’s take it a step further, thanks to some work that Seth Masket has done. He looks at Colorado vote returns by county. The question is: does having an Obama field office in the county increase his vote share in that county, relative to Kerry’s in 2004? The “relative to 2004” is important, because Obama does better in virtually every Colorado county relative to Kerry. (See Seth’s post. This is the “uniform swing’ that Andy has discussed here and here.) Here is what Seth finds:


Having a field office in the county increased Obama’s vote share (relative to 2004) by about 2 additional points. Seth also accounts for a potential confounding factor: Obama tended to put field offices in liberal or Democratic-leaning counties. But the effect of field offices is robust, even when controlling for the partisan complexion of counties.

As Seth would no doubt agree, this finding is still provisional because he may not have accounted for other attributes of counties that might be correlated with both field offices and votes. Nevertheless, this is a promising start.

I hope to have more posts on this theme in the weeks ahead.

7 Responses to The Hunt for Campaign Effects in 2008

  1. Andrew November 8, 2008 at 10:53 pm #


    I don’t think the plus signs are so horrible. But I’d label the y-axis at every 5% and the x-axis at every 10%. Your labels (and Nate’s) are too busy.

    And boy are those boxplots ugly. Buy, borrow, or steal a scatterplot, dude!

  2. chrismealy November 8, 2008 at 11:52 pm #

    A quibble: of course Obama’s organizing mattered — the question is whether it showed up in the pre-election polls.

  3. John Sides November 9, 2008 at 8:46 am #

    Andy, I respectfully disagree. Everyone knows that numbers without plus signs are positive numbers. And, for my taste, the labels are not too busy. Most importantly: that’s a gutsy thing to say about the boxplots. Thou darest blaspheme Tukey?

    Chris: Your question is correct. But we cannot say “Of course Obama’s organizing mattered” until we can actually prove it!

  4. Andrew November 9, 2008 at 3:36 pm #

    Tukey had many good ideas and many bad ones as well. Remember stem-and-leaf plots? Hanging rootograms? His wacky multiple comparisons procedures? I’d prefer to follow his principles than to use all his methods.

  5. Eric L. November 9, 2008 at 5:47 pm #

    As Nick Cox has noted several times on Statalist, Tukey in essence re-invented the box plot. The Stata manual cites a 1933 paper from the Scottish Geographical Magazine as the earliest source, though that MS used the more boring name of “dispersion diagram.”

    So let’s keep Tukey out of this.

  6. Seth Masket November 10, 2008 at 2:23 pm #

    Say what you want about me, but leave my boxplots out of this. Anyway, I’ve redone the figure as a scatterplot. Hope this helps.

  7. Doug H. November 11, 2008 at 4:55 pm #

    Interesting stuff (not the debate about box plots). It would be worthing knowing just how rapidly these counties changed between 2000 and 2004 (demographically, but also just movement in and out). I look at demographic data about some Mountain and Southwestern counties and the population growth and turnover is amazing.