After the 2008 primary season, I analyzed Ron Paul’s performance in each state to see how institutional factors such as caucus and primary form affected his electoral success. This exercise turned out to be useful for estimating the size of the pro-liberty electorate in each state. In this post, I do the same with the 2012 results.
The dependent variable in this analysis is the percentage of the vote obtained in each state’s statewide primary or caucus. If a state held both a primary and a caucus or convention, I used the primary results. The independent variables are as follows. (All variables were taken from Dave Leip’s Atlas of U.S. Presidential Elections.)
First, I use a dummy variable for whether the state had a caucus or convention versus a primary. As an outsider candidate with a committed band of activist supporters, Paul tended to do much better in caucuses than primaries.
Next, I also included dummy variables for who is eligible to vote: a dummy for an open caucus/primary, in which all voters are allowed to vote, and a dummy for a closed caucus/primary, in which only registered Republicans are allowed to vote. The excluded category consists of “modified-open” elections, in which independents and Republicans are allowed to vote. Since Paul was more popular among self-identified independents than among self-identified Republicans, it stands to reason that he would do best in open primaries and worst in closed primaries.
I also tried variables for the number of other candidates running active races and appearing on the ballot. In this regard, I counted only “significant” candidates that might have had a chance of winning, that is, the following six (other than Paul): Mitt Romney, Rick Santorum, Newt Gingrich, Rick Perry, Jon Huntsman, and Michele Bachmann. Bachmann dropped out of the race after Iowa and Perry and Huntsman after New Hampshire. Thereafter, Santorum withdrew only after the April 10 primaries and Gingrich held on until May 2. However, in many states – almost entirely primary states – candidates that had withdrawn remained on the ballot and often received significant support. We should expect both of these variables to have a negative influence on Paul’s vote share (for instance, in Virginia, where only Romney and Paul were on the ballot, Paul benefited from tactical voting by supporters of other candidates and received 40%), but in fact the number of candidates in the race had a statistically significant, positive association with Paul’s support (candidates on the ballot has the expected negative association, as shown below). This result is theoretically implausible and apparently simply captures Paul’s otherwise unexplained success in Iowa and New Hampshire. Therefore, in the estimations reported here, I have included solely the variable for number of non-Paul candidates on the ballot.
Next, I included a three-point variable for the extent to which Mitt Romney had clinched the Republican nomination when the primary in question was held. After February 28, it became very unlikely that Romney would lose, as Nate Silver noted here. After April 3, it became a near mathematical certainty that Romney would win. Therefore, the “clinching” variable takes on values of 0.5 for contests held after February 28 through April 3 and values of 1.0 for contests held after April 3.
Finally, I include a variable intended to capture home-state effects for active candidates. It is coded zero for all states except Massachusetts, Georgia, and Texas. For the former two, home states of Romney and Gingrich, respectively, it is coded -1, and for the last, Paul’s home state, it is coded 1. (Santorum dropped out of the race before the Pennsylvania vote, for which reason it is coded zero.)
The dependent variable is rather skewed, as the following histogram shows:
As a result, statistical analysis with this variable revealed the presence of heteroskedasticity, which violates the assumptions necessary for Ordinary Least Squares regression. Therefore, I transformed the variable using the natural log, which gave it a roughly normal distribution and eliminated heteroskedasticity in the regression analysis:
The results of the regression analysis are reported below the fold:
Read Full Post »