Regressors
Percentage change in per capita income
Reflects an overall change in well-being since the last election. As the Percentage Change in Per Capita Income decreases, the incumbent becomes more unelectable.
Unemployment
Tells how many people are unable to find work as a percentage of the workforce who cannot find find work. As unemployment increases, the incumbent's chance of re-election decreases.
Control of congress
A variable reflecting whether the incumbent's party controls Congress or not. 1 if the incumbent party controls both the house and the senate, 1/2 if the incumbent's party controls only one chamber, 0 if the incumbent's party controls neither.
Incumbent approval rating
The incumbent's approval from the last month in June. As this number decreases, the incumbent becomes more unelectable.
Incumbent vote in the past three elections
A variable showing how people voted in the past for the incumbent. If the previous incumbent vote is high, chances are people will not go drastically away from that result.
Assumptions
- Percentage Change in Per Capita Income : In 2008 CPI Adjusted dollars [link] , 2012 figures assumes same growth rate as 2009-2010
- Unemployment : 1992-2008 figures are averages for the year, 2012 figures are the unemployment data for May
Results
2008
We used both the regression model on its own as well as the partially filled-in Markov model to predict the 2008 election using data from 2004 and earlier. The results are presented in Table 1 and Figure 3.
| State Prediction | Average Prediction Error (magnitude) | |
| Linear Regression | 52.18% | 2.08% |
| Markov Random Field | 47% | 6.84% |
| Actual Result | 54.57% | - |
2012
In 2012, we predict Mitt Romney to win Colorado with about 60% of the vote. Our regression model predicts 60.36% while our Markov model predicts 60.65%. In figure 4, we present the county-by-county predictions.
Turnout assumption
In order to calculate the statewide prediction based on the individual county predictions, we had to estimate how many people would vote in each county in 2012. The assumption we made was that the percentage of Colorado voters in each county would remain constant.
Conclusions/further research
Based off the results from 2008, the Markov Random Field technique does not seem to perform as well as the regression by itself. This could be due to a lack of data (again, we only had 13 observations to learn the model from), or it could be due to inconsistencies in neighboring counties' relationships with one another. If the latter is true, then our hypothesis was incorrect. That is, the relationships between adjacent counties are not consistent enough to use for election predictions. To answer this question for certain would require many more observations, probably from all types of elections for many years. This is one potential area for future research. Despite the uncertainty regarding our Markov model, we can seemingly conclude success with our regression model. It predicted the correct winner in both 2008 (trained on data from 1992-2004) and 2004 (not shown here, trained on data from 1992-2000). Of course, we used a small set of regressors and there is much room for further research in this area, as well.