Predicting Presidential Elections from Ale to Zinfandel

What variables do you need to consider to accurately predict the outcome of a U.S. presidential election?

The current state of the economy?

International affairs and the threat of terrorism?

The specific plans and policies proposed by each presidential candidate and how well each plan resonates with likely voters?

The amount of money raised by each campaign?

Surely each of these factors, among many others, is an important predictor of who will go on to win the Presidency.

But how about another, perhaps less obvious, variable.

As ridiculous as it might sound, one might be able to predict the winner in a U.S. presidential election based solely on the amount of alcohol people in each state consume.

Data from the Brewer’s Almanac at BeerInstitute.org suggest that residents of Blue States drink about ¾ of a gallon of alcohol more than residents of Red States the year prior to a presidential election (see Table 1).

Table 1: Per capita alcohol consumption (beer, wine, & spirits) for Red States vs. Blue States during the year prior to a presidential election (1996-2012).

Red States vs. Blue States

More specifically, residents of Red States generally consume more beer each year than residents of Blue States, whereas residents of Blue States consistently consume more wine and spirits each year than residents of Red States (see Figures 1-3).

Figure 1: Average per capita consumption of beer for Red States (in Red) and Blue States (in Blue) during the year prior to a presidential election (1996-2012).

Beer Consumptions Red vs. Blue

Figure 2: Average per capita consumption of wine for Red States (in Red) and Blue States (in Blue) during the year prior to a presidential election (1996-2012).

Wine Consumption Red vs. Blue

Figure 3: Average per capita consumption of spirits for Red States (in Red) and Blue States (in Blue) during the year prior to a presidential election (1996-2012).

Spirits Consumption Red vs. Blue

So the question is this: Can we use data on per capita alcohol consumption to predict ahead of time which states will go to a Republican presidential candidate and which ones will go to a Democratic presidential candidate?

Predicting Presidential Elections Based on Per Capita Consumption of Alcohol in Each State

For each of the past five presidential elections (Clinton [D] vs. Dole [R] in 1996; Bush [R] vs. Gore [D] in 2000; Bush [R] vs. Kerry [D] in 2004; Obama [D] vs. McCain [R] in 2008; and Obama [D] vs. Romney [R] in 2012), I used Logistic Regression to try to predict, or rather “post-dict,” the presidential candidate that would win in each state, including the District of Columbia.

The resulting logistic model provides an estimate of the probability of each state being won by the Democratic candidate, based solely on three predictor variables (see Figure 4 for estimated probabilities):

  1. Per capita consumption of beer in each state during the year prior to the general election.
  2. Per capita consumption of wine in each state during the year prior to the general election.
  3. Per capita consumption of spirits in each state during the year prior to the general election.

*Data on per capita consumption of each type of alcohol was gathered from the Brewer’s Almanac at BeerInstitute.org.

Figure 4: Probability of the Democratic candidate winning each state during each election cycle between 1996 and 2012.

Table 2 below shows the number of states each election cycle for which the winning presidential candidate was correctly predicted by the best-fitting logistic model.

Table 2: Number of states (including the District of Columbia) for which the winning presidential candidate was correctly predicted on the basis of per capita consumption of beer, wine, and spirits (1996-2012).

presidential predictions by state (96-12)

On average, the best-fitting logistic model correctly predicted the winning presidential candidate in 79.22% of states, including the District of Columbia.

Moreover, the model correctly predicted which candidate would go on to win 270 votes in the Electoral College (and thus the Presidency) in every single election between 1996 and 2012.

Table 3: Predicted vs. Actual Electoral College (EC) Votes for the Democratic Candidate for President (1996-2012).

Predicted Electoral College Votes (96-12)

Table 4 provides a more detailed breakdown of the model’s accuracy in making presidential predictions each election cycle.

Table 4: Detailed breakdown of model accuracy in predicting presidents based on per capita consumption of beer, wine, and spirits in each state (1996-2012).

SDT measures

“Hits” is the hit rate, and reflects the proportion of states correctly predicted to go to the Democratic candidate.

“FA’s” is the false alarm rate, and reflects the proportion of states incorrectly predicted to go to the Democratic candidate.

“d prime” is a measure of accuracy derived from an influential theory in psychology and psychophysics known as Signal Detection Theory. D prime reflects the model’s accuracy in discriminating between states actually won by a Republican candidate vs. states actually won by a Democratic candidate. A d prime value of 0 reflects chance performance, whereas values greater than 0 reflect superior discrimination.

“c” is also derived from Signal Detection Theory, and it is a measure of the model’s bias in favor of either a Republican candidate or a Democratic candidate. A c value that is less than 0 reflects a bias in favor of the Democratic presidential candidate, whereas a c value that is greater than 0 reflects a bias in favor of the Republican presidential candidate (a c value of 0 reflects no bias).

“EC hits” reflects the number of Electoral College votes that were correctly predicted by the model to go to the Democratic candidate.

“EC FA’s” reflects the number of Electoral College votes that were incorrectly predicted by the model to go to the Democratic candidate.

“EC CR’s” reflects the number of Electoral College votes that were correctly predicted by the model to go to the Republican candidate (CR stands for “Correct Rejections”).

“EC Misses” reflects the number of Electoral College votes that were incorrectly predicted by the model to go to the Republican candidate.

The final column, “Bias,” reflects the difference between the number of Electoral College votes that were incorrectly awarded to the Democratic candidate (“EC FA’s”) vs. those that were incorrectly awarded to the Republican candidate (“EC Misses”).

As you can see in Table 4, the model consistently performs quite well in discriminating between Red States and Blue States each election cycle (d prime was greater than 0 each year).

Nonetheless, some states were consistently difficult to predict based on per capita alcohol consumption alone. For the four states listed below, the model made incorrect predictions 80% of the time (i.e., in 4 out of the 5 elections):

  1. Alaska – correctly predicted to go to George W. Bush (R) in 2000 and incorrectly predicted to go to the Democratic candidate every other year.
  2. Idaho – correctly predicted to go to George W. Bush (R) in 2004 and incorrectly predicted to go to the Democratic candidate every other year.
  3. Iowa – correctly predicted to go to George W. Bush (R) in 2004 and incorrectly predicted to go to the Republican candidate every other year.
  4. Pennsylvania – correctly predicted to go to Bill Clinton (D) in 1996 and incorrectly predicted to go to the Republican candidate every other year.

Furthermore, although the model discriminates Red States from Blue States fairly well, it exhibits a slight bias in favor of Republican candidates. In general, Republican candidates benefit when the model incorrectly allocates Electoral College votes (“EC FA’s” and “EC Misses”), as more votes are mistakenly allocated each year to Republicans than to Democrats, with the exception of 1996.

Per Capita Wine Consumption as a Predictor of Presidential Elections

Although the election predictions presented here were based on per capita consumption of beer, wine, and spirits in each state, the strongest and most consistent predictor each election cycle was per capita consumption of wine.

Figures 5 through 9 show the relationship between per capita wine consumption in each state and the model’s estimates of the probability of each state being won by the Democratic candidate.

The red data points represent the states actually won by the Republican candidate and the blue data points represent the states actually won by the Democratic candidate.

Figure 5: Probability of each state being won by a Democrat in 1996 vs. per capita wine consumption in 1995.

Wine and Election prob_1996

Figure 6: Probability of each state being won by a Democrat in 2000 vs. per capita wine consumption in 1999.

Wine and Election prob_2000

Figure 7: Probability of each state being won by a Democrat in 2004 vs. per capita wine consumption in 2003.

Wine and Election prob_2004

Figure 8: Probability of each state being won by a Democrat in 2008 vs. per capita wine consumption in 2007.

Wine and Election prob_2008

Figure 9: Probability of each state being won by a Democrat in 2012 vs. per capita wine consumption in 2011.

Wine and Election prob_2012

As you can see by comparing Figures 5 through 9, per capita wine consumption has been an increasingly strong predictor of presidential election outcomes since 1996.

This is especially clear from the estimated odds ratios from each election cycle, shown below in Figure 10.

The odds ratio represents the multiplicative change in the odds of a Democratic candidate winning a given state for every gallon of wine a resident of that state consumed the year prior to the general election.

Figure 10: Odds ratio estimates for per capita consumption of beer and wine for each election cycle (1996-2012).

Odds Ratio Estimates (96-12)

In 1996, Bill Clinton’s odds of winning a given state increased by a factor of 5.34 for every gallon of wine a resident consumed in 1995.

Meanwhile in 2012, Barack Obama’s odds of winning a given state increased by a factor of 13.93 for every gallon of wine a resident consumed in 2011.

In general, the odds of a Democrat winning a given state increase by a factor of 8.74 for every gallon of wine a resident consumes the year before the general election. And the odds of a Democrat winning a state increase by 2.37 for every gallon of spirits a resident of the state consumes.

Meanwhile, the odds of a Republican winning a given state increase by a factor of 1.13 for every gallon of beer a resident consumes the year before the general election.

Conclusion

So what should we make of all this?

Does drinking wine and spirits make a person more Democratic and liberal? That’s pretty doubtful. Likely, individual alcohol preferences relate to the political leanings of a state because preferences for beer, wine, or spirits relate to traits, characteristics, attitudes, and opinions that are more highly predictive of whether someone votes Democrat or Republican.

But regardless of why per capita alcohol consumption relates to the political leanings of a state, can we predict who is going to win the 2016 Presidential election once we know how much alcohol people have consumed in 2015?

Maybe, but I’d caution against jumping to even this conclusion. After all, we only know that the model performed fairly well in five presidential elections, so our sample size is rather small.

Nonetheless, in my next post I’ll apply the model to projections of per capita alcohol consumption for 2015 and make a tentative prediction as to whether the eventual Democratic or Republican candidate will win the Presidency in 2016.

If the model’s predictions turn out to be accurate come November 2016, well then that will be worth celebrating with a fine bottle of wine…or a fine bottle of craft beer.

Leave a Reply

Your email address will not be published. Required fields are marked *