The Role of the Economy in Elections
Samuel Thau
9/19/2020
Introduction
Many election forecasting models rely at least in part on so called “fundamentals,” that are outside of the candidates control. The classic example of a “fundamental” is the economy. The incumbent challengers (and theoretically, a candidate from the incumbent party) could have some influence on the economy, but in general it is well outside of the control of any one person or campaign. For example, some believe that the economy was the deciding factor in the 2012 election, given then President Barack Obama the edge over Mitt Romney, with the economy growing just enough to overcome the typical anit-incumbent sentiment in United State elections1.
In doing so, a number of assumptions are made about voters. There are three main factors to consider when attempting to model voter choice as a function of the economy:
- How direct is the relationship between voting and the economy?
- Do voters have complete information?
- Do voters have a sociotropic or individual focus?
Questions 2 and 3 have been studied empirically, and will inform the model that I build during this post. The answer to question 1 is more philosophical: does seeing higher GDP numbers make voters more inclined to vote one way or another? Or is it more indirect, where a higher GDP leads to more money in people’s pockets, making them happier which makes them more inclined to vote for the status quo? Figuring out the answer is tricky.
Choosing Model Inputs
To model voter behavior, we must first think about the inputs into a model. Let’s begin with the issue of time frame. It is a well documented fact that people have short memories when it comes to abstract ideas like the economy, and that remains true when looking at voting outcomes. Not only are people much more responsive to the election year economy2, placing nearly 75% of the weight on the election year economy, they place the majority of their weight on the final two quarters before the election3. Because of this, I will use data from the second quarter of 2020 (as third quarter data does not yet exist).
To deal with the problem of figuring out which variables to choose, consider both questions 1 and 3 from the introduction. I believe that it is a reasonable assumption that people care about their own income, but also are influenced by the national economy (either indirectly through the job market, or just through the news). Because of this, I will use two economic indicators: national level GDP growth in quarters 1 and 2 of the election year, and personal income growth in quarters 1 and 2. In both cases, the data I use is compared to the previous quarter and controls for the time of year. I will vary the scope (e.g. national vs state) of the personal income data, depending on the exact model.
Varying the scope of personal income is intentional: I believe it is a reasonable assumption to make that for more direct economic indicators, people may be influenced a more local level. After examining the results, we will return to if this is actually a reasonable assumption to make.
A National Model
We can begin with a national level model, looking at data over time. We can regress national personal income growth in the first and second quarters, along with national GDP growth on the incumbents two party vote share in each election year4.
##
## Call:
## lm(formula = vote_margin ~ 1 + GDP_growth_qt_1 + GDP_growth_qt_2 +
## RDI_growth_1 + RDI_growth_2, data = nat_econ_df)
##
## Residuals:
## Min 1Q Median 3Q Max
## -6.478 -2.683 1.478 2.616 5.366
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 48.343 2.621 18.447 4.72e-09 ***
## GDP_growth_qt_1 2.254 1.467 1.536 0.156
## GDP_growth_qt_2 2.149 1.426 1.507 0.163
## RDI_growth_1 -1.675 2.149 -0.779 0.454
## RDI_growth_2 1.416 1.928 0.735 0.479
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 4.388 on 10 degrees of freedom
## Multiple R-squared: 0.5003, Adjusted R-squared: 0.3004
## F-statistic: 2.503 on 4 and 10 DF, p-value: 0.1091This model does an all together mediocre job of prediction.
It has an R^2 value of
0.50, which means that the model explains almost exactly half of the variance in the dataset.There is a negative coefficient on the first quarter personal income of
-1.675, while there is a positive coefficient on the second quarter personal income of1.416. This would suggest that voters really have a “what have you done for me lately” attitude. This means that for an increase of one percent personal income growth in the first quarter, holding all else constant, the vote share for the incumbent party is expected to decrease.In general, the standard errors for the coefficients are quite high, relative to the magnitude of the coefficients, suggesting that these results cannot be taken particuarly seriously. However, this may just be a problem with sample size.
All together, this model seems like it could be useful, but we must first map it onto the state level. Something to check is if the predicted outcomes makes sense. This model predicts that Trump will win 28.69 percent, an exceedingly low fraction of the popular vote. This is likely tied to the severe decline in GDP in the second quarter due to COVID.
A State Level Model
As discussed in last week’s post, the states are what really matters for winning elections. We can move to a state by state model, first estimating the results for each state indepedently. To do so, we draw on personal income data from the first and second quarters of each election year dating back to 19605. Unfortunately, the state by state data for quarter 2 will not be released until later this week, so for the time being I used the national personal income growth for quarter 2 in every state. We run 50 indepdent regressions, one for each state and estimate the results of this year’s election by doing so.
This election map is somewhat surprsing: again, it would appear that a massive GDP drop off in the second quarter dooms President Trump. This would be one of the worst results in history of the United States. We can look at the vote margins for Trump:
## State Trump Vote Share Vote Margin Winner
## 1 Alabama 26.403849 -47.192301 Biden
## 2 Alaska 42.090632 -15.818736 Biden
## 3 Arizona 5.013441 -89.973118 Biden
## 4 Arkansas 18.532859 -62.934282 Biden
## 5 California 28.805607 -42.388786 Biden
## 6 Colorado 17.237667 -65.524667 Biden
## 7 Connecticut 21.051791 -57.896417 Biden
## 8 Delaware 19.302765 -61.394470 Biden
## 9 Florida 21.019253 -57.961494 Biden
## 10 Georgia 28.036505 -43.926991 Biden
## 11 Hawaii 34.493926 -31.012149 Biden
## 12 Idaho -8.778350 -117.556699 Biden
## 13 Illinois 29.752929 -40.494143 Biden
## 14 Indiana 24.464451 -51.071097 Biden
## 15 Iowa 43.976066 -12.047869 Biden
## 16 Kansas 23.640135 -52.719730 Biden
## 17 Kentucky 32.260160 -35.479679 Biden
## 18 Louisiana 5.473828 -89.052345 Biden
## 19 Maine 28.321509 -43.356982 Biden
## 20 Maryland 36.251570 -27.496860 Biden
## 21 Massachusetts 28.295551 -43.408898 Biden
## 22 Michigan 27.068831 -45.862339 Biden
## 23 Minnesota 49.285636 -1.428729 Biden
## 24 Mississippi 5.763629 -88.472741 Biden
## 25 Missouri 30.524398 -38.951204 Biden
## 26 Montana 16.542776 -66.914448 Biden
## 27 Nebraska 3.466203 -93.067594 Biden
## 28 Nevada -11.082928 -122.165855 Biden
## 29 New Hampshire 5.685177 -88.629645 Biden
## 30 New Jersey 18.881777 -62.236445 Biden
## 31 New Mexico 19.027538 -61.944925 Biden
## 32 New York 32.251470 -35.497059 Biden
## 33 North Carolina 23.879665 -52.240671 Biden
## 34 North Dakota 4.273159 -91.453682 Biden
## 35 Ohio 29.517961 -40.964079 Biden
## 36 Oklahoma 20.836628 -58.326744 Biden
## 37 Oregon 43.607713 -12.784573 Biden
## 38 Pennsylvania 31.253629 -37.492743 Biden
## 39 Rhode Island 27.016549 -45.966903 Biden
## 40 South Carolina 15.756994 -68.486011 Biden
## 41 South Dakota 73.173053 46.346105 Trump
## 42 Tennessee 37.509403 -24.981195 Biden
## 43 Texas 22.564355 -54.871289 Biden
## 44 Utah -23.921214 -147.842427 Biden
## 45 Vermont 45.519178 -8.961644 Biden
## 46 Virginia 23.229998 -53.540004 Biden
## 47 Washington 29.065603 -41.868794 Biden
## 48 West Virginia 27.949838 -44.100324 Biden
## 49 Wisconsin 43.614864 -12.770272 Biden
## 50 Wyoming -8.058901 -116.117802 Biden
## 51 District of Columbia 203.792194 307.584387 TrumpBased on this model, Trump is not only consistently losing, but consistently getting routed. This includes several states with negative vote shares, and the District of Columbia with a nonsensical 203.79 vote share. One potential problem with this model is that each state is independent: it could be that in years where the economy is in bad shape, certain states respond in similar ways across time which could reduce some of the more drastic effects.
Incorporating State Fixed Effects
Instead of running 50 seperate regressions, we can instead run a single, much larger regression but with dummy variables for the states. This controls for particular states behaving similarly across time6.
##
## Call:
## lm(formula = vote_share ~ Q1 + Q2 + gdp_1 + gdp_2 + state, data = state_reg)
##
## Residuals:
## Min 1Q Median 3Q Max
## -43.325 -6.317 0.219 6.497 49.152
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 47.64501 2.92181 16.307 < 2e-16 ***
## Q1 0.14035 0.06158 2.279 0.0229 *
## Q2 -0.09386 0.05943 -1.579 0.1147
## gdp_1 2.07015 0.49006 4.224 2.71e-05 ***
## gdp_2 2.57721 0.38422 6.708 4.05e-11 ***
## stateAlaska 0.35305 3.97471 0.089 0.9292
## stateArizona -0.23693 3.97564 -0.060 0.9525
## stateArkansas -1.53453 3.97404 -0.386 0.6995
## stateCalifornia -0.36613 3.97611 -0.092 0.9267
## stateColorado -0.24455 3.97579 -0.062 0.9510
## stateConnecticut 1.10445 3.97457 0.278 0.7812
## stateDelaware 0.24263 3.97565 0.061 0.9514
## stateDistrict of Columbia -1.41817 4.04246 -0.351 0.7258
## stateFlorida 0.40891 3.97685 0.103 0.9181
## stateGeorgia -1.39878 3.97452 -0.352 0.7250
## stateHawaii 2.98462 3.97598 0.751 0.4531
## stateIdaho -1.50155 3.97445 -0.378 0.7057
## stateIllinois -0.44256 3.97451 -0.111 0.9114
## stateIndiana -0.66008 3.97450 -0.166 0.8681
## stateIowa -1.50825 3.97491 -0.379 0.7045
## stateKansas -1.15899 3.97510 -0.292 0.7707
## stateKentucky -0.85581 3.97477 -0.215 0.8296
## stateLouisiana -1.39417 3.97428 -0.351 0.7258
## stateMaine 2.08528 3.97431 0.525 0.6000
## stateMaryland 0.50613 3.97563 0.127 0.8987
## stateMassachusetts 0.24969 3.97437 0.063 0.9499
## stateMichigan 0.49366 3.97410 0.124 0.9012
## stateMinnesota -1.34342 3.97588 -0.338 0.7355
## stateMississippi -1.48956 3.97494 -0.375 0.7080
## stateMissouri -1.15179 3.97413 -0.290 0.7720
## stateMontana -2.12416 3.97598 -0.534 0.5933
## stateNebraska -0.45873 3.97422 -0.115 0.9081
## stateNevada -1.01995 3.98179 -0.256 0.7979
## stateNew Hampshire 1.04226 3.97466 0.262 0.7932
## stateNew Jersey 1.64157 3.97466 0.413 0.6797
## stateNew Mexico -0.83239 3.97407 -0.209 0.8342
## stateNew York 0.87950 3.97435 0.221 0.8249
## stateNorth Carolina -0.65306 3.97477 -0.164 0.8695
## stateNorth Dakota -2.72779 3.98947 -0.684 0.4944
## stateOhio -0.13395 3.97468 -0.034 0.9731
## stateOklahoma -0.40766 3.97544 -0.103 0.9184
## stateOregon -1.40166 3.97756 -0.352 0.7247
## statePennsylvania -0.74369 3.97715 -0.187 0.8517
## stateRhode Island 1.39328 3.97444 0.351 0.7260
## stateSouth Carolina -1.37423 3.97595 -0.346 0.7297
## stateSouth Dakota -1.88598 4.00152 -0.471 0.6376
## stateTennessee -1.25400 3.97622 -0.315 0.7526
## stateTexas -0.11298 3.97593 -0.028 0.9773
## stateUtah -0.16320 3.97423 -0.041 0.9673
## stateVermont 1.57425 3.97492 0.396 0.6922
## stateVirginia 0.14062 3.97537 0.035 0.9718
## stateWashington 0.28444 3.97969 0.071 0.9430
## stateWest Virginia -1.84890 3.97510 -0.465 0.6420
## stateWisconsin -1.18991 3.97404 -0.299 0.7647
## stateWyoming -1.58783 3.97480 -0.399 0.6897
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 10.69 on 708 degrees of freedom
## Multiple R-squared: 0.13, Adjusted R-squared: 0.06364
## F-statistic: 1.959 on 54 and 708 DF, p-value: 8.309e-05Taking a careful look at this regression reveals some interesting trends.
Q1andQ2are the personal income in the first and second quarters before an election year, and based on the coefficients, they do not seem to matter for the incumbent’s vote share nearly as much as GDP or the state fixed effects.One way to interpret the state fixed effects is as a measure of willingness to vote for an incumbent. With that interpretation in mind, it is not surprising that many of the coefficients are negative, given the well documented bias against incumbents.
Before making predictions or looking at the electoral map, again this would seem to tilt heavily in Biden’s favor for the upcoming election, given the severe drop in GDP during the second quarter of this year.
The standard error is huge, at
10.69, meaning that this entire regression may not be of any real use.
We can use this model to generate an electoral map, along with electoral college results.
## State Trump Vote Fraction Vote Margin Winner
## 1 Alabama 23.52186 -52.95628 Biden
## 2 Alaska 23.65034 -52.69931 Biden
## 3 Arizona 23.34108 -53.31785 Biden
## 4 Arkansas 21.84697 -56.30605 Biden
## 5 California 23.07152 -53.85696 Biden
## 6 Colorado 23.16503 -53.66994 Biden
## 7 Connecticut 24.55614 -50.88772 Biden
## 8 Delaware 23.79256 -52.41487 Biden
## 9 District of Columbia 22.00545 -55.98911 Biden
## 10 Florida 23.94481 -52.11038 Biden
## 11 Georgia 22.15116 -55.69769 Biden
## 12 Hawaii 26.08542 -47.82916 Biden
## 13 Idaho 22.16066 -55.67867 Biden
## 14 Illinois 23.00912 -53.98175 Biden
## 15 Indiana 22.62317 -54.75365 Biden
## 16 Iowa 22.08379 -55.83243 Biden
## 17 Kansas 22.22251 -55.55497 Biden
## 18 Kentucky 22.56780 -54.86440 Biden
## 19 Louisiana 22.02945 -55.94111 Biden
## 20 Maine 25.63521 -48.72957 Biden
## 21 Maryland 23.92974 -52.14052 Biden
## 22 Massachusetts 23.70137 -52.59726 Biden
## 23 Michigan 23.56638 -52.86723 Biden
## 24 Minnesota 22.13634 -55.72733 Biden
## 25 Mississippi 21.93405 -56.13189 Biden
## 26 Missouri 22.25779 -55.48443 Biden
## 27 Montana 21.20121 -57.59758 Biden
## 28 Nebraska 23.06314 -53.87373 Biden
## 29 Nevada 22.23524 -55.52953 Biden
## 30 New Hampshire 24.39570 -51.20860 Biden
## 31 New Jersey 25.00904 -49.98191 Biden
## 32 New Mexico 22.97018 -54.05964 Biden
## 33 New York 24.09259 -51.81483 Biden
## 34 North Carolina 22.88283 -54.23433 Biden
## 35 North Dakota 20.72390 -58.55221 Biden
## 36 Ohio 23.23352 -53.53295 Biden
## 37 Oklahoma 23.04403 -53.91194 Biden
## 38 Oregon 21.99389 -56.01223 Biden
## 39 Pennsylvania 22.62378 -54.75243 Biden
## 40 Rhode Island 24.85900 -50.28199 Biden
## 41 South Carolina 22.18974 -55.62053 Biden
## 42 South Dakota 21.50957 -56.98087 Biden
## 43 Tennessee 22.14155 -55.71691 Biden
## 44 Texas 23.50713 -52.98574 Biden
## 45 Utah 23.48498 -53.03004 Biden
## 46 Vermont 25.06804 -49.86391 Biden
## 47 Virginia 23.74669 -52.50661 Biden
## 48 Washington 23.76420 -52.47161 Biden
## 49 West Virginia 21.57472 -56.85057 Biden
## 50 Wisconsin 22.27581 -55.44838 Biden
## 51 Wyoming 21.91999 -56.16002 BidenBiden wins quite literally every single state, in what would be the most lopsided election in United States history. There are also some particularly odd results: Biden would be expected to win Missouri by more than he would win Rhode Island, whcih is directly contrary to reality. Biden would be expected to win every state by roughly the same margin as in the national prediction; this suggests that particular states do not vote differently based on economic variables.
Earlier, when I assumed that there would be heterogeneity between states, I was clearly wrong. These results also suggest that trying to use the economy at the state level as a predictor is much more difficult than at the national level - it may be that other factors like demographics, voting restrictions, social issues, the Supreme Court, and other issues are much more important at the local level. While this is an unsatisfying conclusion to a blog post, it is also a valuable lesson for future predictions and modelling: what works at a national level may not work at a state level.
What does this mean for the election?
All three versions of the economics based model I ran give Trump essentially no shot at winning the election, but that does not mean his re-election campaign is completely doomed. Reputable (and much more sophisticated) models like the ones from FiveThirtyEight and The Economist give Trump a serious chance of winning, ranging from 10 to 25 percent.
What my models say is that Trump is going up against a terrible national enviornmnet, but not that he has no shot. A more sophisticated model (similar to what FiveThirtyEight and The Economist do) would use the national enviornment as a starting point and then use other factors like polling, accounting for turnout, and shocks to come up with an eventual winner. This could happen in the form of Bayesian updating7, using a prior based on the economy, demographics and previous election results, and then using polling and current events to update. Once we begin to incorporate polling, this will become a viable strategy for modelling the election.
Sides, John, et al. The Gamble: Choice and Chance in the 2012 Presidential Election - Updated Edition. Princeton University Press, 2014. Project MUSE muse.jhu.edu/book/64467.↩
Healy, Andrew, and Lenz, Gabriel S. “Substituting the End for the Whole: Why Voters Respond Primarily to the Election-Year Economy.” American Journal of Political Science, vol. 58, no. 1, 2014, pp. 31–47.↩
Achen, Christopher H, and Bartels, Larry M. Democracy for Realists. REV - Revised ed., Princeton University Press, 2017.↩
Data for this regression comes from a number of sources. Real disposable income comes from the Bureau of Economic Analysis (BEA). GDP growth comes from the BEA and the Department of Commerce.↩
This data comes from the U.S. Bureau of Economic Analysis, “Quarterly Personal Income By State.”↩
A careful examination of the code reveals that I am not conducting a fixed effects regression in the standard way. To get consistent and unbiased estimates for the coefficients, I would normally subtract the mean value across time for each state. However, with this dataset, doing so leads to problems with collinearity (as GDP and personal income are correlated) when trying to make predictions. For the sake of having predictions, I decided to just include dummy variables instead.↩
More on this later - this is probably the end goal for my overall prediction model.↩