The 2016 Presidential Election and Automation in the Mountain West

Ember Smith, advised by Dr.Djeto Assane

University of Nevada, Las Vegas.

DOI: http://dx.doi.org/10.15629/6.7.8.7.5_6-1_S-2020_3

Citation: E. Smith, “The 2016 Presidential Election and Automation in the Mountain West” Nevada State Undergraduate Research Journal. V6:I1 Spring-2020. (2020). http://dx.doi.org/10.15629/6.7.8.7.5_6-1_S-2020_3

Abstract

The purpose of this study is to determine the effect of a county’s average automation potential, the probability that an occupation will be automated, on the likelihood that a county voted for Donald Trump in the 2016 presidential election in the Mountain West region (Nevada, Arizona, Utah, New Mexico, and Colorado). Although automation potential was statistically insignificant in several ordinary least squares models, factors like education and location both contributed significantly. This research builds on a growing body of analysis investigating the 2016 electoral anomaly and inspires room for further evaluation of the impact of automation in election outcomes.

Introduction

In one of the greatest political upsets of all time, Republican Donald Trump won the 2016 United States presidential election. Pundits, Trump’s opponents, and even his own campaign staff expected his Democratic opponent, Hillary Clinton, to secure the election. As the results from democratic-leaning Rust Belt states like Michigan and Pennsylvania rolled in, her all-but-certain victory slipped away. Trump’s battleground gains proved enough to overcome Clinton’s popular vote victory and leave much of the country wondering how he pulled it off (1).

The driving factors that propelled Trump to victory remain contested, but experts recognize two central themes to his success (2); the first contends that sexism and “racial attitudes” motived support for a candidate sympathetic to hostile attitudes (2–5) and the second suggests that exposure to “economic dislocation” from automation, globalization, or outsourcing motivated populist support, or at least aversion to the Democratic Party (4, 6, 7). While both of these narratives may be true to some degree, Autor and his coauthors argue that both work in tandem (6). Less-educated white voters, especially men, are more likely to be affected by “trade shocks” and feel a heightened sense of “in-group/out-group identification,” both of which may contribute to a surge in the support of rightwing populists (4, 6). Although there seems to be merit to this theory in the Rust Belt (8), automation’s impact in more demographically diverse, economically distinct regions is largely unstudied. As a result, this paper will evaluate if the average automation potential of a county in the Mountain West (Nevada, Arizona, Utah, New Mexico, and Colorado) impacted the vote distribution in the 2016 election.

Race, education, income, gender, and other demographic characteristics are among the traits related to voting behavior. A 2018 Pew Research (9) examination of voters in the 2016 presidential election described voting trends in different groups. Relevant to this study, Pew found that Trump voters in the general election were: more likely to be white (88% of Trump voters were white compared to 60% of Clinton voters); less likely to be college graduates (29% compared to 43%); and more likely to be men (54% compared to 39%) than people that voted for Clinton. When these factors overlap, they compound the likelihood of voting for Trump; white voters without a college degree, 44% of all voters, were over twice as likely (64% to 28%) to vote for Trump than Clinton (9). As is typical for Republican supporters, the median Trump supporter in the 2016 Republican primary election had a household income over $10,000 more than the median Clinton supporter at the same time (10). Other factors, like religion or other educational attainment levels, also impact voter affiliation. Figure 1 displays how several demographic groups are distributed among Trump and Clinton voters using 2018 Pew Research findings.

Figure 1 - Demographic Characteristics among Donald Trump and Hillary Clinton’s Supporters in the 2016 Election.
Figure 1:Demographic Characteristics among Donald Trump and Hillary Clinton’s Supporters in the 2016 Election

Automation potential appeared to have a nontrivial impact on voting attitudes in the Rust Belt, which heightened pressure from automation and offshoring preceding the election (8). Carl Frey and his coauthors argue that, if not for exposure to the adoption of robots in the years preceding the election, Michigan, Pennsylvania, and Wisconsin would not have voted for Trump (8). Workers at high risk of losing their jobs to automation or offshoring are often friendly to Trump’s promise to take on the fight of the “forgotten Americans,” which pushed him to victory in counties with abnormally high automation or offshoring risk (11). Perhaps as a result, Republican districts face an average automation potential — the portion of job tasks at risk of being completed by a robot— of 47.5% compared with 44.7% in Democratic districts (12). Automation may play a role in some families’ voting behaviors; all but one of the ten states most exposed to automation voted for Trump and all but one of the ten states least exposed to automation voted for Clinton (12). In the Mountain West, a region defined to include Nevada, Utah, Arizona, New Mexico, and Colorado, the average county faces an occupation and population-weighted automation potential value of 45.43% (13).

Despite the apparent relationship between automation potential and electoral outcomes in 2016, there may or may not be merit to the claim that workers accurately perceive their job’s measured risk. On one hand, evidence suggests that automation can result in more polarized wages for lower and medium-skilled workers and that increasing the number of robots reduces both wages the employment-to-population ratio (14). On the other hand, if automation complements labor, an increase in automation may make labor more valuable and result in higher wages for workers (12).

As observed in the voting patterns of the 2016 presidential election, it is not necessarily whether a position is going to be automated itself that motivates a worker to vote for a Republican or populist candidate, but the “anxiety” associated with a changing economy (11). Workers may fail to accurately perceive their risk of losing their job to a robot. For example, workers in the manufacturing industry (which as an automation potential of 59%) may expect their job to be at risk to automation, but workers in accommodation and food services may not, even though their job’s automation risk is over ten percentage points higher (12). As a result, county automation potential may only have increased Trump’s support in regions with many workers in industries that are perceived to be under threat.

Although automation disrupted several swing state economies that were consequential in deciding the outcome of the election (8, 15), the correlation may not hold throughout the United States. One counterargument to the automation anxiety hypothesis is that jobs with higher wages typically have a lower automation potential, so the fact that Trump’s supporters make thousands of dollars more than Clinton’s on average may indicate that people with jobs at risk of automation could be more likely to vote for a Democrat. The Rust Belt also faced a substantially steeper decline in manufacturing employment than other regions in the U.S. (15), which may indicate that manufacturers’ response to an increasingly automated economy influenced the election, rather than an industry non-specific reaction to automation potential. Unique industry concentrations may prevent other regions in the country from the same fate, either because other industries have a lower automation potential, or because their workers are less likely to perceive it and vote accordingly.

To investigate whether automation anxiety played a role in the 2016 election in regions without a high concentration of positions in manufacturing, this study will attempt to evaluate automation potential in Mountain West (Arizona, Colorado, Nevada, New Mexico, and Utah) counties. If Autor et al. (6) and Frey et al. (8) are correct about automation and offshoring shifting voting preferences to the right, this project should find that the higher the average automation potential in a county, the larger the share of votes cast for Trump after controlling for demographic factors. This project models automation potential’s impact on the portion of a county that voted for Trump to examine the effect of automation anxiety on the 2016 presidential election results in the Mountain West. The next section will describe each model, the following section will describe the data and descriptive statistics, the next section investigates and discusses empirical results, and the final section will consider limitations and future directions for study.

Model

In order to represent the electoral results of the 2016 presidential election in counties in the Mountain West, the models include data from the MIT Elections Lab, the Brookings Institution, the Census Bureau, and the McKinsey Global Institute. Electoral statisticians frequently use proxy variables for education, race, socioeconomic status to predict elections. Each category is represented in the models.

The data are represented at the county level across the five states in the Mountain West. The dependent variable, retrieved from the MIT Elections Lab (16), is either the proportion of the county that voted for Trump in the 2016 election or whether he received more votes than Clinton, depending on the model. The independent variable of interest is the automation potential of each county, calculated by the Brookings Institution based on a weighted average of automation potential assigned to each major occupational category by the McKinsey Global Institute by employment in each county described by the Bureau of Labor Statistics (12, 13). In their report “A Future that Works: Automation, Employment, and Productivity,” the McKinsey Global Institute developed an exhaustive list of automation potential by “technical potential for automation” for each of over 800 occupations in the U.S. economy (17). Each value represents the “share of current task content that could be automated by 2030 or in next decades based on currently demonstrated technologies,” referred to as “automation potential” (12). To reflect the regional distribution of automation potential, this study uses an employment-weighted automation potential for each county estimated by Muro and his coauthors.

As a result, counties with employment concentrated in industries with a high automation potential will have a high weighted average of automation potential. For example, Eureka County, Nevada has an automation potential of 59.8% (13), the highest county average in the Mountain West. Eureka’s employment in Farming, Fishing, & Forestry Occupations is 12.8 times higher than expected in the country (18); an informative characteristic given that the same industry category has an automation potential of 57% (12). Nearly 85% of the county’s votes went to Trump (16). Los Alamos, New Mexico, on the other hand, had an automation potential of only 34.8% (12), the region’s lowest. Over 20 times more Los Alamos County residents work in Life, Physical, and Social Science occupations than the rest of the U.S. (18), an occupational category with a low automation potential of only 34% (12). Only about 31% of Los Alamos county’s vote share went to Donald Trump (16).

Other independent variables sourced from the 2016 American Community Survey 5-year estimates include the percent of people with a bachelor’s degree in each county to describe average education, the median age to control for generational political tendencies, median income, median house value, the percent of the county that is white, and population size. The models attempt to account for the unique increase in the share of votes for Trump, an abnormal, populist candidate, so two models include the percent that each county voted for Republican Mitt Romney in 2012 (16) and all models include binary variables to control for state fixed effects. Table 1 describes each variable included in the models.

Table 1 - Variable Descriptions

The following three models attempt to describe the presidential election results in 2016 by county in the Mountain West:

Model 1: The first model (M1) includes the county and state attributes that influence the probability a county voted for Trump except for the portion of the county that voted for Romney in 2012 and is described by the following function:Percent Trump = F(automation potential, county controls excluding percent Romney, state controls)

Model 2: The second model (M2) includes the county and state attributes that influence the probability a county voted for Trump, including percent Romney, and is described by the following function: Percent Trump = F(automation potential, county controls including percent Romney, state controls)

Model 3: The third model (M3) is a linear probability model that includes a dependent variable equal to one if a higher portion of the county voted for Trump than voted for Clinton and equal to zero if not. It includes the same county and state controls and is described by the following function: Trump win = F(automation potential, county controls including percent Romney, state controls)

Because its outcome was often inaccurately predicted, typical indicators may not yield the best predictive results for the 2016 presidential election. Nevertheless, because automation potential is the highest for jobs held by people with similar characteristics to Trump’s base, automation potential should be positively correlated with the percentage of people who voted for Trump in a county.

Recent election results indicate that there should be a negative relationship between the portion of a county’s population with a bachelor’s degree or higher and voting Republican (19). Similarly, age is likely to have a positive relationship with voting for conservative candidates, a trend influenced by varying degrees of social conservatism across generations (19). Median income and housing value are also likely positively correlated with voting Republican because wealthier individuals are less likely to vote for a party that attempts to expand social programs to avoid paying high tax rates for programs they do not benefit from. It is possible that either the median income or housing value could have a positive correlation with voting for a Democrat in the modern election era because the areas with the highest home prices in the United States, like New York City and San Francisco, are also more liberal than average and have high wages to match the cost of living. In the Mountain West, it is unclear which theory of housing prices and income will prevail.

Because white people tend to be more socially conservative and Republicans are more socially conservative than Democrats, the larger the portion of white residents a county has, the more likely the county is to have voted for Donald Trump in 2016 (9). This is especially true in 2016 in response to Trump’s nationalist rhetoric. Because the United States is facing extreme hyper-partisanship, it is likely that the proportion of a county that voted for Republican Mitt Romney in 2012 has a strong, positive relationship with the proportion voting for Trump in 2016. This is likely the case even though Trump was a much less conventional candidate than Romney. On the campaign trail in 2016, Trump focused heavily on the urban-rural social divide in the United States. As a result, it seems likely that the lower the population of a county, the more likely it is to vote for Trump.

The state control variables are all compared to Utah voters. Because Utah typically votes Republican, counties in Colorado, Nevada, and New Mexico are expected vote for Trump less than Utah and counties in Arizona are expected to vote for Trump less, but by a smaller margin than the other states.

Data and Descriptive Statistics

Counties in the Mountain West voted on average 55.70% for Donald Trump, close to the amount that each voted for Romney in 2012 (58.79%), as displayed in Table 2. Taos County, New Mexico had the lowest Trump vote share (17.87%) in the region and Trump won Piute County, Utah by the greatest margin with 85.87% of the vote share. It is important to note that, although the average county voted for Trump, the average county in the sample also had a population of 126,317 people. Even though the mean of the sample is higher than 50%, three of the states (Nevada, Colorado, and New Mexico) voted for Democrat Hillary Clinton because their more populated counties overwhelmed the number of votes contributed by rural areas. The population-weighted mean percent Trump value for the sample was 45.03%.

The average county in the Mountain West had a median age of about 40 years old with a median household income of $50,088 and a median house value of $184,022. The average county was 85.71% white, an unsurprising figure given that the majority of the counties included are incredibly rural and the summary statistics are not weighted by counties’ population size. Of the counties in the sample, 9.50% are in Arizona, 40.51% are in Colorado, 20.89% are in New Mexico, 10.76% are in Nevada, and 18.85% are in Utah.

Table 2 - Descriptive Statistics

Empirical Results

Automation potential and the margin by which counties in the Mountain West voted for Trump over Clinton appear to have a positive relationship. Figure 2 demonstrates the weak positive relationship between the two. Each county in the sample is represented by a bubble and the size of each bubble reflects the county’s population size. The color represents the size of the margin a candidate won a county by; the darker the color, the larger the margin represented on the vertical axis. For example, a bubble on the top right side of the figure represents a county with a high automation potential that Trump won by a large margin.

In contrast to the apparent positive relationship between automation potential and the margin by which a county voted for Trump, two of the models in this study present an unexpected result: a negative relationship between automation potential and the percent each county voted for Trump after controlling for a variety of demographic and economic considerations. Table 3 illustrates the empirical results of the three election models used to predict the portion of a county in the Mountain West that voted for Donald Trump. A discussion of each model and a discussion follow.

Figure 2 - Automation Potential vs Margin Trump over Clinton
Figure 2: Automation Potential vs Margin Trump over Clinton (percentage points).

Model 1: Reduced

The reduced model regresses automation potential, county characteristics (percent of people with a bachelor’s degree, median age, the natural log of the median income, the natural log of the median house value, the percent of the population that is white, and the natural log of the population size) as well as state variables (binary variables included for Arizona, Colorado, New Mexico, and Nevada) with the percent that each county voted for Trump. This model does not include the portion of each county that voted for Romney in 2012.

As the percentage of people with a bachelor’s degree in a given county in the Mountain West increases by one percentage point, the percent voting for Trump is expected to decrease by 0.91 points. Conversely, a one percent increase in the number of white residents in a county results in a 0.459 point increase in Trump’s vote share. The portion of the county that voted for Trump increases by 0.29 percentage points as median income increases by one percent and as median house value increases by one percentage point, the portion of a county that voted for Trump decreases by 0.12 percentage points. If a county is in New Mexico, it is expected to negatively impact Trump’s vote share by 8.124 points compared to counties in Utah.

Automation potential, median age, population size, and every state binary variable except New Mexico are all insignificant at the 5% level. The model is statistically significant and accounts for 64% of the electoral variance among counties in the Mountain West.

Model 2: Full

The full model regresses automation potential, county characteristics (percent of people with a bachelor’s degree, median age, the natural log of the median income, the natural log of the median house value, the percent of the population that is white, the percent of the 2012 votes cast for Republican Mitt Romney, and the natural log of the population size) as well as state variables (binary variables included for Arizona, Colorado, New Mexico, and Nevada) with the percent that each county voted for Trump.

As a county’s population with a bachelor’s degree increases by one point, Trump’s vote share decreases by 0.44 percentage points. Similarly, as the median age of a county increases by one year, the portion a county votes for Trump increases by 0.305 percentage points. Every additional percentage point a county voted for Romney in 2012 predicts a 0.79 percentage point increase in Trump’s vote share. The binary variable coefficients suggest that if a county is in Colorado, New Mexico, or Nevada, Trump’s vote share is expected to be 14.06, 8.69, and 11.76 percentage points higher than counties in Utah, respectively.

The model is statistically significant and accounts for 89% of the variance in the portion each of the Mountain West counties voted for Trump, but automation potential, median income, median house value, percent white, population size, and the Arizona binary variable are all statistically insignificant at conventional levels. It is especially interesting that these variables are also jointly insignificant. Given that income, house value, and percent white are significant in the reduced model, it is possible that including Romney’s vote share explained their individual effects because they are collinear, but the variables’ variance inflation factors suggest otherwise.

Tab. 1 - Empirical Results

Model 3: Linear Probability Model (LPM)

The linear probability model regresses automation potential, county characteristics (percent of people with a bachelor’s degree, median age, the natural log of the median income, the natural log of the median house value, the percent of the population that is white, the percent of the 2012 votes cast for Republican Mitt Romney, and the natural log of the population size) as well as state variables (binary variables included for Arizona, Colorado, New Mexico, and Nevada) with a binary dependent variable equal to one if the county voted more for Trump than Clinton and zero otherwise.

As the portion of a county with a bachelor’s degree or higher increases by one percentage point, the probability that the county voted more for Trump than Clinton decreases by 0.87 percentage points at a significance level of 5%. A one-year increase in a county’s median age predicts a 1.47 percentage point increase in the odds that the county voted for Trump more than Clinton at all conventional levels. As in the other models, the percent that a county voted for Romney had a sizable effect on the portion of the county that voted for Trump. For every percentage point increase in the county’s vote for Romney, the probability that the county voted for Trump increases by 2.02 percentage points. If a county is in Colorado, it is 17.32 percentage points more likely to have voted for Trump at the 5% significance level.

The automation potential, median income, median house value, the portion of a county that is white, the population size, Arizona binary, New Mexico binary, and Nevada binary variables are all statistically insignificant at all conventional levels. Although insignificant, the coefficient of the automation potential variable is positive, consistent with the hypothesis that counties with higher average automation potential have a higher Trump vote share.

Discussion

The work of Autor et al. (6 ) and Frey et al. (8 ) combined with Figure 2 suggest that automation potential may positively impact the portion of a county voting for Trump. While statistically insignificant at conventional levels, the reduced and full models produced the opposite result. The coefficient for automation potential appeared negative in both, leaving room for additional study and interpretation. Given that the Rust Belt states are demographically and electorally quite different from the states included in this study, it is possible that automation potential does not, by itself, have much of an impact on voting behavior in every industry.

Many counties in the Mountain West have a high average automation potential but dissimilar employment and demographic concentrations to the Rust Belt. Clark County, Nevada, for example, had 1.86 times more Food Preparation & Serving Related Occupations than the country in 2016 (18 ). The Accommodation and Food Services industry has an automation potential of 73% (12 ). An average Accommodation and Food Service worker in Las Vegas will likely have a different perspective on their job’s automatability than a manufacturing worker in Michigan. The Las Vegan is much less likely to be white, and also likely less likely to predict that their service-oriented position will be automated in the near future than a manufacturer that has already seen robots’ effect in their workplace. As a result, voters in highly automatable positions in Las Vegas may prioritize other electoral considerations, like immigration or health care, over a seemingly distant automation risk. Rust Belt workers in highly automatable positions may be more homogenous in demographic and voting behavior and thus prioritize economic dislocation over other factors than their more diverse counterparts in the Mountain West.

A negative automation potential coefficient may also reflect voters’ prediction of which jobs could be automated versus which jobs voters believe will be. Highly automatable Rust Belt positions, those most emphasized in the literature, are concentrated in non-service-oriented industries like manufacturing. Highly automatable positions in Mountain West counties, by contrast, are far more client-facing than their Rust Belt counterparts. Perhaps Rust Belt workers perceive the changing economy as a threat to their job security and Mountain West workers are confident that their face-to face, service-oriented jobs are secure.

Conclusion

Automation’s effect on elections will likely continue to grow as technology advances. Voters’ response to increasing automation will depend on regional occupation concentration as well as the perception of workers. This project adds to the literature by examining automation’s implication on the 2016 election in a seldom-discussed region. While the other studies find that Rust Belt swung in favor of Donald Trump in 2016 partially because of automation (6, 8), this study does not find that automation potential had a statistically significant effect on the outcome of the presidential election of 2016 in the Mountain West. However, more research should be done given that this study found unexpected (albeit insignificant) automation potential coefficients.

There are several limitations to this study. Using county-level results in only a handful of states produced a relatively small sample size, and also made it difficult to determine how individual voters in varying industries respond to automation. It may be beneficial to expand the geographic scope to include other regions in the United States or use smaller geographic areas to increase the sample size. This study also failed to account for the effect of offshoring or globalization on workforce anxiety, as previous studies have (8 ). While county-level observations may inform the role of some demographic characteristics on elections, they are not able to separate the demographic characteristics of voters and non-voters. Less-educated people, those more likely to hold highly automatable jobs, are also less likely to vote, but this study did not account for the potential disconnect between a shift in non-voters’ political preferences caused by automation and the outcome of an election.

Similarly, the automation potential variable is limited. Other papers using measures to empirically describe robot integration may do a more accurate job of describing the trend at the time (8 ). Using the automation potential variable is likely better at capturing some workers’ anxiety about the future of their jobs rather than just the present situation, but there are also drawbacks to its predictive nature. A litany of assumptions went in to creating this variable, including current technological capacity, task concentration, and a prediction of which technology will be available. It cannot confidently describe which technology will be adopted nor which sectors which will adopt new technology, only which tasks have the potential to be automated. Even if the measurement was able to accurately predict automation in the future, it does not describe which jobs were already in the process of being automated prior to the 2016 election. While automation potential is assumed to capture some of voters’ perception about the future of their jobs, it also weakly assumes voters rationally respond to an intangible estimation of their job tasks’ potential to be automated in the coming decade. How workers will perceive risk will depend on their industry and the type of information they receive about automation.

Although statistically insignificant, two models in this study found an unexpected relationship between Mountain West counties average automation potential and Trump vote share. Rather than boosting Trump’s support in a county, automation potential seems to have had a slight negative effect on his vote share in the Mountain West. This outcome suggests that regions respond differently to automation anxiety and may consequently shift their political preferences. Because 25% of jobs are categorized as being at “high risk” of automation by 2030 (12 ), federal and state governments should consider how regional responses to automation differ to adjust their political and campaign strategies accordingly.

Acknowledgements

I am incredibly grateful for my family’s encouragement, Dr. Djeto Assan´e’s guidance, and Jeffrey Wheble’s invaluable feedback.

References

[1] J. Sides, M. Tesler, L. Vavreck, Journal of Democracy 28(2), 34–44 (2017).
[2] V. Williamson and I. Gelfand, The Brookings Institute, (August 2019). [Retrieved from]: http://www.brookings.edu/blog/fixgov/2019/08/14/trumpand-racism-what-do-the-data-say/.
[3] B. F. Schaffner, M. Macwilliams, T. Nteta, Political Science Quarterly 133(1), 9–34 (2018).
[4] T. B. Edsall, New York Times, (January 2018). [Retrieved from]: http://www.nytimes.com/2018/01/11/opinion/trump- robots- electoral- college.html.
[5] G. Lopez, Vox, (December 2017). [Retrieved from]: http://www.vox.com/identities / 2017/12/15/16781222/trump-racism-economic-anxietystudy.
[6] D. Autor, D. Dorn, G. Hanson, K. Majlesi, (December 2017). [Retrieved from]: http://www.ddorn.net/papers/ADHM-PoliticalPolarization.pdf.
[7] M. McQuarrie, The British Journal of Sociology 68, S120–S152 (November 2017).
[8] CC. B. Frey, T. Berger, C. Chen, Oxford Review of Economic Policy 34(3), 418–442 (July 2018).
[9] Pew Research Center, (August 2018)[Retrieved from]: http://www.people-press.org/2018/08/09/an- examination- of- the- 2016- electoratebased-on-validated-voters.
[10] N. Silver, FiveThirtyEight, (May 2016). [Retrieved from]: https://fivethirtyeight.com/features/the-mythology-of-trumps-working-class-support/.
[11] J. Kolko, FiveThirtyEight, (November 2016). [Retrieved from]: http://fivethirtyeight.com/features/trump – was – stronger – where – the – economy – is – weaker.
[12] M. Muro, J. Whiton, R. Maxim, The Brookings Institute. (March 2019). [Retrieved from]: http://www.brookings.edu/blog/the-avenue/2019/03/25/automation-perpetuates-the-red-bluedivide.
[13] R. Molla, Vox, (December 2019). [Retrieved from]: http ://www.vox.com/2019/1/28/18185061/robot-automation-jobs-employment-revolutionworse-men-brookings.
[14] D. Acemoglu, P. Restrepo, Journal of Political Economy 6, 2188–2244 (2020).
[15] Tech. rep., S. Alder, D. Lagakos, L. Ohanian, National Bureau of Economic Research, (September 2014). [Retrieved from]: www.ridge.uy/wp-content/uploads/2018/12/Alder-Simeon-Labor-MarketConflict-and-the-Decline-of-the-Rust-Belt.pdf.
[16] MIT Election Data and Science Lab, Harvard Dataverse 6 (2018).
[17] J. Manyika, M. Chui, et al., McKinsey Global Institute,(January 2017). [Retrieved from]: http://www.mckinsey.com/global-themes/digital-disruption/harnessing-automation-for-a-futurethat-works.
[18] U.S. Census Bureau, (2016). Occupation by Sex for the Civilian Employed Population 16 Years and Over. 2016 American Community Survey 5-Year Estimates, Table S2401. Generated from http://data.census.gov.
[19] Pew Research Center, (March 2018). [Retrieved from]: http://www.people- press.org/2018/03/20/1-trends-in-party-affiliation-among-demographicgroups.