Globalization’s Concentrated Costs (and Benefits): Import Competition,
Globally Engaged Firms, and U.S. Presidential Voting*
J. Bradford Jensen
Dennis P. Quinn
Stephen Weymouth
25 September 2015
Preliminary and Incomplete.
Abstract
We study how integration in the global economy influences U.S. presidential elections. We
propose that increased wages and employment from export growth increases support for the
incumbent president’s party, while lower wages, higher unemployment and increased
employment volatility from import competition diminishes incumbent support. We therefore
expect that voting will reflect the spatial distribution of economic activity among firms that win
and lose from increased trade. We examine this proposition using Census Bureau data on the
universe of U.S. firms to construct measures of employment in high- and low-wage tradable
activities, which we link to U.S. presidential vote share at the county-level. Our results indicate
increases (decreases) in incumbent vote shares in counties characterized by the concentration of
high-skilled (low-skilled) tradable goods and services. We extend this work to the national-level
models, and show for the first time that increasing imports negatively affect incumbent vote
shares and increasing exports are associated with increasing vote shares for incumbents over the
post-war (1952-2012) and post-RTAA period (1936-2012). The estimated effects are large and
politically consequential. We conclude with a preliminary forecast for the 2016 election based on
current domestic and international conditions.
*
We thank James Vandeventer and Pierre Boudet for excellent research assistance. Kent Carlson, Brian
Saba, and Jennifer Yan participated in a GUROP working group on economic voting in spring 2012, and
we thank them as well. Alexis Antoniades and Quynh Nguyen offered helpful comments on drafts. We
thank participants in the Mortara Georgetown Political Economy seminar series, an APSA 2015 panel, an
ETH/UniZurich CIS seminar, and a World Trade Institute (Bern) seminar for helpful comments on an
earlier draft. We especially thank Stefanie Walter. We thank Annalisa Quinn for editorial assistance. All
remaining errors of fact and analysis are our own.
1
Economic voting models of presidential elections are well developed in American
political science and economics scholarship, and offer strong support for the proposition that
national economic performance affects the fate of incumbent Presidents or their partisan wouldbe successors. National-level (“macro”) voting studies of presidential elections have shown
large effects of economic variables on incumbent party vote shares: adverse economic outcomes
especially diminish the electoral prospects of incumbent Presidents and parties.1
We propose that exposure to globalization has an independent effect – separate from
trade’s longer term effects on economic growth or other macro-economic variables – on voting
in U.S. presidential elections. Of particular interest to this paper is the U.S. economy’s growing
integration into the world economy in flows of goods, services and finance, with the resulting
direct and indirect effects on growth, prices, employment, wages and job security. In an
important contribution, Margalit (2011) shows in county-level (“meso-“) analyses that the
aggregate effect of offshored jobs, measured through applications for Trade Adjustment
Assistance (TAA), on the incumbent vote share is negative.2 With the exception of Margalit,
however, the effect of economic integration on U.S. Presidential voting has yet to be studied.
1
Initial works in the area include Fair 1978 and Tufte 1978. Recent contributions include the papers in
the 2008 Special Issue of the Journal of International Forecasting (Campbell and Lewis-Beck (eds.)
2009), Fair 2009, and Erikson 2009a. A comprehensive review of the literature is Lewis-Beck and Tien
2008. See also the special issue of PS, April 2014, “US Presidential Election Forecasting.”
2
Antoniades and Calomiris 2014, in an interesting study of county-level Presidential voting, find that
constrained credit conditions hurt incumbent vote shares.
2
Changes in trade exposure will affect voter preferences because trade exposure shapes the
terms of employment and wages, which are imprecisely captured in aggregate economic
measures of growth, disposable income, and even aggregate unemployment or trade-related job
losses. Multilateral institutions such as the WTO, preferential reductions in tariffs, reduced
transportation and information technology costs, and the resulting changes in firm activities have
fundamentally altered the American economic landscape. To take one germane example, many
service products, once thought be ‘non-tradable,’ have become among America’s leading
exports. In comparatively disadvantaged sectors, like low-skilled manufacturing, trade has
increased unemployment and reduced wages. One prominent study finds that import competition
resulting from China’s integration into the world trade system explains a quarter of the decline in
U.S. manufacturing employment since 1990 (Autor, Dorn, and Hanson 2013).
These changes have shifted the composition, terms, security, and levels of employment
and income. Individuals in firms facing trade shocks will experience the effects of these shocks
more immediately than aggregate economic indicators will reveal. That is, voters/employees
will experience the ‘inside lag’ from changes in import competition or exports successes before
the ‘outside lag’ of trade are reflected in changes in growth or disposable income. Levels of
employment, the quality and composition of employment, and wage levels will be variously
affected by exports and imports. Employees in import-competing, export-oriented and
multinational firms therefore confront diverse economic realities that national indicators of
unemployment or growth cannot reflect.
While globalization benefits the employees of the largest and most productive firms that
engage in international trade and investment, employees of firms that compete with imports may
be harmed by economic integration that increases competition from abroad. We expect voting in
3
presidential elections to reflect these distributional consequences of globalization. Voters
employed in firms in declining sectors will respond to employment volatility, slower wage
growth, and import competition by voting against incumbents. Conversely, employees of
‘winning’ firms (i.e., globalized firms, generally in high wage, comparative advantage sectors)
are more likely to vote for incumbents. Thus, we expect spatial concentrations of globally
engaged firms to be associated with increasing incumbent vote shares, and concentrations of
comparatively disadvantaged firms associated with declining incumbent vote shares.
We examine these arguments at the macro (national) level, adapting the core macro
models and methods prevailing in the literature.3 The macro studies are, however, unable –
owing to limited numbers of observations – to discriminate precisely between and among the
effects of competing economic forces, particularly those that have emerged from globalization.
We therefore also examine our argument using data at the county-level, where nearly 19,000
observations over 3,111 counties and six elections allow for greater precision in estimates. The
data upon which these results rest, however, only enable us to examine U.S. elections from 1992
(or 1996) onward, and do not contain county-level data on import competition. In light of this
time vs. cross-sectional tradeoff, we consider the joint results to be more informative than the
separate results independently.
3
Relative to the literature, the paper does not seek to identify a single ‘right’ model of economic voting.
Rather, the assumption of the paper is that each of the main scholarly models of economic voting has
merit, but that much can be gained from examining the role of international trade and considering
subnational variation in exposure to globalization.
4
The common practice in the macro-level economic voting literature has been to treat
economic performance as a “valence” issue that affects both Democrats and Republicans alike.
In meso-level studies, in contrast, some scholars have allowed for economic performance to be a
partisan issue. Wright 2012 explores whether these economic effects are generalized to
incumbents of both or, whether the Democratic or Republican Parties “own” certain issues.
Wright finds that unemployment is an issued “owned” by the Democrats (2012, 699).4 We will
examine the “valence” vs. “partisan” issue at both the macro and meso levels.
We explore first the empirical implications of the argument that subnational voting will
reflect the distribution of economic activity in globally-engaged firms using county-level data. In
particular, we estimate the effects of employment volatility, industry location quotients, and
other measures of internationally-engaged employment to identify how globalization affects
voting behavior.
In addition to making use of publicly available information on the industrial composition
of counties, we also use confidential, establishment-level data from the Census Bureau’s
Longitudinal Business Database (LBD). The LBD contains information on the universe of
establishments in the scope for the Census Bureau’s County Business Patterns (CBP) program.
4
Wright (2012) shows convincing evidence at the county level that Republican incumbent vote shares are
harmed by rising unemployment, but rising unemployment does not harm Democratic incumbent vote
shares. We replicate and confirm that finding below.
5
The CBP program covers most of the country's economic activity.5 The data allow us to measure,
with great precision at the county level, the number of employees engaged in comparative
advantage, tradable activities as well as those employed in positions vulnerable to import
competition. We estimate how variation in exposure to globalization at the county level affects
voting in U.S. Presidential elections.
One important virtue of the county-level data is that we are able to disaggregate results
by state groupings. We compare the estimated results in swing states to those of non-swing
states using several categorizations of swing-states.
Next, building on extant macro models of economic voting, we add changes in the U.S.
international economic position as a variable of interest in national-level voting models. Our
analysis indicates that trade is unique from measures of growth and employment. We find that
rising import (export) shares are associated with decreasing (increasing) incumbent vote shares
in Presidential elections.
The paper proceeds with a literature review followed by a theoretical section describing
the rationale for treating trade separately from aggregate economic measures. Standard voting
models using time-series methods, as well as exploratory factor analyses, are employed to test
our hypotheses. The core findings of the paper are: at the county-level: 1) high unemployment
volatility and a high concentration of economic activity in comparative disadvantage sectors
(low-skilled manufacturing) hurt incumbent vote shares; but 2) high concentration of economic
5
The only major exclusions are self–employed individuals, employees of private households, railroad
employees, agricultural production employees, and most government employees. Thus, the data represent
the practical totality of the private non–agricultural sector of the economy.
6
activity in comparative advantage sectors (such as tradable services, e.g., NAICS 56 - business
services) increase incumbent vote shares; and 3), at the macro level, rising imports and declining
trade balances (rising exports) as a percentage of GDP hurt (help) incumbent vote shares. We
find no evidence of partisan effects in the macro-level studies, but do find support for Wright’s
(2012) hypothesis that rising unemployment, especially, hurts incumbent Republicans. We also
confirm Margalit’s (2011) finding that rising Trade Adjustment Assistance claims are associated
with decreasing incumbent vote shares.
Literature Review
Prior studies provide overwhelming evidence that economic outcomes affect Presidential
voting. Starting with the “Fair Model” and Tufte’s “political control of the economy” model,
scholars have recognized that positive economic performance strongly improves the likelihood
of either incumbent (or the incumbent party) reelection prospects.6 (See reviews of the “macro”
models in Campbell 2008, Kayser and Leininger 2015, and Lewis-Beck and Tien 2008.)
The macro models are necessarily parsimonious because of the few degrees of freedom
involved in estimating models, which generally date from either 1948 or 1952.7 Some models
offer an even shorter sample because of the theorist’s beliefs that the structure of the economy
has changed over time. (See, e.g., Erikson 2009.). Invariably, macro models see some aspect of
economic performance as key: economic growth (Fair 2009), disposable income, employment,
job growth (Lewis-Beck and Tien 2008), and business sentiment (Erikson 2009) are contending
6
Fair 1978; Tufte 1980.
7
Fair 2009 is an exception, estimating models using data starting in 1916.
7
variables. Quinn and Woolley (2001) show that economic volatility (measured as the standard
deviation of economic growth) drives down votes for the incumbents in a comparative, crossnational setting, a finding that Hibbs 2000 suggested did not apply to the U.S. setting.
In addition, some non-economic considerations often enter the models. Erikson 1989
used “net candidate advantage.” The July/June Gallup Presidential Approval indicators have
been previously entered. (See Abramowitz 2008, and Lewis-Beck and Tien 2004). Abramowitz
2008’s “Time for Change” model also incorporates how long a party has governed in terms of
the “costs of governing.”
This impressive progress in the literature notwithstanding, points of scholarly difference
remain in terms of modeling strategies. One consideration is whether voters judge economic
performance retrospectively (Abramowitz 2008; Fiorina 1981; Hibbs 2008), prospectively
(Mackuen, Erikson, and Stimson 1992; Erikson 2009b; Erikson and Wleizen 2009; Norpoth
1996), or with some combination of forward- and backward-looking lenses.8 Another point of
difference in investigative strategies is how to (or whether to) include information about prior
electoral outcomes. Fair (1978, 2009) and Powell and Whitten (1993) included prior incumbent
vote shares as an independent variable whereas Abramowitz (1988; 2008) included the number
of terms of incumbency for the current President’s party in his “Time for Change” Model. Other
8
Another difference among investigators is what type of information about the economy is included.
Economic growth, employment, changes in personal income, inflation, consumer expectations, consumer
evaluation of business conditions, and inflation are candidate variables. This issue will be discussed in
greater detail below. To foreshadow: We will use various indicators of economic performance to
establish the robustness of the results.
8
scholars offer parsimonious models without including prior electoral outcomes (e.g., Erikson
1989; Hibbs 2008). Another point of difference is whether additional non-economic variables
enter the models, such as incumbent President approval.
Largely missing from the economic voting literature to date are the estimated effects of
the global trade integration of the U.S. economy on voting. Our contribution emphasizes the
effects of changing exposure to international economic competition. Our motivation is that
globalization, particularly changes in exposure to import competition, reflects changes in voter
economic experiences that are not immediately captured by changes in growth or personal
income. Economic growth will likely lag because of policy stabilizers (government policies that
intentionally slow trade adjustment effects), and slow reactions by import competing firms.
Furthermore, wages and jobs are sticky due to a number of factors (e.g., unemployment
insurance, buyouts, and the lag between import surges and how firms respond), which imply a
lag between the immediate effects on firms and the longer term effects on growth.
Furthermore, as we argue below, the effect of increased trade will likely be very
heterogeneous across firms and workers, with concentrations of globally engaged firms
associated with greater voter satisfaction, and concentrations of comparatively disadvantaged
firms associated with voting against the incumbent. That is, the acceleration of imports leads to
changes in the composition of work and wages depending on the firm’s position in the global
economy. We will propose and demonstrate empirically that the effects of trade are not simply
subsumed by growth or aggregate unemployment.
While trade integration and exposure are largely absent from the economic voting
studies, we know that globalization has had strong effects on political outcomes in various other
9
(non-voting) settings. Milner (1988) demonstrated that whether (and how) firms were
internationally engaged accounted for their political engagement on trade issues. Scheve and
Slaughter (2004) find that inward FDI, and globalization more generally, increases wage and
employment volatility, leading workers to express feelings of economic insecurity. Similarly,
Walter (2010) finds in a study of Swiss respondents that individuals hurt by trade are more likely
to report economic insecurity. Rickard (2015) demonstrates the linkage between export success
in a Congressional House member’s district and his or her likelihood to support Trade
Adjustment Assistance. Jensen, Quinn, and Weymouth (2015) address the puzzle of declining
trade dispute filings in an era of foreign currency undervaluation and increased import
competition. They argue that firms’ investments in undervalued countries explain their lack of
support for antidumping filings as a non-market strategy. Owen and Quinn (2016) find that
changes in international integration influence the U.S. policy mood (Stimson 1991), with
increasing import (export) shares shifting public opinion in a ‘leftward’ (‘rightward’) direction.
Owen (2015) finds evidence that the ‘off-shorability’ of jobs in a Congress member’s district
decreased the likelihood of voting in favor of a trade agreement. If trade has the distributional
consequences posited above and developed in the ensuring section, voters harmed (benefitted) by
trade will likely shift away (toward) from the incumbent or the incumbent’s party.
In summary, investigators have found past economic performance and stability and voter
expectations about future economic activity explain incumbent party Presidential vote shares.
However, with the important exception of Margalit (2011), there is little work on the direct
effects of trade on U.S. Presidential voting. Our aim is to build on the established models and to
examine the direct effects of international trade indicators in the context of these extant models.
10
Theory
Because the U.S. is a relatively high-skill-abundant country, the U.S. will have a
comparative advantage in high-skill activities, and a comparative disadvantage in low-skill
activities. Thus, firms in labor-intensive tradable goods industries tend to face greater import
competition. As a result, unskilled and low-skilled workers are threatened by rising imports from
low wage countries, which have dramatically expanded exports in low-wage, labor-intensive
industries (such as apparel). U.S. firms in import-competing industries that do not shift
production to low-wage countries are likely to fail or to change industries as a result of this
competition (Jensen, Quinn, and Weymouth 2015). The result of plant closures and/or the
redeployment of production abroad is lower employment and slower wage growth in the U.S.
among less competitive firms in comparative disadvantage industries (Bernard, Jensen, and
Schott 2006). In contrast, firms using high-skilled labor in tradable goods and services industries
are more likely to expand employment and pay higher wages (Ibid.).
Scholarly evidence suggests that workers in import-competing firms may be particularly
harmed by globalization. In a survey of displaced workers, Kletzer (2001) finds that
manufacturing workers displaced by import competition are less likely to be re-employed than
manufacturing workers displaced for other reasons. Furthermore, exporting industries are not
likely to absorb these workers, due to different factor demands. Among workers fired from
import-competing sectors, those re-employed have weekly earnings that are 13 per cent lower on
11
average, and approximately 25 percent report wage losses of 30 percent or more.9 See also
Autor, Dorn, and Hanson (2013), and Autor et al. (2014).
While comparative advantage based on relative factor endowments surely explains some
of the distributional effects of trade in the U.S., recent research demonstrates substantial withinindustry variation in firm participation in the global economy. Dubbed “new, new trade” or
“heterogeneous firms” theory, the body of work explains why some firms in an industry trade
internationally and undertake FDI, while others do not (Melitz 2003, Helpman, Melitz, and
Yeaple 2004). The literature explains how firms with different characteristics respond differently
to changes in the economic environment (for example, some firms expand their exports in
response to trade liberalization, while others shut down). Firm-level characteristics—and not
industry or factor advantages per se—are the fundamental determinant of trade and international
investment activity.
New, new trade theory leads us to expect that the gains from trade will not align strictly
according to factors of production or industry. For instance, while manufacturing activity has
shrunk in the U.S. over the past two decades, some manufacturing firms have been able to
expand and compete, often through the offshoring of production to low-wage countries. Indeed,
the globalization of production through vertical foreign direct investment and intrafirm trade in
intermediate goods is common across U.S. industries. Ramondo, Rappoport and Ruhl (2013)
provide evidence that in all industries there are subsidiaries set up for vertical production and
9
Kletzer 2001, 4.
12
trade purposes, and the intrafirm trade intensities among MNC affiliates are quite similar across
industries. This research informs the expectation that at least some firms (and their employees)
will have gained from globalization across all sectors, including those in comparatively
disadvantaged sectors.
Thus, one of the novel political economy implications of new, new trade theory is that the
distributional consequences of trade depend on the relative competitiveness of firms within their
industry. The microeconomic conditions facing employees will vary depending on their firms’
value chains, product markets and global strategies (Jensen, Quinn, and Weymouth, 2015), along
with their geographic location within the United States (Autor et al. 2014). Since firms that
export generally have superior economic performance on many dimensions (higher employment,
sales, wages, productivity and investment) than firms that do not export (Bernard and Jensen
1999), rising exports tend to benefit employees of exporting firms. Thus, the gains from trade
will be focused among the employees of exporting firms and multinationals, which on average
use more skilled labor. The burdens of trade and trade adjustment tend to be focused on importcompeting firms in the tradable sector. These expected heterogeneous effects of trade integration
among firms within industries are a complement to standard approaches, which interpret the
distributional consequences of trade as manifested across factors (Heckscher-Ohlin) or sectors
(Ricardo-Viner), or both. Our empirical approach attempts to more precisely identify these
effects by using microdata data to capture changes in employment in firms paying high wages
and producing tradable goods.
Our argument suggests that the effect of the increasing importance of international trade
in the U.S. economy on voting cannot be fully understood theoretically or econometrically by
analyzing unemployment or growth rates alone. First, the distributional or welfare consequences
13
of trade extend beyond employment levels. Trade changes the composition of firms in the
economy; it is not just the quantity of work in the import-competing sectors that declines, but
also its quality and composition in terms of wages, benefits, job security and skills. Much of this
variation can only be captured at the subnational level. Secondly, trade is likely to affect voting
before the effects of rising imports or exports are reflected in the unemployment rate. Even
before import-competing firms cut jobs, workers can anticipate the effects of reduced production
orders (for example, shorter hours, reduced wages, eventual lay-offs). Employees of firms in the
import-competing sector might retain their jobs or find new ones, but the terms of employment
are likely to worsen in quality. Employees of firms trading in industries in which the U.S. has a
comparative advantage (e.g., business services) are more likely satisfied with U.S. integration,
and less likely to voice dissatisfaction by voting against the incumbent. The distributional
consequences of trade will be felt across firms within industries.
What is less evident from theory is whether partisanship colors voter attributions to
incumbents on some of these international economic factors. While Presidential voting has, in
the main, focused on attribution to incumbents regardless of party, Wright (2012) shows that, for
unemployment, Republican incumbents only are harmed.10
We will explore this supposition
below.
In summary, we expect that these varied consequences of globalization across firms will
affect citizen voting, with the principle beneficiaries of globalization (those employed in
10
This result is consistent also with findings in Margalit (2011, 175, Table 2.4), though Margalit
examines only the 2004 election, in which a Republican incumbent’s vote share was negatively
associated with rising unemployment.
14
productive, exporting firms, most often in sectors of U.S. comparative advantage) more likely to
vote for the incumbent. In contrast, (lower skilled) employees of import-competing firms, often
in the low-skilled manufacturing sector are more likely to express their dissatisfaction by voting
against the incumbent.
Empirical Implications
We examine the empirical implications of our argument at the county- and macro/country-level.
At the county-level, we expect:
1. Unemployment volatility will be associated with decreased support for the incumbent.
2. Concentration of economic activity in winning industries (high wage/tradable) will be
associated with increasing support for the incumbent.
3. Concentration of economic activity in losing industries (low wage/non-tradable) will be
associated with decreasing support for the incumbent.
4. Employment concentrations in high (low) exporting firms, regardless of industry, will be
associated with increasing (decreasing) support for the incumbent.
At the country-level, we expect:
5. Imports (exports) will be associated with decreased (increased) support for the
incumbent.
County-level Analysis
We examine the determinants of incumbent party vote share at the county-level. The
baseline OLS, year- and county-fixed effects model is:
15
Incumbent 2-Party Vote Sharei,t = ß0 + ß1 (Incumbent 2-Party Vote Sharei,t-1) + ß2
(Unemployment Ratei,t) + ß3 (Δ Unemployment (1-yeari,t)) + ß4 (Unemployment
Volatilityi,t-1) + ß5(LnAveragePayi,t ) + ß6 Δ Average Pay (1-year)i,t +
ß7(LnManufacturing Concentration i,t) + ß8(LnTradable Services Concentrationi,t) +
ß9(LnNon-Tradable Services Concentrationi,t) + ß10(Democratic Incumbent) +
ß11(Democratic Incumbent x Δ Unemployment (1-year)i,t) + φi + τt + εt,i
1996, 2000, 2004, 2008, 2012
t=1992,
(1)
The dependent variable, Incumbent 2-Party Vote Shareit, is the incumbent party vote as a
share of the total Democratic and Republican votes in county i year t. The models begin in 1992
because of data limitations on economic concentration, which the Bureau of Labor Statistics
reports at the county-level beginning in 1992. We also include county φi and election year τt
dummies.11 In addition to its plausibility as a determinant of incumbent vote shares (Fair 2009;
Powell and Whitten 1993), prior incumbent vote share (Incumbent 2-Party Vote Shareit-1) is
entered as a lagged endogenous variable to attenuate possible omitted variable bias. Following
Margalit (2011), some of our models control for aggregate job losses due to globalization: the
lagged sum of the estimated number of workers filing for trade adjustment assistance (TAA) as a
share of the labor force.12 Following Wright (2012), some of our models take Democratic party
vote share as the dependent variable, and others include a Democratic Incumbent interaction with
11
A Hausman test of random vs. fixed-effects rejects the random effects model with a χ2 test producing a typical
value of over 500
12
The TAA data come from Public Citizen. http://www.citizen.org/Page.aspx?pid=4536 (accessed March
2, 2015).
16
changes in unemployment. In models estimated at the Census Bureau (disclosure pending), we
are further able to distinguish between employment in high wage, highly traded (and exported)
manufacturing and employment in low wage, low traded (and not exported) manufacturing.
We examine the preference for stability hypothesis using county-level data on
unemployment and wages. The variable Unemployment Volatility is the standard deviation of the
unemployment rate in county i over the four-year period inclusive of the election year.13 The
unemployment data are from the U.S. Bureau of Labor Statistics (BLS). The income data are
from the Quarterly Census of Employment and Wages, conducted by the U.S. BLS. (The two
indicators are highly positively correlated such that higher unemployment rates are substantially
collinear with increased unemployment volatility.) We also enter change in unemployment from
the year prior to the election and change in income.
We incorporate a number of different measures of voters’ exposure to globalization. Our
objective is to investigate whether voting behavior is influenced by the international exposure of
local industries. First, we capture the concentration of workers in different sectors using location
quotients calculated by the U.S BLS. The BLS reports sectoral employment location quotients. A
location quotient is a ratio used to compare the concentration of employment in different sectors
among different counties in the United States. For instance, the manufacturing employment
location quotient is the number of workers employed in manufacturing in county i as a share of
total workers in county i, as a ratio of the share of manufacturing workers to total workers in the
United States.
13
For example, in 1996, Employment Volatility is the standard deviation of the unemployment rate in
county i among the years 1993, 1994, 1995, and 1996.
17
The manufacturing employment location quotient provides a proxy for the relative
concentration of workers that might be harmed by globalization. A higher manufacturing
location quotient in a county may make it more susceptible to import competition and
outsourcing. In contrast, the employment location quotient of an industry in which the United
States has a comparative advantage, such as business or other tradable services, will capture the
concentration of workers that may gain from globalization in a particularly county. We
incorporate the (logged) employment location quotients of Manufacturing, Tradable Services,
and Non-tradable Services into our models to test our hypotheses about the differential economic
voting responses of the losers and winners of globalization, respectively.14
We also develop a number of measures of international exposure using confidential data
from the U.S. Census Bureau.15 We take the median household income in the relevant year as the
threshold for “high-wage.” We identify “high-wage” activities at two different levels of
aggregation. First, we construct estimates of employment in a county in high-wage industries
using the average industry wage as the determining measure and classify all workers in that
industry as being high- or low-wage. The second method is to identify workers in high-wage
activities based on the average wage paid within individual establishments. Here, workers are
14
Tradable services are Trade, Transportation and Utilities, Information, Finance, Business and
Professional Services and Leisure and Hospitality. Non-tradable services are Education and Healthcare
and Other Services. The BLS does not distinguish among Manufacturing industries in reporting the
Manufacturing location quotients. See http://www.bls.gov/cew/datatoc.htm <accessed June 21, 2015>.
The analysis of firm-level data at the Census Bureau allows us to capture high-wage, traded
manufacturing.
15
The analysis based on data from the U.S. Census Bureau is ongoing, and not reported here.
18
classified as high-wage if the establishment in which they work has average wages above the
median household income.
Our goal is to examine the international exposure of the entire local economy – not
merely to assume that the manufacturing sector is trade-exposed and that all other sectors are not.
We make use of trade cost estimates to identify tradable industries.16 We assign an industry level
indicator variable for whether an industry has sufficiently low trade costs to suggest that it is
tradable.
We construct measures of the share of workers in each county that are in tradable
industries with above-median income in the manufacturing sector, the share of workers in
tradable industries with below-median income in the manufacturing sector, the share of workers
in tradable industries with above-median income wages in the service sector, and the share of
workers in tradable industries with below-median income wages in the service sector. Our
expectation is that low-wage workers in tradable industries (whether in the manufacturing sector
or the service sector) are more exposed to international competition and, therefore, more likely to
be dissatisfied with globalization.
We also construct an ‘exporter’ indicator at the plant level. Using confidential Census
records, we observe which plants in which industries export, and aggregate the county-level
employment for exporting establishments.
The analysis includes all 3,111 U.S. counties for which complete voting data are
available for our period of study (1992-2012). Consistent with Margalit (2011) and Wright
16
Gervais and Jensen (2014) use the geographic distribution of producers and consumers in the U.S. to
estimate trade costs at a detailed 6-digit NAICS industry level. We use as the trade cost threshold for
tradability their estimated trade cost that implies 90 percent of manufacturing employment is tradable.
19
(2012), we exclude Alaska because the voting data are reported in districts that cannot be
mapped to specific counties.
We assess the statistical adequacy of the models using (and reporting) Durbin’s M
statistics, as well as relying on residuals plots. This is necessary given that the models contain a
lagged endogenous variable. We also use Bayesian Information Criterion (bic) to develop
preferred models.
National-level Analysis
The dependent variable in this part of the investigation is postwar Incumbent Party share
of the two main party Presidential votes (Incumbent 2-Party Vote Share,t), 1952 until 2012. The
sample is determined by the availability of quarterly data on economic growth.17 We also
estimate a model, 1936-2012, with more limited covariates using data from Fair 2009. The
passage of the Reciprocal Trade Agreements Act (RTAA) of 1934 repealed the Smoot-Hawley
Tariff, and is widely seen as marking the modern era of U.S. trade integration. (See Bailey,
Goldstein, and Weingast 1997, and Hiscox 1999 for discussions of the RTAA.)18
The standard approach in the literature has been to estimate OLS time-series models with
a necessarily parsimonious set of explanatory variables. While investigators differ in
specifications, the most commonly used approach contains one or more measures of economic
17
Quarterly data for the four quarters prior to the election (Q12 through Q15) are used rather than annual
growth data (Q13 through Q16). The latter indicator includes information for the 53 to 59 days of
economic activity after the election (depending on the date of the election in a particular year).
18
As Goldstein 1994 notes, U.S. trade policy post-RTAA contained important legacies of prior
protectionist policies and programs, which attenuated slowly over time. We expect therefore (and find)
weaker estimated effects in earlier periods. Results available from the authors.
20
performance, a measure of voter sentiment or public opinion, and some information about prior
incumbency or vote share. We adopt that approach here, adding trade variables to the models.
Because the list of plausible measures of the explanatory variables of Incumbent Vote
Share,t exceeds the plausible degrees of freedom given at most 16 (or 20) observations, the issue
of omitted variable bias in the estimations arises. That is, omitting either some hard-to-observe
factor (e.g., candidate quality) or a broad array of plausible explanatory variables might
systematically influence the dependent variable, and the influence of the omitted variables will
be attributed to the error term. As noted above, prior incumbent vote share (IncVoteShare t-1) is
both a plausible correlate of current vote share, and is entered as a lagged endogenous variable to
attenuate possible omitted variable bias. 19
Additional independent variables proposed in prior studies include retrospective
indicators of economic performance: per capita real economic growth; changes in personal
disposable income; job growth during a Presidential term (Lewis-Beck and Tien 2004); inflation
during the 12 months prior to the election; and changes in unemployment. Among the variables
scholars propose to represent citizen perceptions of either future economic performance or
approval of the President are: perceived business confidence in quarter 15 (Erikson 2009); and
Presidential approval in the first July Gallup poll.20
19
As an alternative, number of incumbent party presidential terms (#IncTerms) from Abramowitz 1988,
2008 is also alternatively entered in some models.
20
Campbell 2012 used the Gallup polls in early September of the election year instead of the July Gallup;
Abramowitz 2012 used “net candidate advantage” derived from the June Gallup poll. We also rotate
these variables in, and the results are substantively identical.
21
The variables are highly like to contain overlapping information, and cannot all be
entered in any event owing to limited degrees of freedom. To explore both the identifying and
common-pool variances contained in the proposed explanatory variables, we undertake factor
analysis. 21 Given the joint need to avoid omitted variable bias while maintaining a parsimonious
model, factor analysis can help to determine which measures of the base-line model provide
unique identifying variances. To ensure the reader of the robustness of the results, models with
alternative measures of the variables in the baseline model will also be reported. Models and
variables from leading studies will also be estimated, with trade variables entering the
specification.
The variables, their timing, and their sources are described in the Data Appendix. The
timing of the variables is such that information after the Presidential elections in November is
excluded. In most of our models, the investigation starts with the 1952 data, with data from the
1948 election used as a lagged endogenous variable in the model.
To assess the statistical adequacy of the models used in the regression analyses, a number
of diagnostic tests are reported. This is especially important in the context of a small number of
observations with potentially correlated errors (see Grant and Lebo forthcoming). 22
21
Factor analysis explores how much of the variability in a set of variables is due to one or more
underlying ‘latent’ variables, or factors. Once the number of components or factors is
established, factor analysis shows which indicators among the proposed determinants of
incumbent vote share indicators are associated with a given factor, and how strong that
association is. (See Kim and Mueller (1978) for a discussion of factor analysis.)
22
Because the models include some form of lagged endogenous variable, the classical Durbin-Watson
statistic is replaced with a Lagrange-multiplier test for 1st and 2nd order residual autocorrelation – the AR
22
Correlations and Descriptive Statistics. The pairwise correlations among the dependent and
independent variables are presented in Appendix Table A1a (macro data) and Table A1b
(county-level data). The descriptive statistics for the variables are reported in Table A2a and
A2b (macro- and county-level respectively).
At the macro-level, the correlations between and among the indicators of economic
growth, perceived Business Sentiment, change in unemployment, Presidential job approval, the
number of prior Incumbent Terms, and the dependent variable, Incumbent Vote Share, t are
large, positive, and statistically significant. Economic growth, change in the trade balance, and
change in exports are highly correlated. Change in imports and change in the number of jobs are
weakly correlated with the main dependent variable and the other regressors.
In contrast to the macro-level data, when using county-level data in the factor analysis,
the main variables of interest have much smaller correlation coefficients. The main exception is
1-2 test. (The LM test nests a Durbin’s H statistic. See Doornick and Hendry 2009, 286.) The inclusion
of prior incumbent vote share raises the prospect of correlation between the lagged endogenous variable
and the error term. To account for possible error correlation and heteroskedasticity, which can bias
standard errors, Newey-West heteroskedasticity and autocorrelation consistent standard errors (HACSE)
are reported. A “Normality test for residuals” based upon Jarque-Bera with the Doornik-Hansen small
sample correction is also reported. Finally, an ARCH test for conditional first order heteroscedasticity is
reported. All three tests assume a null hypothesis in the diagnostic statistics of no assumption violations.
The AR and ARCH tests are F-tests, and the normality test is a χ2 test: statistically significant p-values
signify assumption violations.
23
that the standard deviation of unemployment and the unemployment rate are correlated at .5***.
That is, increased employment volatility is associated with increased unemployment.
Factor Analysis. Factor analysis allows for the intercorrelations among the variables and an
assessment of whether an underlying or latent variable nests the explanatory variables. Given
the high pairwise correlations between variables in the macro data, it is empirically possible that
different variables represent a facet of one or more common underlying variables, which is tested
below and reported in Table Appendix Table A3a (macro) and Table A3b (county-level).
At the macro level, the principal factor analysis shows three latent variables undergird the
macro data, accounting for 85 percent of the variance in the data. Business Sentiment, July
Gallup Presidential Approval, change in unemployment, economic growth, and change in
exports all load strongly on the first factor.23 Two variables (change in consumer prices and job
growth - ΔJobs) load on multiple dimensions. Incumbent Party terms and Incumbent Prior votes
load on separate factors. Only ΔImports/GDP fails to load on a factor: it consequently has a high
uniqueness score.24
23
As noted above, Campbell 2012 used the Gallup polls in early September of the election year instead of
the July Gallup; Abramowitz 2012 used “net candidate advantage” derived from the June Gallup poll.
Both variables load on the same first factor as July Gallup and other variables listed above.
24
“Uniqueness” refers to the information overlap between and among variables. In principal component
analysis, the assumption is that variables have a high ‘communality’ of information. In principal factor
analysis, used here, that assumption is tested. The higher the 0-1 ‘unique’ score, the more the variable is
measuring a phenomenon different from that being measured in other variables. Scores above .6
considered to be ‘high’ and a sign that the variable is a reliably different measure from other variables.
24
Among the county-level variables, in contrast, the intercorrelations among the variables
are weak to negligible. One underlying factor is marginally present, and only one variable loads
on the marginal factor. All other variables have high uniqueness scores, suggesting that the
information overlap among the variables is very low.
Implications. Many of the macro variables load on the first factor, including the prospective and
retrospective indicators of economic performance, suggesting that the economic performance,
the sentiment, and incumbent vote share variables are different facets of a single common
underlying variable. It is not evident, therefore, that the retrospective and prospective indicators
contain enough unique identifying variance to fully distinguish in this investigation between
retrospective and prospective economic effects. Given the results of the factor analysis, it is
likely that whichever of the various economic performance indicators are used, substantively
similar results will be obtained. This is likely true for the sentiment/approval indicators as well,
a supposition which is explored below.
In contrast, the evidence suggests that ΔImports/GDP is not subsumed in the other
factors, and, as indicated by its ‘uniqueness score,’ contains useful identifying variance. The
ΔExports/GDP indicator, in contrast, is likely to overlap in information with the economic
performance and voter sentiment variables. (The ΔTrade Balance/GDP does not load on any
factor.) Hence, the trade variables partly contain independent information.
In contrast, the county-level results show far less overlap in information. Each of the
variables of interest has a high ‘uniqueness score’ and can be considered to be independent of
any underlying latent joint construct.
The measures at the county level, however, do not include some important variables of
interest. Trade balances are not computed at county levels, and neither are the public opinion
25
variables. We propose therefore that a joint consideration of the macro-level and country-level
results is best suited for this investigation.
County-level Results
Table 1 reports models of incumbent two-party vote shares at the county level during the
U.S. presidential elections years 1992 through 2012. Unless otherwise noted, all models include
year and county fixed effects, and the standard errors are clustered at the county-level. In the
Appendix, we include a replication of Table 1 including a number of additional county-level
control variables, and the results are virtually unchanged. (See Table A4.)
The results reported in column 1 of Table 1 represent strong evidence that voters respond
negatively to employment volatility and positively to increases in wages. Voters are particularly
unlikely to vote for the incumbent in counties where unemployment is highly volatile. The
coefficient estimate corresponding to Unemployment Volatility25 is negative and statistically
significant. A one-standard-deviation increase in unemployment volatility is associated with a
.83% decrease in incumbent vote share. A one-standard-deviation jump in unemployment is
associated with a .52% decrease in vote shares. An increase in average pay over the previous
year (Δ Average Pay (1-year)) equal to one standard deviation increases the incumbent vote
share by .33%.
Our results strongly support Wright’s (2012) finding that increases in unemployment
harm Republican incumbents and help Democratic incumbents. A one-percent increase in
25
Unemployment volatility is measured as the standard deviation of the unemployment rate over
the 4 years prior to and including the election year.
26
unemployment is associated with a .4% decrease in vote share for Republican incumbents, and
with a .13% increase for Democratic incumbents.26
Model 2 introduces the location quotient of manufacturing employment, which measures
the concentration of workers in the manufacturing sector in each county. The estimated
coefficient indicates that manufacturing concentration is associated with lower vote shares for
the incumbent, even after controlling for income, unemployment, and employment volatility. In
terms of substantive impact, a one-standard-deviation increase in the manufacturing location
quotient is associated with a .82% decline in the incumbent vote share.
Models 3-5 assess whether the expected winners from globalization are more likely to
vote for the incumbent party than are the expected losers from globalization. In Model 3, we
introduce the employment location quotient for tradable services, a comparative advantage
sector, and non-tradable sectors. Employees of firms in the tradable services sector are likely to
experience expanding wage and employment opportunities from increased exports and
investment. The estimated effect of this variable on incumbent vote share contrasts sharply with
that of manufacturing. The estimates reported in column 3 indicate that a one-standard deviation
increase of concentration of employment in tradable services is associated with a .81% increase
26
In appendix Table A5, following Wright (2012), we report models of Democratic presidential
vote share. We find that unemployment is associated with increased Democratic vote share result
over the period 1992-2012 in models that include either county or state fixed effects. The table
also shows that the interaction terms Democratic Incumbent x Δ Unemployment (1-year) or
Democratic Incumbent x Unemployment enter positive and statistically significant, which is
consistent with Wright’s result.
27
in the incumbent vote share. Our results hold when we include manufacturing concentration in
columns 4-5. Furthermore, manufacturing concentration remains negative and statistically
significant, suggesting that dislocations in the manufacturing sector independently help explain
political discontent.27 To ensure that the results are not sensitive to weighting, we include the
unweighted model estimates in column 5. Our main variables retain statistical significance; we
note that the magnitude of the manufacturing coefficient increases, while the magnitudes of the
estimated coefficients for unemployment volatility and of tradable services decrease.28
Figure 1 displays the estimated effects of the explanatory variables, based on the
coefficient estimates reported in column 4 of Table 1. The marginal effects correspond to onestandard-deviation changes in each of the explanatory variables.
The degree to which our framework can help explain the outcome of national-level
presidential contests largely depends on whether our indicators of economic insecurity explain
variation in incumbent vote shares in counties in the so-called “swing” states. These are states
that, owing to the fact that the two major political parties have similar levels of support among
voters, are crucial to determining U.S. presidential elections.
27
In Appendix Table A6, following Margalit (2011), we show that the number of workers filing
for Trade Adjustment Assistance (TAA), is negatively correlated with incumbent vote shares.
28
In unreported models, we include the variable Swing as used in Powell and Whitten 1993. The
coefficient and standard errors corresponding to our variables of interest are nearly identical to
those reported in column 4.
28
We report coefficient estimates among the sample of swing- and non-swing states in
Table 2.29 The results reported in column 1 indicate that Unemployment Volatility is associated
with lower incumbent vote shares in both swing and non-swing states, but the estimated
coefficient is larger in the swing-states sample. Higher unemployment is also statistically
significantly associated with lower vote shares in swing and non-swing states alike.
Manufacturing Concentration is strongly negatively associated with changes in incumbent vote
share, particularly in swing states. The estimated coefficient corresponding to Manufacturing
Concentration is nearly twice as large in the swing states as compared to our previous estimates
using the full sample of counties. A one-standard-deviation increase in manufacturing
concentration is associated with a 1.9 percent decline in incumbent vote share in swing states. In
column 2, we report estimates of counties outside the swing states. Here we find that higher
levels of tradable services concentration and higher levels of average income are associated with
greater support for the incumbent. In columns 3 and 4, we probe the robustness to a smaller
sample of swing states, dropping Nevada and Wisconsin, and the results outside of this smaller
sample. The estimates are little changed.
Our results from the county-level analysis can be summarized as follows. We find strong
evidence that employment volatility and rising unemployment significantly reduce support for
the incumbent. Volatility in employment varies substantially across counties in the U.S., and has
yet to be linked to voting behavior. Our results indicate that voters strongly prefer stability and
that they voice dissatisfaction with instability at the ballot box. We also find that the sectoral
positions of voters, and the concentration of economic activity in comparative advantage and
29
These are Colorado, Florida, Iowa, North Carolina, New Hampshire, Ohio, Pennsylvania, Virginia,
Nevada, and Wisconsin (“The 10 Closet States in Election 2012, Washington Post, November 8, 2012.).
29
comparative disadvantage activities, helps explain voting behavior. Voters appear to be more
inclined to support the incumbent in comparatively advantaged sectors such as tradable services.
We also find some preliminary evidence that concentrations of economic activity in
manufacturing influence support for the incumbent party. Typically viewed as comparatively
disadvantaged in the U.S. and subject to high levels of competition from abroad, we find support
for the proposition that concentrations of employment in this sector are correlated with political
discontent. However, a significant amount of high value added manufacturing is conducted in the
U.S., and a significant number of workers are employed in globally competitive manufacturing
firms. Thus, there are clear winners within this sector. We expect some portion of manufacturing
employees – those in high wage jobs in firms producing tradable goods – to be relatively content
with their place in the global economy.
We expect that our analysis of the Census data will provide greater resolution and a better
opportunity to identify heterogeneity in economic opportunity within sectors. In particular, these
data allow us to capture the share of employment in high-skilled traded manufacturing and
services.
Macro Regression Models. In light of prior theory and statistical testing, the base macro model
to be estimated is:
Incumbent 2-Party Vote Sharet = ß0 + ß1(Incumbent Vote Sharet-1) + ß2(Economic Growth t-1)
+ εt
t=1952-2012
(2)
To this model will be added change in the trade indicators:
either ß3(ΔTradeBal/GDPt-1 ) or + ß3′(ΔImport/GDPt-1) and ß34′(ΔExports/GDPt-1),
plus an indicator of either Business Sentiment or July Approval - ß5(Sentiment/Approvalt-1 ).
30
Scholars have tended to use either consumer/business sentiment or Presidential approval
variables: both have strong grounding in both theory and empirical results as useful predictors of
incumbent vote shares.
As an initial indicator of economic performance, we use real per capita GDP from
Quarter 12 through Quarter 15. As will be demonstrated below, the choice of the economic
growth variable is not influential in the results. To represent voter sentiment (and prospective
economic activity), we follow Erikson (2009) and use Business Conditions Q15. An alternative
indicator is the widely used July Gallup (Q15) Presidential approval ratings (Abramowitz 2008;
Lewis-Beck and Tien 2008). The trade variables use data from Quarter 12 through Quarter 15.
To assure the reader that the results do not derive from choice of indicators, additional results
using other indicators (#Incumbent Terms in lieu of prior incumbent vote share, e.g.) are
reported. Other models proposed by other scholars are also estimated with the trade terms
entered. Models with ΔImports/GDPt-1 and ΔJobst-1 jointly and separately entered are
reported.
Regression Results. Table 3 reports the main macro results. In Model 3.1, prior incumbent vote
share and economic growth are entered. The estimation properties of the model are good, and
the results are consistent with prior findings. In these models and all models reported, economic
growth has an estimated coefficient that is positive, highly statistically significant, and
substantively large. The lagged endogenous variable has a negative and highly statistically
significant coefficient, which is consistent with the theories about the “costs of governing” and
the standard findings of a decline in incumbent vote margins in subsequent elections. Taking the
mean of two party Incumbent Vote Sharet-1 (53.7% for the 1948-2008 elections), multiplying it
by the parameter estimate, and adding to its product the estimate of the constant produces an
31
estimate of Incumbent Vote Sharet of 46.8%, or a -6.9% change from the prior election,
assuming a zero increase in economic growth.
Multiplying the estimated coefficient of growth
times the sample mean (a growth rate of 2.2%) and adding it to the above calculation produces
an estimate of Incumbent Vote Sharet of 52%. A growth rate of one and a quarter percent or
lower brings the estimate of Incumbent Vote Sharet below 50%. This very simple model has an
adjusted r-squared of .61.
In Models 3.2 and 3.3, the trade indicators are entered. Change in the trade balance
(Model 3.2) has a statistically significant positive coefficient that is substantively large and
consistent with the theory developed above. A one unit increase (decrease) in the U.S. trade
balance is associated with a 4% estimated increase (decrease) in incumbent vote shares. Change
in imports (Model 3.3) has a statistically significant negative coefficient, which is substantively
large and consistent also with theory. A one unit increase (decrease) in imports as a percentage
of GDP is associated with a 4% decrease (increase) in incumbent vote shares. Change in exports
as a percentage of GDP has a positive and statistically significant coefficient that is substantively
large: a one unit increase is associated with a six percent increase in Presidential vote shares.
The explanatory power of the models, judged via adjusted R-squared indicators, rises 19 points
with the inclusion of the trade variables.30
Models 3.4 and 3.5 add the indicators of Business Sentiment Q15 and July Gallup
(respectively) to the model with ΔTradeBal/GDPt-1. Both indicators have positive, statistically
significant, and substantively large estimated coefficients that are consistent with prior theory
30
Change in exports loads on Factor 1 (Table A3.a) along with the growth indicators. The variable
therefore contains overlapping information with the indicators of economic growth.
32
and findings.31 The inclusion of the Business Sentiment and July Gallup variables leads to a
marked increase in the explanatory power of the models: from .8 without either to .99 for
Business Sentiment Q15 and .9 for July Gallup. The estimated coefficients of ΔTradeBal/GDPt1 remain positive and highly statistically significant. Its coefficient estimate is larger in the
model including Business Sentiment (3.4) compared to the model with July Gallup (3.5). Model
3.5, however, shows evidence of statistically significant heteroskedasticity in the residuals.
In Models 3.6 and 3.7, ΔTradeBal/GDPt-1 is replaced by ΔImport/GDPt-1 and
ΔExports/GDPt-1: Business Sentiment and July Gallup (respectively) are included. The models
have good estimation properties and explanatory power. The estimated coefficient of
ΔImport/GDPt-1 remains negative and highly statistically significant, and the estimated
coefficient of ΔExport/GDPt-1 remains positive and statistically significant.
Other variables are not the main focus, but the estimated coefficient of economic growth,
while always positive and statistically significant at beyond the .01 confidence level, is
diminished in size with the inclusion of either business sentiment or July Gallup. The coefficient
31
The Business Sentiment Q15 data are available only from 1954 onward, making the 1956 election the
first election in the sample. The July Gallup variable is available from the 1940s onward. In order to
compare the estimated effects of change in imports across the different specification, the 1956-20012
sample is used. The results for the models with July Gallup in the 1952-20012 sample are nearly identical
to the models reported.
33
estimates of Business Sentiment and July Gallup are always positive, statistically significant and
substantively important.32
In Table 4, we use Model 3.3 as the base model and add additional indicators. Models 4.1
uses the number of Incumbent Party prior terms. Model 4.2 adds “Jobs” from Lewis-Beck and
Tien. Models 4.3 and 4.4 add indicators of change in unemployment and inflation, used in the
Stimson, Erikson, and Mackuen 2004 “Macropolity” model. Model 4.6 adds a time trend. In all
cases, the export and import coefficient estimates retain the expected sign, and the estimated
coefficients are statistically significant at the .1 level or better.
As a further experiment, we extend the sample back to the 1936 election, which is postRTAA, using data and models from Fair 2009. Model 4.8 enters change in the trade balance;
model 4.9 enters changes in imports and exports. The coefficient estimates retain similar signs
and levels of statistical significance.
Conclusion
Popular accounts in the press and prior academic research indicate that globalization –
characterized by increases in financial integration, rising import competition, and the offshoring
of production – shapes politics through its effects on employment, wages and economic
insecurity. The effects of globalization on the most fundamental political activity – voting – is
not well understood, as most models of economic voting to date have ignored how trade shapes
voting behavior. Our paper demonstrates that changes in trade flows and the concentration of
economic activity in winning and losing industries and firms are unique contributors to
32
Models that include both Business Sentiment and July Gallup leave neither with statistically significant
estimated coefficients, though the trade variable estimates are substantively similar. Details available.
34
explaining national Presidential voting. The results demonstrate that the relative decline in the
U.S. trade position over the past two decades has been harmful to incumbent presidents, and
suggests that an array of related macroeconomic conditions – currency valuation, China’s
accession to the WTO, and the increasing export competitiveness of other emerging markets –
are potentially consequential for U.S. presidential elections.
We consider variation in the spatial distribution of economic activity in the U.S. to better
understand how globalization affects the vote. Our argument is that increases in trade and
investment principally benefit the employees of the largest and most productive firms that engage
in these activities, while employees of firms that compete with imports may be dissatisfied with
economic integration that increases competition from abroad. We examine this proposition at the
county level by considering the volatility of local employment and wages, and the concentration
of economic activity among firms in import competing and high-skill, exporting sectors. We find,
for the first time, that unemployment volatility influences U.S. presidential voting, as citizens are
more likely to demonstrate their insecurity by voting against the incumbent. We confirm the
Margalit (2011) finding that aggregate trade-related job losses decrease the incumbent vote share.
Controlling for job losses and employment volatility, we find evidence that the concentration of
economic activity in manufacturing diminishes the incumbent vote, while counties experiencing
increases in human capital intensive, exportable activities demonstrate greater satisfaction with the
incumbent.
Based on our analyses and using the most recent data available for all indicators, we
make a prediction of which party is more likely to win the 2016 Presidential elections. Using
models 3.4 and 3.5, we forecast that, were the election to be held in August of 2015, the nominee
of the Democratic Party would comfortably win the two party popular vote totals. (We do not
35
forecast the Electoral College results.) The results, however, are fragile to relatively modest
changes in economic conditions. For example, if the economic conditions of the fall of 2000
prevailed – growth at ‘only’ 3% and an import surge of 1% of GDP linked to a sharp rise in the
value of the U.S. dollar the year before – the Democratic Party incumbent is forecast to receive
slightly less than 50% of the two party vote shares.33 Recent BEA trade data indicates rising
U.S. imports, especially from the Eurozone. We will update as the election cycle proceeds.
33
Between November 1999 and November 2000, the trade-weighted value of the U.S. Dollar increased
by 9%. Between August of 2014 and August of 2015, the trade-weighted value of the U.S. Dollar
increased by 15.6%. (Source: U.S. Federal Reserve Bank of St. Louis (FRED), series TWEXB.
36
Table 1. County-level Determinants of Incumbent Two-Party Vote Shares, 1992-2012
Presidential Elections
Incumbent Vote Share (lagged)
Unemployment
Δ Unemployment (1-year)
Democratic Incumbent
Democratic Incumbent x Δ Unemployment (1-year)
Unemployment Volatility
Average Pay
Δ Average Pay (1-year)
(1)
1.035***
(0.005)
-0.002***
(0.001)
-0.004***
(0.001)
0.085***
(0.003)
0.005***
(0.002)
-0.014***
(0.002)
0.020
(0.014)
0.092***
(0.020)
Manufacturing Concentration
(2)
1.034***
(0.005)
-0.002***
(0.001)
-0.004***
(0.001)
0.085***
(0.003)
0.005***
(0.002)
-0.013***
(0.002)
0.022
(0.015)
0.098***
(0.022)
-0.010***
(0.003)
Non-Tradable Services Concentration
(3)
1.035***
(0.005)
-0.002***
(0.001)
-0.004***
(0.001)
0.085***
(0.003)
0.006***
(0.002)
-0.014***
(0.002)
0.022
(0.013)
0.095***
(0.020)
Constant
-0.260*
(0.135)
-0.285**
(0.144)
0.005
(0.005)
0.020***
(0.005)
-0.271**
(0.130)
Observations
R-squared
Counties
18663
0.951
3111
16167
0.951
2896
17969
0.951
3072
Tradable Services Concentration
(4)
1.034***
(0.005)
-0.002***
(0.001)
-0.004***
(0.001)
0.085***
(0.003)
0.005***
(0.002)
-0.014***
(0.002)
0.023
(0.014)
0.099***
(0.022)
-0.007**
(0.003)
0.004
(0.005)
0.018***
(0.005)
-0.283**
(0.139)
16095
0.951
2889
(5)
1.013***
(0.003)
-0.002***
(0.000)
-0.000
(0.001)
0.063***
(0.001)
0.002**
(0.001)
-0.007***
(0.001)
-0.014*
(0.008)
0.022*
(0.011)
-0.013***
(0.002)
0.004
(0.003)
0.008***
(0.003)
0.086
(0.074)
unweighted
16095
0.927
2889
Note: The dependent variable is the incumbent two-party vote share. Democratic party incumbent election years are
1996, 2008, and 2012; Republican party incumbent election years are 1992, 2004, and 2008. All models include
county and year fixed effects. Unemployment volatility is the standard deviation of the county-level unemployment
rate calculated over the 4-year period including the 3 years prior to the election and the election year. All other
regressors are also measured at the county level, and correspond to the election year. Average pay and the
concentration measures enter as natural logs. The concentration measures are employment location quotients from
the U.S. Quarterly Census of Employment and Wages and computed by the U.S. Bureau of Labor Statistics.
Estimates in columns 1-4 are weighted by population size in 1990. The robust standard errors (reported in
parentheses) are adjusted for clustering at the county-level. * p-value < .1; ** p-value < .05; *** p-value < .01.
37
Table 2. County-level Determinants of Incumbent Two-Party Vote Shares, 1992-2012
Presidential Elections (Swing and non-Swing States)
Incumbent Vote Share (lagged)
Unemployment
Δ Unemployment (1-year)
Democratic Incumbent
Democratic Incumbent x Δ Unemployment (1-year)
Unemployment Volatility
Average Pay
Δ Average Pay (1-year)
Manufacturing Concentration
Non-Tradable Services Concentration
Tradable Services Concentration
Constant
Observations
R-squared
Counties
(1)
10 Swing
States
0.999***
(0.013)
-0.005***
(0.001)
-0.004**
(0.001)
0.075***
(0.005)
0.006**
(0.003)
-0.009***
(0.002)
-0.036*
(0.020)
0.049
(0.037)
-0.023***
(0.006)
-0.001
(0.008)
0.010
(0.008)
0.316
(0.194)
3951
0.941
693
(2)
non-Swing
States
1.043***
(0.006)
-0.001*
(0.001)
-0.004**
(0.002)
0.087***
(0.003)
0.004**
(0.002)
-0.015***
(0.003)
0.040**
(0.016)
0.109***
(0.026)
-0.004
(0.004)
0.004
(0.006)
0.019***
(0.006)
-0.463***
(0.158)
12144
0.954
2196
(3)
8 Swing
States
1.001***
(0.014)
-0.004***
(0.001)
-0.005***
(0.002)
0.079***
(0.006)
0.009***
(0.003)
-0.008***
(0.002)
-0.043**
(0.019)
0.072*
(0.038)
-0.016***
(0.005)
-0.001
(0.008)
0.010
(0.007)
0.379**
(0.189)
3466
0.945
609
(4)
non-Swing
States
1.042***
(0.005)
-0.002**
(0.001)
-0.004**
(0.002)
0.086***
(0.003)
0.004**
(0.002)
-0.015***
(0.003)
0.041**
(0.016)
0.106***
(0.026)
-0.005
(0.004)
0.005
(0.006)
0.019***
(0.006)
-0.469***
(0.158)
12629
0.953
2280
Note: The dependent variable is the incumbent two-party vote share. All models include county and year fixed
effects. The 10 swing states are Colorado, Florida, Iowa, North Carolina, New Hampshire, Ohio, Pennsylvania,
Virginia, Nevada, and Wisconsin. The sample of 8 swing states drops Nevada and Wisconsin. Unemployment
volatility is the standard deviation of the county-level unemployment rate calculated over the 4-year period
including the 3 years prior to the election and the election year. All other regressors are also measured at the county
level, and correspond to the election year. Average pay and the concentration measures enter as natural logs. The
concentration measures are employment location quotients from the U.S. Quarterly Census of Employment and
Wages and computed by the U.S. Bureau of Labor Statistics. Estimates are weighted by population size in 1990. The
robust standard errors (reported in parentheses) are adjusted for clustering at the county-level. * p-value < .1; ** pvalue < .05; *** p-value < .01.
38
Table 3. Base Models – Dependent Variable is National Incumbent Party Vote Shares (Two Party), 1952-2012; 1936-2012
Model 1
Model 2
Model 3
Model 4
Model 5
Model6
Model 7
Model 8
(1936-
Model 9
(1936-
Prior Incumbent Vote t-1
-0.74***
-0.773***
-0.764***
-0.811***
-0.538***
-0.746***
-0.542***
-0.443
-0.441
(0.216)
(0.155)
(0.152)
(0.06)
(0.57)
(0.122)
(0.157)
(0.255)
(0.265)
Growth Q12_15
0.022***
0.031***
0.032***
0.018***
0.021***
0.018**
0.023***
0.018***
0.018***
(0.005)
(0.004)
(0.004)
(0.005)
(0.005)
(0.006)
(0.005)
(0.004)
((0.004)
0.037***
0.032**
(0.009)
(0.012)
ΔTradeBal Q12_15
0.04***
(0.012)
ΔImportsQ12_15
ΔExportsQ12_15
0.026**
(0.01)
-0.04**
-0.036***
-0.028**
-0.025*
(0.013)
(0.011)
(0.012)
(0.015)
0.06***
0.038*
0.044**
0.028*
(0.019)
(0.017)
(0.014)
(0.018)
BusSentimentQ15
0.0006***
0.001**
(0.0001)
(0.0002)
July Gallup
0.0015***
0.0014**
(0.0006)
(0.0005)
War
Constant
Obs.
0.007
0.009
(0.034)
(0.02)
0.865***
0.869***
0.853***
0.818***
0.672***
0.817***
0.69***
0.722***
0.719***
(0.113)
(0.081)
(0.081)
(0.03)
(0.079)
(0.069)
(0.095)
(0.139)
(0.144)
16
16
16
15
16
15
16
20
20
0.61
0.798
0.802
0.988
0.9
0.915
0.91
0.41
0.44
AR 1-2 test [p-value]
[0.87]
[0.37]
[0.94]
[0.37]
[0.94]
[0.42]
[0.92]
[0.25]
[0.28]
ARCH 1-1 test [p-value]
[0.51]
[0.85]
[0.96]
[0.93]
[0.04**]
[0.98]
[0.16]
[0.93]
[0.99]
[0.87]
[0.85]
[0.88]
[0.40]
[0.38]
Adj. R2
Normality test [p-value]h
[0.85]
[0.53]
[0.85]
[0.84]
Notes: * p-value < .1; ** p-value < .05; *** p-value < .01. Data for the 1936, 1940, 1944,
and 1948 elections are from Fair 2009 and the U.S. Bureau of Economic Analysis.
39
Table 4. Alternative Measures of Incumbency, Growth, Sentiment, and Trade in Macro Models
Prior Incumbent Vote t-1
Growth Q12_15
ΔImportsQ12_15
ΔExportsQ12_15
#Incumbent Party Terms
Model 1
Model 2
Model 3
Model 4
Model 5
-0.673***
-0.831**
-0.745***
-0.746***
-0.752***
(0.15)
(0.159)
(0.169)
(0.015)
0.027***
(0.163)
0.031***
0.029***
0.032***
0.032***
(0.005)
(0.004)
(0.006)
(0.004)
(0.004)
-0.039***
-0.037**
-0.038***
-0.038**
-0.034**
(0.012)
(0.014)
(0.014)
(0.015)
(0.014)
0.038*
0.053**
0.061***
0.061***
0.064***
(0.021)
(0.019)
(0.019)
(0.019)
(0.018)
-0.011
(0.006)
ΔJobs (Q15-Q1)/Q1
0.003
(0.002)
ΔUnemployment
-0.011
(0.014)
Inflation (CPI)
-0.001
(0.003)
Time Trend
-0.002
(0.002)
Constant
Obs.
Adj. R2
AR 1-2 test [p-value]
ARCH 1-1 test [p-value]
0.843***
0.868***
0.848***
0.847***
0.486***
(0.076)
(0.081)
(0.083)
(0.087)
(0.024)
16
16
16
16
16
0.83
0.81
0.79
0.78
0.81
[0.85]
[0.67]
[0.96]
[0.72]
[0.6]
[0.13]
[0.55]
[0.42]
[0.65]
[0.76]
[0.82]
[0.56]
[0.66]
[0.64]
[0.18]
Normality test [p-value]
40
Figure 1. Marginal Effects (based on estimates in Table 1, column 4)
Unemployment
Unemployment Volatility
Average Income
Average Income (one-year change)
Manufacturing Concentration
Non-Tradable Services Concentration
Tradable Services Concentration
-.01
-.005
0
.005
.01
Note: The figure shows the estimated effect of a 1 standard deviation change in the explanatory
variables on the incumbent vote share. The results are based on model 4 of Table 1, with
regressors standardized to mean of 0 and standard deviation of 1.
41
References
Abramowitz, Alan I. 1988. "An Improved Model for Predicting Presidential-Election
Outcomes." Ps-Political Science & Politics 21 (4): 843-7.
Abramowitz, Alan I. 2001. "The Time for Change Model and the 2000 Election." American
Politics Research 29 (3): 279-82.
Abramowitz, Alan I. 2008. "It's about Time: Forecasting the 2008 Presidential Election with the
Time-for-Change Model." International Journal of Forecasting 24 (2): 209-17.
Acemoglu, Daron, and David Autor. 2010. “Skills, Tasks and Technologies: Implications for
Employment and Earnings.” In Handbook of Labor Economics, Volume 4, Orley
Ashenfelter and David E. Card (Eds.) Amersterdam: Elsevier.
Acemoglu, Daron, and David Autor. 2011. “Skills, Tasks and Technologies: Implications for
Employment and Earnings.” In Handbook of Labor Economics, Volume 4, edited by
Orley Ashenfelter and David E. Card. ,pp 1043-1171. Amsterdam: Elsevier.
Aldrich, John, Claire Kramer, Peter Lange, Renan Levine, Jennifer Merrolla, Laura Stephenson,
and Elizabeth Zechmeister. ND. In Pursuit of the Missing Link. Presented at the 2002
Annual Meeting of the Political Science Association, Boston, Massachusetts.
Althaus, Scott. 2003. Collective Preferences in Democratic Politics. Cambridge: Cambridge
University Press.
Antoniades, Alexis, and Charles W. Calomiris. 2014. “Mortgage Market Credit Conditions and
U.S. Presidential Elections.” Paper presented at the Mortara Political Economy Seminar
Luncheon Series.
Autor, David H., David Dorn, and Gordon H. Hanson. 2013. “The China Syndrome: Local Labor
Market Effects of Import Competition in the United States.” American Economic Review
103(6): 2121-2168.
Autor, David, H., David Dorn, Gordon H. Hanson, and Jae Song. 2014. Trade Adjustment:
Worker Level Evidence. No. w19226. National Bureau of Economic Research, 2013.
Forthcoming, Quarterly Journal of Economics.
Bailey, Michael. 2001. “Quiet Influence: the Representation of Diffuse Interests on Trade
Policy, 1983-94.” 26(1):45-80.
Bailey, Michael A., Judith Goldstein, and Barry R. Weingast. 1997. “The Institutional Roots of
American Trade Policy.” World Politics 49(3):309-338.
Bernard, Andrew B., and J. Bradford Jensen. 1999. “Exceptional Exporter Performance: Cause,
Effect, or Both?” Journal of International Economics 47:1–25.
42
Bernard, Andrew B., J. Bradford Jensen, Stephen J. Redding and Peter K. Schott. 2007. “Firms
in International Trade.” The Journal of Economic Perspectives 21(3):105-130.
Bernard, Andrew, J. Bradford Jensen and Peter K. Schott. 2006. “Survival of the Best Fit:
Exposure to Low-Wage Countries and the (Uneven) Growth of U.S. Manufacturing
Plants.” Journal of International Economics 68:219-237.
Bureau of Economic Analysis (BEA). 2011. National Economic Accounts: Table 3. Available
from http://www.bea.gov/newsreleases/national/gdp/gdpnewsrelease.htm.
Burstein, Ariel, and Jonathan Vogel. 2010. “Globalization, Technology, and the Skill Premium:
A Quantitative Analysis.” NBER Working Paper No. 16459. Department of Economics,
UCLA, Los Angles, California.
Burstein, Paul. 2003. “The Impact of Public Opinion on Public Policy: A Review and an
Agenda.” Political Research Quarterly 56:29–40.
Campa, Jose M. and Linda S. Goldberg. 2001. “Employment versus Wage Adjustments and the
U.S. Dollar.” Review of Economics and Statistics 83(3):477-489.
Campbell, James E. 2008. "Evaluating U.S. Presidential Election Forecasts and Forecasting
Equations." International Journal of Forecasting 24 (2): 257-69.
Campbell, James E., and Kenneth A. Wink. 1990. "Trial-Heat Forecasts of the Presidential
Vote." American Politics Quarterly 18 (3): 251-69.
Campbell, James E., and Michael S. Lewis-Beck. 2008. "US Presidential Election Forecasting:
An Introduction." International Journal of Forecasting 24 (2): 189-92.
DeBoef, Suzanna, and Luke Keele. 2008. “Taking Time Seriously.” American Journal of
Political Science 52 (1):184–200.
Doornick and Hendry. 2009
Duch, Raymond M., and Randy Stevenson. 2008. Voting in Context: How Political and
Economic Institutions Condition the Economic Vote. New York: Cambridge University
Press.
Erikson, Robert S. 1989. "Economic-Conditions and the Presidential Vote." American Political
Science Review 83 (2): 567-73.
Erikson, Robert S. 2009a. "The American Voter and the Economy, 2008." Ps-Political Science
& Politics 42 (3): 467-71.
Erikson, Robert S. 2009b. "Do Voters Look to the Future? Economics and Elections." Political
Science Quarterly 124 (2): 361-2.
43
Erikson, Robert S., and Christopher Wlezien. 2008. "The Economy and the Presidential Vote:
What Leading Indicators Reveal Well in Advance." International Journal of Forecasting
24 (2): 218-26.
Erikson, Robert S., and Christopher Wlezien. 2009. "The Economy and the Presidential Vote in
2008." Ps-Political Science & Politics 42 (1): 22-.
Fair, Ray C. 1978. "Effect of Economic Events on Votes for President." Review of Economics
and Statistics 60 (2): 159-73.
Fair, Ray C. 2009. "Presidential and Congressional Vote-Share Equations." American Journal of
Political Science 53 (1): 55-72.
Feenstra, Robert C., and G. H. Hanson. 1996. "Globalization, Outsourcing, and Wage
Inequality." American Economic Review 86 (2): 240-5.
Fiorina, Morris P. 1981. Retrospective Voting in American National Elections. New Haven:
Yale.
Goldstein, Judith. 1994. Ideas, Interests, and American Trade Policy. Ithaca: Cornell University
Press.
Grant, Taylor, and Matthew J. Lebo. Forthcoming. “Error Correction Methods with Political
Time Series.” Political Analysis: In Press.
Guisinger, Alexandra. 2009. “Determining Trade Policy: Do Voters Hold Politicians
Accountable?” International Organization 63 (3):553–57.
Hays, Jude. 2009. Globalization and the New Politics of Embedded Liberalism. New York:
Oxford University Press.
Hibbs, Douglas A. 2000. "Bread and Peace Voting in U.S. Presidential Elections." Public Choice
104 (1-2): 149-80.
Hibbs, Douglas A., Jr. 2008. "Implications of the 'Bread and Peace' Model for the 2008 U.S.
Presidential Election." Public Choice 137 (1-2): 1-10.
Hiscox, Michael J. 1999. “The Magic Bullet? The RTAA, Institutional Reform, and Trade
Liberalization.” International Organization 53(4):669-698.
Holbrook, Thomas M. 2009. "Economic Considerations and the 2008 Presidential Election." PsPolitical Science & Politics 42 (3): 473-8.
Jacoby, William. 2000. “Issue Framing and Public Opinion on Government Spending.”
American Journal of Political Science 44:750–67.
Jensen, J. Bradford, Dennis P. Quinn, and Stephen Weymouth. 2015. “The Influence of Firm
Global Supply Chains and Foreign Currency Undervaluations on U.S. Trade Disputes.”
44
International Organization. Forthcoming. Published online at:
http://journals.cambridge.org/action/displayAbstract?fromPage=online&aid=9876418&f
ulltextType=RA&fileId=S0020818315000247
Kang, Shin-Goo, and G. Bingham Powell Jr. 2010. “Representation and Policy Responsiveness:
The Median Voter, Election Rules, and Redistributive Welfare Spending.” Journal of
Politics 72 (4):1014–28.
Kayser, Mark A., and Arndt Leininger. 2015. “Vintage Errors: Do Real-Time Economic Data
Improve Election Forecasts?” Research and Politics DOI: 10.1177/2053168015589624.
Kim, J. -O., & Mueller, C. W. (1978b). Factor Analysis: Statistical methods and practical issues.
Newbury Park, CA: Sage Publications
Kim, So Young. 2007. “Openness, External Risk, and Volatility: Implications for the
Compensation Hypothesis.” International Organization 61 (1):181–216.
Kletzer, Lori. 2001. Job Loss from Imports: Measuring the Costs. Institute for International
Economics: Washington, DC.
Lewis-Beck, Michael S., and Charles Tien. 2008. "Forecasting Presidential Elections: When to
Change the Model." International Journal of Forecasting 24 (2): 227-36.
Lewis-Beck, Michael S., and Mary Stegmaier. 2000. "Economic Determinants of Electoral
Outcomes." Annual Review of Political Science 3 : 183-219.
Mackuen, Michael B., Robert S. Erikson, and James A. Stimson. 1992. "Peasants Or Bankers the American Electorate and the United-States-Economy." American Political Science
Review 86 (3): 597-611.
Margalit, Yotam. 2011. “Costly Jobs: Trade-Related Layoffs, Government Compensation, and
Voting in U.S. Elections.” American Political Science Review 105 (1):166–88.
Mayda, Anna Maria, and Dani Rodrik. 2005. “Why are Some People (and Countries) More
Protectionist than Others?” European Economic Review 49:1393–430.
Melitz, Marc J. 2003. "The Impact of Intra-industry Reallocations and Aggregate Industry
Productivity." Econometrica 71(6):1695-1725
Milner, Helen. 1988. “Trading Places: Industries for Free Trade.” World Politics 40 (3):350–76.
Nadeau, R., and Michael S. Lewis-Beck. 2001. "National Economic Voting in U.S. Presidential
Elections." Journal of Politics 63 (1): 159-81.
Norpoth, Helmut. 1996. "Presidents and the Prospective Voter." Journal of Politics 58 (3): 77692.
45
Owen, Erica. 2015. “Exposure to Offshoring and the Politics of Trade Liberalization: Debates
and Votes on Free Trade in the 108th Congress.” Paper Presented at the 2015 American
Political Science Association Convention, San Francisco.
Owen, Erica, and Dennis P. Quinn. 2016. “Does Economic Globalization Influence the U.S.
Policy Mood?: A Study of U.S. Public Sentiment, 1954-2011.” British Journal of
Political Science: In Press. Published online at:
http://dx.doi.org/10.1017/S000712341400009X.
Page, Benjamin I., and Robert Y. Shapiro. 1983. “Effects of Public Opinion on Policy.”
American Political Science Review 77:175–90.
Page, Benjamin L., and Robert Y. Shapiro. 1992. The Rational Public: Fifty Years of Trends in
Americans’ Policy Preferences. Chicago: University of Chicago Press.
Powell, G. Bingham, and Guy D. Whitten. 1993. "A Cross-National Analysis of Economic
Voting - Taking Account of the Political Context." American Journal of Political Science
37 (2): 391-414.
Quinn, Dennis P., and John T. Woolley. 2001. "Democracy and National Economic
Performance: The Preference for Stability.” American Journal of Political Science
45(July): 634-57.
Ramondo, Natalia, Veronica Rappoport, and Kim J. Ruhl. 2013. Horizontal Versus Vertical
Foreign Direct Investment: Evidence from U.S. Multinationals. UC San Diego Working
Paper.
Rehm, Philipp. 2010. “Risk Inequality and the Polarized American Electorate.” British Journal
of Political Science 41 (2):363–87.
Rickard, Stephanie J. 2015. “Compensating the Losers: an Examination of Congressional Votes
on Trade Adjustment Assistance.” International Interactions 41(1):46-60.
Ruhl, Kim J. 2013. An Overview of U.S. Intrafirm-Trade Data Sources. New York University
Working Paper.
Scheve, Kenneth, and Matthew J. Slaughter. 2004. “Economic Insecurity and the Globalization of
Production.” American Journal of Political Science 48(October):662-674.
Scheve, Kenneth, and Matthew J. Slaughter. 2001. “What Determines Individual Trade-Policy
Preferences?” Journal of International Economics 54 (2):267–92.
Shapiro, Robert Y. 2011. “Public Opinion and American Democracy.” Public Opinion Quarterly
75 (5):982–1015.
Stevenson, Randolph. 2001. “The Economy and Policy Preference: A Fundamental Dynamic of
Democratic Politics.” American Journal of Political Science 45 (3):620–33.
46
Stimson, James A., Michael B. MacKuen, and Robert S. Erikson. 1995. “Dynamic
Representation.” American Political Science Review 89:543–65.
Tufte, Edward R. 1978. Political Control of the Economy. Princeton: Princeton University
Press.
Walter, Stephanie. 2010. “Globalization and the Welfare State: Testing the Microfoundations of
the Compensation Hypothesis.” International Studies Quarterly 54 (2):403–26.
Wlezien, Christopher, and Robert S. Erikson. 2004. “The Fundamentals, the Polls, and the
Presidential Vote.” Ps-Political Science & Politics 37 (4): 747-51.
Wright, John R. 2012. “Unemployment and the Democratic Electoral Advantage.” American
Political Science Review 106(4):685-702,
Yeaple, Stephen Ross. 2009. “Firm Heterogeneity and the Structure of U.S. Multinational
Activity.” Journal of International Economics 78 (2):206–15.
Zeile, William J. 2003. Trade in Goods Within Multinational Companies: Survey-Based Data
and Finding for the United States of American. Paper prepared for the Working Party on
Statistics, OECD Committee on Industry and Business Environment, 3-4 November.
47
Appendix Table A1a
Incumb. Growth
Vote
Q12Q15
Share
1
Vote Share
1
.58**
Growth
.68***
.84***
Business Sent.
-.02
-.56**
ΔTradebal
-.09
.16
ΔImports
-.08
-.55**
ΔExports
-.02
-.48*
Inc. Terms
.16
.18
ΔJobs
.83***
.42*
July_gallup
-.8***
ΔUnemployment -.57***
Bus
SentQ15
1
-.43
.05
-.59**
-.27
.3
.73***
-.84***
ΔTradeBal
Q12_15
1
-.53**
.8***
-.42*
.12
.1
.4
ΔImports ΔExports #Incumbent ΔJobs
Q12_15 Q12_153 Party
(Q15Terms
Q1)/Q1
1
.09
-.16
-.1
-.09
-.01
1
-.61***
.05
.05
.53**
1
-.1
-.46*
.14
1
.06
-.03
Gallup
(Q15)
1
-.54**
Appendix Table A1b
Incumb.
Incumb.
Vote
Terms
Share
Vote Share
Inc. Terms
ΔUnemployment
ΔUnemployment
– Standard Dev.
Income
TAA Covered
Traded Services
Non-Traded Ser.
Manufacturing
ΔUnemploy
ΔUnemp ment –
Income
Rate
Standard
Dev.
1
0.01
-0.03***
1
0.04***
1
-0.07***
0.01
0.02***
-0.01**
-0.02***
0.00
0.11***
-0.1***
0.09***
0.02***
0.01
0.01
0.49***
-0.04***
0.11***
-0.16***
-0.04***
0.06***
1
-0.04***
0.16***
-0.12***
-0.07***
0.09***
1
-0.03***
0.14***
0.00
0.00
TAA
Covered
1
-0.11***
-0.05***
0.18***
48
Traded
Services
1
0.27***
-0.38***
NonTraded
Services
1
-0.18***
Manuf.
1
ΔUnempl
1
ΔUnemployment
A2a. Descriptive Statistics – Macro-Level
Mean
Std. Deviation
0.52
2.08
103.3
-0.12
0.48
0.32
1.81
8.59
47.6
-0.28
Vote Share
Inc. Terms
ΔUnemployment
ΔUnemployment – Standard Dev.
Income
TAA Covered
Traded Services
Non-Traded Ser.
Manufacturing
A2b. Descriptive Statistics – County-Level
Mean
Std. Deviation
0.51
1.67
6.13
0.80
9.57
0.01
-0.44
-0.20
0.04
Vote Share
Growth Q12_Q15
Business Sent. Q15
ΔTradebal Q12_Q15
ΔImports Q12_Q15
ΔExports Q12_Q15
Inc. Terms
ΔJobs (Q15-Q1)/Q15
July_gallup Q15
49
N
0.06
2.44
37.9
0.63
0.52
0.47
1.28
2.99
14.17
.81
16
16
15
16
16
16
16
16
16
16
N
0.15
0.75
2.80
0.61
0.21
0.02
0.40
0.41
0.84
27,990
18,660
18,660
18,659
18,659
15,550
18,337
18,035
16,162
Table A3a. Factor Analysis, Macro, National Data
Component
VARIABLE
Bus Sentiment Q15
ΔUnemployment
Growth Q12_Q15
July Gallup Approval
ΔExports
CPI
Incumbent Prior Votes
ΔJobs
Incumbent Party Terms
ΔImports
1
2
0.95
-0.89
0.87
0.74
-0.67
-0.54
3
0.52
0.84
0.57
Uniqueness
0.56
0.56
-0.66
0.06
0.20
0.16
0.18
0.46
0.13
0.15
0.32
0.31
0.95
TS Squared
Loadings
3.89
1.88
1.32
% of Total Variance
47
23
16
Cumulative % of
Variance
47
69
85
Notes: Number of obs=15. Unrotated matrix with Eigen Values > 1.00. Principal Factor Analysis is used.
The factor loading scores are the correlation between the variable and the Factor. Factor loading scores
below |.5| can be considered to be substantively insignificant, and are omitted. The square of the factor
loading score is the size of the variable’s total variance represented by the factor. For example, Business
Sentiment Quarter 15 loads roughly 90% on Factor 1. “Uniqueness” represents the proportion of a
variable’s variance that is not explain by one or more common factors. The higher the 0-1 ‘unique’ score,
the more the variable is measuring a phenomenon different from that being measured in other variables.
Scores above .6 considered to be ‘high’ and a sign that the variable is a reliably different measure from
other variables.
50
Table A3b. Factor Analysis, Micro, County-level Data
Component
VARIABLE
Location Concentration Tradable
Services
1
Uniqueness
-.56
Incumbent Prior Votes
Incumbent Party Terms
Income
Location Concentration Non-Tradeable
Services
Location Concentration Manufacturing
Unemployment Rate
Standard Deviation of Unemployment
0.68
1.00
0.99
0.94
0.90
0.85
0.78
0.78
TS Squared
Loadings
1.07
% of Total Variance
13
Cumulative % of
Variance
13
Notes: Number of observations = 16,089. Unrotated matrix with Eigen Values > 1.00. Principal Factor
Analysis is used. The factor loading scores are the correlation between the variable and the Factor.
Factor loading scores below |.5| can be considered to be substantively insignificant, and are omitted. The
square of the factor loading score is the size of the variable’s total variance represented by the factor.
“Uniqueness” represents the proportion of a variable’s variance that is not explain by one or more
common factors. The higher the 0-1 ‘unique’ score, the more the variable is measuring a phenomenon
different from that being measured in other variables. Scores above .6 considered to be ‘high’ and a sign
that the variable is a reliably different measure from other variables. If TAA covered workers are
included, the indicator does not load on any factor and has a uniqueness score of .94. Because of a large
number of ‘negative’ eigen values in the initial factor estimates, the ‘altdivisor’ option is used to
computer % of total variance explained.
51
Table A4. County-level Determinants of Incumbent Two-Party Vote Shares, 1992-2012
Presidential Elections (Table 1 Replication with Additional County-Level Regressors)
Incumbent Vote Share (lagged)
Unemployment
Δ Unemployment (1-year)
Democratic Incumbent
Democratic Incumbent x Δ Unemployment (1-year)
Unemployment Volatility
Average Pay
Δ Average Pay (1-year)
(1)
1.035***
(0.005)
-0.002***
(0.001)
-0.004***
(0.001)
0.084***
(0.003)
0.006***
(0.002)
-0.014***
(0.002)
0.027*
(0.014)
0.082***
(0.020)
Manufacturing Concentration
(2)
1.034***
(0.005)
-0.002***
(0.001)
-0.004***
(0.001)
0.085***
(0.003)
0.005***
(0.002)
-0.014***
(0.002)
0.033**
(0.015)
0.085***
(0.021)
-0.012***
(0.003)
Non-Tradable Services Concentration
-0.152**
(0.068)
0.510***
(0.145)
0.027
(0.041)
-0.052
(0.037)
-0.164***
(0.055)
0.017**
(0.008)
-0.762***
(0.163)
-0.148**
(0.071)
0.583***
(0.159)
0.032
(0.041)
-0.049
(0.037)
-0.196***
(0.059)
0.014*
(0.008)
-0.821***
(0.174)
0.006
(0.005)
0.021***
(0.005)
-0.162**
(0.069)
0.511***
(0.142)
0.039
(0.039)
-0.047
(0.036)
-0.178***
(0.056)
0.015*
(0.008)
-0.759***
(0.161)
18663
0.951
3111
16167
0.951
2896
17969
0.951
3072
Tradable Services Concentration
Retired
Female
African American
Hispanic
Bachelor's Degree
Population
Constant
Observations
R-squared
Counties
(3)
1.035***
(0.005)
-0.002***
(0.001)
-0.004***
(0.001)
0.084***
(0.003)
0.006***
(0.002)
-0.014***
(0.002)
0.030**
(0.014)
0.083***
(0.020)
(4)
1.034***
(0.005)
-0.002***
(0.001)
-0.004***
(0.001)
0.085***
(0.003)
0.006***
(0.002)
-0.014***
(0.002)
0.034**
(0.015)
0.085***
(0.021)
-0.009***
(0.003)
0.005
(0.006)
0.019***
(0.005)
-0.160**
(0.072)
0.574***
(0.155)
0.042
(0.040)
-0.043
(0.037)
-0.199***
(0.059)
0.013
(0.008)
-0.798***
(0.172)
16095
0.951
2889
(5)
1.014***
(0.003)
-0.002***
(0.000)
-0.001
(0.001)
0.063***
(0.001)
0.003***
(0.001)
-0.007***
(0.001)
-0.012
(0.008)
0.025**
(0.012)
-0.012***
(0.002)
0.003
(0.003)
0.007**
(0.003)
-0.107***
(0.041)
0.150***
(0.056)
-0.084***
(0.026)
-0.088***
(0.022)
-0.315***
(0.036)
0.039***
(0.005)
-0.361***
(0.094)
unweighted
16095
0.927
2889
Note: The dependent variable is the incumbent two-party vote share. Democratic party incumbent election years are
1996, 2000, and 2012; Republican party incumbent election years are 1992, 2004, and 2008. All models include
county and year fixed effects. Unemployment volatility is the standard deviation of the county-level unemployment
rate calculated over the 4-year period including the 3 years prior to the election and the election year. All other
regressors are also measured at the county level, and correspond to the election year. Average pay and the
concentration measures enter as natural logs. The concentration measures are employment location quotients from
52
the U.S. Quarterly Census of Employment and Wages and computed by the U.S. Bureau of Labor Statistics. The
variables Retired, Female, African American, Hispanic, and Bachelor’s Degree enter as a share of the population.
The variable Population is logged. Estimates in columns 1-4 are weighted by population size in 1990. The robust
standard errors (reported in parentheses) are adjusted for clustering at the county-level. * p-value < .1; ** p-value <
.05; *** p-value < .01.
53
Table A5. Unemployment and Democratic Presidential Vote Share
Unemployment
Δ Unemployment (1-year)
Unemployment Volatility
Average Pay
Δ Average Pay (1-year)
Democratic Party Share (lagged)
(1)
Democratic
Party Share
1996-2008
0.001**
(0.000)
0.001***
(0.000)
0.004***
(0.001)
0.037***
(0.008)
-0.010
(0.013)
0.637***
(0.012)
(2)
Democratic
Party Share
1996-2008
0.001***
(0.000)
0.001***
(0.000)
0.000
(0.001)
0.041***
(0.003)
-0.002
(0.012)
0.962***
(0.004)
(3)
Democratic
Party Share
1992-2012
0.002***
(0.000)
-0.000
(0.000)
-0.002***
(0.001)
0.037***
(0.002)
0.018**
(0.009)
0.949***
(0.004)
(4)
Democratic
Party Share
1992-2012
0.003***
(0.000)
-0.001
(0.000)
-0.001
(0.001)
0.040***
(0.006)
-0.007
(0.011)
0.691***
(0.007)
Incumbent Vote Share (lagged)
Democratic Incumbent
Democratic Incumbent x Δ Unemployment (1-year)
(5)
Incumbent
Party Share
1992-2012
-0.002***
(0.000)
-0.001
(0.001)
-0.007***
(0.001)
-0.016**
(0.006)
0.006
(0.011)
(6)
Incumbent
Party Share
1992-2012
-0.004***
(0.000)
0.001***
(0.000)
-0.007***
(0.001)
-0.012*
(0.006)
0.006
(0.011)
1.014***
(0.003)
0.056***
(0.001)
0.002**
(0.001)
1.006***
(0.003)
0.040***
(0.003)
Democratic Incumbent x Unemployment
Constant
Observations
Unit FE
R-squared
Counties
-0.179**
(0.081)
12443
County
0.661
3111
-0.393***
(0.024)
12443
State
0.914
3111
-0.290***
(0.020)
18663
State
0.912
3111
-0.200***
(0.057)
18663
County
0.711
3111
0.103*
(0.060)
18663
County
0.930
3111
0.002***
(0.000)
0.080
(0.059)
18663
County
0.930
3111
Note: The dependent variable in columns 1-4 is Democratic presidential votes as a share of Democratic and
Republican votes; in columns 5-6 the dependent variable is the incumbent two-party vote share. All models include
county (or state) and year fixed effects. Unemployment volatility is the standard deviation of the county-level
unemployment rate calculated over the 4-year period including the 3 years prior to the election and the election year.
All other regressors are also measured at the county level, and correspond to the election year. Average pay and the
concentration measures enter as natural logs. The concentration measures are employment location quotients from
the U.S. Quarterly Census of Employment and Wages and computed by the U.S. Bureau of Labor Statistics. The
variables Retired, Female, African American, Hispanic, and Bachelor’s Degree enter as a share of the population.
The variable Population is logged. The robust standard errors (reported in parentheses) are adjusted for clustering at
the county-level. * p-value < .1; ** p-value < .05; *** p-value < .01.
54
Trade Adjustment Assistance
In Table A.3, we estimate models that include the number of workers filing for trade
adjustment assistance, which Margalit showed were negatively associated with changes in the
incumbent (George W. Bush) vote share between 2000 and 2004. TAA is measured as number of
workers filing for trade adjustment assistance in the 4 years leading up to and including the
election year (as a share of the total workforce). Our models include all presidential elections
between 1996 and 2012.35 The results confirm Margalit’s finding over the longer period of our
study: TAA is associated with decreases in incumbent vote shares. In column 1, following
Margalit, we model the change in incumbent vote shares, and include state fixed effects. The
standard errors clustered at the state level. Model 2 includes county fixed effects and countylevel clustering and the negative coefficient corresponding to TAA workers remains negative and
statistically significant at the 90% level. We find aggregate job losses associated with
globalization, as captured by the TAA variable, reduce incumbent vote share, consistent with
Margalit (2011).
In column 3 of Table 4, we introduce TAA Workers into our preferred specification,
which models the incumbent party vote share, and includes the lagged vote share as an
explanatory variable. We include our main control variables, and we note that TAA Workers
retains a negative coefficient, but is no longer statistically significant.
In column 4 we introduce the manufacturing and services concentration measures.
Unfortunately, due to limitation in the TAA data, the models in Table 4 do not include the 1992
election, which was included in all models reported in Tables 1 and 2. Thus, the results are not
directly comparable to our prior estimates. As before, tradable services concentration enters
35
The TAA data are not available prior to 1994, so we cannot include the 1992 election in the models
reported in this table.
55
positive and highly statistically significant. In columns 3-4, the estimated coefficient
corresponding to Manufacturing concentration is positive and statistically significant, in contrast
to our previous estimates.
Table A.6. Trade Adjustment Assistance and Incumbent Presidential Vote Share
Unemployment
Average Pay
TAA Workers
(1)
(2)
Δ Incumbent Δ Incumbent
Party Share Party Share
State FE,
County FE,
State-level County-level
clustered SE clustered SE
0.001*
-0.001*
(0.001)
(0.001)
0.018***
0.057***
(0.004)
(0.014)
-0.092**
-0.074*
(0.043)
(0.041)
Incumbent Vote Share (lagged)
Δ Unemployment (1-year)
Democratic Incumbent
Democratic Incumbent x Δ Unemployment (1-year)
Unemployment Volatility
Δ Average Pay (1-year)
(3)
Incumbent
Party Share
County FE,
County-level
clustered SE
-0.001*
(0.001)
0.028*
(0.015)
-0.068
(0.041)
1.043***
(0.004)
-0.007***
(0.002)
0.010***
(0.002)
0.010***
(0.002)
-0.010***
(0.002)
0.140***
(0.027)
Manufacturing Concentration
Non-Tradable Services Concentration
Tradable Services Concentration
Constant
Observations
R-squared
Counties
-0.141***
(0.043)
15553
0.4848
50
-0.535***
(0.138)
15553
0.456
3111
-0.274*
(0.151)
15553
0.958
3111
(4)
Incumbent
Party Share
County FE,
County-level
clustered SE
-0.001*
(0.001)
0.026*
(0.016)
-0.063
(0.044)
1.041***
(0.004)
-0.007***
(0.002)
0.010***
(0.002)
0.009***
(0.002)
-0.010***
(0.002)
0.139***
(0.027)
0.008**
(0.004)
0.000
(0.005)
0.021***
(0.007)
-0.246
(0.155)
13516
0.958
2889
Note: The dependent variable in columns 1-2 is the change in the incumbent two-party vote share; in columns 3-4,
the dependent variable is the incumbent two-party vote share. The 1992 election is not included because TAA data
are not available prior to 1994. Model 1 includes state and year fixed effects and the standard errors are adjusted for
state-level clustering. All other models include county and year fixed effects and the standard errors are clustered at
the county level. TAA workers are the total number of workers covered by Trade Adjustment Assistance over the 4year period including the 3 years prior to the election and the election year, as a share of total employed workers in
the county in the election year. Unemployment volatility is the standard deviation of the county-level unemployment
56
rate calculated over the 4-year period including the 3 years prior to the election and the election year. All other
regressors are also measured at the county level, and correspond to the election year. Average pay and the
concentration measures enter as natural logs. The concentration measures are employment location quotients from
the U.S. Quarterly Census of Employment and Wages and computed by the U.S. Bureau of Labor Statistics.
Estimates are weighted by population size in 1990. * p-value < .1; ** p-value < .05; *** p-value < .01.
57
Table Appendix A7. Base Models – Dependent Variable is National Incumbent Party Vote Shares (Two Party), 1952-2012
Model 4
Model 5
Data
Sentiment
Gallup
Prior Incumbent Vote t-1
-0.811***
-0.538***
(0.06)
(0.57)
Growth Q12_15
0.018***
0.021***
(0.005)
(0.005)
ΔTradeBal Q12_15
0.037***
0.032**
(0.009)
(0.012)
51.96%
-0.42
-0.3
3.8%
0.068
0.08
-0.2%
-0.004
-0.003
133
0.0798
ΔImportsQ12_15
ΔExportsQ12_15
BusSentimentQ15
0.0006***
(0.0001)
July Gallup
0.0015***
46%
0.069
(0.0006)
War
Constant
0.818***
0.672***
(0.03)
(0.079)
Incumbent Vote share
2016
58
0.818
0.672
54.1%
53.8%
But, what if we have economic conditions as of 2000?
Growth ‘only’ 3%? An import surge from US $ Appreciation
(ΔTradeBalance -1% of GDP)?
Table Appendix A8. Base Models – Dependent Variable is National Incumbent Party Vote Shares (Two Party), 1952-2012
Prior Incumbent Vote t-1
Growth Q12_15
ΔTradeBal Q12_15
Model 4
Model 5
Data
Sentiment
Gallup
-0.811***
-0.538***
51.96%
-0.42
-0.3
(0.06)
(0.57)
0.018***
0.021***
3.0%
0.053
0.062
(0.005)
(0.005)
0.037***
0.032**
-1.0%
-0.038
-0.033
(0.009)
(0.012)
133
0.0798
ΔImportsQ12_15
ΔExportsQ12_15
BusSentimentQ15
0.0006***
(0.0001)
July Gallup
0.0015***
46%
0.069
(0.0006)
War
Constant
0.818***
0.672***
(0.03)
(0.079)
Incumbent Vote share
2016
59
0.818
0.672
49.2%
49.1%
Data Appendix B. Measures and Data Sources
Economic Growth. Multiple measures of economic growth are used. Quarterly per capita real GDP
growth is taken from the Bureau of Economic Analysis, Survey of Current Business. The growth rate
from Q12 through Q15 is used, as is the growth rate in Q14 (following Abramowitz 2008). The results
are not sensitive to the indicator of growth or its timing. Quarterly growth available only from 1952
onward
Exports and Imports. Exports and Imports are as a percentage of GDP, and are Q12 through Q15, both
in levels and changes. The data come from the Bureau of Economic Analysis, Survey of Current
Business. The data prior to 1952 are also from BEA, but are annual changes.
Consumer Sentiment and Business Conditions. The data from both series are taken from the
University of Michigan/Reuters series. Historical data on Consumer Sentiment is also available from the
St. Louis Federal Reserve (UMCSENT1 series). The Consumer Sentiment data are available from late
1952 on. The Business Conditions series (used in Erikson 2009) is available from 1954 onward.36 The
indicator on the closest date prior to the Presidential election is used.
Inflation, Jobs, and Unemployment. The unemployment series is the unemployment rate of
economically active individuals from the Bureau of Economic Analyses web-site; the CPI is from the
Bureau of Labor Statistics. “Jobs” is percentage change in numbers employed during a Presidential term
measured from January in Q1 to June of the election years, and is from Lewis-Beck and Tien 2004, with
an update by the authors.
July Gallup Poll and Net Candidate Advantage. July Gallup Poll is the Presidential popularity in the
first Gallup Poll in July of an election year, and is from Lewis-Beck and Tien 2008 (with an update by the
author). An alternative used in Abramowitz 2008 is to take the difference between the incumbent’s
36
In the University of Michigan survey, respondents are asked, “And how about a year from now, do you
expect that in the country as a whole business conditions will be better, or worse than they are at present,
or just about the same.” The “relative” score is used, which is (better-worse)+100.
60
favorable and unfavorable ratings. The empirical results are almost identical. Net Candidate Advantage
is incumbent candidate (or party) advantage from the American National Election Survey: items
VCF0403, VCF0407, and VCF0409.
Number of Incumbent Terms. This indicator is the number of consecutive terms an incumbent party
has controlled the Presidency. The variable is suggested by Abramowitz (1988, 2008) and is also known
as the “Time for Change” variable.
61