Costly Jobs February 2011 different unemployment variables is not particularly Iyengar 2000;Pew Research Center 2010).Summary high (ranges from 0.16 to 0.42).Therefore,multi- statistics for all these control variables are provided in collinearity between the unemployment covariates is Appendix Table A1. not a serious concern in this case.25 Income controls are included in the model because previous research finds a strong association between RESULTS economic standing and vote choice (e.g.,Gelman. Park,and Shor 2008;McCarty,Poole,and Rosenthal A central result of the analyses is a consistent and sta- 2006).Measures include the counties'per capita in- tistically significant association between the share of come in the year of the election and the percent change county workers hurt by trade-related job losses and a in per capita income from the previous elections. decline in the level of support for the president between These data are obtained from the Regional Economic the 2000 and 2004 presidential elections.This result Information System at the U.S.Bureau of Economic holds while including a range of controls for the em- Analysis(BEA).26 ployment shifts in the county,demonstrating the strong Demographic controls include measures of the racial electoral impact of trade-related job losses above and composition,the age distribution,and percent of home beyond the electoral effect of changes in the unem- ownership in the counties'population based on the ployment rate.The first set of analyses is presented in U.S.Census.27 Using data from the Religious Congre- Table 2.The dependent variable in the analysis is the gations Membership Study (RCMS),the model also change in Bush's vote share between the two elections controls for the breakdown of religious denominations The measure of trade-related job losses is based on in the county.The analysis also controls for the share the share of all county workers that applied to receive of small employers,small-to-medium enterprises,and TAA certifications in the 4-year period between the large corporations among the county's business estab- elections.As the results indicate,decline in support lishments using figures from the BLS's Business Pat- for President Bush was greater the higher the share of terns Data.28 These controls serve as a proxy for the adversely affected workers in the county ease with which the government could preemptively To examine the sensitivity of the findings,Table 2 assist local businesses to avoid layoffs.29 They also help shows the results from estimating a number of differ- reduce the possibility that the electoral effect associ- ent specifications.In the first column,I estimate Equa- ated with trade-related job losses is instead capturing a tion (1)without controls.The bivariate relationship response to the impact of government policies geared between trade-related layoffs and change in support toward different types of businesses.30 Finally,to exam- for Bush is negative yet statistically insignificant.In ine the potential impact of the media coverage of local column(2),adding to the model controls for level and job losses,some specifications control for the counties' change in the county's income,the coefficient of trade- designated market areas (DMAs).commonly referred related layoffs is more than doubled in size,remains to as"media markets."31 Accounting for the media cov- negative,and is statistically significant.The results also erage is important because despite the spread of cable show that Bush gained votes in counties that experi- and Internet services,Americans still receive a major enced higher rates of growth between the two elec- share of their news from local broadcasts(Gilliam and tions.In column(3),when adding controls for shifts in the counties'employment situation,the magnitude of 25 See Table A3 in the online Appendix for the correlations between the estimated effect of trade-related lavoffs decreases the unemployment variables. and the estimate is marginally significant in statistical 26 In 2001.the BEA redefined the boundaries of counties in Virginia, terms.As expected,the results show that an increase and thus its data do not apply to the exact same geographic units in in a county's unemployment rate is negatively asso- 2000 and 2004.I therefore exclude these counties from analyses that rely on the BEA income data,in which case the total sample is ciated with support for the incumbent.In column(4), reduced to 3,054 counties.See the online Appendix for detail on the which also includes the demographic controls,the point excluded counties. estimate of trade-related job losses is slightly larger 27 http://www.census.gov/support/DataDownload.htm (accessed and the estimate is more precise (as indicated by the October 24.2010). smaller standard error).In column (5),I add state fixed 28 Large corporations are businesses with 500 employees or more. Small businesses have less than 5 employees. effects because these may capture unobserved time- 29 It is presumably easier to provide assistance to a few very large invariant effects at the state level.Inclusion of the fixed local employers than to many small "mom and pop"shops.To deal effects substantially increases the variation explained with the possibility that the government preemptively assisted cer- by the model (from 33%to 64%).and the coefficient tain industries,I also test a model that includes as controls the share of trade-related job losses decreases somewhat.yet re- of county workers employed in each industry.These two estimations mains significant in statistical terms (p=.02).Finally. help address the potential limitation that the prevalence of trade- related job losses may not be entirely exogenous to the electoral to examine whether a longer-term trend of decline in process. counties'support for the Republicans accounts for the 0 During theelection campaign,Bush repeatedly touted his policies observed effects,column (6)also includes a control as pro"small business"and criticized the Democrats'tax policy as a for the voting trend in the preceding election cycle threat to small business owners.See,for example,Elizabeth Olson, (between 1996 and 2000).Controlling for this trend "Courting the Small-business Owner,"New York Times,September 23.2004. only marginally decreases the estimated effect of trade- 31 DMAs are constructed by Nielsen Media Research(Broadcasting related job losses,which remains negative and highly and Cable Yearbook 2003). significant. 174Costly Jobs February 2011 different unemployment variables is not particularly high (ranges from 0.16 to 0.42). Therefore, multi￾collinearity between the unemployment covariates is not a serious concern in this case.25 Income controls are included in the model because previous research finds a strong association between economic standing and vote choice (e.g., Gelman, Park, and Shor 2008; McCarty, Poole, and Rosenthal 2006). Measures include the counties’ per capita in￾come in the year of the election and the percent change in per capita income from the previous elections. These data are obtained from the Regional Economic Information System at the U.S. Bureau of Economic Analysis (BEA).26 Demographic controls include measures of the racial composition, the age distribution, and percent of home ownership in the counties’ population based on the U.S. Census.27 Using data from the Religious Congre￾gations Membership Study (RCMS), the model also controls for the breakdown of religious denominations in the county. The analysis also controls for the share of small employers, small-to-medium enterprises, and large corporations among the county’s business estab￾lishments using figures from the BLS’s Business Pat￾terns Data.28 These controls serve as a proxy for the ease with which the government could preemptively assist local businesses to avoid layoffs.29 They also help reduce the possibility that the electoral effect associ￾ated with trade-related job losses is instead capturing a response to the impact of government policies geared toward different types of businesses.30 Finally, to exam￾ine the potential impact of the media coverage of local job losses, some specifications control for the counties’ designated market areas (DMAs), commonly referred to as “media markets.”31 Accounting for the media cov￾erage is important because despite the spread of cable and Internet services, Americans still receive a major share of their news from local broadcasts (Gilliam and 25 See Table A3 in the online Appendix for the correlations between the unemployment variables. 26 In 2001, the BEA redefined the boundaries of counties in Virginia, and thus its data do not apply to the exact same geographic units in 2000 and 2004. I therefore exclude these counties from analyses that rely on the BEA income data, in which case the total sample is reduced to 3,054 counties. See the online Appendix for detail on the excluded counties. 27 http://www.census.gov/support/DataDownload.htm (accessed October 24, 2010). 28 Large corporations are businesses with 500 employees or more. Small businesses have less than 5 employees. 29 It is presumably easier to provide assistance to a few very large local employers than to many small “mom and pop” shops. To deal with the possibility that the government preemptively assisted cer￾tain industries, I also test a model that includes as controls the share of county workers employed in each industry. These two estimations help address the potential limitation that the prevalence of trade￾related job losses may not be entirely exogenous to the electoral process. 30 During the election campaign, Bush repeatedly touted his policies as pro “small business” and criticized the Democrats’ tax policy as a threat to small business owners. See, for example, Elizabeth Olson, “Courting the Small-business Owner,” New York Times, September 23, 2004. 31 DMAs are constructed by Nielsen Media Research (Broadcasting and Cable Yearbook 2003). Iyengar 2000; Pew Research Center 2010). Summary statistics for all these control variables are provided in Appendix Table A1. RESULTS A central result of the analyses is a consistent and sta￾tistically significant association between the share of county workers hurt by trade-related job losses and a decline in the level of support for the president between the 2000 and 2004 presidential elections. This result holds while including a range of controls for the em￾ployment shifts in the county, demonstrating the strong electoral impact of trade-related job losses above and beyond the electoral effect of changes in the unem￾ployment rate. The first set of analyses is presented in Table 2. The dependent variable in the analysis is the change in Bush’s vote share between the two elections. The measure of trade-related job losses is based on the share of all county workers that applied to receive TAA certifications in the 4-year period between the elections. As the results indicate, decline in support for President Bush was greater the higher the share of adversely affected workers in the county. To examine the sensitivity of the findings, Table 2 shows the results from estimating a number of differ￾ent specifications. In the first column, I estimate Equa￾tion (1) without controls. The bivariate relationship between trade-related layoffs and change in support for Bush is negative yet statistically insignificant. In column (2), adding to the model controls for level and change in the county’s income, the coefficient of trade￾related layoffs is more than doubled in size, remains negative, and is statistically significant. The results also show that Bush gained votes in counties that experi￾enced higher rates of growth between the two elec￾tions. In column (3), when adding controls for shifts in the counties’ employment situation, the magnitude of the estimated effect of trade-related layoffs decreases and the estimate is marginally significant in statistical terms. As expected, the results show that an increase in a county’s unemployment rate is negatively asso￾ciated with support for the incumbent. In column (4), which also includes the demographic controls, the point estimate of trade-related job losses is slightly larger and the estimate is more precise (as indicated by the smaller standard error). In column (5), I add state fixed effects because these may capture unobserved time￾invariant effects at the state level. Inclusion of the fixed effects substantially increases the variation explained by the model (from 33% to 64%), and the coefficient of trade-related job losses decreases somewhat, yet re￾mains significant in statistical terms (p = .02). Finally, to examine whether a longer-term trend of decline in counties’ support for the Republicans accounts for the observed effects, column (6) also includes a control for the voting trend in the preceding election cycle (between 1996 and 2000). Controlling for this trend only marginally decreases the estimated effect of trade￾related job losses, which remains negative and highly significant. 174