Trade and Labor Market Outcome Data

We estimate the effects of the German trade shock on economic and political outcomes in England and Wales, using parliamentary constituencies as the unit of analysis.Footnote ³ We measure the change in import penetration at the local level using the empirical strategy developed by Autor, Dorn, and Hanson (Reference Autor, Dorn and Hanson2013)—that is, we compute

$$ \Delta {\mathrm{IPW}}_{it}=\sum \limits_j^n\frac{L_{ij}}{L_i}\frac{\Delta {M}_{jt}}{L_j}, $$

where $ {L}_{ij}/{L}_i $ is the share of employment in industry j in constituency i in the base year, 1881, and $ \Delta {M}_{jt}/{L}_j $ is the change in imports for industry j in year t, relative to total employment in that industry in 1881. We index the change in imports relative to different years in different specifications: In long first-difference specifications, $ \Delta {M}_{jt} $ is the change in imports relative to the previous period. In other models that use constituency fixed effects, we index relative to the first year used in the analysis. We Winsorize the industry-level change in imports per worker at plus or minus 500 pounds per worker, equivalent to the 97th percentile.

We use the full-count 1881 census of England and Wales (Minnesota Population Center 2020; Schürer and Higgs Reference Schürer and Higgs2014) to compute the sizes and distributions of different industries and combine this with product-level data on imports from the Annual Statement of the Trade of the United Kingdom. Occupational categories in the nineteenth-century census contain a high degree of specificity about industries, distinguishing, for instance, “Ironfounders” from “Iron clasp, buckle, and hinge makers” and “Brass founders.” We group occupational categories and product-level import data into 94 industries, with the goal of identifying the finest level of variation present in both the trade statistics over the total period and the occupational categories.

British parliamentary constituencies do not coincide with administrative units, which has prevented scholars from computing economic variables at the constituency level. We resolve this problem by allocating parishes—the finest level of aggregation in the census—to constituencies. For the 1881 census, we use crosswalk files constructed by Jusko (Reference Jusko2017), who manually assigned parishes to constituencies based on contemporary reports by the boundary commission and maps. For other years, we first link the census data to a consistent GIS based on parishes in the 1851 census (Satchell et al. Reference Satchell, Kitson, Newton, Shaw-Taylor and Wrigley2016), using crosswalk files constructed by Day (Reference Day2016). We then assign parishes to constituencies using shapefiles from the Great Britain Historical GIS Project (2004). Where parishes fall into multiple constituencies, we weight the fraction assigned to each constituency by the fraction of the parish falling into that constituency multiplied by the relative population density of the constituency.

We compute two measures of the economic effects of the trade shock—the percentage of vagrants and the percentage employed in unskilled occupations—at the constituency level, using full-count data from the 1881, 1891, 1901, and 1911 censuses. We classify vagrants as those whose occupation was listed as “No specified occupation—vagrants, unemployed.” This measure plausibly captures labor-market disruption in the form of increased unemployment and the unemployed migrating in search of work. Using the limited time-series data collected by Poor Law administrators, Boyer (Reference Boyer2019, 111–2) finds that rates of vagrancy and unemployment closely tracked one another.

We classify unskilled occupations using the Seventy-Fourth Annual Report of the Registrar General, 1913, which allocated census occupations to eight social classes. The percentage of people in occupations in class 5 (“occupations including mainly unskilled men” [xli]) has been used in the historical geography literature to measure poverty at the local level (Gregory, Dorling, and Southall Reference Gregory, Dorling and Southall2001). These occupations are primarily various forms of unskilled laborers such as “shipyard labourers,” “navvies,” “bill posters,” and workers in “scavenging and disposal of refuse.” The fraction employed in unskilled jobs would plausibly increase in response to import competition if there was a reduction in higher-skilled employment, leading unemployed skilled workers to take on casual labor.

In many of our regression specifications, we control for 1881 manufacturing employment interacted with year dummies in order to separate the effects of the German trade shock from time-variant effects related to manufacturing. We compute this measure using the fraction of people employed in secondary occupations—those in which raw materials were converted into finished products—according to the classification system developed by Wrigley (Reference Wrigley2010) and Bennett et al. (Reference Bennett, Smith, van Lieshout and Newton2017). Figure 4 shows the geographic distribution of import competition in 1910, with and without this control.

Figure 4.

Geographic Distribution of Change in German Imports per Worker, 1885–1910

Election Data

Our primary measure of the political effects of import competition is the share of the vote won by Conservative and Unionist parliamentary candidates. This variable captures the main left–right division in British politics over this period. The Labour Party only contested elections after 1900 and did so in an electoral pact with the Liberal Party. We use data from Eggers and Spirling (Reference Eggers and Spirling2014) and compute the share of the vote won by different parties in the eight general elections from 1885 to 1910. Constituency boundaries and the electoral franchise were consistent over this period. The franchise was also relatively broad: around two-thirds of adult men could vote. Exclusion was somewhat arbitrary, based primarily on residency criteria, leading one historian to conclude that “the overall occupational structure [of the franchise] does not differ vastly from what one would have expected from a fully inclusive franchise” (Brodie Reference Brodie2004, 52). Table 1 reports descriptive statistics for the economic and political variables.

Table 1.

Summary Statistics

Newspaper Measures of Local Concerns

We use data from the British Newspaper Archive to estimate the prevalence of different local concerns. The British Newspaper Archive is a project that seeks to digitize the British Library’s extensive historical newspaper collections. Over the 1885–1910 period, it contains text for 480 newspapers, which we geocode and link to parliamentary constituencies.Footnote ⁴ We compute the number of references to specific terms made in a given year by a given newspaper, divide by the number of issues of the newspaper in the British Newspaper Archive in that year, and then subtract the mean and divide by the standard deviation of that variable to aid interpretation. We use newspaper fixed effects in all specifications to control for time-invariant linguistic or topical features of specific newspapers.

Our intuition in using these measures is that if an issue became more prevalent in a given constituency in a given year, one would expect newspapers to devote greater attention to it. Although newspapers might reflect the opinions of their owners and editors rather than those of their readers, theoretical and empirical studies of media bias suggest that newspapers tend to cater to their readers’ views and concerns (Gentzkow and Shapiro Reference Gentzkow and Shapiro2010). Incentives for newspapers to provide more representative opinion are stronger when demand for media and potential advertising revenues are high, so the returns to providing popular news that will appeal to readers is large (Petrova Reference Petrova2011). These theoretical predictions should apply during the period we study: by the 1880s the removal of newspaper taxes and developments in printing technology had made possible a business model for newspapers based on large circulations and advertising revenues (Lee Reference Lee1976).Footnote ⁵

Other Data

We additionally use an unpublished dataset of parliamentary candidates’ manifestos compiled by Laura Bronner and Daniel Ziblatt. From the late-nineteenth century onward, candidates could distribute one leaflet for free via Royal Mail to inform voters of their views. Bronner and Ziblatt collected and digitized manifestos for all parliamentary candidates in general elections from 1892 to 1910. We use these data in a way similar to that for the newspaper data. We divide the number of references to a given term by the number of words in the manifesto and then standardize that measure.Footnote ⁶

Model Specification

Our estimation strategy examines the effects of within-constituency changes in imports per worker on a set of outcome variables: labor market distress, voting for particular parties, and the prevalence of different issues in newspapers and campaign manifestos. We use two main model specifications. For the economic outcome variables, using decadal data from the census, we estimate regressions of the form

$$ \Delta {Y}_{it}\hskip1.5pt =\hskip2pt {\beta}_1\Delta {\mathrm{IPW}}_{it}+{\mathbf{X}}_{it}^{\prime }{\beta}_2+{\gamma}_t+{\varepsilon}_{it}, $$

where $ \Delta {Y}_{it} $ is the change in a given outcome variable in constituency i relative to the previous census, $ \Delta {IPW}_{it} $ is the change in the trade shock measure relative to the previous census, $ {\gamma}_t $ is a year fixed effect, and $ {\mathbf{X}}_{it}^{\prime } $ is a vector of controls. We estimate these models in stacked first differences, consistent with other economic studies of the effects of trade shocks (Autor, Dorn, and Hanson Reference Autor, Dorn and Hanson2013).

We estimate the majority of regressions with political dependent variables in levels. This practice is consistent with empirical studies of the effects of trade shocks on voting (Colantone and Stanig Reference Colantone and Stanig2018b; Reference Colantone and Stanig2018c; Feigenbaum and Hall Reference Feigenbaum and Hall2015). We are interested in the effects of long-term changes in import penetration, not the effects of year-to-year variation. This focus makes 10-year census-to-census first-differences appropriate, but election-to-election first-differences inappropriate, given the short gap between some elections in our sample.Footnote ⁷ We estimate regressions of the form

$$ {Y}_{it}\hskip1.5pt =\hskip2pt {\beta}_1\Delta {\mathrm{IPW}}_{it}+{\mathbf{X}}_{it}^{\prime }{\beta}_2+{\gamma}_t+{\delta}_i+{\varepsilon}_{it}, $$

where $ {Y}_{it} $ is some political outcome variable, $ \Delta {\mathrm{IPW}}_{it} $ is the change in imports per worker for constituency i in year t relative to the start year, $ {\mathbf{X}}_{it}^{\prime } $ is a vector of controls, $ {\gamma}_t $ is a year fixed effect, and $ {\delta}_i $ a constituency fixed effect. Note that the differenced dependent variables and constituency fixed effects account for time-invariant confounders.

Identification

Estimates from these regressions can be interpreted causally within the difference-in-differences framework. Although our measure of imports per worker is computed according to a shift-share formula, our identification strategy does not rely on the use of exogenous variation in the form of exports from Germany to a third party. Goldsmith-Pinkham, Sorkin, and Swift (Reference Goldsmith-Pinkham, Sorkin and Swift2020) argue that shift-share designs rely on the assumption that the initial shares used to construct the shift-share variable are exogenous to the outcome variable. This assumption is more plausibly satisfied in research designs like ours, which control for unit fixed effects and for which the equivalent identifying assumption is that these shares are exogenous to changes in the outcome variables. Thus for our estimates to be interpreted causally, one must believe that, apart from the effects of changes in imports, constituencies with greater employment in affected industries would have followed trajectories similar to those of constituencies with less employment in those industries.

We address this assumption in three ways. First, we include controls for initial manufacturing interacted with year dummies across all our specifications. We thus allow more industrial constituencies to follow different nonlinear trajectories to less industrial constituencies and implicitly compare constituencies affected by German imports in a given year with less-affected industrial constituencies. Second, we follow the procedure outlined by Goldsmith-Pinkham, Sorkin, and Swift (Reference Goldsmith-Pinkham, Sorkin and Swift2020) to identify the industry-year combinations for which our estimated coefficients are most sensitive to misspecification and show that our results are robust to controlling for these initial industry shares interacted with year dummies and to controlling for the first three principal components of the 1881 industry shares interacted with year dummies. These robustness checks suggest it is unlikely that differential trends relating to specific industries or clusters of industries are driving our results. Third, we employ traditional difference-in-differences robustness tests: controlling for constituency time trends and checking that leads of the trade shock variable do not affect outcomes.

The shift-share design is important to our empirical strategy as an accounting method and as a way to avoid bias from posttreatment economic changes. It is important to emphasize that our primary use of the Autor, Dorn, and Hanson (Reference Autor, Dorn and Hanson2013) trade shock formula is simply to measure the incidence of import competition at the local level. Using the 1881 industry shares, as opposed to subsequent shares, has the additional benefit of separating our measure of exposure to German imports from changes in local economies that may themselves be affected by German imports.

Standard Errors

We cluster standard errors at the county level rather than at the more granular constituency level. This is a conservative choice to account for potential spatial autocorrelation in the error term due to local spillover effects. In the Extended Appendix, we reestimate all regressions in the article using the aggregation method recommended by Borusyak, Hull, and Jaravel (Reference Borusyak, Hull and Jaravel2022). Adão, Kolesár, and Morales (Reference Adão, Kolesár and Morales2019) note that in shift-share designs, conventional standard errors fail to account for correlation in the error structure between units with similar shares. Aggregating the relevant variables to the industry level gives “exposure robust” standard errors that account for errors correlated across units with similar shares in the same way that one can avoid problems with within-cluster correlations by aggregating to the level of the cluster.

Voter Concern about German Trade

Before directly studying these mechanisms, we examine whether the trade shock increased attention to trade in newspapers. A theory in which the direct economic effects alone accounted for the political changes—unemployed voters supported the welfare state—would not require voters to necessarily pay more attention to trade. However, increased attention to trade is an important part of mechanisms in which trade affected beliefs about the risk of economic upheaval—perhaps by tapping into fears about national decline and global competition—and the moral desert of the unemployed, perhaps because foreign industrialization is unrelated to the effort of domestic workers. We regress a standardized measure of the per-issue references to different trade-related terms on $ \Delta \mathrm{IPW} $ , with newspaper and year fixed effects and time-varying manufacturing controls. Table 4 shows the results of these regressions. Over the whole period, import competition was associated with increased references in newspapers to trade and imports. The coefficient magnitudes suggest that a 1-pound increase in imports per worker was associated with a 0.1-standard-deviation increase in coverage. The effect is driven by the 1900–1910 period—models (3), (4), (7), and (8)—when we find trade had a political effect.

Table 4.

Effects of Import Competition on Newspaper References to Trade

Note: Newspaper-level regressions. Dependent variable is number of uses of specified term per newspaper issue, standardized. All models include newspaper and year fixed effects. For newspapers in cities, ΔIPW is calculated at the city-level, not the constituency-level. Standard errors clustered by county in parentheses. Table F-3 of the Full Tables Appendix reports full output. *p < 0.10, ^** p < 0.05, ^*** p < 0.01.

Support for the Neo-Welfare State

At the constituency level, the contents of parliamentary candidates’ appeals provide evidence that import competition led to increased demand for the neo-welfare state. We expect that candidates could observe some signal of local demand for particular policies and would emphasize policies that were more popular with voters in their constituencies. If candidates emphasized a policy more in a given area, it was presumably in part because that policy was more popular there. We regress a normalized measure of references to specific policy-related terms in Liberal manifestos on $ \Delta \mathrm{IPW} $ . We focus on three terms, “social reform,” which was used to refer broadly to social policy, “poor law,” the punitive system of welfare that Liberal governments in the 1900s promised to reform, and “labour exchange,” a proposed policy to deal with unemployment due to economic fluctuations. These policies sought to address hardships endured by adult unemployed workers, those affected by import competition. Table 5 shows a consistent positive association between import competition and Liberal candidates mentioning these phrases.

Table 5.

Effect of Local Trade Shocks on References to Social Reform in Liberal Campaign Manifestos

Note: Manifesto-level regressions. Dependent variable is number of uses of specified term relative to total length of manifesto, by Liberal candidates, standardized. All models include constituency and election fixed effects. Standard errors clustered by county in parentheses. Table F-4 of the Full Tables Appendix reports full output. *p < 0.10, ^** p < 0.05, ^*** p < 0.01.

Qualitative newspaper evidence suggests in addition that voters understood that voting Liberal meant voting for the welfare state. Conservative campaigners in January 1910 argued that unemployment “was ‘the’ issue” in the election.Footnote ¹¹ Responding to this Conservative challenge, the Liberal chancellor Lloyd George argued that the Liberals’ proposed budget “makes a larger provision for mitigating the evils of unemployment than any measure ever introduced,” drawing emphasis in particular to labor exchanges and unemployment insurance.Footnote ¹²

There is also evidence from historians and primary sources that import competition led Liberal politicians to prioritize welfare state reforms. Green (Reference Green1995, 230) notes that economic dislocation lent credence to Conservative promises of tariff reform, which promised to “deal with the causes as well as the symptoms of social distress.” Searle (Reference Searle1992) argues that the Liberal party adopted an expanded policy of social reform in response to this electoral threat. In 1910, the Labour MP Philip Snowden argued that supporters of Free Trade had to promote “social reforms which will so improve the conditions of the working classes that they will not be victims of the sophistries and plausibilities of Tariff ‘Reform.’”Footnote ¹³

Changing Attitudes toward the Unemployed

The results presented thus far—that German import competition induced a shift toward the Lib-Lab pact proposing the early welfare state—could be explained by a direct compensation effect (Rodrik Reference Rodrik1998). We also find evidence consistent with a different mechanism in which trade-induced economic turmoil, because it was unrelated to the behavior of those affected, changed beliefs about the moral desert of the unemployed. A new concept of unemployment emerged in this period, and we find evidence that its emergence was linked to the incidence of the trade shock. We also see this concept of unfair misfortune linked to economic fluctuations in Liberal campaign rhetoric.

There is qualitative evidence that a shift in attitudes toward unemployment occurred in early twentieth-century Britain. Beveridge (Reference Beveridge1910), later the architect of the welfare state, argued that unemployment, “the problem of the adjustment of the supply of labour and the demand for labour” (4), was the product of technical change, “fluctuations of industrial activity” (13), and the need for excess labor for industries to hire in boom periods. While acknowledging that the least productive workers may be more likely to be unemployed, Beveridge noted that “The best and most regular of workmen may in a changing world find himself exceptionally unemployed” (142). The prevalence of unemployment was thus distinct from the moral character of the unemployed. The concept of “unemployment” as distinct from vagrancy entered common usage at this time. This sharp break can be seen in Figure 6, which plots references to “unemployment,” “vagrancy,” and “pauperism” in the Times newspaper over the period.

Figure 6.

References to Unemployment, Vagrancy, and Pauperism in the Times

This attitudinal shift was linked to the incidence of the trade shock. Table 6 examines the link between import competition and the use of terms related to this new concept of unemployment in newspapers. It shows the results of newspaper-level regressions in which the dependent variable is the number of references to “unemployment,” “employment,” and the “unemployed,” minus the number of references to “pauper(s),” “pauperism,” “vagrant(s),” and “vagrancy,” standardized. Positive coefficients across specifications suggest that coverage of the economic effects of the trade shock focused on the morally neutral phenomenon of unemployment, not morally charged notions of vagrancy and pauperism. In Appendix D, we employ a more principled approach and use natural language processing methods to identify terms more associated with the new concept of unemployment relative to older notions of pauperism. We find a similar effect of import competition on newspaper usage of terms connected to this new concept of unemployment in Table A-17.Footnote ¹⁴

Table 6.

Effects of Import Competition on Newspaper References to Unemployment, Vagrancy, and Pauperism

Note: Newspaper-level regressions. Dependent variable is the number of references to “unemployed,” “unemployment,” and “employment,” minus the number of references to “vagrants,” “vagrancy,” “pauper,” and “pauperism,” standardized. All models include newspaper and year fixed effects. For newspapers in cities, $ \Delta $ IPW is calculated at the city-level, not the constituency level. Standard errors clustered by county in parentheses. Table F-5 of the Full Tables Appendix reports full output. *p < 0.10, ^** p < 0.05, ^*** p < 0.01.

The new concept of unemployment featured in Liberal arguments for the early welfare state. Campaigning in 1910, Lloyd George claimed, “Unemployment entails great suffering on the part of people who do not deserve it… They are not responsible for the fluctuations in trade. They are purely its victims, and I think that it is a duty of any country within the limits of its resources to see that that suffering is mitigated.”Footnote ¹⁵ The idea that economic volatility meant that people out of work were not responsible for their misfortune was thus part of the argument used to convince voters to support the neo-welfare state.

Alternative Theories of the Effects of Import Competition

Existing research highlights an alternative set of political effects of trade exposure. Scholarship on the China trade shock finds that voters who are negatively affected by increased trade want less trade and turn to protectionist candidates and parties (Che et al. Reference Che, Lu, Pierce, Schott and Tao2016), punish incumbent politicians (Jensen, Quinn, and Weymouth Reference Jensen, Quinn and Weymouth2017), and experience a shift in values toward authoritarianism and xenophobia (see for instance Ballard-Rosa et al. Reference Ballard-Rosa, Malik, Rickard and Scheve2021). A shift toward protectionism cannot explain our results, as the Liberals remained committed to free trade, whereas the Conservatives, who were perceived to be more supportive of tariffs, partially embraced protectionism in 1906 and doubled down on that policy in the 1910 elections. We similarly find no evidence that German imports prompted voters to punish incumbent politicians whether we define incumbency at the individual, party-constituency, or national level (Table A-16).

We do find evidence of increased attention to immigration in both newspapers and the election addresses of Conservative candidates. Anti-immigrant politics could be the manifestation of in-group favoritism and xenophobia caused by import competition, as suggested by scholarship on the China trade shock, or it could reflect voters’ changed economic priorities. Protectionism and immigration restriction can be substitutes: restricting the supply of foreign workers who would compete in the labor market offers politicians a different way of limiting the harm to workers affected by rising imports (Peters Reference Peters2017).

In the 1900s, the British government began to regulate immigration. The Conservative government in 1905 introduced the Aliens Act, which defined categories of undesirable immigrants and gave the state power to exclude them. The act mainly excluded Jewish immigrants from Eastern Europe. In Table A-20, we report a positive effect of import competition on Conservative candidates referring to immigrants, aliens, and Jews. Import competition may have created demand for xenophobic policies, which Conservative candidates sought to capitalize on. We also find a positive effect on coverage of immigration in newspapers (Table A-21).

The net effect of anti-immigrant politics on electoral outcomes during this period is of secondary importance relative to the rise of the welfare state. The Aliens Act was a policy that Conservative MPs campaigned for and a Conservative government implemented, so an increase in anti-immigrant politics cannot explain the shift toward the Liberals. We leave a study of when trade-induced xenophobia is electorally dominant, which would require more than one case study, for future research. A possible explanation for why the electoral effects of xenophobia in early twentieth-century Britain were relatively muted is that the scale of immigration, although historically unprecedented, was relatively small and immigrants were concentrated in a handful of parliamentary constituencies in east London (Pelling Reference Pelling1967).