Populism and the Skill-Content of Globalization
What this paper finds — and why it matters
Layer 1: Overview
This paper investigates how the skill structure of globalization shocks — rather than globalization per se — drives the long-run evolution of populism across countries, making a unified empirical case that what gets imported or who immigrates matters as much as how much.
Research question and motivation. The literature has documented that trade exposure and immigration fuel populist voting, but prior work has studied these channels separately, used narrow time windows, and relied on binary party classifications that cannot capture shifts in populism across the full party landscape. Rodrik’s (2018) widely-cited hypothesis holds that trade shocks drive left-wing populism (as in Latin America) and immigration drives right-wing populism (as in Europe). The authors examine whether this hypothesis survives when skill content is explicitly disaggregated and both channels are studied jointly in a unified long-panel setting.
Data, sample, and empirical strategy. The authors construct a new continuous, time-varying populism score for 3,860 party-election pairs covering 1,206 unique parties across 628 national elections in 55 countries from 1960 to 2018. The score is built from the Manifesto Project Database (MPD) using two dimensions identified in the political-science literature: an anti-establishment stance (AES) and a commitment-to-protect stance (CTP). A two-stage polychoric PCA extracts synthetic indices for each dimension and then combines them into a single populism score. The paper defines populist parties as those scoring more than one standard deviation above the mean (threshold validated by comparison with four external databases — Van Kessel, Swank, PopuList, GPop 1 — with ratios of accurate forecasts ranging from 80 to 91 percent). Two dependent variables are studied: (i) the volume margin of populism, the vote share of classified populist parties, estimated with PPML given many zero observations (about 60 percent of the full sample); and (ii) the mean margin of populism, the vote-weighted average populism score of all parties, estimated with OLS. Globalization regressors are skill-specific: imports of low-skill and high-skill labor-intensive goods (as shares of GDP, sourced from Feenstra et al. 2005 and UN Comtrade) and immigration inflows of low-skill and high-skill workers (from Abel 2018, skill-level imputed from dyadic migrant-stock selection ratios). To address reverse causality — populist governments restrict trade and immigration, biasing OLS downward — the authors implement a gravity-based IV strategy: a zero-stage PPML regression predicts bilateral flows using time-invariant dyadic fixed effects interacted with a post-1990 dummy and origin-country-year fixed effects, then aggregates to the destination level; these predicted flows serve as instruments. For the volume margin, a reduced-form IV approach replaces actual with predicted flows (to avoid the incidental-parameter problem in PPML with fixed effects). For the mean margin, standard 2SLS is used; the Kleibergen-Paap F-statistic is around 10–12, reasonable given four instruments.
Main quantitative findings. (All claims below are with country and year fixed effects throughout; IV results reinforce baseline OLS/PPML results.)
Low-skill labor-intensive imports raise total and right-wing populism along both the volume margin and the mean margin. In the OLS mean-margin specification the coefficient on low-skill imports is approximately 4, implying a 1 percentage-point increase in the import-to-GDP ratio for low-skill goods is associated with a 0.04 increase in the mean margin of populism (scaled in standard deviations of the populism score). The 2SLS coefficient on the total mean margin is approximately 5.0 (significant at 5%), and on the right-wing mean margin approximately 4.1 (significant at 5%). For the volume margin, the reduced-form IV coefficient on low-skill imports is 0.91 (significant at 10%) for total and 1.82 (significant at 5%) for right-wing populism. These effects are larger by a factor of approximately 1.3 when IV is used relative to OLS/PPML, consistent with downward bias from reverse causality. Low-skill imports do not significantly affect left-wing populism in baseline estimates; a left-wing response cannot be ruled out during severe crises, when shocks are persistent, or among EU countries specifically.
High-skill labor-intensive imports reduce the volume of populism, especially right-wing populism. In the reduced-form IV specification the coefficient on high-skill imports is -1.22 (significant at 10%) for total volume and -2.14 (significant at 5%) for right-wing volume. The mean-margin effect of high-skill imports is insignificant.
Low-skill immigration induces a transfer of votes from left-wing to right-wing populist parties, leaving total volume and the mean margin unchanged. The baseline PPML coefficient on low-skill immigration is 1.52 (significant at 1%) for right-wing volume and -1.78 (significant at 1%) for left-wing volume. In the reduced-form IV the right-wing volume coefficient is 1.97 (significant at 1%) and the left-wing coefficient is -1.70 (significant at 10%). The mean margin of total populism is not significantly affected by low-skill immigration in any specification.
High-skill immigration reduces the volume of right-wing populism (PPML coefficient -1.32, significant at 1%; IV coefficient -2.02, significant at 5%) and generates a weak substitution toward left-wing populism in the baseline.
Descriptive findings: populism fluctuated since the 1960s, peaking after major economic crises (the oil shocks of the 1970s, deep crises of the 1990s, and after 2008). Right-wing populism reached an all-time high in the EU after 2005. The share of elections with at least one right-wing populist party rose from about 5 percent to more than 50 percent in EU member states over the study period.
Mechanisms. Decomposing the volume margin into extensive (number of populist parties) and intensive (average vote share per party) sub-margins reveals that: the trade channel operates primarily through the intensive margin (existing populist parties gaining more votes); the immigration channel operates through the extensive margin (new right-wing populist parties with moderate scores entering parliament). Low-skill trade and immigration never increase the populism score of parties that have never been classified as populist, indicating that globalization shifts the composition of the party system rather than radicalizing mainstream parties.
Amplifiers and heterogeneity. The right-wing populism response to low-skill imports is amplified during periods of de-industrialization and when internet coverage is high. Diversity in the origin mix of imported goods dampens the right-wing response. The populism response to low-skill immigration is not amplified by cultural distance between natives and immigrants; if anything, high cultural distance slightly reduces the centrist and left-wing populist responses. The effects on volume margin are primarily driven by EU28 countries.
Scope conditions and caveats. Analysis is at the country level; party-level repositioning dynamics are left for further research. The unified trade-plus-immigration framework is new, but the long panel setting, unbalanced sample, and aggregate data impose limits on identifying specific mechanisms. The finding that globalization does not affect never-populist parties’ scores limits concerns about contamination through party contagion in the short run. These results only partially confirm Rodrik’s (2018) hypothesis — left-wing populism is not robustly driven by trade shocks at the aggregate level, and trade’s effects are not confined to non-European contexts.
Layer 2: Deep Dive
What is the identification strategy and what are the main threats to it?
The identification relies on a two-stage approach. In the first stage (zero-stage gravity model), the authors predict bilateral flows of low- and high-skill goods and migrants using (i) time-invariant dyadic fixed effects interacted with a post-1990 structural-break dummy and (ii) origin-country-year fixed effects capturing time-varying push factors at the source. Critically, destination-country-time characteristics are excluded from the zero-stage, so the predicted aggregated flows capture only supply-side variation and bilateral connectivity — not demand-side populism dynamics in the destination. These predicted flows are then used as instruments. For the mean margin, standard 2SLS is implemented; for the volume margin, a reduced-form IV approach replaces actual flows with predicted flows to avoid the incidental-parameter problem in a PPML model with many fixed effects. The main threats are: (1) correlated origin shocks — if a push shock in origin country j simultaneously triggers populism in destination i through channels other than trade/migration (e.g., financial contagion), the exclusion restriction is violated; the authors cannot fully rule this out but note that including year fixed effects absorbs common global shocks; (2) the post-1990 structural break is used as an additional source of variation for bilateral dyadic ties, but the Berlin Wall dummy simultaneously captures many unobserved structural changes; (3) imputation of the skill structure of migration flows from census-round selection ratios (1990, 2000, 2010) introduces measurement error, though the authors show robustness to using only the year-2000 ratio; (4) Kleibergen-Paap F-statistics are around 10–12 when all four endogenous variables are instrumented simultaneously, which is modest; the authors show values are substantially larger when instrumenting one or two variables at a time.
How are trade and immigration distinguished empirically, and how is the skill content measured?
Trade data come from Feenstra et al. (2005) for 1962–2000 and UN Comtrade for 2001–2015. Product categories at the SITC 3-digit level are classified by skill and technology intensity following the Trade and Development Report (2002), yielding five categories: primary commodities, labor-intensive/resource-based, and manufacturing with low-, medium-, and high-skill labor intensity. The baseline uses only the low-skill and high-skill manufacturing ends; medium-skill goods are tested in robustness (their inclusion causes collinearity that kills volume-margin significance while preserving mean-margin results). Migration data come from Abel (2018) — five-year bilateral migration flow estimates interpolated to annual frequency. The skill level of migration flows is imputed by applying census-round skill-selection ratios (ratio of college graduates in the dyadic migrant stock to the native pre-migration population, from the closest available census round of 1990, 2000, or 2010) to the interpolated flows. Both trade and immigration variables enter as percentages — imports as share of GDP, immigration as share of destination population — averaged over the election year and the preceding year.
What is the difference between the volume margin and the mean margin of populism, and why does it matter?
The volume margin is the aggregate vote share of parties classified as populist (using a binary threshold of one standard deviation above mean in the populism score); it equals zero in elections with no populist party (about 60 percent of observations). The mean margin is the vote-weighted average populism score of all parties — populist and non-populist alike — so it is always defined and continuous. The mean margin captures the average ideological ’exposure’ of voters to populist ideas in a given election, including the spillover of populist ideas into mainstream parties. The distinction matters because globalization can affect the political landscape through multiple channels: it may shift votes toward existing populist parties (intensive margin of the volume margin), it may encourage new populist parties to enter (extensive margin), or it may shift the policy positions of all parties toward more populist stances (captured by the mean margin). The paper finds that low-skill trade raises both margins, but through different mechanisms — the volume effect operates through the intensive margin while the mean-margin effect partly reflects score increases among centrist populist parties. Low-skill immigration raises only the volume margin (through extensive-margin changes, not the mean margin).
How is the populism score constructed, and how is it validated?
The score is built from the Manifesto Project Database, which counts quasi-sentences associated with specific political topics as shares of party manifestos. Six MPD variables are selected, grouped into two dimensions: anti-establishment stance (AES — political corruption mentions and anti-pluralism/political authority mentions) and commitment-to-protect stance (CTP — protectionism, internationalism, EU institutions, and nationalization). A polychoric PCA within each dimension extracts the first principal component (by Kaiser criterion — eigenvalues above one). The two synthetic indices are then combined into a single populism score by equal weighting. A party is classified as populist if its score exceeds one standard deviation above the mean. This threshold maximizes the partial correlation with three of four external databases and maximizes accurate-forecast rates across all four databases. Probit regressions of existing binary classifications (Van Kessel 2015, Swank 2018, PopuList 2019, GPop 1 2020) on the continuous score yield ratios of accurate forecasts between 80 and 91 percent. OLS correlations with continuous external measures (GPop 2 leader-speech scores, CHES expert survey) are positive and significant. Unsupervised k-means clustering on the (AES, CTP) space confirms that parties above the one-SD threshold cluster distinctly in a well-separated region of the two-dimensional space. Extended scores using more MPD variables do not improve fit, confirming parsimony.
What heterogeneity across left-wing and right-wing populism is documented?
The paper systematically decomposes results by political orientation (terciles of the RILE left-right index from MPD). Key heterogeneities: (1) Low-skill imports raise total and right-wing populism but not left-wing populism along the volume margin — this holds in baseline PPML and reduced-form IV. The mean-margin result is also concentrated in total and right-wing. (2) Low-skill immigration shifts votes from left-wing to right-wing populism (with opposing-sign PPML coefficients of 1.52 and -1.78, both significant at 1%), leaving total populism unchanged. High-skill immigration reverses this — it reduces right-wing and weakly increases left-wing populism. (3) High-skill imports reduce right-wing populism particularly (PPML -1.30, IV -2.14) and weakly shift votes toward left-wing populism. (4) Descriptively, the average populism score of right-wing populist parties increased since 2005 and reached 1.7 (2.1 standard deviations) in 2018, while left-wing populist parties’ average score declined to 1.4 (1.75 standard deviations) — for the first time since the 1960s, radical-right populism is more intense than radical-left. (5) The volume-margin effects of globalization are primarily driven by EU28 countries. Among non-EU countries or when Latin America is excluded, results are directionally preserved but sometimes less precisely estimated.
What robustness checks are run?
The authors conduct an extensive battery documented in Appendix D: (1) Lag structure — the globalization variables are redefined using flows at t, t-1, t-2, average of t and t-1 (baseline), and the sum between elections; results on immigration are robust across lags; trade significance holds except at very short (election year) or very long (between elections) windows. (2) Populism threshold — results are preserved at the lax (0.9 SD) threshold and mostly preserved at the strict (1.1 SD) threshold, though some become insignificant when well-known parties like Syriza, M5S, and La France Insoumise exit the classification. (3) Skill imputation for immigration — using only year-2000 selection ratios yields similar results; interactions with migrant-stock quartile dummies are mostly insignificant. (4) Skill content of imports — adding labor-intensive and medium-skill imports does not disturb the baseline; collinearity from medium-skill imports kills volume-margin trade significance. (5) Origin-country income level — positive populism responses are concentrated in flows from low-income countries on the volume margin, but the mean-margin positive response is more driven by North-North movements. (6) Sub-samples — results are not driven by post-1990 years alone (interaction with post-1990 dummy attenuates but does not eliminate effects), not by Latin American countries (exclusion leaves results unchanged), and not by the unbalanced panel structure (restricting to countries present since 1970 confirms results). (7) Turnout — globalization variables do not significantly predict turnout, and results are robust to controlling for turnout. (8) Electoral system — results hold when controlling for electoral system; proportional representation systems show a significant effect of low-skill imports on left-wing populism volume. (9) Exports and emigration — including skill-specific export and emigration flows does not substantially alter the main coefficients; export and emigration effects are less significant and robust than import and immigration effects. (10) Vote-share normalization — results are robust to normalizing vote shares to sum to 100 percent.
How does this paper relate to and differ from closely related prior work, especially Autor et al. (2020) and the immigration literature?
Autor, Dorn, Hanson, and Majlesi (2020) study the electoral consequences of the China trade shock in the US, documenting polarization effects concentrated in a specific trade shock and a narrow time frame. The present paper extends this by: (1) spanning 60 years and 55 countries (vs. US-focused short panels); (2) studying trade and immigration jointly in one specification; (3) using continuous populism scores rather than party platforms; (4) distinguishing left- vs. right-wing populism responses; (5) examining skill content rather than origin-country GDP growth. On immigration, Edo et al. (2019) and Moriconi et al. (2022, 2019) document that the skill structure of immigration matters for voting — high-skill immigration reduces far-right votes while low-skill immigration raises them. The present paper confirms these findings in a much larger multi-decade panel and adds the novel result that low-skill immigration does not affect total populism but merely shuffles votes between left-wing and right-wing populism. On Rodrik’s (2018) taxonomy, the paper only partially confirms his hypothesis: left-wing populism is not robustly driven by trade shocks in the cross-country aggregate (only under specific amplifying conditions), and trade’s effects are not confined to non-European settings. A key novelty vs. the entire prior literature is the simultaneous inclusion of skill-specific trade and immigration flows — no prior cross-country long-panel study had done this.
What are the policy implications and their scope conditions?
The skill-content result implies that globalization’s effect on populism depends critically on whether economic integration predominantly involves low-skill or high-skill goods and workers. Policies that shift the composition of globalization toward high-skill activities — skill-upgrading policies, investment in education and retraining, managed migration policies that attract high-skill workers — could mechanically reduce populist pressures. The finding that low-skill immigration transfers votes from left to right without increasing total populism has a nuanced implication: reducing low-skill immigration may primarily benefit left-wing parties at the expense of right-wing ones rather than reducing aggregate political instability. The amplification by de-industrialization and internet access suggests that the populist dividend of adverse trade shocks is largest precisely when affected regions are also losing manufacturing jobs and when social media spreads grievance discourse. The attenuation by diversity in imported goods suggests that more geographically diversified trade may reduce the cultural-threat salience of any single origin. Scope conditions: the volume-margin effects are largely driven by EU28 countries, so the quantitative magnitudes may not generalize to other institutional contexts with different electoral systems; the analysis is at the country level and abstracts from regional labor-market dynamics; party-level repositioning of mainstream parties is not modeled.
How does the paper handle the measurement challenge of comparing populism scores across countries and time?
This is a central methodological concern. The authors use party manifestos, which are available consistently across the 55 countries and the full 1960–2018 period in the Manifesto Project Database, allowing a principled content-based scoring without relying on expert surveys (which are available only for limited periods) or dichotomous external classifications (which are time-invariant in some datasets and country-limited in others). The two-stage PCA with polychoric principal components ensures that the dimensions are extracted from the structure of the data without imposing cardinal interpretations on ordinal quasi-sentence counts. The populism score has zero mean by construction with a standard deviation of 0.81, making cross-country and cross-time comparisons meaningful within the sample. The authors validate cross-country comparability by showing that the GPop 1 classification (which spans 1960–2018 for 36 countries) is well predicted by the score even though the score was not calibrated to that dataset specifically. An unsupervised clustering algorithm (k-means on the two dimensions) independently recovers the same set of parties as those above the one-SD threshold, without using any external label. The authors acknowledge that deliberate exclusion of immigration and multiculturalism variables from the score construction prevents mechanical correlation between the populism measure and the globalization regressors, which is an important design choice for the causal analysis.
What are the trends in the right-left decomposition of populism over the study period?
Descriptively (Section 3): the number of left-wing populist parties (as counted by the extensive margin) increased more than right-wing populist parties in the most recent period, partly because centrist parties are entering the populist bucket. However, the vote share gains (intensive margin) are dominated by right-wing populist parties. The share of elections with at least one left-wing populist party rose from about 15 to 30 percent globally over the study period. The share of elections with at least one right-wing populist party rose from about 5 to more than 50 percent in the EU and from about 10 to 25 percent in the rest of the world. The average populism score of right-wing populist parties increased since 2005, reaching 1.7 (about 2.1 standard deviations) in 2018, while the average score of left-wing populist parties declined to 1.4 (about 1.75 standard deviations). This means that for the first time since the 1960s, right-wing populist parties are on average more populist (by their own score) than left-wing populist parties. The gap between populist and non-populist parties’ average scores has widened since 2008, consistent with the within-country Theil inequality increase after the financial crisis.
Key Concepts
Volume margin of populism: The aggregate vote share obtained by parties classified as populist (those with a populism score exceeding one standard deviation above the mean). Estimated with PPML given the large share of zero observations (about 60 percent of the sample). Captures whether populist parties win more votes.
Mean margin of populism: The vote-weighted average populism score of all parties that obtained at least one seat in an election, regardless of whether they are classified as populist. Captures the average ideological ’exposure’ of voters to populist ideas, including spillovers into mainstream parties. Estimated with OLS.
Anti-establishment stance (AES): One of two dimensions underlying the paper’s populism score. Measured from Manifesto Project Database quasi-sentences on political corruption and anti-pluralism (political authority), capturing the core populist premise that the people are virtuous and the ruling class corrupt, leaving no room for pluralism or minority protection.
Commitment-to-protect stance (CTP): The second dimension underlying the populism score. Measured from Manifesto Project Database quasi-sentences on protectionism, internationalism, EU institutions, and nationalization, capturing populists’ claim to shield ’the people’ from external or alien economic and cultural threats.
Skill-content of globalization: The decomposition of import flows into goods intensive in low-skill vs. high-skill labor (using the SITC 3-digit classification from the Trade and Development Report 2002), and of immigration inflows into low-skill and high-skill workers (using dyadic skill-selection ratios from census rounds). The key empirical innovation of the paper: it is the skill content, not the size, of globalization flows that determines the direction and ideological valence of populist responses.
Gravity-based IV strategy: An instrumentation approach that predicts bilateral skill-specific flows of goods and migrants using a zero-stage PPML regression with time-invariant dyadic fixed effects (interacted with a post-1990 structural-break dummy) and origin-country-year fixed effects, then aggregates predicted flows to the destination level. Excludes destination-country-time characteristics to purge reverse causality (populist governments restricting trade and immigration) and omitted variable bias.
Extensive vs. intensive margin of the volume margin: The decomposition of the total vote share for populist parties into the number of populist parties running (extensive margin) and the average vote share per populist party (intensive margin). Low-skill imports primarily affect the intensive margin (existing populist parties gain more votes); low-skill immigration primarily affects the extensive margin (new right-wing populist parties enter parliament).
Vote-transfer mechanism of low-skill immigration: The paper’s finding that low-skill immigration reallocates votes between left-wing and right-wing populist parties without changing total populism. The authors interpret this as low-skill immigration enabling new right-wing populist parties with moderate populism scores to gain at least one seat in parliament (an extensive-margin effect), while simultaneously reducing the vote share and/or number of left-wing populist parties.