📈 economics

Shift-Share Designs in Political Science

This paper introduces shift-share designs to political scientists, reviews their application and recent methodological advancements, critiques the field's frequent reliance on share exogeneity over the underutilized shifter exogeneity framework, and illustrates the latter with new theoretical results to improve causal inference in political science research.

Original authors: Peter Kyungtae Park

Published 2026-03-03

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Peter Kyungtae Park

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: What is a "Shift-Share" Design?

Imagine you are trying to figure out if rain causes mudslides in a neighborhood.

The Problem: You can't just look at one house. Some houses are on a hill (high risk), some are on flat ground (low risk). If it rains, the hill gets muddy, but the flat ground stays dry. If you just compare "muddy houses" to "dry houses," you might think the house itself caused the mud, not the rain.
The Solution (Shift-Share): You need a way to measure how much "rain exposure" each house has, based on its location, and then see if that exposure predicts the mud.

In political science and economics, researchers often face a similar problem. They want to know if a big national event (like a trade shock from China or a wave of immigration) changes local politics (like voting for a populist party). But every town is different. Some towns rely heavily on manufacturing; others rely on tourism.

A Shift-Share Design is a mathematical recipe to mix two things together:

The Shift (The Rain): A big change happening everywhere (e.g., Chinese imports go up nationwide).
The Share (The Hill): How much a specific local area depends on that thing (e.g., Town A has 50% of its jobs in manufacturing; Town B has 0%).

You multiply the Shift by the Share to create a custom "dose" of the shock for every town. Then, you see if that dose changes the political outcome.

The Two Ways to Cook the Meal (The Two Frameworks)

The paper argues that for this recipe to work, you have to be very careful about which part of the recipe is "clean" (exogenous) and which part might be "dirty" (endogenous). There are two main ways to do this:

1. The "Fixed Recipe" Approach (Share Exogeneity)

The Idea: Imagine the "Shares" (how much a town relies on manufacturing) are like a fixed family recipe passed down for generations. It doesn't change based on what happens today.
The Logic: If the recipe is fixed and random, then the only thing changing is the "Shift" (the rain). If the rain is random, we can trust the results.
The Catch: In reality, towns often choose their recipes. If a town is already struggling, they might stop manufacturing. So, the "Share" isn't actually fixed; it's reacting to the economy. If you assume it's fixed when it's not, your results are like baking a cake with spoiled ingredients—you might get a result, but it's not trustworthy.
Current Status: Most political science papers use this approach, often without realizing the ingredients might be spoiled.

2. The "Many Random Drops" Approach (Shift Exogeneity)

The Idea: Imagine the "Shifts" (the rain) are like thousands of tiny, independent raindrops falling from the sky.
The Logic: Even if the "Shares" (the hills) are messy and correlated, if you have enough independent raindrops, the errors cancel each other out (like the Law of Large Numbers). The noise averages out, and you can see the true signal.
The Catch: You need a lot of shifts. If you only have one big storm, this doesn't work. You need to break the problem down into many small, independent pieces.
Current Status: This is the more modern, rigorous approach used in top economics journals, but political scientists are just starting to catch on.

The Detective Story: The "China Shock" Replication

To prove his point, the author (Peter Park) acts like a detective and re-investigates a famous case: Colantone and Stanig (2018).

The Original Case: They claimed that Chinese imports caused a rise in nationalism and far-right voting in Europe. They used the "Fixed Recipe" approach (Share Exogeneity).
The Re-investigation: Peter Park took their data and applied the "Many Random Drops" approach (Shift Exogeneity), which is stricter. He also fixed some data errors (like "dividing by zero" when a country had almost no workers in a specific industry).

The Result?
When he used the stricter, more modern math:

The statistical significance disappeared.
The strong link between Chinese imports and far-right voting became much weaker or non-existent.

The Lesson: It wasn't that the original authors were "bad" scientists. It's that they used a tool (Share Exogeneity) that wasn't quite sharp enough for the specific job. They needed the newer, sharper tool (Shift Exogeneity).

Why Should Political Scientists Care?

The paper has three main messages for the field:

Stop Guessing: You can't just plug a "Shift-Share" variable into a computer and hope for the best. You have to ask: Is my "Share" really fixed? Or are my "Shifts" really independent?
The "American Politics" Gap: Most of these studies are about trade and immigration in Europe. But the US has massive amounts of data (donors, lobbyists, local elections). The author argues that the US is actually the perfect place to use these methods, but political scientists there are barely using them.
Don't Ignore the Math: The paper provides a "cheat sheet" (checklists and tests) to help researchers verify if their assumptions hold up. If you don't check your assumptions, you might be publishing a story that is just a statistical illusion.

The Takeaway Metaphor

Think of political science research like building a bridge.

Shift-Share Designs are a very powerful type of steel beam.
For a long time, everyone used the "Old Steel" (Share Exogeneity) because it was easy to find.
Peter Park is saying: "Hey, the Old Steel has a hidden rust problem. If you build your bridge on it, it might collapse when the wind blows (when you test the assumptions). We have a new, stronger 'New Steel' (Shift Exogeneity). It's harder to work with, but it's the only way to build a bridge that won't fall down."

The paper is a call to action: Upgrade your tools, check your assumptions, and don't be afraid to use the harder math to get the truth.

1. Problem Statement

Shift-share designs (often utilizing "Bartik instruments") are increasingly popular in political science for estimating causal effects of economic shocks (e.g., trade, immigration, aid) on political outcomes. However, the field faces a critical methodological gap:

Lack of Awareness: While political scientists frequently apply these methods, many are unaware of recent, rigorous econometric developments regarding identification assumptions.
Misapplication: Most studies rely on the Share Exogeneity framework without justification, or they apply the Shift Exogeneity framework incorrectly (e.g., failing to account for non-zero means or correlations in shifts).
Inference Errors: Standard error estimators (e.g., cluster-robust SEs) often fail in shift-share contexts, leading to over-rejection of null hypotheses.
Underutilization: The Shift Exogeneity framework, which is prevalent in modern economics, is underutilized in political science, particularly in American politics.

2. Methodological Framework

The paper synthesizes the literature into two distinct identification frameworks, defining the shift-share measure as $X_i = \sum_{j=1}^m w_{ij} D_j$ , where $w_{ij}$ are shares and $D_j$ are shifts.

A. Share Exogeneity Framework (Goldsmith-Pinkham et al., 2020)

Assumption: Shares ( $w_{ij}$ ) are exogenous (uncorrelated with the error term) conditional on covariates, while shifts ( $D_j$ ) can be endogenous.
Mechanism: This treats the design as a pooled Difference-in-Differences (DiD) where units are comparable because their exposure to common shocks is determined by fixed, exogenous shares.
Inference: Valid standard errors can be obtained via Generalized Method of Moments (GMM) treating shares as instruments.
Limitations: Requires shares to be exogenous. If shares are equilibrium outcomes (e.g., determined by past economic performance correlated with current political shocks), this framework fails. It also struggles with spatially correlated errors.

B. Shift Exogeneity Framework (Adão et al., 2019; Borusyak et al., 2022)

Assumption: Shifts ( $D_j$ ) are exogenous (mean-independent of errors and shares), while shares ( $w_{ij}$ ) can be endogenous.
Mechanism: Relies on a "Law of Large Numbers" logic. If there are many shifts, the biases introduced by endogenous shares cancel out, provided no single share dominates.
Key Requirements:
1. Demeaning: Shifts must be mean-zero (or controlled for) to remove systematic variation.
2. Negligibility: No single shift cluster should dominate the sample (low Herfindahl-Hirschman Index).
3. Independence: Shifts must be independent (or weakly correlated) across clusters.
Inference: Standard cluster-robust errors are invalid. Researchers must use inverted regression (aggregating data to the shift level) or specific estimators (e.g., BHJ estimator) that account for the correlation structure induced by shares.

C. Randomization Inference (Borusyak & Hull, 2023)

An alternative for finite samples where shifts are assumed to be exchangeable within groups. This avoids asymptotic assumptions and allows for exact inference without demeaning.

3. Key Contributions

Theoretical Synthesis

The paper unifies the literature using consistent notation, clarifying the distinction between Share Exogeneity (comparability of units) and Shift Exogeneity (comparability of shocks).
It introduces a three-dimensional framework for designing shift-share studies:
1. Type: Measurement vs. Instrumentation (Bartik vs. Propensity Score).
2. Definition: How units, shifts, and shares are constructed (e.g., handling predicted populations or incomplete shares).
3. Exogeneity Source: Choosing the correct identification strategy based on the data structure.

Methodological Innovations

Shift Transformation & Residualization: The author proposes specific procedures to handle shifts that are not mean-zero or have high variance (e.g., due to small denominators in trade data). This involves regressing shifts on fixed effects (country-industry and time) to isolate the stochastic component ( $\hat{\eta}_{cjt}$ ) before estimation.
Shift Replacement: A scheme to replace outlier shifts (e.g., those with very small denominators or low industry presence) with zero to protect the "identical mean" assumption and reduce measurement error.
Diagnostic Tests: The paper outlines rigorous tests for both frameworks, including:
- Placebo Tests: Checking if shifts predict pre-shock unit-level or shift-level characteristics.
- Share Negligibility: Checking the distribution of cluster shares (HHI).
- Overidentification Tests: Testing if shift-level controls are uncorrelated with demeaned shifts.

Empirical Replication (The "China Shock" in Europe)

Study: Replicates Colantone and Stanig (2018b), which linked Chinese imports to the rise of nationalism in Europe.
Findings:
- The original study used a design similar to Autor et al. (2013) but applied to a different context (Europe vs. US).
- The replication reveals that the original design violates key Shift Exogeneity assumptions due to:
  1. Divide-by-zero issues: Small European economies had industries with near-zero employment, creating volatile shifts.
  2. Non-zero means: Shifts had different means across countries due to economic size differences.
- Result: After applying shift transformation (residualization) and shift replacement, and using the correct BHJ (inverted regression) estimator, the statistical significance of the original findings disappears. The standard errors increase significantly, indicating that the original results were driven by correlated errors and misspecified shifts rather than a causal effect.

4. Results

Literature Review: A survey of 35 political science articles shows that while usage is growing (especially in trade and immigration), only ~33% of recent papers acknowledge endogeneity issues or cite the relevant econometric literature. Most rely on Share Exogeneity, often inappropriately.
Replication: The replication of Colantone and Stanig (2018b) demonstrates that:
- Standard 2SLS with clustered SEs (NUTS-2 region-year) underestimates uncertainty.
- Correctly accounting for shift heterogeneity and using the inverted regression estimator yields null results.
- The original findings were likely artifacts of unaddressed identification issues (correlated shifts and non-zero means).

5. Significance and Implications

For Political Science: The paper serves as a critical "wake-up call" for the field. It argues that many published causal claims using shift-share designs may be spurious due to incorrect identification strategies or standard error estimation.
Methodological Rigor: It provides a "how-to" guide for political scientists to correctly implement Shift Exogeneity, including data reconstruction, residualization, and the use of specific estimators (e.g., ssaggregate, ShiftShareSE).
Future Research Directions:
- American Politics: The author argues that American politics is uniquely suited for Shift Exogeneity due to the granular data available (donor networks, lobbying, district-level data), which provides the "many independent shifts" required for the framework.
- Network Analysis: Suggests shift-share designs can be adapted for network interference studies where treatment assignment depends on neighbor exposure.
Legacy: As a posthumous work, the paper aims to preserve and advance the intellectual legacy of Peter Park, urging researchers to move beyond "black box" applications of shift-share methods toward rigorous, assumption-aware causal inference.

Conclusion

The paper concludes that shift-share designs remain a powerful tool for political science but require a paradigm shift in how they are applied. Researchers must move away from defaulting to Share Exogeneity and instead carefully evaluate whether Shift Exogeneity is appropriate. If Shift Exogeneity is chosen, they must rigorously test for shift independence, mean-zero conditions, and use appropriate inference procedures to avoid false positives. The replication exercise serves as a stark reminder that methodological precision is paramount to substantive validity.