The Heterogeneous Earnings Impact of Job Loss Across Workers, Establishments, and Markets

Susan Athey, Lisa K. Simon, Oskar Nordström Skans, Johan Vikström, Yaroslav YakymovycharXiv 2023 · DOI · arXiv

TL;DR

This paper utilizes causal machine learning (Generalized Random Forests) on decades of Swedish administrative data to reveal that the economic impact of job loss is far more "layered" and heterogeneous than previously understood. The researchers demonstrate that while average earnings losses are significant, the hardest-hit workers lose nearly half their income, and effective policy targeting requires looking beyond single indicators like age or education to account for complex interactions between worker, firm, and market characteristics.

What Problem Does This Paper Address?

When a factory closes or a firm goes under, workers lose more than just a paycheck; they often lose "rents" (higher-than-average wages) associated with that specific job and face a difficult transition back into the labor market. Economists have long known that these losses are persistent, but previous research typically examined one variable at a time—asking, for example, "how does age affect loss?" or "how does industry affect loss?"

The authors argue that this "one-at-a-time" approach misses the bigger picture. In reality, a worker's outcome is determined by a complex web of factors: their personal skills, the specific firm they left, the health of their local economy, and the industry's national trend. This paper seeks to provide a unified, multidimensional framework to identify exactly who is most vulnerable and how policymakers can better target interventions to improve both economic efficiency and social equity.

Methods

The researchers employ Generalized Random Forests (GRF), a state-of-the-art causal machine learning method.

  • Causal ML vs. Standard ML: Unlike standard machine learning, which predicts outcomes (e.g., "what will this person earn?"), GRF estimates the Conditional Average Treatment Effect (CATE)—the specific "delta" or impact caused by the job loss itself for a person with specific characteristics.
  • Cross-Fitting: To prevent the model from "memorizing" the data (overfitting), they use a 5-fold cross-fitting procedure. They train the model on one portion of the data and generate predictions for a "held-out" portion.
  • Policy Trees: After identifying the most vulnerable workers using the full model, they use optimal policy trees to find "simplified" rules that a government agency could actually implement (e.g., "target workers over 50 in routine occupations").
  • Identification Strategy: They compare workers displaced by establishment closures to a matched control group of workers at surviving firms, ensuring the "shock" of job loss is exogenous (not due to the worker's own performance).

Datasets

The study relies on the Swedish RAMS data, an exceptionally rich administrative dataset covering:

  • Timeline: All annual transactions from 1985 to 2017.
  • Scope: Approximately 180,000 displaced workers across 21,000 closure events, compared against a pool of over 4 million potential control workers.
  • Granularity: Detailed information on demographics, family structure, schooling, firm-level wage premia, industry dynamism (churning rates), and local labor market conditions (population density, unemployment).

Key Insights and Ideas

  • The "Within" vs. "Across" Paradox: Surprisingly, the variation in earnings loss within groups (e.g., among all 50-year-old high school grads) is just as large as the variation between different groups. This means that simple demographic buckets are poor predictors of individual hardship.
  • The Multi-Layered Nature of Loss: No single factor explains the damage. Some losses come from losing firm-specific "rents," while others stem from a total lack of local re-employment opportunities.
  • Vulnerability is Compounded: The workers predicted to lose the most from a closure are the same ones who already had lower earnings and a higher baseline risk of job loss. Job displacement thus acts as an inequality accelerator.
  • Low Geographic Mobility: Despite suffering massive losses, the most vulnerable workers do not move to different regions for work. This suggests that "place-based" policies (investing in specific struggling towns) may be more effective than "people-based" policies that assume workers will simply move to where the jobs are.

Key Results

  • Average Impact: On average, displaced workers see a 24% drop in annual earnings one year after displacement. Ten years later, they are still earning 8% less than their non-displaced peers.
  • The Extremes: The GRF model successfully identified a "hardest-hit" decile that loses 46% of their earnings, while the "least-affected" decile loses only 5.9%.
  • Targeting Efficiency: Targeting based on the full GRF model is significantly more effective than any single-variable rule. However, if a simple rule is needed, a combination of age and job routineness performs best, capturing much of the vulnerability associated with declining industries like manufacturing.
  • Insurance Buffer: Sweden’s tax and transfer system (unemployment insurance, etc.) mitigates about 71% of the loss for the most vulnerable, though this "insurance" has weakened over time as benefit caps failed to keep pace with wage growth.

Limitations and Future Work

  • Context Specificity: The results are based on Swedish data. While the causal ML methods are universal, the specific "vulnerability profiles" might differ in countries with less robust social safety nets or different industrial structures (like the US).
  • Worker-Firm Sorting: The paper does not explicitly model the "sorting" process—how high-ability workers end up at high-wage firms in the first place—which could influence the estimated "rents" lost.
  • Future Pathways: The authors suggest that future research should use these ML tools to study different "back-to-work" pathways, such as whether retraining programs or self-employment subsidies work better for specific subgroups identified by the model.

Summary updated

Abstract

Using rich Swedish administrative data, we apply causal machine learning methods to study how earnings losses after job displacement vary with observable characteristics that may be relevant for targeting policy interventions for workers. Heterogeneity in effects is as large within as across worker groups defined by age and schooling, and as large within as across establishments. A substantial portion of cross-establishment heterogeneity can be explained by industry and local labor market characteristics, suggesting a role for place- and industry-based targeting. The largest losses are concentrated among already vulnerable workers, indicating that well-designed targeting policies can improve both efficiency and equity.
BibTeX
@misc{athey2023heterogeneous,
  author = {Susan Athey and Lisa K. Simon and Oskar Nordström Skans and Johan Vikström and Yaroslav Yakymovych},
  title = {{The Heterogeneous Earnings Impact of Job Loss Across Workers, Establishments, and Markets}},
  year = {2023},
  eprint = {2307.06684},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2307.06684},
}
Figure