Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

To Adopt or Not to Adopt: Heterogeneous Trade Effects of the Euro

Harry Aytug

TL;DR

This paper develops a causal-forest framework with double machine learning to estimate the full distribution of euro-adoption effects on bilateral trade, revealing substantial heterogeneity across country pairs. By applying this method to EU15 data (1995–2015), it shows an average trade increase around 24% (15% after fixed-effects correction) with a wide cross-pair range, and a core-periphery pattern where core eurozone links benefit most. The analysis also demonstrates that pre-euro trade intensity and GDP largely drive heterogeneity, and it provides counterfactual predictions for non-eurozone members (e.g., UK +23.9%, Sweden +19.9%, Denmark +19.4%) under euro adoption. These findings reconcile the long-standing 4–30% puzzle by showing that effects are not uniform but conditional on pair characteristics, informing policy decisions for prospective adopters and clarifying the mechanics of currency unions in trade integration.

Abstract

Two decades of research on the euro's trade effects have produced estimates ranging from 4% to 30%, with no consensus on the magnitude. We find evidence that this divergence may reflect genuine heterogeneity in the euro's trade effect across country pairs rather than methodological differences alone. Using Eurostat data on 15 EU countries from 1995-2015, we estimate that euro adoption increased bilateral trade by 24% on average (15.0% after fixed effects correction), but effects range from -12% to +68% across eurozone pairs. Core eurozone pairs (e.g., Germany-France, Germany-Netherlands) show large gains, while peripheral pairs involving Finland, Greece, and Portugal saw smaller or negative effects, with some negative estimates statistically significant and interpretable as trade diversion. Pre-euro trade intensity and GDP explain over 90% of this variation. Extending to EU28, we find evidence that crisis-era adopters (Slovakia, Estonia, Latvia) pull down naive estimates to 5%, but accounting for fixed effects recovers estimates of 14.0%, consistent with the EU15 fixed-effects baseline of 15.0%. Illustrative counterfactual analysis suggests non-eurozone members would have experienced varied effects: UK (+24%), Sweden (+20%), Denmark (+19%). The wide range of prior estimates appears to be largely a feature of the data, not a bug in the methods.

To Adopt or Not to Adopt: Heterogeneous Trade Effects of the Euro

TL;DR

This paper develops a causal-forest framework with double machine learning to estimate the full distribution of euro-adoption effects on bilateral trade, revealing substantial heterogeneity across country pairs. By applying this method to EU15 data (1995–2015), it shows an average trade increase around 24% (15% after fixed-effects correction) with a wide cross-pair range, and a core-periphery pattern where core eurozone links benefit most. The analysis also demonstrates that pre-euro trade intensity and GDP largely drive heterogeneity, and it provides counterfactual predictions for non-eurozone members (e.g., UK +23.9%, Sweden +19.9%, Denmark +19.4%) under euro adoption. These findings reconcile the long-standing 4–30% puzzle by showing that effects are not uniform but conditional on pair characteristics, informing policy decisions for prospective adopters and clarifying the mechanics of currency unions in trade integration.

Abstract

Two decades of research on the euro's trade effects have produced estimates ranging from 4% to 30%, with no consensus on the magnitude. We find evidence that this divergence may reflect genuine heterogeneity in the euro's trade effect across country pairs rather than methodological differences alone. Using Eurostat data on 15 EU countries from 1995-2015, we estimate that euro adoption increased bilateral trade by 24% on average (15.0% after fixed effects correction), but effects range from -12% to +68% across eurozone pairs. Core eurozone pairs (e.g., Germany-France, Germany-Netherlands) show large gains, while peripheral pairs involving Finland, Greece, and Portugal saw smaller or negative effects, with some negative estimates statistically significant and interpretable as trade diversion. Pre-euro trade intensity and GDP explain over 90% of this variation. Extending to EU28, we find evidence that crisis-era adopters (Slovakia, Estonia, Latvia) pull down naive estimates to 5%, but accounting for fixed effects recovers estimates of 14.0%, consistent with the EU15 fixed-effects baseline of 15.0%. Illustrative counterfactual analysis suggests non-eurozone members would have experienced varied effects: UK (+24%), Sweden (+20%), Denmark (+19%). The wide range of prior estimates appears to be largely a feature of the data, not a bug in the methods.
Paper Structure (45 sections, 11 equations, 30 figures, 25 tables)

This paper contains 45 sections, 11 equations, 30 figures, 25 tables.

Table of Contents

  1. Introduction
  2. Literature Review
  3. Methodology
  4. Existing Approaches: Gravity and Synthetic Control
  5. From PSM to Causal Forests
  6. Intuition: How Causal Forests Discover Heterogeneity
  7. Causal Forest with Double Machine Learning
  8. Constructing Counterfactuals
  9. Treatment Definition and Identification
  10. Threats to Identification
  11. Data
  12. Results
  13. Gravity Benchmark
  14. Selection and Identification
  15. Heterogeneous Treatment Effects
  16. ...and 30 more sections

Figures (30)

  • Figure 1: Illustration of causal tree splitting. The tree recursively partitions country pairs into subgroups with different treatment effects. At each node, the split is chosen to maximize the difference in euro effects between subgroups. Terminal nodes (leaves) show estimated effects ranging from +77% for high-trade, high-GDP pairs (e.g., France--Luxembourg) to $-3\%$ for low-trade, low-GDP pairs (e.g., Finland--Ireland). Colors indicate effect magnitude: green (high), yellow (medium), red (low). A causal forest averages predictions across many such trees.
  • Figure 2: Distribution of Log GDP Product for eurozone pairs (solid) and non-eurozone pairs (dashed) across key years. In 1990, no eurozone pairs exist as the euro had not yet been adopted.
  • Figure 3: Distribution of Log GDP Per Capita for eurozone pairs (solid) and non-eurozone pairs (dashed) across key years.
  • Figure 4: Distribution of Log Bilateral Trade for eurozone pairs (solid) and non-eurozone pairs (dashed) across key years.
  • Figure 5: Propensity score distribution by treatment status. Left panel shows logistic regression estimates; right panel shows random forest estimates. Substantial overlap between treated (eurozone) and control (non-eurozone) pairs supports the positivity assumption.
  • ...and 25 more figures