Difference-in-differences with as few as two cross-sectional units -- A new perspective to the democracy-growth debate

Basic Information

• Paper ID: 2408.13047
• Title: Difference-in-differences with as few as two cross-sectional units -- A new perspective to the democracy-growth debate
• Authors: Gilles Boevi Koumou, Emmanuel Selorm Tsyawo
• Classification: econ.EM (Econometrics)
• Publication Date: October 14, 2025
• Paper Link: https://arxiv.org/abs/2408.13047

Abstract

Pooled panel analysis often obscures heterogeneity in unit-specific treatment effects. For instance, in studies examining the impact of democracy on economic growth, results vary significantly due to differences in country composition. To address this challenge, this paper introduces a difference-in-differences (DiD) estimator that leverages temporal variation in data to estimate unit-specific average treatment effects (ATT) with as few as two cross-sectional units. Under weak identification and time-dependence conditions, the proposed DiD estimator is shown to be asymptotically normal. The method is equipped with an identification test that, unlike pre-trend tests, is more powerful and can detect violations of parallel trends in post-treatment periods. Empirical results using the DiD estimator suggest that if Benin had not democratized, its economy would have been on average 6.3% smaller during 1993-2018.

Research Background and Motivation

Problem Identification

1. Heterogeneity Masking Problem: Traditional pooled regression analysis becomes uninformative when treatment effects exhibit significant heterogeneity, particularly when only a limited number of cross-sectional units are available.
2. Limitations of Existing Methods:
   • Traditional DiD requires numerous cross-sectional units
   • Synthetic control (SC) methods require multiple candidate control units
   • Existing methods perform poorly in fixed N, large T settings

Research Significance

The impact of democracy on economic growth remains a long-standing controversial topic, with contradictory findings in existing literature:

• Some studies find positive effects (e.g., Acemoglu et al., 2019)
• Some studies find negative effects (e.g., Gerring et al., 2005)
• Some studies find no significant effects (e.g., Murtin and Wacziarg, 2014)

These contradictions primarily stem from modeling differences, data limitations, and heterogeneity in regime performance.

Research Motivation

This paper proposes to revisit the democracy-growth relationship from the perspective of unit-specific effects, developing new methods applicable to extremely limited cross-sectional units (minimum two) but with abundant time-series observations.

Core Contributions

1. Methodological Innovation: Proposes the T-DiD estimator capable of estimating unit-specific average treatment effects with only two cross-sectional units
2. Theoretical Development: Establishes asymptotic normality theory under weak identification conditions, allowing for time dependence and cross-sectional dependence
3. Identification Testing: Develops overidentification restriction tests capable of detecting parallel trends violations in post-treatment periods
4. Empirical Application: Applies the method to assess the impact of Benin's democratization on economic growth

Methodological Details

Task Definition

Estimate the parameter: $ATT_{\omega,T} = \sum_{t=1}^T w_T(t)ATT(t)$

where $ATT(t) = E[Y_t(1) - Y_t(0)|D = 1]$ is the treatment effect in period $t$, and $w_T(t)$ is a convex weighting scheme.

Model Architecture

1. Basic Setup

• One treated unit, one control unit
• Large T pre-treatment periods, large T post-treatment periods
• Define $X_t := Y_{1,t} - Y_{0,t}$ as the difference in outcomes between treated and control units

2. T-DiD Estimator

$\hat{ATT}_{\omega,T} = \sum_{t=1}^T w_T(t)(Y_{1,t} - Y_{0,t}) - \sum_{-\tau=1}^T \psi_T(-\tau)(Y_{1,\tau} - Y_{0,\tau})$

where $\psi_T(-\tau)$ are weights for pre-treatment periods.

3. Regression Form

The model can be expressed as a weighted linear regression: X_t = \beta_0 + \tilde{ATT}_{w,\psi}_n \mathbf{1}\{t \geq 1\} + U_t

Identification Conditions

Asymptotic Parallel Trends Assumption (Assumption 1)

$\frac{1}{T\bar{T}}\sum_{-\tau=1}^T\sum_{t=1}^T(E[Y_t(0)-Y_\tau(0)|D=1]-E[Y_t(0)-Y_\tau(0)|D=0]) = O((T \wedge \bar{T})^{-(1/2+\gamma)})$

Asymptotic Finite Expectations Assumption (Assumption 2)

$\frac{1}{\bar{T}}\sum_{-\tau=1}^T E[Y_\tau(1)-Y_\tau(0)|D=1] = O(\bar{T}^{-(1/2+\delta)})$

Technical Innovations

1. Weak Identification Conditions: Allows controlled violations of standard parallel trends and no-anticipation assumptions
2. Time Dependence: Addresses time-series dependence through near-epoch dependence (NED) structures
3. Overidentification Testing: Utilizes overidentification restrictions when multiple control units are available
4. Non-stationarity Handling: Can accommodate unit roots, deterministic trends, and other complex time-series features

Experimental Setup

Dataset

• Outcome Variable: Per capita GDP of Benin and Togo (2015 constant USD) from 1960-2018
• Treatment Variable: Democracy index from V-Dem project (average of five high-level democracy indices)
• Treatment Period: 1990-1992 (Benin's democratization transition)
• Data Source: World Bank Development Indicators Database

Control Unit Selection

Togo is selected as the control unit for Benin based on:

• Shared colonial history (France)
• Geographic proximity (651 km border)
• Same monetary union (UEMOA)
• Similar cultural and institutional backgrounds
• Similar political institutions before 1990

Comparison Methods

• Synthetic Control (SC)
• Synthetic Difference-in-Differences (S-DiD)
• Augmented Synthetic Control (ASC)
• Before-After (BA) estimator

Experimental Results

Main Results

Impact on Log Per Capita GDP

ATT estimates under different transition windows:

• 1990-1992: 0.063 (Standard Error: 0.017)
• 1990: 0.078 (Standard Error: 0.018)
• 1991: 0.074 (Standard Error: 0.020)
• 1992: 0.059 (Standard Error: 0.017)

All estimates are significant at the 1% level, indicating that Benin's economy grew on average 5.9%-7.8% following democratization.

Impact on Per Capita GDP in Levels

• 1990-1992: $45.511 (Standard Error:$16.577)
• Corresponding to 4.4% of Benin's 2015 per capita GDP

Method Comparison Results

Comparison using 20 African countries as candidate control units:

• T-DiD: 0.038 (Standard Error: 0.010)
• S-DiD: 0.249 (Standard Error: 0.282)
• SC: 0.090 (p-value: 0.700)
• ASC: 0.291 (p-value: 0.201)

The T-DiD estimate is most precise and statistically significant.

Identification Testing

The overidentification restriction test yields a p-value of 0.798, failing to reject the identification hypothesis and supporting the validity of the method.

Diagnostic Tests

• ADF test: Rejects unit root hypothesis (p<0.05)
• KPSS test: No trend non-stationarity detected (p=0.100)
• Durbin-Watson test: No autocorrelation detected (p>0.10)

Related Work

Traditional DiD Methods

• Classical DiD by Card (1990), Angrist and Pischke (2008)
• Multi-period DiD by Callaway and Sant'Anna (2021)
• These methods require numerous cross-sectional units

Synthetic Control Methods

• Original SC method by Abadie and Gardeazabal (2003)
• Requires multiple candidate control units for pre-treatment fitting
• Performs poorly when control units are limited

Panel Data Methods

• Panel data methods by Hsiao et al. (2012)
• Counterfactual analysis for non-stationary data by Masini and Medeiros (2022)
• Typically require factor structure assumptions

Conclusions and Discussion

Main Conclusions

1. Method Validity: T-DiD can reliably estimate treatment effects with only two units
2. Empirical Findings: Benin's democratization significantly promoted economic growth; without democratization, the economy would have been on average 6.3% smaller
3. Theoretical Contribution: Establishes asymptotic theory under weak conditions, extending the applicability of DiD methods

Limitations

1. Time Requirements: Requires substantial pre-treatment and post-treatment periods
2. Heterogeneity Trade-off: Sacrifices time-heterogeneity analysis through weighted averaging
3. Control Unit Selection: Still requires at least one valid control unit

Future Directions

1. Extension to multiple treated units
2. Development of more flexible weighting schemes
3. Handling more complex time-series features

In-Depth Evaluation

Strengths

1. Methodological Innovation: Fills a gap in causal inference methods for fixed N, large T settings
2. Theoretical Rigor: Establishes comprehensive asymptotic theory under weak conditions
3. Practical Value: Applicable to many real-world scenarios, particularly for unit-specific effect studies
4. Test Completeness: Provides identification tests more powerful than pre-trend tests

Weaknesses

1. Applicability Constraints: Still requires substantial time observations
2. Effect Aggregation: Cannot estimate period-specific treatment effects
3. Control Unit Dependence: Method validity highly depends on control unit selection quality

Impact

1. Academic Contribution: Provides new theoretical tools for causal inference literature
2. Policy Significance: Offers new methods for evaluating institutional change effects
3. Practical Value: Applicable to many real-world policy evaluation scenarios

Applicable Scenarios

• Unit-specific policy effect evaluation
• Institutional change impact analysis
• Firm-level intervention effect studies
• Regional policy impact assessment

References

1. Abadie, A., & Gardeazabal, J. (2003). The economic costs of conflict: A case study of the Basque Country. American Economic Review, 93(1), 113-132.
2. Acemoglu, D., Naidu, S., Restrepo, P., & Robinson, J. A. (2019). Democracy does cause growth. Journal of Political Economy, 127(1), 47-100.
3. Callaway, B., & Sant'Anna, P. H. (2021). Difference-in-differences with multiple time periods. Journal of Econometrics, 225(2), 200-230.
4. Hsiao, C., Ching, H. S., & Wan, S. K. (2012). A panel data approach for program evaluation. Journal of Applied Econometrics, 27(5), 705-740.

This paper makes important methodological contributions to econometrics, providing effective tools for causal inference problems involving extremely limited cross-sectional units but rich temporal dimensions. It possesses significant value at both theoretical and applied levels.