📊 statistics

Seemingly Unrelated Cointegrating Regressions with Autoregressive Distributed Lag Dynamics: Estimation, Bounds Testing, and Cross-Equation Inference

This paper proposes a Seemingly Unrelated Autoregressive Distributed Lag (SUR-ARDL) framework for small-N panels that enhances estimation efficiency and cointegration testing power by leveraging cross-equation error correlations while avoiding the finite-sample distortions of long-run covariance matrix estimators, as demonstrated through theoretical proofs, Monte Carlo simulations, and an application to the renewable energy-growth nexus in ASEAN countries.

Original authors: Muhammad Afnan Arif, Fumitaka Furuoka

Published 2026-08-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Muhammad Afnan Arif, Fumitaka Furuoka

Original paper licensed under CC BY 4.0 (https://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

Economics often feels like a study of isolated islands. When researchers look at how a country's economy grows, they typically build a model for that single nation, treating its history as a story that unfolds in a vacuum. They ask: does more energy lead to more wealth? Does trade open the door to prosperity? But in the real world, nations are not islands. They are part of a vast, interconnected archipelago. A shock to oil prices in the Middle East ripples through Asia; a financial crisis in one capital sends tremors to its neighbors; shared weather patterns and trade routes bind their fates together. For decades, statisticians have struggled to capture this web of connections without losing the unique character of each country. The standard tools often forced researchers to either ignore the connections between nations or to build models so complex they would collapse under their own weight when applied to a small group of countries.

This is the problem tackled by a new statistical framework developed by Muhammad Afnan Arif and Fumitaka Furuoka of the University of Malaya. Their work focuses on a specific type of economic relationship called cointegration. In plain terms, cointegration describes a long-term bond between two or more economic variables that move together over time, even if they wander apart in the short term. Think of a dog walking on a leash with its owner. The dog might dart ahead, lag behind, or circle around, but it never strays too far because the leash keeps it tethered to the owner. In economics, variables like energy consumption and gross domestic product often behave this way, drifting together over decades despite daily fluctuations. The challenge has been testing for this "leash" when looking at a group of countries simultaneously, especially when the data is messy and the connections between countries are strong.

The authors propose a new way to look at this data, which they call a "Seemingly Unrelated Autoregressive Distributed Lag" system. The name is a mouthful, but the concept is straightforward. They take a proven method for analyzing a single country's economic history and expand it to handle a whole group of countries at once. The key innovation is how they handle the "noise" in the data. In previous methods, trying to account for the connections between many countries required calculating a massive, unwieldy map of every possible relationship, a task that often led to mathematical errors or unreliable results, especially when the number of years of data was not huge compared to the number of countries. The new approach simplifies this by focusing only on the immediate, day-to-day connections between the countries' economic errors, rather than trying to map out every possible long-term ripple effect. This reduction in complexity allows the model to remain stable and accurate even with smaller datasets, a common reality in economic research where decades of data are precious.

To see if this new tool actually works, the researchers put it through a rigorous series of tests using computer simulations. They created thousands of fake economic histories for groups of countries, knowing exactly what the true relationships were, and then asked their new model to find them. They compared the results against the old, standard methods. The findings were clear: the new system was significantly more precise. When the countries in the group were closely linked, the new method reduced the margin of error in its estimates by as much as eighty percent compared to the old way of looking at each country in isolation. It also proved much better at detecting the "leash" of cointegration. In many simulated scenarios where the old method failed to see a long-term relationship, the new system spotted it immediately. This is a crucial advantage because it means researchers are less likely to miss important economic bonds simply because they were looking at the data one country at a time.

The researchers also tested the new method on real-world data, applying it to six Southeast Asian nations: Indonesia, Malaysia, the Philippines, Singapore, Thailand, and Vietnam. They examined the relationship between renewable energy consumption and economic growth from 1990 to 2024. The results offered a striking example of the method's power. When using the traditional approach, the data for Vietnam suggested there was no long-term link between renewable energy use and economic growth. However, when the new system analyzed all six countries together, accounting for how their economies influenced one another, it found a strong, significant connection for Vietnam that the old method had missed. Similarly, for Malaysia, the new method revealed that the speed at which the economy adjusted to changes was statistically significant, whereas the old method had left the result uncertain.

Beyond just finding new connections, the system allowed the researchers to ask questions that were previously impossible to answer. They could now test whether the speed of economic adjustment was the same across different countries. The analysis revealed a distinct pattern: the countries fell into two groups. The fossil fuel producers adjusted their economies slowly, while the energy importers adjusted quickly. This kind of insight, which relies on comparing the inner workings of different nations simultaneously, was invisible to researchers using the old, single-country tools. The study concludes that by treating these nations as a connected system rather than isolated cases, economists can build a clearer, more accurate picture of how energy and wealth interact in the real world. The method does not just offer a slight improvement; it opens the door to seeing economic dynamics that were previously hidden in the noise of isolated analysis.

Technical Summary: Seemingly Unrelated Cointegrating Regressions with Autoregressive Distributed Lag Dynamics

Problem Statement
The paper addresses the estimation and inference of cointegrating relationships in panel data settings where the cross-sectional dimension ( $N$ ) is small relative to the time series dimension ( $T$ ). While the literature has developed Seemingly Unrelated Regression (SUR) frameworks for cointegration—specifically Fully Modified (FM) estimators like FM-SOLS and FM-SUR (Park & Ogaki, 1991; Wagner et al., 2020) and FM-GLS (Lin & Reuvers, 2025)—these approaches share a critical practical limitation: the "curse of dimensionality." Existing methods require the consistent estimation of large $2N \times 2N$ long-run covariance matrices. This necessitates kernel-based methods with bandwidth selection, leading to severe finite-sample distortions, high rates of matrix invertibility failure (exceeding 50% in some simulations), and significant size distortions in hypothesis testing when $N$ increases relative to $T$ .

Methodology
The authors propose a Seemingly Unrelated Autoregressive Distributed Lag (SUR-ARDL) framework. This approach embeds the single-equation ARDL error-correction specification of Pesaran et al. (2001) within a Zellner (1962) SUR system.

Model Specification: The framework models a system of $N$ equations where each unit $i$ follows an ARDL( $p_i, q_{1,i}, \dots, q_{K,i}$ ) process. The model is reparameterized into a conditional error-correction form (ECM) that separates long-run level parameters from short-run dynamic adjustments.
Estimation Strategy: Unlike FM estimators that rely on semi-parametric corrections for serial correlation and endogeneity, the SUR-ARDL approach utilizes the parametric lag structure of the ARDL model to absorb these issues. Consequently, the estimator only requires the $N \times N$ contemporaneous error covariance matrix ( $\Omega$ ), rather than the $2N \times 2N$ long-run covariance matrix ( $\Omega_{LR}$ ).
Feasible Generalized Least Squares (FGLS): The estimation proceeds in two steps:
1. Equation-by-equation OLS to obtain residuals.
2. Construction of the estimated contemporaneous covariance matrix $\hat{\Omega}$ and application of FGLS to the stacked system.
Inference:
- Bounds Testing: The paper derives the asymptotic null distribution of the SUR-based bounds $F$ -statistic. It demonstrates that the statistic inherits the "bounds property" (critical values bounded between pure $I(0)$ and pure $I(1)$ cases) regardless of the integration order of regressors.
- Cross-Equation Inference: The framework enables Wald tests for cross-equation hypotheses, including coefficient poolability, long-run coefficient equality, and speed-of-adjustment homogeneity, utilizing the joint variance-covariance matrix of the system.

Key Contributions
The paper makes five specific contributions to the SUR cointegration literature:

Asymptotic Theory: It establishes that the SUR-ARDL FGLS estimator achieves a mixed-normal limiting distribution free of second-order bias terms, as the ARDL lag structure parametrically removes biases without semi-parametric corrections.
Bounds Testing Validity: It proves that the SUR bounds $F$ -statistic retains the Pesaran et al. (2001) bounds property and is monotonically increasing in cross-equation correlation, providing a formal basis for the power advantages of system-based testing.
Cross-Equation Hypothesis Testing: It develops chi-squared Wald tests for poolability and coefficient equality across equations, capabilities absent in single-equation methods.
Dimensionality Reduction: By reducing the nuisance parameter estimation from $O(N^2)$ long-run covariances to $O(N^2)$ contemporaneous covariances (specifically $N \times N$ vs. $2N \times 2N$ ), the method eliminates the primary source of finite-sample distortion found in FM-SUR and FM-GLS approaches.
Comprehensive Simulation Evidence: A Monte Carlo study (7 experiments, 5,000 replications) validates the finite-sample properties, showing reliable covariance estimation with zero invertibility failures even in configurations where FM-SUR fails.

Results
The Monte Carlo simulations and empirical application yield the following results:

Efficiency: The SUR-ARDL estimator achieves Root Mean Squared Error (RMSE) reductions of 15–80% for the error-correction parameter ( $\lambda$ ) relative to single-equation OLS, with gains increasing in $N$ , $T$ , and cross-equation correlation ( $\rho$ ).
Size and Power: The bounds test exhibits good size when using bootstrap critical values, though tabulated PSS critical values can be liberal for $N \geq 5$ with strong correlation. The test demonstrates substantial power gains (up to 30 percentage points) over single-equation methods, particularly in detecting cointegration when the speed of adjustment is moderate.
Covariance Estimation: The $N \times N$ contemporaneous covariance matrix is reliably estimable with zero invertibility failures across all configurations, contrasting sharply with the high failure rates of long-run covariance estimation in existing FM-SUR frameworks.
Empirical Application: Applied to the renewable energy-economic growth nexus in six ASEAN countries (1990–2024), the SUR-ARDL approach:
- Reduced standard errors by 17–41% compared to OLS.
- Detected cointegration in Vietnam where single-equation methods failed.
- Revealed a two-cluster structure of adjustment speeds (slow-adjusting fossil fuel producers vs. fast-adjusting energy importers) through cross-equation hypothesis tests.

Significance and Claims
The paper claims that the SUR-ARDL framework extends the practical applicability of system-based cointegration analysis to settings where $T$ is moderate relative to $N$ , a regime where existing FM-SUR methods are often infeasible due to matrix invertibility issues. By leveraging the parametric efficiency of the ARDL specification, the method avoids the need for semi-parametric long-run covariance estimation, thereby eliminating the principal source of finite-sample distortion in current SUR cointegration literature. The authors emphasize that the approach offers a robust alternative for small- $N$ panels, providing both efficiency gains in estimation and the unique ability to test cross-equation restrictions that are untestable in equation-by-equation analyses.

The paper notes limitations, specifically that delta-method standard errors for long-run coefficients (ratios of parameters) can be unreliable in small samples or when the speed of adjustment is weak, suggesting the need for alternative inference procedures like Fieller confidence sets or bootstrapping in future work. It also acknowledges that while the method handles small to moderate $N$ , shrinkage estimators may be required for larger $N$ .

More like this