Common Factor Restrictions in ARDL Models: Reconnecting the Time Series and Panel Literatures

From Sargan and Hendry to Cook and Webb and Beyond

Robert Walker

2026-07-27

Motivation

Two literatures studying the same algebra

The literature on common factor restrictions (CFRs) evolved along two largely separate paths:

Time Series Econometrics

  • Sargan (1964)
  • Hendry and the LSE tradition
  • Dynamic specification
  • Error-correction models
  • General-to-specific modelling

Panel Econometrics

  • Dynamic fixed effects
  • GMM estimators
  • Mean Group estimation
  • Pooled Mean Group estimation
  • Common Correlated Effects

Central claim: These literatures study closely related dynamic models but rarely engage with one another.

Core Question

Suppose we observe persistence in a dependent variable.

Is that persistence due to:

  1. genuine lagged effects of explanatory variables?

or

  1. serially correlated disturbances?

This distinction is fundamental for:

  • causal interpretation,
  • long-run multipliers,
  • forecasting,
  • policy analysis.

Common factor restrictions provide a testable answer.

The Basic Result

Consider

y_t = \alpha y_{t-1} + \beta x_t + u_t

with

u_t = \rho u_{t-1} + \varepsilon_t.

Substituting yields

y_t = (\alpha+\rho)y_{t-1} -\alpha\rho y_{t-2} +\beta x_t -\rho\beta x_{t-1} +\varepsilon_t.

The model can be written as an ADL(2,1).

But only a subset of ADL(2,1) models satisfy this representation.

Common Factor Restrictions

The lag polynomial

(1-\rho L)

appears as a common factor in both

  • the autoregressive polynomial,
  • the distributed lag polynomial.

Therefore

\delta_1 = -\rho \beta

and

\gamma_2 = -\rho \alpha.

These nonlinear parameter restrictions are the common factor restrictions.

Hendry–Sargan

Estimate:

ADL(p,q)

first.

Then:

  1. Test common factor restrictions.
  2. Impose them only if supported by data.

Interpretation:

  • residual serial correlation may indicate omitted dynamics;
  • AR-error models are not automatically appropriate;
  • rejected restrictions imply genuine lagged structural effects.

Error-Correction Models

Every ADL can be rewritten as an ECM.

When common factor restrictions hold:

  • parsimonious ECM representation exists;
  • adjustment occurs through a single dynamic mechanism.

When restrictions fail:

  • additional short-run dynamics remain;
  • richer lag structures are required.

Common factor restrictions therefore connect:

ARDL models, ECMs, and serial correlation models.

Cook and Webb (2021)

Main contribution

Cook and Webb revisit this issue in political methodology.

Their central argument:

An unrestricted ADL contains both genuine lagged effects and AR-error representations.

Therefore:

  • estimate unrestricted dynamics;
  • test common factor restrictions;
  • do not assume lagged regressors merely proxy serial correlation.

This is fundamentally a specification-testing argument.

Key Observation: Nothing in the algebra requires (N=1)

The Cook and Webb derivation is not intrinsically a time-series result.

The same substitution works immediately in panel data settings.

This raises an important question:

Why has the panel literature paid so little attention to common factor restrictions?

Extending the Result to Panel Data

Consider

y_{it} = \alpha_i y_{i,t-1} + \beta_i x_{it} + u_{it}

with

u_{it} = \rho_i u_{i,t-1} + \varepsilon_{it}.

Substitution gives

y_{it} = (\alpha_i+\rho_i)y_{i,t-1} - \alpha_i\rho_i y_{i,t-2} + \beta_i x_{it} - \rho_i\beta_i x_{i,t-1} + \varepsilon_{it}.

The same common factor restrictions emerge. A trivial paper no one has written.

Four Major Panel Approaches

Dynamic Fixed Effects (DFE)

Assumes:

  • common slope coefficients,
  • common dynamic parameters,
  • unit-specific intercepts.

Effectively pools all units into one dynamic model.

Suitable when dynamics are believed to be homogeneous.

Four Major Panel Approaches

Mean Group (MG)

Pesaran and Smith (1995)

Procedure:

  1. Estimate a separate time-series model for each unit.
  2. Average estimated coefficients.

Allows:

  • heterogeneous short-run dynamics,
  • heterogeneous long-run effects.

Most flexible panel ARDL estimator.

Four Major Panel Approaches

Pooled Mean Group (PMG)

Pesaran, Shin, and Smith (1999)

Allows:

  • heterogeneous short-run dynamics,
  • heterogeneous adjustment speeds,

but imposes:

  • common long-run coefficients.

Motivation:

  • countries, firms, or regions may adjust differently,
  • but share the same equilibrium relationship.

Widely used in macroeconomic panel ARDL applications.

Four Major Panel Approaches

Common Correlated Effects (CCE)

Pesaran (2006)

Addresses:

  • cross-sectional dependence,
  • omitted common shocks.

Adds cross-sectional averages to absorb latent factors.

Examples:

  • global business cycles,
  • common technology shocks,
  • financial crises.

“Common factors” here means latent cross-sectional factors—not common lag polynomials.

Why the Literatures Diverged

Time-Series Tradition

Main concerns:

  • specification testing,
  • encompassing,
  • dynamic completeness,
  • model reduction.

Representative authors:

  • Sargan
  • Hendry
  • Mizon
  • Davidson

Why the Literatures Diverged

Panel Tradition

Main concerns:

  • Nickell bias,
  • dynamic panel estimation,
  • heterogeneity,
  • cointegration,
  • cross-sectional dependence.

Representative authors:

  • Arellano
  • Bond
  • Pesaran
  • Shin
  • Smith

The specification-testing agenda largely disappeared.

Historical Timeline

1964  Sargan
      Common factor restrictions

1970s–1980s
      Hendry–Pagan–Sargan
      Dynamic specification testing

1987
      Engle–Granger ECM representation

1990s
      Cointegration revolution

1995
      Mean Group (MG)

1999
      Pooled Mean Group (PMG)

2000s
      Dynamic heterogeneous panels

2006
      Common Correlated Effects (CCE)

2021
      Cook and Webb
      Revival of common factor restrictions

Two Meanings of “Common Factors”

Time-Series Literature Panel Literature
Common lag polynomial Latent cross-sectional factor
Dynamic specification Cross-sectional dependence
AR errors vs lagged effects Omitted common shocks
Sargan-Hendry Pesaran CCE

The terminology is similar.

The underlying concepts are different.

A Unified Perspective

All of the following may be viewed as variants of an unrestricted panel ARDL:

  • DFE
  • MG
  • PMG
  • CCE-ARDL

The fundamental question remains:

Are observed lag structures genuine dynamics, or merely representations of serially correlated errors?

This is precisely the question addressed by common factor restrictions.

A Possible Research Program

Estimate unrestricted panel ARDL models and:

  1. Test common factor restrictions.
  2. Compare MG, PMG, and DFE implementations.
  3. Extend tests to CCE specifications.
  4. Evaluate finite-sample properties.
  5. Study consequences for long-run multipliers.

This appears largely absent from the modern literature.

Main Conclusions

  1. Common factor restrictions are fundamentally algebraic.
  2. The Cook and Webb argument extends naturally to panel data.
  3. The extension applies to MG, PMG, DFE, and related estimators.
  4. Time-series and panel econometrics evolved into largely separate literatures.
  5. Reintegrating these traditions offers a promising research agenda.

References

Cook, S. J., & Webb, M. D. (2021). Lagged outcomes, lagged predictors, and lagged errors: A clarification on common factors. Political Analysis.

Engle, R. F., & Granger, C. W. J. (1987). Co-integration and error correction: Representation, estimation and testing. Econometrica.

Hendry, D. F. (1995). Dynamic Econometrics. Oxford University Press.

Hendry, D. F., Pagan, A. R., & Sargan, J. D. (1984). Dynamic specification.

Pesaran, M. H. (2006). Estimation and inference in large heterogeneous panels with a multifactor error structure. Econometrica.

Pesaran, M. H., Shin, Y., & Smith, R. P. (1999). Pooled Mean Group estimation of dynamic heterogeneous panels. JASA.

Pesaran, M. H., & Smith, R. P. (1995). Estimating long-run relationships from dynamic heterogeneous panels. Journal of Econometrics.

Sargan, J. D. (1964). Wages and prices in the United Kingdom: A study in econometric methodology.

Wilkins, A. S. (2018). To lag or not to lag? Re-evaluating the use of lagged dependent variables in regression analysis.