Factor-driven two-regime regression

Sokbae Lee, Yuan Liao, Myung Hwan Seo, Youngki ShinThe Annals of Statistics 2021 · DOI

TL;DR

This paper introduces a new "factor-driven two-regime regression" model where the switch between two distinct behaviors (regimes) is governed by unobservable factors. The authors develop novel computational methods using mixed integer optimization to estimate this model and provide a theoretical framework for understanding how the estimation accuracy changes based on the data's structure, revealing an interesting "phase transition."

What Problem Does This Paper Address?

Traditional threshold regression models often rely on a single observable variable to determine regime switches. However, in many real-world scenarios, the underlying factors driving these switches are complex and unobservable, potentially influenced by a multitude of variables. The authors address this limitation by proposing a model where regime switching is driven by a vector of latent factors. Estimating such a model presents significant computational and theoretical challenges, particularly when the number of factors is large or their exact nature is unknown. This paper aims to provide a robust and computationally feasible solution for analyzing such complex, unobservable-driven regime changes.

Methods

The core of the paper lies in its novel approach to estimating the factor-driven two-regime regression model. Key methods include:

  • Latent Factor Estimation: When factors are unobservable, they are estimated using Principal Component Analysis (PCA) on a panel dataset. This allows the model to "learn" the driving factors from the data.
  • Mixed Integer Optimization (MIO): The estimation problem, which involves identifying regime switches, is reformulated as a Mixed Integer Optimization problem. This is a powerful technique for handling discrete choices (like regime switches) alongside continuous parameters.
  • Computational Algorithms: Two alternative algorithms are presented to solve the MIO problem:
  • Mixed Integer Quadratic Programming (MIQP): A direct approach to find the global optimum.
  • Block Coordinate Descent (BCD): A faster, iterative algorithm that combines MILP and MIQP steps, designed for scalability in large-dimensional problems.
  • Asymptotic Theory: The paper derives the asymptotic distribution of the estimators, particularly focusing on a "diminishing threshold effect" where the regime switch becomes less pronounced as the sample size grows. A key theoretical contribution is the identification of a "phase transition" in the estimator's behavior, depending on the relative sizes of the cross-sectional and time-series dimensions of the data.
  • Bootstrap Inference: To address the non-pivotal nature of the estimator for the threshold parameter, a wild bootstrap method is developed for inference.

Datasets

The paper uses simulated data for its Monte Carlo experiments. For the empirical application, it utilizes monthly unemployment rates for males aged 20 and over, along with data from 132 macroeconomic variables to construct a "real factor." The empirical application revisits the dataset used by Hansen (1997) and Ludvigson and Ng (2009).

Key Insights and Ideas

  • Unobservable Factors Drive Regimes: The model's core idea is that complex, unobservable factors, rather than single observable variables, can drive regime shifts, offering a more realistic representation of many economic phenomena.
  • MIO for Regime Switching: Reformulating the regime-switching estimation problem as MIO is a significant methodological advancement, enabling efficient and accurate estimation.
  • Phase Transition in Estimation: The most striking theoretical finding is the "phase transition" in the estimator's performance. The relative sizes of the cross-sectional and time-series dimensions of the data dictate whether the estimator achieves a "strong oracle" property (near-perfect estimation) or a "weak oracle" property (slower convergence), with a continuous transition in between. This highlights the crucial role of data structure in factor model estimation.
  • Bootstrap for Robust Inference: The development of a bootstrap procedure for inference on the threshold parameter is important for practical applications, as it doesn't require prior knowledge of the factor estimation error rates.

Key Results

  • Computational Efficiency: The proposed MIQP and BCD algorithms are shown to be effective in solving the complex optimization problem, with BCD demonstrating superior performance in large-dimensional settings.
  • Asymptotic Properties: The paper provides theoretical guarantees for the estimators, including consistency and rates of convergence. The phase transition phenomenon is rigorously established, showing how the estimator's performance varies with data dimensions.
  • Empirical Application: The model is applied to US unemployment data, demonstrating its ability to capture regime dynamics more effectively than traditional threshold models, particularly when using estimated factors. Specification (3) in the empirical example, which uses both a lagged unemployment rate and the estimated real factor, shows the lowest misclassification error and best explains unemployment dynamics.
  • Bootstrap Test Performance: The bootstrap test for the null hypothesis of no threshold effect provides strong evidence for the existence of two regimes in the unemployment data.

Limitations and Future Work

  • Computational Complexity: While MIO and BCD are proposed, the computational burden can still be significant for very high-dimensional problems.
  • Factor Identification: The paper acknowledges the "labeling problem" common in hidden regime models, where the sign of estimated factors might be arbitrary. While addressed by assumptions, it's a point of consideration.
  • Nonparametric Extension: The authors suggest an interesting avenue for future research: extending the model to nonparametric regime switching, where the switching indicator is replaced by a nonparametric function of observable variables. This could lead to even more flexible and powerful models.

Summary updated

Abstract

We propose a novel two-regime regression model where regime switching is driven by a vector of possibly unobservable factors. When the factors are latent, we estimate them by the principal component analysis of a panel data set. We show that the optimization problem can be reformulated as mixed integer optimization, and we present two alternative computational algorithms. We derive the asymptotic distribution of the resulting estimator under the scheme that the threshold effect shrinks to zero. In particular, we establish a phase transition that describes the effect of first-stage factor estimation as the cross-sectional dimension of panel data increases relative to the time-series dimension. Moreover, we develop bootstrap inference and illustrate our methods via numerical studies.
BibTeX
@article{lee2021factor,
  author = {Sokbae Lee and Yuan Liao and Myung Hwan Seo and Youngki Shin},
  title = {{Factor-driven two-regime regression}},
  journal = {The Annals of Statistics},
  year = {2021},
  doi = {10.1214/20-aos2017},
}
Figure