Least Squares-Based Permutation Tests in Time Series

Technical Summary: Least Squares-Based Permutation Tests in Time Series

Problem Statement
The paper addresses the challenge of conducting valid hypothesis tests for regression parameters and monotone trends in time series settings where observations exhibit weak dependence. Classical Ordinary Least Squares (OLS) inference relies on assumptions of independent and identically distributed (i.i.d.) errors and often Gaussianity. In time series, these assumptions frequently fail due to serial dependence, rendering standard parametric tests (e.g., $t$-tests, $\chi^2$-tests) potentially invalid with inflated Type I error rates.

While permutation tests offer a nonparametric alternative that provides exact finite-sample inference under the "randomization hypothesis" (exchangeability of errors), they are traditionally inapplicable to time series. Permuting covariates in a time-dependent setting destroys the temporal structure and the relationship between covariates and responses. Consequently, the fundamental assumption of exchangeability required for exactness often fails, and the permutation distribution may not approximate the true sampling distribution, leading to asymptotic invalidity.

Methodology
The authors propose a framework for constructing permutation tests that are asymptotically valid for a wide class of stationary, weakly dependent processes (specifically $\alpha$-mixing sequences) while retaining exactness when the randomization hypothesis holds (i.e., when errors are i.i.d.).

The core methodology involves two distinct problem settings, both utilizing studentized test statistics to achieve asymptotic pivotality:

1. Testing Regression Coefficients ($H_0: \beta = 0$):

   • Setup: The paper considers a stationary, $\alpha$-mixing sequence $\{(Y_i, X_i)\}$ where $X_i \in \mathbb{R}^p$. The goal is to test if the unconditional regression coefficient $\beta$ is zero.
   • Issue: The unstudentized statistic $\sqrt{n}\hat{\beta}_n$ has a limiting Gaussian distribution with covariance $\Sigma_X^{-1}\Gamma\Sigma_X^{-1}$, whereas the permutation distribution converges to a Gaussian with covariance $\sigma_Y^2 \Sigma_X^{-1}$. These covariances generally differ under dependence, causing the unstudentized permutation test to be invalid.
   • Solution: The authors construct a studentized statistic by dividing the regression coefficient by a consistent estimator of the long-run variance. Specifically, they define a variance estimator $\hat{\Gamma}_n$ (analogous to Newey-West estimators) that accounts for serial correlation in the residuals. The test statistic is constructed as:
     $$Z_n = \left( \hat{\Gamma}_n^{-1/2} \hat{\Sigma}_X \right) \sqrt{n}\hat{\beta}_n$$
   • Result: Under the null hypothesis, both the true sampling distribution of $Z_n$ and its permutation distribution converge to the standard multivariate normal distribution $N(0, I_p)$, regardless of the underlying dependence structure (provided mixing and moment conditions are met).

2. Testing for Monotone Trend ($H_0^{(m)}$: No monotone trend):

   • Setup: The paper considers a triangular array model $Y_{n,i} = X_{n,i} + \lambda_{n,i}$, where $X_{n,i}$ is a row-wise stationary $\alpha$-mixing array and $\lambda_{n,i} = g(i/n)$ represents a potential monotone trend.
   • Issue: The authors demonstrate that without restrictions on the trend function $g$, the permutation distribution can degenerate (e.g., if the trend grows too rapidly, a single high-leverage observation dominates the test statistic). Furthermore, the variance of the unstudentized test statistic and the permutation distribution generally mismatch under dependence.
   • Solution: The authors impose a restriction that the trend function $g$ is bounded and continuous, preventing extreme leverage issues. They then construct a studentized statistic using a variance estimator $\hat{\tau}_n^2$ that estimates the long-run variance of the error process $X_{n,i}$. The test statistic is:
     $$T_n = \frac{n^{3/2} \hat{\beta}_n}{\sqrt{12} \hat{\tau}_n}$$
   • Result: Similar to the regression case, the studentized statistic and its permutation counterpart both converge to the standard normal distribution $N(0, 1)$ under the null hypothesis, ensuring asymptotic validity.

Key Contributions and Results

• Asymptotic Validity under Dependence: The primary theoretical contribution is proving that permutation tests, when properly studentized, are asymptotically valid for stationary $\alpha$-mixing processes. This extends the applicability of permutation tests beyond the i.i.d. setting to general weakly dependent time series.
• Exactness Preservation: The proposed tests retain the property of exact finite-sample inference when the randomization hypothesis (exchangeability) holds, a feature standard parametric tests lack without Gaussian assumptions.
• Variance Estimation: The paper provides specific constructions for consistent estimators of the long-run variance ($\hat{\Gamma}_n$ and $\hat{\tau}_n^2$) that are robust to the permutation operation, ensuring the permutation distribution matches the true limiting distribution.
• Simulation Evidence: Monte Carlo simulations confirm that the studentized permutation tests control Type I error rates at the nominal level (e.g., $\alpha=0.05$) for various dependent processes (including $m$-dependent and ARMA processes), whereas classical Chi-squared tests and unstudentized permutation tests often fail, exhibiting severe over-rejection or under-rejection depending on the dependence structure and tail behavior.
• Software Implementation: The methods are implemented in the R package permixOLS, which includes worked examples and handles numerical issues such as matrix inversion via singular value decomposition with truncation.

Significance and Claims
The paper claims that its framework provides a robust solution for confirmatory inference in time series regression and trend analysis where standard assumptions fail. The significance lies in bridging the gap between the exactness of randomization tests and the flexibility required for dependent data.

The authors modestly note that while the tests are asymptotically valid, the rate of convergence to the nominal level depends on the tail behavior of the data. In cases of very heavy-tailed distributions, larger sample sizes may be required for the studentized variance estimators to stabilize. However, even in these difficult regimes, the studentized permutation test is shown to be superior to classical alternatives, which can exhibit rejection probabilities far exceeding the nominal level as sample size increases.

Ultimately, the paper establishes that permutation tests can be a valid approach for inference in time series, provided the test statistics are studentized to account for the specific dependence structure, thereby recovering asymptotic pivotality.