Inference for High-Dimensional Regression with a Possible Threshold

arXiv:2404.08105 · econ.EM · Submitted 2024-04-11 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.

Ruth: We're looking at the paper: "Inference for High-Dimensional Regression with a Possible Threshold".

Warren: Here's the gist of the paper. The paper develops a uniform inference theory for high-dimensional slope parameters in threshold regression models, applicable to both cross-sectional and time series data under near-epoch dependence assumptions.

Ruth: The core contribution is establishing a debiased Lasso estimator that achieves uniform convergence over a large class of parameters without needing to pre-specify the existence of a threshold effect, even when the number of covariates grows much faster than the sample size (p>n).

Warren: This contrasts with existing literature primarily focused on high-dimensional linear regression frameworks.

Ruth: The theory is extended to accommodate heteroskedastic non-subgaussian error terms and non-subgaussian covariates in cross-sectional settings.

Title and authors: Ruth: We're still talking about "Inference for High-Dimensional Regression with a Possible Threshold." The authors are trying to solve the problem of estimating slope parameters in these threshold models when you have more variables than you have data points.

Warren: It’s about establishing a uniform inference theory, meaning they’re proving that their estimation and testing methods work consistently across many different parameter settings, not just one specific scenario.

Iris: The paper focuses on the main contribution being the debiased Lasso estimator itself, which is designed to handle this high-dimensional setting without needing to know if a threshold exists.

Ruth: That’s right, they show that this estimator can achieve uniform convergence over a large class of parameters, even when you have p covariates and only n observations where p is much larger than n.

Warren: They also tackle the complexity of the error terms by accommodating heteroskedastic non-subgaussian errors and non-subgaussian covariates in cross-sectional data.

Iris: And they extend this work to time series data, specifically under the near-epoch dependence assumption, which is a pretty advanced requirement for this kind of analysis.

Ruth: This paper really sets a new benchmark because existing literature has mostly focused on debiased methods within high-dimensional linear regression frameworks.

Warren: So, the implication is that you can now apply these powerful debiasing techniques to models where the underlying relationship between variables isn't just straight lines.

Iris: It opens up possibilities for analyzing complex systems where multiple regimes or states might be switching based on a few key variables.

The paper's summary: Ruth: To summarize what they found, the main takeaway is that their debiased Lasso estimator works uniformly across a broad set of parameters without needing to specify the existence of a threshold effect beforehand.

Warren: They show that the asymptotic distributions of tests involving an increasing number of slope parameters are identical whether there is no threshold effect or with a fixed threshold effect present.

Iris: That’s a neat result because it means you can perform uniform inference without needing to commit to one model structure or another, which is really powerful for model selection.

Ruth: Beyond that, they derive oracle inequalities for prediction errors and one estimation errors under those more general conditions we mentioned earlier <ref:2404.08105#pg1,oracle inequalities for prediction errors and>.

Warren: They also prove that the debiased Lasso estimator is asymptotically normal when you look at a linear combination of parameters, which gives us confidence intervals that are asymptotically honest and contract at the optimal rate.

Iris: The time series extension means we can apply this uniform inference theory to high-dimensional time series threshold regression under near-epoch dependence conditions.

Ruth: So, they’ve built a framework that is robust enough to handle high dimensions, complex errors, and non-linear relationships in both cross-sectional and time series data.

Warren: It means we can get reliable estimates and tests even when the number of covariates explodes relative to our sample size.

Iris: The paper really solidifies the idea that we don't need to pre-specify a threshold structure to use these powerful debiasing tools for high-dimensional problems.

The paper's improvements: Ruth: One of the major improvements they propose is moving away from just focusing on linear regression frameworks and applying debiased methods in those simpler settings.

Warren: They show that their method can work when you have non-subgaussian covariates and heteroskedastic error terms, which is a big step in terms of generality for cross-sectional data.

Iris: And they extend this to the time series setting under the near-epoch dependence assumption, which is a necessary extension for those kinds of dependent data structures.

Ruth: The core improvement is establishing that the debiased Lasso estimator achieves uniform convergence over a large class of parameters without needing to pre-specify a threshold effect.

Warren: That’s key because it contrasts with previous work that was heavily focused on linear regression settings, so this broadens the applicability significantly.

Iris: They also improve oracle inequalities for prediction errors and one estimation errors by allowing for those more general conditions on the error terms and covariates <ref:2404.08105#pg1,oracle inequalities for prediction errors and>.

Ruth: So, they’re not just solving one problem; they're making the tools more versatile to handle a wider variety of real-world data complexities.

Conclusion: Ruth: To wrap up, this paper introduces inference for high-dimensional regression with a possible threshold by showing a debiased Lasso estimator that performs uniformly across many parameter settings without needing to pre-specify the threshold.

Warren: The key results are the identical asymptotic distributions of tests in both no threshold and fixed threshold cases, and the uniform normality of the debiased Lasso estimator over a large zero ball <ref:2404.08105#pg2>.

Iris: The time series extension is also important, allowing us to apply this theory to high-dimensional time series models under near-epoch dependence conditions.

Ruth: This paper suggests that we can perform reliable inference in high dimensions even when the data structure is complex and non-linear, without needing a simple pre-defined model type.

Warren: The authors have shown how to handle more general error structures, which means the results apply to a wider range of real-world applications than previously thought.

Iris: Overall, this work provides a solid framework for analyzing complex systems where you have many parameters and complex data dependencies in both cross-sectional and time series settings.

Huazhong University of Science and Technology · Arizona State University

econ.EM

Submitted: 2024-04-11

Updated: 2026-11-12

License: http://creativecommons.org/publicdomain/zero/1.0/

Importance score: 92/100

The gist: The paper develops a uniform inference theory for high-dimensional slope parameters in threshold regression models…

Key concepts

Debiased Lasso Estimator (DTLasso)
This is an improved version of the standard Lasso method used to estimate slope parameters in complex models. It works by 'debiasing' the initial Lasso estimate, which helps ensure that the resulting estimator converges correctly to the true parameter value, even when many variables are involved.
Uniform Inference Theory
This is a mathematical framework that allows researchers to make reliable statistical inferences (like hypothesis tests) about many parameters simultaneously. The theory here ensures these tests behave consistently and predictably across a wide range of possible parameter values.
Near-Epoch Dependence
This assumption is used when analyzing time series data, meaning the current observation is strongly related only to the most recent past observations, rather than the entire history. This assumption helps in developing inference methods specifically for sequential data structures.
Threshold Regression Models
These are statistical models where the relationship between predictors and an outcome changes abruptly once a certain threshold value is crossed. The paper applies its methods to these models, showing they can be handled even when the number of predictors exceeds the sample size.

Terminology

Summary

The paper develops a uniform inference theory for high-dimensional slope parameters in threshold regression models, applicable to both cross-sectional and time series data under near-epoch dependence assumptions. p. 2

Where the paper says it

Specifically, we derive oracle inequalities for prediction errors and ℓ1 estimation errors for the Lasso estimator of the slope parameters and the threshold parameter under more general conditions, allowing for heteroskedastic non-subgaussian error terms and non-subgaussian covariates when studying cross-sectional data. Moreover, we further extend the framework to high-dimensional time series threshold regression models and establish uniform inference theory under the near-epoch dependence assumption.

The core contribution is establishing a debiased Lasso estimator that achieves uniform convergence over a large class of parameters without needing to pre-specify the existence of a threshold effect, even when the number of covariates grows much faster than the sample size (p>n). p. 2
Where the paper says it

The main contribution of this paper is to develop a uniform inference procedure for an increasing number of slope parameters in high-dimensional threshold regression models, a class of parsimonious nonlinear regression models. To the best of our knowledge, we are the first to establish that the debiased Lasso estimator achieves uniform convergence over a large class of parameters without the need to pre-specify the existence of a threshold effect, even when the number of covariates grows much faster than the sample size.

This contrasts with existing literature primarily focused on high-dimensional linear regression frameworks. p. 2
Where the paper says it

To the best of our knowledge, we are the first to establish that the debiased Lasso estimator achieves uniform convergence over a large class of parameters without the need to pre-specify the existence of a threshold effect, even when the number of covariates grows much faster than the sample size. In contrast, the existing literature has primarily focused on applying debiased methods within highdimensional linear regression frameworks.

The theory is extended to accommodate heteroskedastic non-subgaussian error terms and non-subgaussian covariates in cross-sectional settings. p. 0

Where the paper says it

Uniform Inference in High-Dimensional Threshold Regression Models* Jiatong Li† Hongqiang Yan ‡ September 16, 2025 Abstract We develop a uniform inference theory for high-dimensional slope parameters in threshold regression models, allowing for either cross-sectional or time series data. We first establish oracle inequalities for prediction errors, and ℓ1 estimation errors for the Lasso estimator of the slope parameters and the threshold parameter, accommodating heteroskedastic non-subgaussian error terms and non-subgaussian covariates.

Furthermore, it extends to time series data under the near-epoch dependence assumption. p. 0
Where the paper says it

We show that the asymptotic distributions in both cases are the same, allowing us to perform uniform inference without specifying whether the model is a linear or threshold regression. Additionally, we extend the theory to accommodate time series data under the near-epoch dependence assumption.

The results are shown to hold uniformly over the parameter space B 0(s 0), defined by constraints on the 0 norm of parameters (| alpha 0| infinity at most C) and a bound on their magnitude (alpha 0 0 1 at most s 0).

Key findings include:

  1. Uniform Inference: The asymptotic distributions of tests involving an increasing number of slope parameters are shown to be identical in cases with no threshold effect and with a fixed threshold effect, allowing for uniform inference without specifying the model type (linear vs. threshold regression). p. 0

    Where the paper says it

    Uniform Inference in High-Dimensional Threshold Regression Models* Jiatong Li† Hongqiang Yan ‡ September 16, 2025 Abstract We develop a uniform inference theory for high-dimensional slope parameters in threshold regression models, allowing for either cross-sectional or time series data. We first establish oracle inequalities for prediction errors, and ℓ1 estimation errors for the Lasso estimator of the slope parameters and the threshold parameter, accommodating heteroskedastic non-subgaussian error terms and non-subgaussian covariates.

  2. Oracle Inequalities: Oracle inequalities for prediction errors and 1 estimation errors are derived under more general conditions, accommodating heteroskedastic non-subgaussian error terms and non-subgaussian covariates in cross-sectional data.

  3. Asymptotic Normality: The debiased Lasso estimator is shown to be asymptotically normal under various conditions, demonstrating that the convergence of a linear combination of parameters to the standard normal distribution is uniformly valid over the 0-ball with a radius at most s 0. These confidence intervals are asymptotically honest and contract at the optimal rate. p. 3

    Where the paper says it

    We also provide a uniformly consistent covariance matrix estimator in both cases. 6 We further construct asymptotically valid confidence intervals for the interest of the slope parameter, which are uniformly valid and contract at the optimal rate.

  4. Time Series Extension: Uniform inference theory for the debiased Lasso estimator is established in high-dimensional time series threshold regression under near-epoch dependence, supported by oracle inequalities derived using concentration inequalities for dependent data (Lemma A.27) and mixingale concentration inequalities (Lemma A. p. 3

    Where the paper says it

    We use the concentration inequality 4 for the partial sum of random variables that we propose in Lemma A. 2 to derive oracle inequalities for both the prediction error and ℓ1 estimation error for coefficients, which are qualitatively same as those in Lee et al.

    28).

The paper proposes a debiased Lasso estimator (DTLasso) for high-dimensional slope parameters in threshold regression models, applicable to both cross-sectional and time series data. p. 0

Where the paper says it

Uniform Inference in High-Dimensional Threshold Regression Models* Jiatong Li† Hongqiang Yan ‡ September 16, 2025 Abstract We develop a uniform inference theory for high-dimensional slope parameters in threshold regression models, allowing for either cross-sectional or time series data. We first establish oracle inequalities for prediction errors, and ℓ1 estimation errors for the Lasso estimator of the slope parameters and the threshold parameter, accommodating heteroskedastic non-subgaussian error terms and non-subgaussian covariates.

Monte Carlo simulations compare DTLasso against linear regression estimators, showing comparable performance even when a fixed threshold effect does not exist. p. 3
Where the paper says it

to empirical applications, due to sample splitting, the total number of parameters may be larger than the sample size in the regime with the fewest observations, particularly when multiple threshold points exist, leading to poor estimation and out-of-sample prediction in finite samples. However, traditional estimation and inferential methods, such as OLS and MLE, are no longer valid even in high-dimensional linear regression models.

Empirical applications are provided related to multiple steady states in economic growth models and the effects of military news shocks on government spending and GDP. p. 1
Where the paper says it

In this paper, we confirm the existence of multiple steady states in cross-country economic growth by showing that some threshold-effect coefficients are significantly different from zero. In addition, we apply the high-dimensional local projection threshold model to find a “data-driven” threshold point that defines the state of the economy and reestimate the impulse response to a military spending news shock in government spending and GDP.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Uniform Inference Theory for Threshold Regression Models: The AI system can perform uniform inference for high-dimensional slope parameters in threshold regression models, allowing for either cross-sectional or time series data without specifying whether the model is linear or threshold regression.

  2. High Dimensional Parameter Estimation with Sparsity: The system can estimate the Lasso estimator of slope parameters and the threshold parameter under more general conditions, accommodating heteroskedastic non-subgaussian error terms and non-subgaussian covariates, even when the number of covariates grows much faster than the sample size.

  3. Asymptotic Distribution of Tests: The AI system can derive the asymptotic distribution of tests involving an increasing number of slope parameters by debiasing (or desparsifying) the Lasso estimator in cases with no threshold effect and with a fixed threshold effect, showing that their asymptotic distributions are identical.

  4. Time Series Threshold Model Inference: The system can develop uniform inference theory for high-dimensional time series threshold regression models under the near-epoch dependence assumption, extending existing methods to handle dependent data structures.

  5. Local Projection Inference for Impulse Response Functions: The system can develop uniform inference theory for debiased impulse response parameters in the high-dimensional local projection threshold model, allowing researchers to estimate specific response functions at different horizons.

Abstract

This paper develops uniform inference for slope coefficients in high-dimensional threshold regressions without pretesting for a threshold effect. Confidence intervals based on a debiased threshold Lasso estimator are uniformly asymptotically valid, allowing for heteroskedasticity and non-sub-Gaussian covariates and errors. We also propose a threshold-effect test and extend inference to stationary time series. Applying the method to racial tipping in U.S. neighborhoods, we find that the relationships between white population change and residential turnover, housing age, commuting schedules, and utility payments differ across minority-share regimes. Threshold effects are detected in fewer metropolitan areas in 2000--2010 than in 1990--2000.

Related papers