Local Projections Inference with High-Dimensional Covariates without Exact Sparsity

summary

The gist

This paper presents a novel framework for local projections that incorporates high-dimensional covariates without relying on sparsity assumptions…

In short

The paper introduces a new method for local projections that handles many covariates without assuming they are sparse (i.e., exactly zero). It uses an adaptation of the Orthogonal Greedy Algorithm with a high-dimensional model selection tool to select the best predictors. This provides a robust framework for economic analysis, improving efficiency and reliability in both sparse and dense datasets.

Key concepts

Local Projections
This is an econometric technique used to estimate how a variable reacts to shocks at specific points in time. Instead of looking at the entire dataset, it focuses on estimating the effect around a particular data point by using local regression models.
High-Dimensional Covariates without Exact Sparsity
This refers to using many predictor variables (covariates) in a regression model where it is not assumed that most of these predictors are zero. The method is designed to work well even when the dataset has a large number of potential predictors, making it flexible for dense data.
OGA+HDAIC
This is a specific model selection method used to choose which covariates should be included in the local projection. It combines the Orthogonal Greedy Algorithm with a High-Dimensional Akaike Information Criterion, helping to select relevant predictors effectively.
Asymptotic Normality and HAC Standard Errors
These are statistical properties that describe how well the estimator behaves as the sample size gets very large. The paper proves that under certain assumptions, the local projection estimates follow a normal distribution, which allows researchers to calculate reliable confidence intervals.

Terminology used across episodes

This episode discusses

The paper

Local Projections Inference with High-Dimensional Covariates without Exact Sparsity · Read on arXiv

Vanderbilt University

This paper develops a high-dimensional local projections framework for estimating impulse responses in the presence of many potential controls. Other approaches often rely on exact sparsity assumptions, but empirical macroeconomic applications may involve nuisance relationships better described by structured coefficient decay. Building on the orthogonal greedy algorithm with high-dimensional AIC (OGA+HDAIC), I propose a double-selection procedure that separately addresses the shock equation and the outcome equation within local projections. The method allows for nonexact sparsity through polynomially decaying nuisance coefficients, with rates indexed by the decay exponent rather than by an effective sparsity level. Under suitable moment, dependence, and approximation conditions, the estimator achieves asymptotic normality with HAC standard errors. Simulation evidence shows reliable coverage under shock contamination and lower point-estimation error under predictable high-dimensional outcome nuisance. I illustrate the method's practical feasibility in a monetary policy application using the FRED-MD dataset.

Transcript

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

Ruth: I'm Ruth.

Warren: And I'm Warren, here with Iris.

Ruth: We're looking at the paper: "Local Projections Inference with High-Dimensional Covariates without Exact Sparsity".

Warren: Here's the gist of the paper. This paper presents a novel framework for local projections that incorporates high-dimensional covariates without relying on sparsity assumptions, offering robustness in both sparse and dense scenarios.

Ruth: The proposed method adapts the Orthogonal Greedy Algorithm with a High-Dimensional Akaike Information Criterion (OGA+HDAIC) model selection method to local projections.

Warren: Pruning redundant reasoning tokens The approach is applicable to various settings, including impulse response analysis and difference-in-differences (DiD) estimation.

Ruth: While methods like LASSO exist, they often rely on sparsity assumptions—most parameters are exactly zero—limiting their effectiveness in dense data generation processes.

Title and authors: Ruth: Now, let's talk about the actual people behind this work. The paper is titled "Local Projections Inference with High-Dimensional Covariates without Exact Sparsity," and the authors are Jooyoung Cha and their team.

Warren: They bring a solid background because they adapted Ing’s OGA+HDAIC method, which is a key part of this new selection technique for local projections.

Iris: Their work builds on previous ideas, extending the OGA+HDAIC approach from cross-sectional double machine learning frameworks to the time series context.

Ruth: So, what’s the main idea they are pushing here? It seems like they're moving away from that strict sparsity requirement that limits LASSO in dense data generation processes.

Warren: They suggest a more flexible model where parameters can be non-zero, but they decay toward zero as the dimensionality grows, which is naturally satisfied by autoregressive processes in time series.

Iris: And theoretically, they leverage the triplex inequality from Jiang and the results from Ing to derive their inference for this new framework.

Ruth: So, we’re seeing a move toward a more versatile tool for analyzing these kinds of economic data sets that aren't perfectly sparse. Where does that leave us when we think about estimating things like impulse responses?

Warren: It means local projections become much more practical because they can handle the complexity of modern data without immediately demanding that every single control variable is zero.

The paper's summary: Ruth: To summarize what this paper is doing, it presents a comprehensive framework for estimating future responses to current shocks using local projections. They are specifically focusing on how to incorporate high-dimensional controls robustly without relying on strict sparsity assumptions.

Warren: Essentially, they propose using the Orthogonal Greedy Algorithm with the OGA+HDAIC model selection method to select those covariates for the local projection regression.

Iris: The training signal for this is built using an h-step ahead local projection regression model, where the parameter of interest, beta h, is defined as the difference between conditional expectations at different data points.

Ruth: That sounds technical—how does that specific definition of beta h help them select covariates? What’s the mechanism there?

Warren: They use a double selection method with a model selection method called OGA+HDAIC by Ing, which involves two steps to select covariates from regression models. The final local projection regression then uses the union set of all those selected covariates as their input.

Iris: This whole process is built on explicit assumptions about the dependence structure, specifically using the near epoch dependence assumption that Adamek and his team employed for time series analysis.

Ruth: So, they’re defining what they need to estimate based on conditional expectations across different time points, and then using a sophisticated selection algorithm to build their final regression set.

The paper's improvements: Warren: The main gain they show is significant efficiency gains over the conventional local projections approach, especially when you're looking at longer horizon estimates for your impulse response analysis.

Ruth: And simulation studies back that up, showing superior performance in dense and persistent scenarios when compared to both standard local projection methods and LASSO-based approaches.

Iris: Empirically, they applied this method to the Acemoglu et al. study on democratization on GDP growth, and it showed efficiency gains plus robustness even when dealing with a large set of controls.

Warren: That’s significant because those empirical results suggest that this framework can handle the real-world complexity of many observable factors better than what we've seen before.

Ruth: Beyond just efficiency, they also address how inflation behaves in different economic states, showing that it’s more rational in good states but pessimistic in bad states due to subjective beliefs.

Iris: And they establish their theoretical foundation using the error bounds from Ing and the triplex inequality from Jiang to prove their statistical properties under assumptions one through four.

Conclusion: Ruth: So, wrapping up this discussion on "Local Projections Inference with High-Dimensional Covariates without Exact Sparsity," the paper shows that we can build a local projection framework that is robust to high-dimensional controls whether your data is sparse or dense.

Warren: The main implication for us as practitioners is that this offers a more practical tool for both impulse response analysis and difference-in-differences estimation, especially when dealing with complex datasets.

Iris: Theoretically, they give us root T asymptotic normality with HAC standard errors under their stated assumptions, which provides a solid foundation for the inference they’re making about these estimates.

Ruth: And empirically, the results suggest that local projection is a preferred alternative to VARs because of its robustness to model misspecification and ease of implementation in many settings.

Warren: The method advances econometric modeling by offering a practical solution for high-dimensional datasets where we used to be limited by sparsity assumptions.

Iris: To add one final thought, the authors flag that this method relies on assumptions one through four, so you need to check those before applying it broadly.

Ruth: That's all on this paper from "Local Projections Inference with High-Dimensional Covariates without Exact Sparsity." We’ll be looking at other papers next week.

More episodes

← Home