Linear Multidimensional Regression with Interactive Fixed-Effects

This paper proposes a Neyman-orthogonal estimator for linear multidimensional panel data with interactive fixed-effects that combines factor model methods with a weighted-within transformation to achieve parametric consistency and asymptotic normality, demonstrated through an application to beer demand elasticity.

The Gist

Imagine you are trying to figure out how much people's beer drinking habits change when the price goes up. This is a classic economics problem: Demand Elasticity.

But here's the catch: you aren't just looking at a simple list of prices and quantities. You have a massive, multi-layered dataset. You have data for:

1. Different Products (Miller Lite, Budweiser, etc.)
2. Different Stores (Whole Foods, local gas stations, etc.)
3. Different Times (Week 1, Week 2, etc.)

This is a 3D (or more) puzzle.

The Problem: The "Ghost" in the Machine

In this world, there are invisible forces affecting your data. Let's call them "The Ghosts."

• Maybe a huge NBA playoff game happened in Chicago.
• This game made people in Store A buy more Miller Lite, but people in Store B buy more Budweiser.
• This effect changed over Time (during the game vs. after).

These "Ghosts" are Interactive Fixed Effects. They are unobserved, complex interactions between products, stores, and time. If you don't account for them, your math will be wrong. You might think the price caused a change in demand, when really it was just the NBA game.

The Old Way (The "Additive" Approach):
Previous methods tried to handle this by looking at each dimension separately. They said, "Okay, let's just subtract the average effect of the product, the average effect of the store, and the average effect of the time."

• Analogy: Imagine trying to clean a muddy window by wiping it with a straight, horizontal stroke, then a vertical stroke. You miss the diagonal smudges. The "Ghosts" that move diagonally (interacting across all three dimensions) remain, ruining your view.

The New Way (The "Interactive" Approach):
This paper introduces a new tool to clean the window perfectly, removing even the most complex, diagonal smudges.

The Solution: The "Smart Filter"

The author, Hugo Freeman, proposes a two-step process to get the true answer.

Step 1: The "Rough Sketch" (The Matrix Method)

First, the author suggests flattening the 3D data into a 2D sheet (like turning a Rubik's cube into a flat puzzle). You can use standard statistical tools on this flat sheet to get a rough guess of what the "Ghosts" are doing.

• The Catch: This rough guess is slow and a bit blurry. It's like looking at a low-resolution photo. It helps, but it's not good enough for a final verdict.

Step 2: The "Weighted-Smart Filter" (The Innovation)

This is the paper's big breakthrough. Instead of just taking a simple average to remove the "Ghosts," the new method uses a Weighted Filter.

• The Analogy: Imagine you are trying to remove a stain from a shirt.
  • Old Method: You scrub the whole shirt with the same amount of soap.
  • New Method: You look at the stain. You realize the stain is stronger in some spots and weaker in others. You apply more soap to the heavy spots and less soap to the light spots, based on how similar the fabric fibers are to the stain.

In math terms, the author uses Kernel Weights. If a specific store's sales pattern looks very similar to a "Ghost" pattern, the filter gives it a heavy weight to remove it. If it looks different, it gives it a light weight.

This allows the model to project out (remove) the complex, interactive "Ghosts" so cleanly that the remaining data shows the true relationship between price and demand.

Why Does This Matter? (The Beer Test)

The author tested this on real beer sales data from Chicago (1991–1995).

1. The Old Way (Factor Models): When they tried to flatten the data into 2D, the results were all over the place. Depending on how they arranged the data (products as rows vs. stores as rows), they got completely different answers. It was like a compass spinning wildly.
2. The New Way (Weighted Filter): The new method gave a very clear, precise answer. It showed that beer demand is highly elastic (people stop buying it quickly if the price goes up), with a result very close to the "gold standard" but with much higher confidence.

The "Double Debias" Trick

To make sure the math holds up perfectly, the author also uses a trick called Neyman Orthogonality (or "Double Debiasing").

• Analogy: Imagine you are trying to measure the speed of a car, but your speedometer is slightly broken. Instead of just fixing the speedometer, you build a second, independent speedometer. You use the first one to guess the error, and the second one to correct it. By combining them, you cancel out the errors, leaving you with a perfect speed reading.

Summary

• The Problem: Multi-dimensional data (Product x Store x Time) has hidden, complex patterns that mess up standard math.
• The Old Fix: Tried to flatten the data, but it was too slow and sensitive to how you arranged the puzzle pieces.
• The New Fix: A Weighted Filter that intelligently removes these hidden patterns by looking at how similar different data points are to each other.
• The Result: We can now measure things like "beer demand" with much higher precision and less confusion, even when the data is messy and multi-layered.

In short, this paper gives economists a sharper, smarter lens to see the truth in a world of complex, multi-dimensional data.

Technical Summary

Here is a detailed technical summary of the paper "Linear Multidimensional Regression with Interactive Fixed-Effects" by Hugo Freeman.

1. Problem Statement

The paper addresses the challenge of estimating linear regression models using multidimensional panel data (data with three or more dimensions, e.g., product $i$, store $j$, and time $t$) in the presence of unobserved interactive fixed-effects.

• Limitations of Existing Methods: Standard additive fixed-effects (e.g., $\alpha_i + \gamma_j + \delta_t$) can only control for heterogeneity varying along subsets of dimensions. They fail to capture interactive heterogeneity that varies jointly across all dimensions (e.g., a shock specific to a product-store-time combination).
• The Model: The paper considers the model:
  $$Y_{ijt} = X'_{ijt}\beta + \sum_{\ell=1}^L \lambda_{i\ell}\delta_{j\ell}\gamma_{t\ell} + \varepsilon_{ijt}$$
  where the interactive term is a sum of outer products of unobserved factors.
• Core Challenges:
  1. Tensor Rank Issues: Unlike matrices, the low-rank approximation problem for tensors (multidimensional arrays) is ill-posed (De Silva & Lim, 2008). There is no unique singular value decomposition (SVD) for tensors, making standard factor model extensions difficult.
  2. Slow Convergence: Simply flattening the tensor into a 2D matrix and applying standard factor methods (e.g., Bai, 2009) leads to consistent estimates but at slow convergence rates (e.g., $N^{-1/6}$), which is insufficient for standard inference.
  3. Endogeneity: The regressors $X$ may be arbitrarily correlated with the interactive fixed effects.

2. Methodology

The author proposes a two-step Neyman-orthogonal approach to achieve parametric convergence rates and asymptotic normality.

Step 1: Preliminary Matrix Estimation (Slow Convergence)

The multidimensional array is "flattened" (matricized) along one dimension to form a standard 2D panel.

• Technique: Apply standard factor model estimation (Bai, 2009; Moon & Weidner, 2015) to the flattened data.
• Result: This yields consistent estimates of the fixed effects but with a slow convergence rate. These estimates serve as proxies for the true unobserved factors.
• Assumption: At least one dimension of the data must admit a low-rank structure (low multilinear rank) for this step to work.

Step 2: Weighted-Within Transformation (Novel Contribution)

To correct the slow convergence and bias, the paper introduces a weighted-within transformation.

• Concept: Instead of subtracting uniform means (standard within transformation), the method subtracts weighted means. The weights are constructed using kernel functions based on the distance between the proxies of the fixed effects estimated in Step 1.
• Mechanism:
  $$\check{Y}_{ijt} = Y_{ijt} - \bar{Y}_{i^*jt} - \bar{Y}_{ij^*t} - \bar{Y}_{ijt^*} + \dots$$
  where $\bar{Y}_{i^*jt} = \sum_{i'} w_{i,i'} Y_{i'jt}$. The weights $w_{i,i'}$ are determined by a kernel $k(\cdot)$ applied to the distance between estimated factor proxies.
• Neyman Orthogonality: The estimator is constructed to be Neyman-orthogonal with respect to the nuisance parameters (the fixed effects). This means that small errors in the preliminary estimation of the fixed effects (Step 1) only affect the final estimator $\hat{\beta}$ at a second-order level, allowing the estimator to achieve the parametric rate of convergence ($N^{-1/2}$).

Inference Correction (Double De-bias)

To ensure asymptotic normality, the paper employs a double de-bias (or inference-corrected) procedure:

1. Estimate $\beta$ using the weighted-within transformation.
2. Correct for the remaining bias arising from the estimation of the fixed effects using sample splitting or cross-fitting techniques to break dependence between the nuisance parameter estimates and the error terms.

3. Key Contributions

1. Theoretical Framework for Multidimensional Data: The paper extends the interactive fixed-effects literature from 2D panels to $d$-dimensional tensors, addressing the mathematical ill-posedness of tensor rank approximation.
2. Weighted-Within Estimator: It introduces a novel estimator that uses kernel-weighted means to project out interactive fixed effects. This method is robust to the specific dimension chosen for flattening and does not require the analyst to know which dimensions are low-rank, only that at least one is.
3. Parametric Rate and Asymptotic Normality: The paper proves that combining the preliminary matrix estimator with the weighted-within transformation and a double-debias correction achieves the parametric convergence rate ($\sqrt{N}(\hat{\beta} - \beta_0) \xrightarrow{d} N(0, \Sigma)$), a result previously unattainable for 3D+ interactive fixed effects.
4. Robustness to Model Specification: Unlike standard factor models which are highly sensitive to how the tensor is flattened (e.g., treating products as rows vs. stores as rows), the weighted estimator provides stable results regardless of the data arrangement.

4. Results

Theoretical Results

• Proposition 1: Establishes that standard matrix methods on flattened data provide consistent but slow-converging estimates ($O_p(N^{-1/6})$ in 3D).
• Proposition 2: Shows that the kernel-weighted estimator's error is bounded by the convergence rate of the proxy estimates. Under regularity conditions, it achieves the parametric rate.
• Theorems 1-3: Prove asymptotic normality under homoskedasticity, heteroskedasticity, and correlation across dimensions, provided the bias from the fixed-effect estimation is sufficiently small (achieved via the double-debias/sample-splitting).

Simulation Results

• Convergence: Simulations confirm that while standard "Factor" methods (flattening the tensor) converge slowly and exhibit significant finite-sample bias, the Weighted-Within estimator converges rapidly with negligible bias.
• Robustness: When the multilinear rank is heterogeneous (low rank in one dimension, high in others), standard factor methods fail if the wrong dimension is flattened. The weighted estimator remains robust and accurate.
• Coverage: The inference-corrected weighted estimator achieves nominal coverage rates (e.g., 95%) as sample size increases, whereas standard factor methods often undercover.

Empirical Application: Beer Demand Elasticity

• Data: Dominick's supermarket data (Chicago, 1991-1995) with dimensions: Product, Store, Fortnight.
• Goal: Estimate price elasticity of demand for beer.
• Findings:
  • Pooled OLS: Positive elasticity (incorrect).
  • Additive Fixed Effects: Negative but imprecise.
  • Instrumental Variables (IV): Negative (-3.39) but with very large standard errors due to weak instruments (barley price varies only over time).
  • Standard Factor Models: Highly sensitive to data arrangement (estimates ranged from -0.03 to -2.78 depending on which dimension was flattened).
  • Weighted-Within Estimator: Estimated elasticity of -3.12 with much tighter confidence intervals than IV or factor models. The result is consistent with established literature (Hausman et al., 1994) but with superior precision.

5. Significance

This paper provides a critical tool for modern econometrics where data is increasingly multidimensional (e.g., spatial-temporal, product-market-time).

• Solves the "Curse of Dimensionality" in Fixed Effects: It offers a way to control for complex, interactive unobserved heterogeneity without requiring explicit observation of the shocks.
• Improves Precision: By achieving parametric rates, it allows for valid hypothesis testing in settings where previous methods were too slow or biased.
• Practical Utility: The empirical application demonstrates that ignoring the multidimensional interactive nature of shocks leads to either biased estimates (additive effects) or imprecise ones (IV), whereas the proposed method recovers economically meaningful parameters with high precision.

In summary, Freeman bridges the gap between tensor algebra and econometric inference, providing a robust, theoretically grounded, and practically effective estimator for high-dimensional panel data with interactive fixed effects.