Asymptotics for EBLUPs within crossed mixed effect models

Ziyang Lyu, Scott A. Sisson, A. H. WelsharXiv 2025 · DOI · arXiv

TL;DR

This paper provides the first rigorous theoretical foundation for using Empirical Best Linear Unbiased Predictors (EBLUPs) in large-scale crossed mixed effect models. By proving that these predictors follow a predictable asymptotic distribution—even without assuming the data is normally distributed—the authors introduce a computationally efficient way to calculate prediction intervals for complex, high-dimensional datasets like movie ratings or social networks.

What Problem Does This Paper Address?

In statistics, mixed effect models are the workhorse for data with dependencies, such as multiple test scores from the same student or multiple ratings from the same user. While "nested" models (e.g., students within schools) are well-understood, "crossed" models (e.g., every user can rate every movie) are much harder to analyze.

The standard tool for predicting these unobserved "random effects" is the EBLUP. However, most existing theory for EBLUPs assumes the data follows a perfect bell curve (normality) and focuses on nested structures. In crossed designs, calculating the uncertainty (Mean Squared Error or MSE) of these predictions is notoriously difficult. Existing methods, like the Kackar-Harville or Prasad-Rao estimators, require massive matrix inversions that can crash a standard computer when dealing with modern, large-scale datasets. This paper fills that gap by providing a simplified, theoretically justified framework for EBLUPs in crossed designs.

Methods

The authors employ asymptotic analysis, studying what happens to the predictors as the number of rows ($g$), columns ($h$), and observations per cell ($m$) all tend toward infinity.

Key technical contributions include:

  • Linear Approximations: They derive linear approximations for EBLUPs, showing that as the dataset grows, the EBLUP behaves like the true random effect plus a small, normally distributed error.
  • Central Limit Theorems (CLTs): They prove that the difference between the EBLUP and the true random effect satisfies a CLT. This allows them to describe the EBLUP's distribution as a convolution of the true random effect distribution and a normal distribution.
  • The LSW Estimator: They propose a new, simplified MSE estimator (referred to here as LSW). Unlike older methods that require manipulating $N \times N$ matrices, the LSW formulas are simple, interpretable, and computationally "cheap."

Datasets

The paper validates its theoretical findings using two approaches:

  1. Simulated Data: The authors ran 36 scenarios for models without interactions and 32 scenarios for models with interactions. They tested both normal distributions and "mixture" distributions (to simulate non-normal, real-world data) across various grid sizes (from $10 \times 10$ to $100 \times 100$).
  2. MovieLens Dataset: A subset of the famous MovieLens dataset (via Kaggle) consisting of 8,316 ratings from 126 customers across 66 movies. This served as a real-world test case for generating prediction intervals for "customer effects" and "movie effects."

Key Insights and Ideas

  • Normality is Optional: One of the most important "aha" moments is that the Central Limit Theorems hold under "mild moment conditions." This means researchers can safely use these prediction intervals even if their random effects don't follow a Gaussian distribution.
  • Computational Efficiency: The authors highlight that standard MSE estimators often lead to "memory exhaustion" in R or Python because they involve high-dimensional matrix operations. Their proposed LSW estimator avoids this, making it viable for "Big Data" applications.
  • The "Blurring" Effect: The paper elegantly shows that an EBLUP is essentially a "blurred" version of the true random effect. As you get more data, the "blur" (the normal noise component) shrinks, and the EBLUP converges to the true value.
  • Asymmetric Insights: In the movie rating analysis, the authors found that movie random effects were asymmetric. It is much easier to identify the "worst" movies with certainty than it is to distinguish between the "best" movies, which tend to cluster together in the upper tail.

Key Results

  • Theoretical Proofs: The paper successfully derives the joint asymptotic distribution for individual, interaction, and cell-level random effects.
  • Superior Coverage: In simulations, the LSW prediction intervals maintained an empirical coverage very close to the nominal 95% level, even in small samples ($10 \times 10$).
  • Conservative Safety: The LSW intervals were found to be slightly more conservative (wider) than the older Kackar-Harville (KH) and Prasad-Rao (PR) methods, which the authors argue is a safer "good practice" for providing measures of uncertainty.
  • Scalability: The movie rating analysis demonstrated that the method can easily handle thousands of observations and provide clear, visual diagnostic tools (like the prediction intervals superimposed on QQ-plots).

Limitations and Future Work

  • Balanced Designs: The current proofs rely on "balanced" data (where every cell has the same number of observations). The authors note that while their results likely extend to unbalanced data, the math becomes significantly more complex because there is no explicit form for the inverse variance-covariance matrix in those cases.
  • Linearity: The model assumes a linear relationship for the fixed effects. Future work could explore how these asymptotics hold up in generalized linear mixed models (GLMMs) or non-linear structures.
  • Missing Data: The paper does not explicitly address how to handle missingness patterns that are common in crossed designs (e.g., most users only rate a tiny fraction of available movies).

Summary updated

Abstract

In this article, we derive the joint asymptotic distribution of empirical best linear unbiased predictors (EBLUPs) for individual and cell-level random effects in a crossed mixed effect model. Under mild conditions (which include moment conditions instead of normality for the random effects and model errors), we demonstrate that as the sizes of rows, columns, and, when we include interactions, cells simultaneously increase to infinity, the distribution of the differences between the EBLUPs and the random effects satisfy central limit theorems. These central limit theorems mean the EBLUPs asymptotically follow the convolution of the true random effect distribution and a normal distribution. Moreover, our results enable simple asymptotic approximations and estimators for the mean squared error (MSE) of the EBLUPs, which in turn facilitates the construction of asymptotic prediction intervals for the unobserved random effects. We show in simulations that our simple estimator of the MSE of the EBLUPs works very well in finite samples. Finally, we illustrate the use of the asymptotic prediction intervals with an analysis of movie rating data.
BibTeX
@misc{lyu2025asymptotics,
  author = {Ziyang Lyu and Scott A. Sisson and A. H. Welsh},
  title = {{Asymptotics for EBLUPs within crossed mixed effect models}},
  year = {2025},
  eprint = {2502.04723},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2502.04723},
  howpublished = {arXiv},
}
Figure