Marginal asymptotics for the "large p, small n" paradigm: with applications to microarray data

Michael R. Kosorok, Shuangge Ma

Abstract

The "large $p$, small $n$" paradigm is a hallmark of modern microarray studies, where the expression levels of thousands of genes ($p$) are monitored across a relatively small number of subjects ($n$). This dimensionality creates significant challenges for traditional statistical inference. This paper establishes the uniform consistency of marginal distribution functions, sample means, and sample medians under the "large $p$, small $n$" assumption. Furthermore, it demonstrates the uniform consistency of marginal $p$-values based on asymptotic approximations, facilitating robust inference via false discovery rate (FDR) techniques. The study also investigates the impact of data normalization on these asymptotic results, supported by simulation studies and an analysis of breast cancer microarray data.

Introduction

In genomic research, the "curse of dimensionality" is a practical reality. Researchers frequently encounter datasets where the number of variables (genes) far exceeds the number of observations (patients). A typical experiment might involve 10,000 genes but only 20 to 50 samples. Traditional statistical theory often relies on the assumption that $n$ is much larger than $p$, or that $p$ remains fixed as $n$ grows.

This paper addresses a critical question: Can we trust the $p$-values and statistical estimators used to identify differentially expressed genes when $p$ is massive and $n$ is small? By employing "marginal asymptotics," the authors provide a rigorous mathematical foundation for high-dimensional genomic analysis, proving that statistical reliability can be maintained even when $p$ grows almost exponentially relative to $n$.

Theoretical Framework: Marginal Asymptotics

The authors focus on the statistical behavior of individual genes across the entire set, establishing that estimates remain accurate across all genes simultaneously—a property known as uniform consistency. To prove this, the research utilizes several advanced mathematical frameworks:

  • Orlicz Norms and Maximal Inequalities: These tools are used to bound the maximum error across the entire gene set, ensuring that the "worst-case" estimation error vanishes as the sample size increases.
  • Hungarian Construction: This technique demonstrates that the empirical distribution of gene data can be closely approximated by a Brownian bridge (a specific stochastic process), allowing for precise $p$-value approximations.
  • Partial Linear Models: These models are used to account for the normalization process. In microarray experiments, array-specific technical noise (e.g., printing variations) must be separated from actual biological signals. The authors show that the large number of genes actually aids this process; because there are so many data points per array, technical noise can be estimated and removed with high precision.

Key Insights

The "Large p" Advantage

Counter-intuitively, the study finds that having a massive number of genes ($p$) is beneficial for normalization. The abundance of data points allows for a more accurate estimation of the "nuisance parameters" associated with technical variation. The authors prove that as long as $\log p = o(n)$, the errors introduced by normalization do not compromise the consistency of the subsequent gene-specific tests.

Robustness of Non-Parametric Tests

The paper evaluates both parametric tests (like the $t$-test) and non-parametric tests (such as the Wilcoxon rank-sum and Kolmogorov-Smirnov tests). A significant finding is that non-parametric rank tests are "distribution-free" and remain simultaneously consistent regardless of the growth rate of $p$, provided the underlying data distributions are continuous. This makes them highly reliable for genomic data, which often deviates from normality.

Empirical Validation

Simulation Studies

The authors conducted simulations with $p = 2000$ genes and sample sizes of $n = 20$ and $n = 50$. They explored three dependency scenarios: independent genes, genes with strong dependencies, and genes with weak dependencies.

  • At $n = 20$, mean-based approaches successfully identified true positives but suffered from higher false-positive rates.
  • At $n = 50$, median and rank-based tests showed superior accuracy and very low false-positive rates, confirming that the asymptotic properties begin to take effect even at moderate sample sizes.

Real-World Application: Estrogen Data

The methodology was applied to a breast cancer dataset containing expression levels for 7,129 genes across 49 tumor samples (25 lymph-node positive and 24 negative). The analysis compared different statistical methods:

  • Mean-based: 445 significant genes identified.
  • Wilcoxon: 423 significant genes identified.
  • Median-based: 261 significant genes identified.
  • Kolmogorov-Smirnov: 211 significant genes identified.

Despite the variation in the total number of genes identified, there was substantial overlap (e.g., 196 genes common to both mean and median approaches), validating the practical utility of the proposed theoretical framework in identifying biologically relevant markers.

Conclusion and Future Directions

This research provides the necessary theoretical justification for using standard marginal statistics in the "large $p$, small $n$" context. It confirms that uniform consistency is achievable, which is a prerequisite for the valid application of False Discovery Rate (FDR) control methods.

While the paper establishes the asymptotic limits, future work is needed to develop specific guidelines for the minimum sample size $n$ required for various levels of $p$. Additionally, while the current theory allows for arbitrary dependencies, incorporating specific gene-interaction network models could further refine the accuracy of these asymptotic analyses. Finally, as genomic technologies evolve, the impact of more complex, multi-stage normalization processes will require further investigation.

Summary updated

Abstract

The "large p, small n" paradigm arises in microarray studies, where expression levels of thousands of genes are monitored for a small number of subjects. There has been an increasing demand for study of asymptotics for the various statistical models and methodologies using genomic data. In this article, we focus on one-sample and two-sample microarray experiments, where the goal is to identify significantly differentially expressed genes. We establish uniform consistency of certain estimators of marginal distribution functions, sample means and sample medians under the large p small n assumption. We also establish uniform consistency of marginal p-values based on certain asymptotic approximations which permit inference based on false discovery rate techniques. The affects of the normalization process on these results is also investigated. Simulation studies and data analyses are used to assess finite sample performance.
BibTeX
@article{kosorok2007marginal,
  author = {Michael R. Kosorok and Shuangge Ma},
  title = {{Marginal asymptotics for the "large p, small n" paradigm: with applications to microarray data}},
  journal = {The Annals of Statistics},
  year = {2007},
  doi = {10.1214/009053606000001433},
}

Loading related content...

Figure