TL;DR
This paper introduces new statistical tests for high-dimensional generalized linear models, particularly useful for analyzing complex biological data like gene expression. The proposed tests offer improved power and robustness compared to existing methods, especially when dealing with unbounded link functions common in models like Poisson regression.
What Problem Does This Paper Address?
The paper tackles the challenge of performing statistical inference in generalized linear models (GLMs) when the number of predictor variables (covariates) is very large, often exceeding the number of observations. This "high-dimensional" setting is prevalent in fields like bioinformatics, where researchers analyze thousands of gene expression levels for a relatively small number of samples. Traditional statistical tests, like the likelihood ratio or Wald tests, break down in these scenarios because they rely on the invertibility of certain matrices that become singular with too many predictors. While existing methods like Goeman et al.'s test offer a solution for fixed high dimensions, this paper addresses the more challenging case where the dimension diverges with the sample size and also aims to improve the power of these tests, especially for certain types of GLMs.
Methods
The authors build upon the work of Goeman et al. (2011) by analyzing the power properties of their test under diverging dimensions. They identify a weakness in the existing test when the inverse of the link function is unbounded (e.g., in Poisson or Negative Binomial regression), where high dimensionality can reduce the test's power. To overcome this, they propose a new test statistic based on a simpler U-statistic formulation. This new statistic aims to avoid the adverse impact of high dimensionality. They also develop a test for situations where some covariates are of primary interest while others are treated as nuisance parameters. The theoretical analysis involves asymptotic properties of these statistics under different families of alternative hypotheses ("local" and "fixed").
Datasets
The paper uses a publicly available dataset on acute lymphoblastic leukemia, which includes microarray readings for 128 individuals. This dataset is used for a case study to demonstrate the application of the proposed gene-set testing methods.
Key Insights and Ideas
- Unbounded link functions are problematic: The inverse of the link function in GLMs plays a crucial role. When it's unbounded (like in Poisson regression), high dimensionality can significantly reduce the power of existing tests.
- Simpler statistics can be better: A U-statistic formulation, by excluding certain terms from the original test statistic, can lead to improved power and robustness in high-dimensional settings.
- Nuisance parameters matter: Developing tests that can handle nuisance parameters is essential for real-world applications, especially in gene-set analysis where some covariates (like age or gender) might be less of interest than gene expression levels.
- Gene-set analysis benefits: The proposed tests are well-suited for identifying significant gene sets, a critical task in understanding complex biological systems.
Key Results
The paper presents theoretical results showing that the proposed tests maintain asymptotic power equivalent to Goeman et al.'s test under "local" alternatives but offer significantly better power under "fixed" alternatives, especially for models with unbounded link functions. Simulation studies confirm these findings, demonstrating that the proposed global test and the test with nuisance parameters are more powerful than Goeman et al.'s test for Poisson and Negative Binomial models. The case study on leukemia data shows that the proposed tests can effectively identify significant gene sets.
Limitations and Future Work
The authors acknowledge that their analysis is confined to specific families of alternative hypotheses ("local" and "fixed"). They also note that the power analysis for the test with nuisance parameters under "fixed" alternatives was not fully explored due to complexity. Future work could involve extending the analysis to other types of alternatives and further investigating the performance of the nuisance parameter test under various scenarios. The paper also mentions that the proposed tests, when combined with FDR control, are useful for multiple testing in gene-set analysis.
Summary updated