WeiYa's Work Yard

A dog, who fell into the ocean of statistics, tries to write down his ideas and notes to save himself.

Group Inference in High Dimensions

Posted on (Update: )
Tags: High-Dimensional

This post is based on the slides for the talk given by Zijian Guo at The International Statistical Conference In Memory of Professor Sik-Yum Lee

Group Inference

High-dimensional linear regression

where $X_{i\cdot},\beta\in\IR^p$.

  • high dimension: $n » p$
  • sparse model: $\Vert \beta\Vert_0 « n$

For a given set $G\subset \{1,2,\ldots,p\}$, group significance test is

where $\beta_G = \{\beta_j:j\in G\}$.

Group Inference vs. Quadratic Functional

The null $H_0:\beta_G=0$ can be written as

for some positive definite matrix $A\in\IR^{\vert G\vert \times \vert G\vert}$

Here are two special cases:

  • $H_{0,\Sigma}$
  • $H_{0, I}$

Group vs. Individual Significance

For a group of highly correlated variables,

  1. It is ambitious to detect significant single variable $\beta_i$ due to inaccurate estimator of $\beta_i$.
  2. Significance and high correlation: significant variables can treated as non-significant
  3. The group significance

Hierarchical Testing (Meinshausen, 2008)

Divide variables into sub-groups + group significance

  • Variables inside a group tend to be highly correlated
  • Between groups, not highly correlated

Other Motivation I: Interaction Test

Model with interaction:

To test $H_0:\gamma = 0$ is equivalent to test

where

with $W_i = (D_iX_{i\cdot}^T, 1, X_{i\cdot}^T)^T$ and $\eta = (\gamma^T, \gamma_0, \beta^T)^T$.

Other Motivation II: Local Heritability in Genetics

The proportion of variance explained by a subset of genotypes indexed by the group G, which is the set of SNPs located in on the same chromosome. The local heritability is defined as

Goal

Inference for $Q_\Sigma = \beta_G^T\Sigma_{G,G}\beta_G$.

Bias Correction

Initial estimators

  • $\hat \beta = \argmin_{\beta\in\IR^p} \frac{1}{2n}\Vert y-X\beta\Vert_2^2 + \lambda\Vert \beta\Vert_1$
  • $\hat\Sigma = \frac 1nX^TX$

Decompose $\hat\beta_G^T\hat\Sigma_{G,G}\hat\beta_G - \beta_G^T\Sigma_{G,G}\beta_G$ as

Then estimate $\hat\beta_G^T\hat\Sigma_{G,G}(\beta_G-\hat\beta_G)$ and correct $\hat\beta_G^T\hat\Sigma_{G,G}\hat\beta_G$.

It recalls me the decomposition of $\tilde\alpha$ in the post Optimal estimation of functionals of high-dimensional mean and covariance matrix

Construction of Projection Direction

For any $u\in \IR^p$,

  • $\Vert \beta - \hat\beta\Vert_1$ is small
  • Minimize/Constrained $u^T\hat\Sigma u$ and

Initial proposal

s.t.

and

Replace $\cC_0$ with

then it works for any $\vert G\vert$.

The Constrain Variance becomes

Inference Procedure

Estimate the variance of the proposed estimator $\hat Q_\Sigma$ by $\hat V_\Sigma(\tau)$. The decision rule (test) is

and the confidence interval is

Theoretical Justification

Size and Power

References


Published in categories Note