Search

Ultra-high dimensional longitudinal data are increasingly common and the analysis is challenging both theoretically and methodologically. We offer a new automatic procedure for finding a sparse semivarying coefficient model, which is widely accepted for longitudinal data analysis. Our proposed method first reduces the number of covariates to a moderate order by employing a screening procedure, and then identifies both the varying and constant coefficients using a group SCAD estimator, which is subsequently refined by accounting for the within-subject correlation. The screening procedure is based on working independence and B-spline marginal models. Under weaker conditions than those in the literature, we show that with high probability only irrelevant variables will be screened out, and the number of selected variables can be bounded by a moderate order. This allows the desirable sparsity and oracle properties of the subsequent structure identification step. Note that existing methods require some kind of iterative screening in order to achieve this, thus they demand heavy computational effort and consistency is not guaranteed. The refined semivarying coefficient model employs profile least squares, local linear smoothing and nonparametric covariance estimation, and is semiparametric efficient. We also suggest ways to implement the proposed methods, and to select the tuning parameters. An extensive simulation study is summarized to demonstrate its finite sample performance and the yeast cell cycle data is analyzed., Comment: Published in at http://dx.doi.org/10.1214/14-AOS1236 the Annals of Statistics (http://www.imstat.org/aos/) by the Institute of Mathematical Statistics (http://www.imstat.org)

We consider a $l_1$-penalization procedure in the non-parametric Gaussian regression model. In many concrete examples, the dimension $d$ of the input variable $X$ is very large (sometimes depending on the number of observations). Estimation of a $\beta$-regular regression function $f$ cannot be faster than the slow rate $n^{-2\beta/(2\beta+d)}$. Hopefully, in some situations, $f$ depends only on a few numbers of the coordinates of $X$. In this paper, we construct two procedures. The first one selects, with high probability, these coordinates. Then, using this subset selection method, we run a local polynomial estimator (on the set of interesting coordinates) to estimate the regression function at the rate $n^{-2\beta/(2\beta+d^*)}$, where $d^*$, the "real" dimension of the problem (exact number of variables whom $f$ depends on), has replaced the dimension $d$ of the design. To achieve this result, we used a $l_1$ penalization method in this non-parametric setup., Comment: Published in at http://dx.doi.org/10.1214/08-EJS327 the Electronic Journal of Statistics (http://www.i-journals.org/ejs/) by the Institute of Mathematical Statistics (http://www.imstat.org)

Let $Y$ be a Gaussian vector of $\mathbb{R}^n$ of mean $s$ and diagonal covariance matrix $\Gamma$. Our aim is to estimate both $s$ and the entries $\sigma_i=\Gamma_{i,i}$, for $i=1,...,n$, on the basis of the observation of two independent copies of $Y$. Our approach is free of any prior assumption on $s$ but requires that we know some upper bound $\gamma$ on the ratio $\max_i\sigma_i/\min_i\sigma_i$. For example, the choice $\gamma=1$ corresponds to the homoscedastic case where the components of $Y$ are assumed to have common (unknown) variance. In the opposite, the choice $\gamma>1$ corresponds to the heteroscedastic case where the variances of the components of $Y$ are allowed to vary within some range. Our estimation strategy is based on model selection. We consider a family $\{S_m\times\Sigma_m, m\in\mathcal{M}\}$ of parameter sets where $S_m$ and $\Sigma_m$ are linear spaces. To each $m\in\mathcal{M}$, we associate a pair of estimators $(\hat{s}_m,\hat{\sigma}_m)$ of $(s,\sigma)$ with values in $S_m\times\Sigma_m$. Then we design a model selection procedure in view of selecting some $\hat{m}$ among $\mathcal{M}$ in such a way that the Kullback risk of $(\hat{s}_{\hat{m}},\hat{\sigma}_{\hat{m}})$ is as close as possible to the minimum of the Kullback risks among the family of estimators $\{(\hat{s}_m,\hat{\sigma}_m), m\in\mathcal{M}\}$. Then we derive uniform rates of convergence for the estimator $(\hat{s}_{\hat{m}},\hat{\sigma}_{\hat{m}})$ over H\"{o}lderian balls. Finally, we carry out a simulation study in order to illustrate the performances of our estimators in practice., Comment: Published in at http://dx.doi.org/10.1214/08-EJS267 the Electronic Journal of Statistics (http://www.i-journals.org/ejs/) by the Institute of Mathematical Statistics (http://www.imstat.org)

We consider high-dimensional generalized linear models with Lipschitz loss functions, and prove a nonasymptotic oracle inequality for the empirical risk minimizer with Lasso penalty. The penalty is based on the coefficients in the linear predictor, after normalization with the empirical norm. The examples include logistic regression, density estimation and classification with hinge loss. Least squares regression is also discussed., Comment: Published in at http://dx.doi.org/10.1214/009053607000000929 the Annals of Statistics (http://www.imstat.org/aos/) by the Institute of Mathematical Statistics (http://www.imstat.org)

We present a greedy method for simultaneously performing local bandwidth selection and variable selection in nonparametric regression. The method starts with a local linear estimator with large bandwidths, and incrementally decreases the bandwidth of variables for which the gradient of the estimator with respect to bandwidth is large. The method--called rodeo (regularization of derivative expectation operator)--conducts a sequence of hypothesis tests to threshold derivatives, and is easy to implement. Under certain assumptions on the regression function and sampling density, it is shown that the rodeo applied to local linear smoothing avoids the curse of dimensionality, achieving near optimal minimax rates of convergence in the number of relevant variables, as if these variables were isolated in advance., Comment: Published in at http://dx.doi.org/10.1214/009053607000000811 the Annals of Statistics (http://www.imstat.org/aos/) by the Institute of Mathematical Statistics (http://www.imstat.org)

We study a high-dimensional generalized linear model and penalized empirical risk minimization with $\ell_1$ penalty. Our aim is to provide a non-trivial illustration that non-asymptotic bounds for the estimator can be obtained without relying on the chaining technique and/or the peeling device., Comment: Published at http://dx.doi.org/10.1214/074921707000000319 in the IMS Lecture Notes Monograph Series (http://www.imstat.org/publications/lecnotes.htm) by the Institute of Mathematical Statistics (http://www.imstat.org)

We develop minimax optimal risk bounds for the general learning task consisting in predicting as well as the best function in a reference set G up to the smallest possible additive term, called the convergence rate. When the reference set is finite and when n denotes the size of the training data, we provide minimax convergence rates of the form C ([log |G|]/n)^v with tight evaluation of the positive constant C and with exact v in ]0;1], the latter value depending on the convexity of the loss function and on the level of noise in the output distribution. The risk upper bounds are based on a sequential randomized algorithm, which at each step concentrates on functions having both low risk and low variance with respect to the previous step prediction function. Our analysis puts forward the links between the probabilistic and worst-case viewpoints, and allows to obtain risk bounds unachievable with the standard statistical learning approach. One of the key idea of this work is to use probabilistic inequalities with respect to appropriate (Gibbs) distributions on the prediction function space instead of using them with respect to the distribution generating the data. The risk lower bounds are based on refinements of the Assouad's lemma taking particularly into account the properties of the loss function. Our key example to illustrate the upper and lower bounds is to consider the L_q-regression setting for which an exhaustive analysis of the convergence rates is given while q describes [1;+infinity[.

We want to recover the regression function in the single-index model. Using an aggregation algorithm with local polynomial estimators, we answer in particular to the second part of Question~2 from Stone (1982) on the optimal convergence rate. The procedure constructed here has strong adaptation properties: it adapts both to the smoothness of the link function and to the unknown index. Moreover, the procedure locally adapts to the distribution of the design. We propose new upper bounds for the local polynomial estimator (which are results of independent interest) that allows a fairly general design. The behavior of this algorithm is studied through numerical simulations. In particular, we show empirically that it improves strongly over empirical risk minimization., Comment: 36 pages

Let $Y$ be a Gaussian vector whose components are independent with a common unknown variance. We consider the problem of estimating the mean $\mu$ of $Y$ by model selection. More precisely, we start with a collection $\mathcal{S}=\{S_m,m\in\mathcal{M}\}$ of linear subspaces of $\mathbb{R}^n$ and associate to each of these the least-squares estimator of $\mu$ on $S_m$. Then, we use a data driven penalized criterion in order to select one estimator among these. Our first objective is to analyze the performance of estimators associated to classical criteria such as FPE, AIC, BIC and AMDL. Our second objective is to propose better penalties that are versatile enough to take into account both the complexity of the collection $\mathcal{S}$ and the sample size. Then we apply those to solve various statistical problems such as variable selection, change point detections and signal estimation among others. Our results are based on a nonasymptotic risk bound with respect to the Euclidean loss for the selected estimator. Some analogous results are also established for the Kullback loss., Comment: Published in at http://dx.doi.org/10.1214/07-AOS573 the Annals of Statistics (http://www.imstat.org/aos/) by the Institute of Mathematical Statistics (http://www.imstat.org)

The generalized varying coefficient partially linear model with growing number of predictors arises in many contemporary scientific endeavor. In this paper we set foot on both theoretical and practical sides of profile likelihood estimation and inference. When the number of parameters grows with sample size, the existence and asymptotic normality of the profile likelihood estimator are established under some regularity conditions. Profile likelihood ratio inference for the growing number of parameters is proposed and Wilk's phenomenon is demonstrated. A new algorithm, called the accelerated profile-kernel algorithm, for computing profile-kernel estimator is proposed and investigated. Simulation studies show that the resulting estimates are as efficient as the fully iterative profile-kernel estimates. For moderate sample sizes, our proposed procedure saves much computational time over the fully iterative profile-kernel one and gives stabler estimates. A set of real data is analyzed using our proposed algorithm., Comment: 38 pages article

In this paper, we focus on regression estimation in both the inductive and the transductive case. We assume that we are given a set of features (which can be a base of functions, but not necessarily). We begin by giving a deviation inequality on the risk of an estimator in every model defined by using a single feature. These models are too simple to be useful by themselves, but we then show how this result motivates an iterative algorithm that performs feature selection in order to build a suitable estimator. We prove that every selected feature actually improves the performance of the estimator. We give all the estimators and results at first in the inductive case, which requires the knowledge of the distribution of the design, and then in the transductive case, in which we do not need to know this distribution.