We propose a new inferential framework for high-dimensional semiparametric generalized linear models. This framework addresses a variety of challenging problems in high-dimensional data analysis, including incomplete data, selection bias and heterogeneity. Our work has three main contributions: (i) We develop a regularized statistical chromatography approach to infer the parameter of interest under the proposed semiparametric generalized linear model without the need of estimating the unknown base measure function. (ii) We propose a new likelihood ratio based framework to construct post-regularization confidence regions and tests for the low dimensional components of high-dimensional parameters. Unlike existing post-regularization inferential methods, our approach is based on a novel directional likelihood. (iii) We develop new concentration inequalities and normal approximation results for U-statistics with unbounded kernels, which are of independent interest. We further extend the theoretical results to the problems of missing data and multiple datasets inference. Extensive simulation studies and real data analysis are provided to illustrate the proposed approach.

We analyze boosting algorithms [Ann. Statist. 29 (2001) 1189–1232; Ann. Statist. 28 (2000) 337–407; Ann. Statist. 32 (2004) 407–499] in linear regression from a new perspective: that of modern first-order methods in convex optimization. We show that classic boosting algorithms in linear regression, namely the incremental forward stagewise algorithm ($\text{FS}_{\varepsilon}$) and least squares boosting [LS-BOOST$(\varepsilon)$], can be viewed as subgradient descent to minimize the loss function defined as the maximum absolute correlation between the features and residuals. We also propose a minor modification of $\text{FS}_{\varepsilon}$ that yields an algorithm for the LASSO, and that may be easily extended to an algorithm that computes the LASSO path for different values of the regularization parameter. Furthermore, we show that these new algorithms for the LASSO may also be interpreted as the same master algorithm (subgradient descent), applied to a regularized version of the maximum absolute correlation loss function. We derive novel, comprehensive computational guarantees for several boosting algorithms in linear regression (including LS-BOOST$(\varepsilon)$ and $\text{FS}_{\varepsilon}$) by using techniques of first-order methods in convex optimization. Our computational guarantees inform us about the statistical properties of boosting algorithms. In particular, they provide, for the first time, a precise theoretical description of the amount of data-fidelity and regularization imparted by running a boosting algorithm with a prespecified learning rate for a fixed but arbitrary number of iterations, for any dataset.

For long-memory time series, inference based on resampling is of crucial importance, since the asymptotic distribution can often be non-Gaussian and is difficult to determine statistically. However, due to the strong dependence, establishing the asymptotic validity of resampling methods is nontrivial. In this paper, we derive an efficient bound for the canonical correlation between two finite blocks of a long-memory time series. We show how this bound can be applied to establish the asymptotic consistency of subsampling procedures for general statistics under long memory. It allows the subsample size $b$ to be $o(n)$, where $n$ is the sample size, irrespective of the strength of the memory. We are then able to improve many results found in the literature. We also consider applications of subsampling procedures under long memory to the sample covariance, M-estimation and empirical processes.

Much theoretical and applied work has been devoted to high-dimensional regression with clean data. However, we often face corrupted data in many applications where missing data and measurement errors cannot be ignored. Loh and Wainwright [Ann. Statist. 40 (2012) 1637–1664] proposed a nonconvex modification of the Lasso for doing high-dimensional regression with noisy and missing data. It is generally agreed that the virtues of convexity contribute fundamentally the success and popularity of the Lasso. In light of this, we propose a new method named CoCoLasso that is convex and can handle a general class of corrupted datasets. We establish the estimation error bounds of CoCoLasso and its asymptotic sign-consistent selection property. We further elucidate how the standard cross validation techniques can be misleading in presence of measurement error and develop a novel calibrated cross-validation technique by using the basic idea in CoCoLasso. The calibrated cross-validation has its own importance. We demonstrate the superior performance of our method over the nonconvex approach by simulation studies.

This paper studies the optimal tuning of the regularization parameter in LASSO or the threshold parameters in approximate message passing (AMP). Considering a model in which the design matrix and noise are zero-mean i.i.d. Gaussian, we propose a data-driven approach for estimating the regularization parameter of LASSO and the threshold parameters in AMP. Our estimates are consistent, that is, they converge to their asymptotically optimal values in probability as $n$, the number of observations, and $p$, the ambient dimension of the sparse vector, grow to infinity, while $n/p$ converges to a fixed number $\delta$. As a byproduct of our analysis, we will shed light on the asymptotic properties of the solution paths of LASSO and AMP.

We develop a new primal-dual witness proof framework that may be used to establish variable selection consistency and $\ell_{\infty}$-bounds for sparse regression problems, even when the loss function and regularizer are nonconvex. We use this method to prove two theorems concerning support recovery and $\ell_{\infty}$-guarantees for a regression estimator in a general setting. Notably, our theory applies to all potential stationary points of the objective and certifies that the stationary point is unique under mild conditions. Our results provide a strong theoretical justification for the use of nonconvex regularization: For certain nonconvex regularizers with vanishing derivative away from the origin, any stationary point can be used to recover the support without requiring the typical incoherence conditions present in $\ell_{1}$-based methods. We also derive corollaries illustrating the implications of our theorems for composite objective functions involving losses such as least squares, nonconvex modified least squares for errors-in-variables linear regression, the negative log likelihood for generalized linear models and the graphical Lasso. We conclude with empirical studies that corroborate our theoretical predictions.

Functional magnetic resonance imaging (fMRI) is a pioneering technology for studying brain activity in response to mental stimuli. Although efficient designs on these fMRI experiments are important for rendering precise statistical inference on brain functions, they are not systematically constructed. Design with circulant property is crucial for estimating a hemodynamic response function (HRF) and discussing fMRI experimental optimality. In this paper, we develop a theory that not only successfully explains the structure of a circulant design, but also provides a method of constructing efficient fMRI designs systematically. We further provide a class of two-level circulant designs with good performance (statistically optimal), and they can be used to estimate the HRF of a stimulus type and study the comparison of two HRFs. Some efficient three- and four-levels circulant designs are also provided, and we proved the existence of a class of circulant orthogonal arrays.

We investigate Bernstein–von Mises theorems for adaptive nonparametric Bayesian procedures in the canonical Gaussian white noise model. We consider both a Hilbert space and multiscale setting with applications in $L^{2}$ and $L^{\infty}$, respectively. This provides a theoretical justification for plug-in procedures, for example the use of certain credible sets for sufficiently smooth linear functionals. We use this general approach to construct optimal frequentist confidence sets based on the posterior distribution. We also provide simulations to numerically illustrate our approach and obtain a visual representation of the geometries involved.

This article studies the targeted sequential inference of an optimal treatment rule (TR) and its mean reward in the nonexceptional case, that is, assuming that there is no stratum of the baseline covariates where treatment is neither beneficial nor harmful, and under a companion margin assumption.

Our pivotal estimator, whose definition hinges on the targeted minimum loss estimation (TMLE) principle, actually infers the mean reward under the current estimate of the optimal TR. This data-adaptive statistical parameter is worthy of interest on its own. Our main result is a central limit theorem which enables the construction of confidence intervals on both mean rewards under the current estimate of the optimal TR and under the optimal TR itself. The asymptotic variance of the estimator takes the form of the variance of an efficient influence curve at a limiting distribution, allowing to discuss the efficiency of inference.

As a by product, we also derive confidence intervals on two cumulated pseudo-regrets, a key notion in the study of bandits problems.

A simulation study illustrates the procedure. One of the cornerstones of the theoretical study is a new maximal inequality for martingales with respect to the uniform entropy integral.

We propose $L^{p}$ distance-based goodness-of-fit (GOF) tests for uniform stochastic ordering with two continuous distributions $F$ and $G$, both of which are unknown. Our tests are motivated by the fact that when $F$ and $G$ are uniformly stochastically ordered, the ordinal dominance curve $R=FG^{-1}$ is star-shaped. We derive asymptotic distributions and prove that our testing procedure has a unique least favorable configuration of $F$ and $G$ for $p\in [1,\infty]$. We use simulation to assess finite-sample performance and demonstrate that a modified, one-sample version of our procedure (e.g., with $G$ known) is more powerful than the one-sample GOF test suggested by Arcones and Samaniego [Ann. Statist. 28 (2000) 116–150]. We also discuss sample size determination. We illustrate our methods using data from a pharmacology study evaluating the effects of administering caffeine to prematurely born infants.

Principal component analysis (PCA) is a well-known tool in multivariate statistics. One significant challenge in using PCA is the choice of the number of principal components. In order to address this challenge, we propose distribution-based methods with exact type 1 error controls for hypothesis testing and construction of confidence intervals for signals in a noisy matrix with finite samples. Assuming Gaussian noise, we derive exact type 1 error controls based on the conditional distribution of the singular values of a Gaussian matrix by utilizing a post-selection inference framework, and extending the approach of [Taylor, Loftus and Tibshirani (2013)] in a PCA setting. In simulation studies, we find that our proposed methods compare well to existing approaches.

The goal of this paper is to integrate the notions of stochastic conditional independence and variation conditional independence under a more general notion of extended conditional independence. We show that under appropriate assumptions the calculus that applies for the two cases separately (axioms of a separoid) still applies for the extended case. These results provide a rigorous basis for a wide range of statistical concepts, including ancillarity and sufficiency, and, in particular, the Decision Theoretic framework for statistical causality, which uses the language and calculus of conditional independence in order to express causal properties and make causal inferences.

Model averaging has long been proposed as a powerful alternative to model selection in regression analysis. However, how well it performs in high-dimensional regression is still poorly understood. Recently, Ando and Li [J. Amer. Statist. Assoc. 109 (2014) 254–265] introduced a new method of model averaging that allows the number of predictors to increase as the sample size increases. One notable feature of Ando and Li’s method is the relaxation on the total model weights so that weak signals can be efficiently combined from high-dimensional linear models. It is natural to ask if Ando and Li’s method and results can be extended to nonlinear models. Because all candidate models should be treated as working models, the existence of a theoretical target of the quasi maximum likelihood estimator under model misspecification needs to be established first. In this paper, we consider generalized linear models as our candidate models. We establish a general result to show the existence of pseudo-true regression parameters under model misspecification. We derive proper conditions for the leave-one-out cross-validation weight selection to achieve asymptotic optimality. Technically, the pseudo true target parameters between working models are not linearly linked. To overcome the encountered difficulties, we employ a novel strategy of decomposing and bounding the bias and variance terms in our proof. We conduct simulations to illustrate the merits of our model averaging procedure over several existing methods, including the lasso and group lasso methods, the Akaike and Bayesian information criterion model-averaging methods and some other state-of-the-art regularization methods.

We present a new framework on inferring structural similarities and differences among multiple high-dimensional Gaussian graphical models (GGMs) corresponding to the same set of variables under distinct experimental conditions. The new framework adopts the partial correlation coefficients to characterize the potential changes of dependency strengths between two variables. A hierarchical method has been further developed to recover edges with different or similar dependency strengths across multiple GGMs. In particular, we first construct two-sample test statistics for testing the equality of partial correlation coefficients and conduct large-scale multiple tests to estimate the substructure of differential dependencies. After removing differential substructure from original GGMs, a follow-up multiple testing procedure is used to detect the substructure of similar dependencies among GGMs. In each step, false discovery rate is controlled asymptotically at a desired level. Power results are proved, which demonstrate that our method is more powerful on finding common edges than the common approach that separately estimates GGMs. The performance of the proposed hierarchical method is illustrated on simulated datasets.

In the context of a species sampling problem, we discuss a nonparametric maximum likelihood estimator for the underlying probability mass function. The estimator is known in the computer science literature as the high profile estimator. We prove strong consistency and derive the rates of convergence, for an extended model version of the estimator. We also study a sieved estimator for which similar consistency results are derived. Numerical computation of the sieved estimator is of great interest for practical problems, such as forensic DNA analysis, and we present a computational algorithm based on the stochastic approximation of the expectation maximisation algorithm. As an interesting byproduct of the numerical analyses, we introduce an algorithm for bounded isotonic regression for which we also prove convergence.

Consider a large number of detectors each generating a data stream. The task is to detect online, distribution changes in a small fraction of the data streams. Previous approaches to this problem include the use of mixture likelihood ratios and sum of CUSUMs. We provide here extensions and modifications of these approaches that are optimal in detecting normal mean shifts. We show how the (optimal) detection delay depends on the fraction of data streams undergoing distribution changes as the number of detectors goes to infinity. There are three detection domains. In the first domain for moderately large fractions, immediate detection is possible. In the second domain for smaller fractions, the detection delay grows logarithmically with the number of detectors, with an asymptotic constant extending those in sparse normal mixture detection. In the third domain for even smaller fractions, the detection delay lies in the framework of the classical detection delay formula of Lorden. We show that the optimal detection delay is achieved by the sum of detectability score transformations of either the partial scores or CUSUM scores of the data streams.