In the context of minimax theory, we propose a new kind of risk, normalized by a random variable, measurable with respect to the data. We present a notion of optimality and a method to construct optimal procedures accordingly. We apply this general setup to the problem of selecting significant variables in Gaussian white noise. In particular, we show that our method essentially improves the accuracy of estimation, in the sense of giving explicit improved confidence sets in $L_2$-norm. Links to adaptive estimation are discussed.

We consider a new concept of weak dependence, introduced by Doukhan and Louhichi [Stochastic Process. Appl. 84 (1999) 313–342], which is more general than the classical frameworks of mixing or associated sequences. The new notion is broad enough to include many interesting examples such as very general Bernoulli shifts, Markovian models or time series bootstrap processes with discrete innovations.

Under such a weak dependence assumption, we investigate nonparametric regression which represents one (among many) important statistical estimation problems. We justify in this more general setting the “whitening by windowing principle” for nonparametric regression, saying that asymptotic properties remain in first order the same as for independent samples. The proofs borrow previously used strategies, but precise arguments are developed under the new aspect of general weak dependence.

In kernel density estimation, those data values that make a nondegenerate contribution to the estimator (computed at a given point) tend to be spaced well apart. This property has the effect of suppressing many of the conventional consequences of long-range dependence, for example, slower rates of convergence, which might otherwise be revealed by a traditional loss-or risk-based assessment of performance. From that viewpoint, dependence has to be very long-range indeed before a density estimator experiences any first-order effects. However, an analysis in terms of the convergence rate for a particular realization, rather than the rate averaged over all realizations, reveals a very different picture. We show that from that viewpoint, and in the context of functions of Gaussian processes, effects on rates of convergence can become apparent as soon as the boundary between short- and long-range dependence is crossed. For example, the distance between ISE- and MISE-optimal bandwidths is generally of larger order for long-range dependent data. We shed new light on cross-validation, too. In particular we show that the variance of the cross-validation bandwidth is generally larger for long-range dependent data, and that the first-order properties of this bandwidth do not depend on how many data are left out when constructing the cross-validation criterion. Moreover, for long-range dependent data the cross-validation bandwidth is usually perfectly negatively correlated, in the limit, with the optimal stochastic bandwidth.

In many situations regression analysis is mostly concerned with inferring about the conditional mean of the response given the predictors, and less concerned with the other aspects of the conditional distribution. In this paper we develop dimension reduction methods that incorporate this consideration. We introduce the notion of the Central Mean Subspace (CMS), a natural inferential object for dimension reduction when the mean function is of interest. We study properties of the CMS, and develop methods to estimate it. These methods include a new class of estimators which requires fewer conditions than pHd, and which displays a clear advantage when one of the conditions for pHd is violated. CMS also reveals a transparent distinction among the existing methods for dimension reduction: OLS, pHd, SIR and SAVE. We apply the new methods to a data set involving recumbent cows.

In this article, we describe how the theory of sufficient dimension reduction, and a well-known inference method for it (sliced inverse regression), can be extended to regression analyses involving both quantitative and categorical predictor variables. As statistics faces an increasing need for effective analysis strategies for high-dimensional data, the results we present significantly widen the applicative scope of sufficient dimension reduction and open the way for a new class of theoretical and methodological developments.

We consider the problem of estimating the distance from an unknown signal, observed in a white-noise model, to convex cones of positive/monotone/convex functions. We show that, when the unknown function belongs to a Hölder class, the risk of estimating the $L_r$-distance, $1 \leq r < \infty$, from the signal to a cone is essentially the same (up to a logarithmic factor) as that of estimating the signal itself. The same risk bounds hold for the test of positivity, monotonicity and convexity of the unknown signal.

We also provide an estimate for the distance to the cone of positive functions for which risk is, by a logarithmic factor, smaller than that of the “plug-in” estimate.

Nonasymptotic risk bounds are provided for maximum likelihood-type isotonic estimators of an unknown nondecreasing regression function, with general average loss at design points. These bounds are optimal up to scale constants, and they imply uniform $n^{-1/3}$-consistency of the $\ell_p$ risk for unknown regression functions of uniformly bounded variation, under mild assumptions on the joint probability distribution of the data, with possibly dependent observations.

We introduce a new computer-intensive method to estimate the distribution of robust regression estimates. The basic idea behind our method is to bootstrap a reweighted representation of the estimates. To obtain a bootstrap method that is asymptotically correct, we include the auxiliary scale estimate in our reweighted representation of the estimates. Our method is computationally simple because for each bootstrap sample we only have to solve a linear system of equations. The weights we use are decreasing functions of the absolute value of the residuals and hence outlying observations receive small weights. This results in a bootstrap method that is resistant to the presence of outliers in the data. The breakdown points of the quantile estimates derived with this method are higher than those obtained with the bootstrap. We illustrate our method on two datasets and we report the results of a Monte Carlo experiment on confidence intervals for the parameters of the linear model.

This paper introduces a new class of robust estimators for the linear regression model. They are weighted least squares estimators, with weights adaptively computed using the empirical distribution of the residuals of an initial robust estimator. It is shown that under certain general conditions the asymptotic breakdown points of the proposed estimators are not less than that of the initial estimator, and the finite sample breakdown point can be at most $1/n$ less. For the special case of the least median of squares as initial estimator, hard rejection weights and normal errors and carriers, the maximum bias function of the proposed estimators for point-mass contaminations is numerically computed, with the result that there is almost no worsening of bias. Moreover–and this is the original contribution of this paper–if the errors are normally distributed and under fairly general conditions on the design the proposed estimators have full asymptotic efficiency. A Monte Carlo study shows that they have better behavior than the initial estimators for finite sample sizes.