Working Papers

html templates

“Rank-Robust Wald-type tests: a regularization approach,” with Jean-Marie Dufour

link to the paper

Abstract: This paper studies Wald-type tests in the presence of possibly rank-deficient covariance matrices, allowing for singular covariance matrices, either in finite samples or asymptotically. Such difficulties occur in many statistical and econometric problems, such as causality and cointegration analysis in time series, (locally) redundant restrictions, (locally) redundant moment equations in GMM, tests on the determinant of a coefficient matrix (reduced rank hypotheses), tests of linear restrictions on Average Treatment Effects in regression discontinuity designs, etc. Two different types of singularity are considered. First, the estimated 
covariance matrix has full rank but converges to a singular covariance matrix, so the Wald statistic can be computed as usual, but regularity conditions for the standard asymptotic chi-square distribution do not hold. Second, the estimated covariance matrix does not have full 
rank but converges to a population matrix whose rank may differ from the  finite-sample rank. The proposed procedure works in all cases regardless of the finite-sample and asymptotic ranks. To address such difficulties, we introduce a novel mathematical object: the regularized inverse which 
is related to generalized inverses, although different. The  regularized inverse exploits the spectral decomposition of the covariance matrix; its unique representation follows from the Spectral Theorem, see Theorem 1.2a, p.53, Eaton(2007). Results on total eigenprojections (that is the sum of eigenprojections over a subset of the spectral set) are combined with a variance regularizing function; the latter modifies small eigenvalues (using a threshold). The continuity property of the total eigenprojection technique ensures a valid asymptotic theory for the regularized inverse; it always exists and is unique. The proposed class of regularized inverse matrices includes both continuous and discontinuous regularized inverses; the Tikhonov-type inverse is continuous, the spectral cut-off regularized inverse as proposed by Lütkepohl and Burda (1997) is discontinuous, and the full-rank regularized inverse we propose is continuous. Under general regularity conditions, we show that sample regularized inverse matrices converge to their regularized asymptotic counterparts. Regularized Wald statistics are then obtained through 
replacement of the usual inverse of the estimated covariance matrix (or the  generalized inverse) by a regularized inverse. Both Gaussian and non-Gaussian distributions are allowed for the parameter estimates. Two classes of regularized Wald statistics are studied in relative detail. The 
first one admits a nonstandard asymptotic distribution, which corresponds to  a linear combination of chi-square variables when the estimator used is asymptotically Gaussian. In this case, we show that the asymptotic distribution is bounded by the usual (full-rank) chi-square distribution, so standard critical values yield valid tests. In more general cases, we show that the asymptotic distribution can be simulated or bounded by simulation. The second class allows the threshold to vary with the sample size, but additional information is needed. This class of test statistics includes the spectral cut-off statistic proposed by Lütkepohl and Burda (1997, J. Econometrics) as a special case. The regularized statistics are consistent against global alternatives, with a loss of power (in certain directions) for the spectral cut-off Wald statistic. An application to U.S. data illustrates how the procedure works when testing for noncausality between saving, investment, growth and foreign direct investment. 


 

"A quasi-likelihood approach based on eigenfunctions for a Jacobi process", with Christian Gouriéroux.

link to the paper

Abstract:In this paper we consider a discretely sampled Jacobi process appropriate to specify the dynamics of a process with range [0,1], such as a discount coefficient,  a regime probability, or a state price. The discrete time transition of the Jacobi process does not admit a closed form expression and therefore the exact maximum likelihood (ML) is infeasible. We first review a characterization of the transition  function based on nonlinear canonical decomposition. They allow for approximations of the log-likelihood function which can be used to define a quasi-maximum likelihood estimator. The finite sample properties of this estimator are compared with the properties of other  estimators proposed in the literature, such as the Kessler and Sorensen's estimator which is a method of moments based on an approximated score function [see Kessler and Sorensen (1999)] or with a generalized method of moments (GMM) estimator. The quasi-maximum likelihood estimator is further compared with computer-intensive simulation-based estimation techniques such  as the indirect inference estimator or the simulated method of moments (SMM). Indeed, these techniques naturally arise when the ML estimator is infeasible. With respect to computational efficiency, the QML estimator is less time consuming than the simulation-based techniques. We also focus on computational issues for simulating the underlying continuous-time path of  the Jacobi process from a truncated Euler discretization scheme. Finally an empirical application on bond default probability data from MSCI Barra inc. is performed to assess the  performance of the Jacobi process to capturing the dynamic of a probability process.