We propose a new kernel estimator for nonparametric regression with unknown error distribution. We show that the proposed estimator is adaptive in the sense that it is asymptotically equivalent to the infeasible local likelihood estimator (Staniswalis, 1989, Journal of the American Statistical Association 84, 276–283; Fan, Farmen, and Gijbels, 1998, Journal of the Royal Statistical Society, Series B 60, 591–608; and Fan and Chen, 1999, Journal of the Royal Statistical Society, Series B 61, 927–943), which requires knowledge of the error distribution. Hence, our estimator improves on standard nonparametric kernel estimators when the error distribution is not normal. A Monte Carlo experiment is conducted to investigate the finite-sample performance of our procedure.We thank Yuichi Kitamura, Yanqin Fan, Joel Horowitz, Roger Koenker, Jens Perch Nielsen, Peter Phillips, Peter Robinson, Tom Rothenberg, and two referees for helpful comments. Financial support from the NSF and the ESRC (UK) is gratefully acknowledged.

Consider the simple normal linear regression model for estimation/prediction at a new design point. When the slope parameter is not obviously nonzero, hypothesis testing and information criteria can be used for identifying the right model. We compare the performances of such methods both theoretically and empirically from different perspectives for more insight. The testing approach at the conventional size of 0.05, in spite of being the “standard approach,” performs poorly in estimation. We also found that the frequently told story “the Bayesian information criterion (BIC) is good when the true model is finite-dimensional, and the Akaike information criterion (AIC) is good when the true model is infinite-dimensional” is far from being accurate. In addition, despite some successes in the effort to go beyond the debate between AIC and BIC by adaptive model selection, it turns out that it is not possible to share the pointwise adaptation property of BIC and the minimax-rate adaptation property of AIC by any model selection method. When model selection methods have difficulty in selection, model combining is a better alternative in terms of estimation accuracy.This work was completed when the author was on leave from Iowa State University and was a New Direction Visiting Professor at the Institute for Mathematics and its Applications (IMA) at the University of Minnesota. The fundings from both IMA and ISU are greatly appreciated. The work was also partly supported by NSF CAREER grant DMS0094323. The author thanks Xiaotong Shen and Hannes Leeb for very helpful discussions. The paper also benefited from the questions and comments from the participants at the statistics seminars the author gave at the University of Minnesota and Duke University. The author is very grateful to the anonymous reviewers and the co-editor Benedikt Pötscher for carefully reading earlier versions of the paper, bringing my attention to several closely related previous and current results, and making many very valuable suggestions, which significantly improved the paper in both content and presentation.

We discuss the limiting behavior of the serial correlation coefficient in mildly explosive autoregression, where the error sequence is in the domain of attraction of an α-stable law, α ∈ (0,2]. Therein, the autoregressive coefficient ρ = ρn > 1 is assumed to satisfy the condition ρn → 1 such that n(ρn − 1) → ∞ as n → ∞. In contrast to the vast majority of existing literature in the area, no specific form of ρ is required. We show that the serial correlation coefficient converges in distribution to a ratio of two independent stable random variables.The authors thank P.C.B. Phillips and two anonymous referees for a very careful reading of the manuscript, pointing out several mistakes, and providing shorter and simpler proofs. This research was partially supported by NATO grant PST.EAP.CLG 980599 and NSF-OTKA grant INT-0223262. This work was done while the first author was at the University of Utah.

Slowly varying (SV) regressors arise commonly in empirical econometric work, particularly in the form of semilogarithmic regression and log periodogram regression. These regressors are asymptotically collinear. Usual regression formulas for asymptotic standard errors are shown to remain valid, but rates of convergence are affected and the limit distribution of the regression coefficients is shown to be one dimensional. Some asymptotic representations of partial sums of SV functions and central limit theorems with SV weights are given that assist in the development of a regression theory. Multivariate regression and polynomial regression with SV functions are considered and shown to be equivalent, up to standardization, to regression on a polynomial in a logarithmic trend. The theory involves second-, third-, and higher-order forms of slow variation. Some applications to the asymptotic theory of nonlinear trend regression are explored.The author thanks two referees and Pentti Saikkonen for comments and suggestions, Sidney Resnick for references on second-order regular variation, and a Kelly Fellowship and the NSF for partial research support under grants SBR 97-30295 and SES 04-142254. An original draft of the paper was written in June 2000 and circulated under the title “Regression with Slowly Varying Regressors.”

This paper is concerned with the progress of Trygve Haavelmo's research and with his activities in general during his stay in the United States from June 1939 until the publication of his thesis “The Probability Approach in Econometrics” (Haavelmo, 1944, Econometrica 12, Suppl. 1–118) in July 1944. His original intention had been to stay in the United States only until the end of 1940, but the outbreak of World War II and the German occupation of Norway left him stranded. His “Theory and Measurement” treatise (Haavelmo, 1941, “On the Theory and Measurement of Economic Relations,” hectograph), the first version of “The Probability Approach,” was completed by the middle of 1941. From 1942 Haavelmo worked in New York for the Norwegian government in exile and was called upon to present his ideas to econometricians. Throughout his time in the United States he argued that probability should be accepted as an integral part of economic theory and as a basis for verification in economics. This paper considers how much of Haavelmo's approach was the result of his prewar experiences and how much the result of his time in the United States. It elaborates on his contact with Jerzy Neyman, Abraham Wald, and Jakob Marschak and the circumstances leading to the publication of “The Probability Approach.” Haavelmo's activities have been tracked through letters, seminar notes, and a reconstruction of his itinerary. The paper is a sequel to Bjerkholt (2005, Econometric Theory 21, 491–533).The presentation owes very much indeed to the advice of three anonymous referees and to the editor Peter Phillips, as the result of which it was restructured, reformulated, and much improved. I am sincerely grateful to all of these individuals for their generous and constructive advice. The source material for the paper is to a large extent correspondence and documents from the Haavelmo Archive at the Department of Economics, University of Oslo, as organized by Tore Thonstad. It also draws on material from the Frisch correspondence files at the National Library of Norway and from the Frisch Archive at the Department of Economics, University of Oslo. I am most grateful to Wendy Glickman of the Rockefeller Archive Center for providing me with copies of the monitoring sheets on Haavelmo. I owe my colleague Tore Schweder thanks for very helpful advice. I have furthermore benefited from personal communication with Ted Anderson, Kenneth J. Arrow, Leo Hurwicz, Eilev S. Jansen, Lawrence R. Klein, J.J. Polak, and Paul A. Samuelson, for which I am most grateful. NBER's Claudia Goldin, Chris Nagorski, and Robert Lipsey helped me sort out some puzzles. Thanks are also due to Lesney Levene for invaluable linguistic advice and to Inger Bjerkodden for encouragement all along. The responsibility for all remaining errors (surely, a nonempty set) is entirely my own.

This paper considers tests in an instrumental variable (IVs) regression model with IVs that may be weak. Tests that have near-optimal asymptotic power properties with Gaussian errors for weak and strong IVs have been determined in Andrews, Moreira, and Stock (2006, Econometrica 74, 715–752). In this paper, we seek tests that have near-optimal asymptotic power with Gaussian errors and improved power with non-Gaussian errors relative to existing tests. Tests with such properties are obtained by introducing rank tests that are analogous to the conditional likelihood ratio test of Moreira (2003, Econometrica 71, 1027–1048). We also introduce a rank test that is analogous to the Lagrange multiplier test of Kleibergen (2002, Econometrica 70, 1781–1803) and Moreira (2001, manuscript, University of California, Berkeley).Andrews gratefully acknowledges the research support of the National Science Foundation via grant SES-0417911.

Local linear fitting of nonlinear processes under strong (i.e., α-) mixing conditions has been investigated extensively. However, it is often a difficult step to establish the strong mixing of a nonlinear process composed of several parts such as the popular combination of autoregressive moving average (ARMA) and generalized autoregressive conditionally heteroskedastic (GARCH) models. In this paper we develop an asymptotic theory of local linear fitting for near epoch dependent (NED) processes. We establish the pointwise asymptotic normality of the local linear kernel estimators under some restrictions on the amount of dependence. Simulations and application examples illustrate that the proposed approach can work quite well for the medium size of economic time series.We thank Yuichi Kitamura and two referees for helpful comments. This research was partially supported by a Leverhulme Trust research grant, the National Natural Science Foundation of China, and the Economic and Social Science Research Council of the UK.

We consider the modeling of a time series described by a linear regression component whose regressor sequence satisfies the generalized asymptotic sample second moment stationarity conditions of Grenander (1954, Annals of Mathematical Statistics 25, 252–272). The associated disturbance process is only assumed to have sample second moments that converge with increasing series length, perhaps after a differencing operation. The model's regression component, which can be stochastic, is taken to be underspecified, perhaps as a result of simplifications, approximations, or parsimony. Also, the autoregressive moving average (ARMA) or autoregressive integrated moving average (ARIMA) model used for the disturbances need not be correct. Both ordinary least squares (OLS) and generalized least squares (GLS) estimates of the mean function are considered. An optimality property of GLS relative to OLS is obtained for one-step-ahead forecasting. Asymptotic bias characteristics of the regression estimates are shown to distinguish the forecasting performance. The results provide theoretical support for a procedure used by Statistics Netherlands to impute the values of late reporters in some economic surveys.The author thanks two referees and the co-editor for comments and suggestions that led to substantial improvements in the exposition and also thanks John Aston and Tucker McElroy for helpful comments on an earlier draft. Any views expressed are the author's and not necessarily those of the U.S. Census Bureau.

This paper presents the likelihood ratio (LR) test for the number of cointegrating relations in the I(2) vector autoregressive model. It is shown that the asymptotic distribution of the LR test for the cointegration ranks is identical to the asymptotic distribution of the much applied test statistic based on the two-step estimation procedure in Johansen (1995, Econometric Theory 11, 25–59), Paruolo (1996, Journal of Econometrics 72, 313–356), and Rahbek, Kongsted, and Jørgensen (1999, Journal of Econometrics 90, 265–289). By construction the LR test statistic is smaller than the non-LR test statistic from the two-step procedure, and application of the LR test may change rank selection in empirical work. Based on a study of existing empirical applications and related Monte Carlo simulations we conclude that the LR test has much better size properties when compared to the two-step-based test. Overall, we propose use of the LR test for rank determination in I(2) analysis.Discussions with Søren Johansen, Hans Christian Kongsted, and Bent Nielsen are gratefully acknowledged. We also thank the editor and three anonymous referees for very constructive comments that have led to a much improved version of the paper. This research was supported by Danish Social Sciences Research Council grant 2114-04-0001.

This paper by Trygve Haavelmo was never submitted to a journal. The paper has been transcribed from Haavelmo's handwritten notes for a seminar given in 1942. The notes were kept by Haavelmo and were not known to exist by anyone else until Haavelmo died in 1999 and left his scientific remains to the Department of Economics, University of Oslo. The paper was retrieved from these personal archives by Olav Bjerkholt, who transcribed it from the original handwritten manuscript and submitted it to Econometric Theory.

In this paper, we address two important issues in semiparametric survival model selection for censored data generated by the Archimedean copula family: method of estimating the parametric copulas and data reuse. We demonstrate that for selection among candidate copula models that might all be misspecified, estimators of the parametric copulas based on minimizing the selection criterion function may be preferred to other estimators. To handle the issue of data reuse, we put model selection in the context of hypothesis testing and propose a simple test for model selection from a finite number of parametric copulas. Results from a simulation study and two empirical illustrations confirm our theoretical findings.We thank the editor Peter Phillips, three anonymous referees, and Hal White for their comments, which greatly improved the paper. An earlier version of this paper was presented at the 2005 World Congress Meetings of the Econometric Society, the 2005 Joint Statistical Meetings, the University of Waterloo, and the University of Western Ontario. Chen acknowledges support from the National Science Foundation and the C.V. Starr Center at NYU. Fan acknowledges support from the National Science Foundation. We thank Demian Pouzo for excellent research assistance on the numerical work in this paper and Weijing Wang for providing us with the Fortran code for computing the bivariate Kaplan–Meier estimator of Dabrowska (1988).

A new consistent test is proposed for the parametric specification of the diffusion function in a diffusion process without any restrictions on the functional form of the drift function. The data are assumed to be sampled discretely in a time interval that can be fixed or lengthened to infinity. The test statistic is shown to follow an asymptotic normal distribution under the null hypothesis that the parametric diffusion function is correctly specified. Monte Carlo simulations are conducted to examine the finite-sample performance of the test, revealing that the test has good size and power.The author is grateful to Yacine Aït-Sahalia, John Knight, Oliver Linton (the co-editor), Greg Tkacz, Jun Yang, and three anonymous referees for helpful comments and suggestions. He also thanks seminar participants at the Bank of Canada, the 2004 Semiparametrics Conference in Rio de Janeiro, and the 2005 Econometric Study Group in Bristol. The views expressed in this paper are those of the author. No responsibility for them should be attributed to the Bank of Canada.

Adaptations of the classical Wiener–Kolmogorov filters are described that enable them to be applied to short nonstationary sequences. Alternative filtering methods that operate in the time domain and the frequency domain are described. The frequency-domain methods have the advantage of allowing components of the data to be separated along sharp dividing lines in the frequency domain, without incurring any leakage. The paper contains a novel treatment of the start-up problem that affects the filtering of trended data sequences.

For autoregressive moving average (ARMA) models with infinite variance innovations, quasi-likelihood-based estimators (such as Whittle estimators) suffer from complex asymptotic distributions depending on unknown tail indices. This makes statistical inference for such models difficult. In contrast, the least absolute deviations estimators (LADE) are more appealing in dealing with heavy tailed processes. In this paper, we propose a weighted least absolute deviations estimator (WLADE) for ARMA models. We show that the proposed WLADE is asymptotically normal, is unbiased, and has the standard root-n convergence rate even when the variance of innovations is infinity. This paves the way for statistical inference based on asymptotic normality for heavy-tailed ARMA processes. For relatively small samples numerical results illustrate that the WLADE with appropriate weight is more accurate than the Whittle estimator, the quasi-maximum-likelihood estimator (QMLE), and the Gauss–Newton estimator when the innovation variance is infinite and that the efficiency loss due to the use of weights in estimation is not substantial.The authors thank the two referees for their valuable suggestions. The work was partially supported by an EPSRC research grant (UK) and the Natural Science Foundation of China (grant 10471005).

We consider a model for a multivariate time series where the conditional covariance matrix is a function of a finite-dimensional parameter and the innovation distribution is nonparametric. The semiparametric lower bound for the estimation of the euclidean parameter is characterized, and it is shown that adaptive estimation without reparametrization is not possible. Based on a consistent first-stage estimator (such as quasi maximum likelihood), we propose a semiparametric estimator that estimates the efficient influence function using kernel estimators. We state conditions under which the estimator attains the semiparametric lower bound. For particular models such as the constant conditional correlation model, adaptive estimation of the dynamic part of the model is shown to be possible. To avoid the curse of dimensionality one can, e.g., restrict the multivariate density to the class of spherical distributions, for which we also derive the semiparametric efficiency bound and an estimator that attains this bound. A simulation experiment demonstrates the efficiency gain of the proposed estimator compared with quasi maximum likelihood estimation.Rombouts' work was supported by the Centre for Research on e-Finance, HEC Montreal. Hafner gratefully acknowledges financial support by the Fonds Spéciaux de Recherche (FSR 05) of the Université catholique de Louvain. The authors thank three anonymous referees for valuable comments and suggestions and Luc Bauwens, Geert Dhaene, Feico Drost, Wolfgang Härdle, Douglas Hodgson, Jens Peter Kreiss, Oliver Linton, and Bas Werker for helpful discussions. We also thank participants of the CORE econometrics seminar, the York annual meeting in econometrics, the annual econometric study group meeting 2002 in Bristol, the 2003 workshop “The Art of Semiparametrics” in Berlin, and the statistics seminar of the Stockholm School of Economics for valuable comments.

We consider the cointegration tests of Johansen (1988, Journal of Economic Dynamics and Control 12, 231–254; 1991, Econometrica 59, 1551–1580) when a vector autoregressive (VAR) process of order k is used to approximate a more general linear process with a possibly infinite VAR representation. Traditional methods to select the lag order, such as Akaike's information criterion (AIC) or the Bayesian information criterion, often lead to too parsimonious a model with the implication that the cointegration tests suffer from substantial size distortions in finite samples. We extend the analysis of Ng and Perron (2001, Econometrica 69, 1519–1554) to derive a modified Akaike's information criterion (MAIC) in this multivariate setting. The idea is to use the information specified by the null hypothesis as it relates to restrictions on the parameters of the model to keep an extra term in the penalty function of the AIC. This MAIC takes a very simple form for which this extra term is simply the likelihood ratio test for testing the null hypothesis of r against more than r cointegrating vectors. We provide theoretical analyses of its validity and of the fact that cointegration tests constructed from a VAR whose lag order is selected using the MAIC have the same limit distribution as when the order is finite and known. We also provide theoretical and simulation analyses to show how the MAIC leads to VAR approximations that yield tests with drastically improved size properties with little loss of power.We are grateful to two referees for especially useful and constructive comments.

When analyzing time series an important issue is to decide whether the time series is stationary or a random walk. Relaxing these notions, we consider the problem to decide in favor of the I(0) or I(1) property. Fixed-sample statistical tests for that problem are well studied in the literature. In this paper we provide first results for the problem of monitoring sequentially a time series. Our stopping times are based on a sequential version of a kernel-weighted variance-ratio statistic. The asymptotic distributions are established for I(1) processes, a rich class of stationary processes, possibly affected by local nonparametric alternatives, and the local-to-unity model. Further, we consider the two interesting change-point models where the time series changes its behavior after a certain fraction of the observations and derive the associated limiting laws. Our Monte Carlo studies show that the proposed detection procedures have high power when interpreted as a hypothesis test and that the decision can often be made very early.The financial support of the DFG (Deutsche Forschungsgemeinschaft, SFB 475, Reduction of Complexity in Multivariate Data Structures) is gratefully acknowledged. I thank two anonymous referees for constructive and helpful remarks that improved the paper and Dipl.-Math. Sabine Teller for proofreading a revised version.

In this paper we deal with the identification of an autoregressive model for an observed time series and the detection of a unit root in its characteristic polynomial. This is a big issue concerned with distinguishing stationary time series from time series for which differencing is required to induce stationarity. We consider a Bayesian approach, and particular attention is devoted to the problem of the sensitivity of the standard Bayesian analysis with respect to the choice of the prior distribution for the autoregressive coefficients.We thank three anonymous referees for their useful comments, which have improved the final version of the paper.

This paper develops a new semiparametric approach for the estimation of hazard functions in the presence of unobserved heterogeneity. The hazard function is specified parametrically, whereas the distribution of the unobserved heterogeneity is indirectly estimated using the method of kernels. The semiparametric efficiency bounds are derived. The estimator obtains these bounds in large samples.The authors thank Yongmiao Chen, James Heckman, Hidehiko Ichimura, Tony Lancaster, Qi Li, Adrian Pagan, Barry Smith, two anonymous referees, and the co-editor for helpful input. We particularly thank Steven Stern, who prompted us toward this line of research. Any errors are those of the authors. Research funding for Rilstone was provided by the Social Sciences and Humanities Research Council of Canada.

This paper deals with issues of structure and parametrization of VECH models proposed in Bollerslev, Engle, and Wooldridge (1988) and Baba, Engle, Kraft, and Kroner (BEKK) models. Both general models and also restricted versions such as the widely used diagonal VECH (DVECH) and factor generalized autoregressive conditional heteroskedastic (F-GARCH) models are discussed. A simple algorithm is presented that checks whether a given VECH model may be cast as a BEKK model. It is shown that in the bivariate case BEKK models are as general as VECH models. In higher dimensional cases however, VECH models allow for more flexibility. In addition, a parametrization for a generic, i.e., open and dense, class of BEKK models is given, and the frequently cited parametrization by Engle and Kroner (1995, Econometric Theory 11, 122–150) is analyzed. Two shortcomings of the latter are pointed out. Finally, parametrizations for BEKK(p,q,K) models with K ≤ n, including DVECH, F-GARCH, and a generalization of the latter, are discussed.This research was supported by the project P17065 “Identification of multivariate dynamic systems with a focus on dimension reduction” of the Austrian Science Foundation (FWF).