Essays in Honor of Aman Ullah: Volume 36

This paper examines Aman Ullah’s contributions to robust inference, finite sample econometrics, nonparametrics and semiparametrics, and panel and spatial models. His early works on robust inference and finite sample theory were mostly motivated by his thesis advisor, Professor Anirudh Lal Nagar. They eventually led to his most original rethinking of many statistics and econometrics models that developed into the monograph Finite Sample Econometrics published in 2004. His desire to relax distributional and functional-form assumptions lead him in the direction of nonparametric estimation and he summarized his views in his most influential textbook Nonparametric Econometrics (with Adrian Pagan) published in 1999 that has influenced a whole generation of econometricians. His innovative contributions in the areas of seemingly unrelated regressions, parametric, semiparametric and nonparametric panel data models, and spatial models have also inspired a larger literature on nonparametric and semiparametric estimation and inference and spurred on research in robust estimation and inference in these and related areas.

This paper considers the problem of estimating a partially linear varying coefficient fixed effects panel data model. Using the series method, we establish the root N normality for the estimator of the parametric component; and we show that the unknown function can be consistently estimated at the standard nonparametric rate. Furthermore, we extend the model to allow endogeneity in the parametric component and establish the asymptotic properties of the semiparametric instrumental variable estimators.

This paper revisits the joint and conditional Lagrange multiplier tests derived by Debarsy and Ertur (2010) for a fixed effects spatial lag regression model with spatial autoregressive error, and derives these tests using artificial double length regressions (DLR). These DLR tests and their corresponding LM tests are compared using an empirical example and a Monte Carlo simulation.

This paper develops a cross-sectionally augmented distributed lag (CS-DL) approach to the estimation of long-run effects in large dynamic heterogeneous panel data models with cross-sectionally dependent errors. The asymptotic distribution of the CS-DL estimator is derived under coefficient heterogeneity in the case where the time dimension (T ) and the cross-section dimension (N ) are both large. The CS-DL approach is compared with more standard panel data estimators that are based on autoregressive distributed lag (ARDL) specifications. It is shown that unlike the ARDL-type estimator, the CS-DL estimator is robust to misspecification of dynamics and error serial correlation. The theoretical results are illustrated with small sample evidence obtained by means of Monte Carlo simulations, which suggest that the performance of the CS-DL approach is often superior to the alternative panel ARDL estimates, particularly when T is not too large and lies in the range of 30–50.

In this paper, we study a partially linear dynamic panel data model with fixed effects, where either exogenous or endogenous variables or both enter the linear part, and the lagged-dependent variable together with some other exogenous variables enter the nonparametric part. Two types of estimation methods are proposed for the first-differenced model. One is composed of a semiparametric GMM estimator for the finite-dimensional parameter θ and a local polynomial estimator for the infinite-dimensional parameter m based on the empirical solutions to Fredholm integral equations of the second kind, and the other is a sieve IV estimate of the parametric and nonparametric components jointly. We study the asymptotic properties for these two types of estimates when the number of individuals N tends to ∞ and the time period T is fixed. We also propose a specification test for the linearity of the nonparametric component based on a weighted square distance between the parametric estimate under the linear restriction and the semiparametric estimate under the alternative. Monte Carlo simulations suggest that the proposed estimators and tests perform well in finite samples. We apply the model to study the relationship between intellectual property right (IPR) protection and economic growth, and find that IPR has a non-linear positive effect on the economic growth rate.

I derive the finite-sample bias of the conditional Gaussian maximum likelihood estimator in ARMA models when the error follows some unknown non-normal distribution. The general procedure relies on writing down the score function and its higher order derivative matrices in terms of quadratic forms in the non-normal error vector with the help of matrix calculus. Evaluation of the bias can then be straightforwardly conducted. I give further simplified bias results for some special cases and compare with the existing results in the literature. Simulations are provided to confirm my simplified bias results.

This paper considers the finite-sample distribution of the 2SLS estimator and derives bounds on its exact bias in the presence of weak and/or many instruments. We then contrast the behavior of the exact bias expressions and the asymptotic expansions currently popular in the literature, including a consideration of the no-moment problem exhibited by many Nagar-type estimators. After deriving a finite-sample unbiased k-class estimator, we introduce a double-k-class estimator based on Nagar (1962) that dominates k-class estimators (including 2SLS), especially in the cases of weak and/or many instruments. We demonstrate these properties in Monte Carlo simulations showing that our preferred estimators outperform Fuller (1977) estimators in terms of mean bias and MSE.

Although in principle prior information can significantly improve inference, incorporating incorrect prior information will bias the estimates of any inferential analysis. This fact deters many scientists from incorporating prior information into their inferential analyses. In the natural sciences, where experiments are more regularly conducted, and can be combined with other relevant information, prior information is often used in inferential analysis, despite it being sometimes nontrivial to specify what that information is and how to quantify that information. In the social sciences, however, prior information is often hard to come by and very hard to justify or validate. We review a number of ways to construct such information. This information emerges naturally, either from fundamental properties and characteristics of the systems studied or from logical reasoning about the problems being analyzed. Borrowing from concepts and philosophical reasoning used in the natural sciences, and within an info-metrics framework, we discuss three different, yet complimentary, approaches for constructing prior information, with an application to the social sciences.

We examine the potential effect of naturalization on the U.S. immigrants’ earnings. We find the earning gap between naturalized citizens and noncitizens is positive over many years, with a tent shape across the wage distribution. We focus on a normalized metric entropy measure of the gap between distributions, and compare with conventional measures at the mean, median, and other quantiles. In addition, naturalized citizen earnings (at least) second-order stochastically dominate noncitizen earnings in many of the recent years. We construct two counterfactual distributions to further examine the potential sources of the earning gap, the “wage structure” effect and the “composition” effect. Both of these sources contribute to the gap, but the composition effect, while diminishing somewhat after 2005, accounts for about 3/4 of the gap. The unconditional quantile regression (based on the Recentered Influence Function), and conditional quantile regressions confirm that naturalized citizens have generally higher wages, although the gap varies for different income groups, and has a tent shape in many years.

Many Information Theoretic Measures have been proposed for a quantitative assessment of causality relationships. While Gouriéroux, Monfort, and Renault (1987) had introduced the so-called “Kullback Causality Measures,” extending Geweke’s (1982) work in the context of Gaussian VAR processes, Schreiber (2000) has set a special focus on Granger causality and dubbed the same measure “transfer entropy.” Both papers measure causality in the context of Markov processes. One contribution of this paper is to set the focus on the interplay between measurement of (non)-markovianity and measurement of Granger causality. Both of them can be framed in terms of prediction: how much is the forecast accuracy deteriorated when forgetting some relevant conditioning information? In this paper we argue that this common feature between (non)-markovianity and Granger causality has led people to overestimate the amount of causality because what they consider as a causality measure may also convey a measure of the amount of (non)-markovianity. We set a special focus on the design of measures that properly disentangle these two components. Furthermore, this disentangling leads us to revisit the equivalence between the Sims and Granger concepts of noncausality and the log-likelihood ratio tests for each of them. We argue that Granger causality implies testing for non-nested hypotheses.

This paper considers properties of an optimization-based sampler for targeting the posterior distribution when the likelihood is intractable. It uses auxiliary statistics to summarize information in the data and does not directly evaluate the likelihood associated with the specified parametric model. Our reverse sampler approximates the desired posterior distribution by first solving a sequence of simulated minimum distance problems. The solutions are then reweighted by an importance ratio that depends on the prior and the volume of the Jacobian matrix. By a change of variable argument, the output consists of draws from the desired posterior distribution. Optimization always results in acceptable draws. Hence, when the minimum distance problem is not too difficult to solve, combining importance sampling with optimization can be much faster than the method of Approximate Bayesian Computation that by-passes optimization.

Modelling and forecasting interval-valued time series (ITS) have received increasing attention in statistics and econometrics. An interval-valued observation contains more information than a point-valued observation in the same time period. The previous literature has mainly considered modelling and forecasting a univariate ITS. However, few works attempt to model a vector process of ITS. In this paper, we propose an interval-valued vector autoregressive moving average (IVARMA) model to capture the cross-dependence dynamics within an ITS vector system. A minimum-distance estimation method is developed to estimate the parameters of an IVARMA model, and consistency, asymptotic normality and asymptotic efficiency of the proposed estimator are established. A two-stage minimum-distance estimator is shown to be asymptotically most efficient among the class of minimum-distance estimators. Simulation studies show that the two-stage estimator indeed outperforms other minimum-distance estimators for various data-generating processes considered.

This paper considers stationary regression models with near-collinear regressors. Limit theory is developed for regression estimates and test statistics in cases where the signal matrix is nearly singular in finite samples and is asymptotically degenerate. Examples include models that involve evaporating trends in the regressors that arise in conditions such as growth convergence. Structural equation models are also considered and limit theory is derived for the corresponding instrumental variable (IV) estimator, Wald test statistic, and overidentification test when the regressors are endogenous. It is shown that near-singular designs of the type considered here are not completely fatal to least squares inference, but do inevitably involve size distortion except in special Gaussian cases. In the endogenous case, IV estimation is inconsistent and both the block Wald test and Sargan overidentification test are conservative, biasing these tests in favor of the null.

The asymptotic bias and variance of a general class of local polynomial estimators of M-regression functions are studied over the whole compact support of the multivariate covariate under a minimal assumption on the support. The support assumption ensures that the vicinity of the boundary of the support will be visited by the multivariate covariate. The results show that like in the univariate case, multivariate local polynomial estimators have good bias and variance properties near the boundary. For the local polynomial regression estimator, we establish its asymptotic normality near the boundary and the usual optimal uniform convergence rate over the whole support. For local polynomial quantile regression, we establish a uniform linearization result which allows us to obtain similar results to the local polynomial regression. We demonstrate both theoretically and numerically that with our uniform results, the common practice of trimming local polynomial regression or quantile estimators to avoid “the boundary effect” is not needed.

It is known that model averaging estimators are useful when there is uncertainty governing which covariates should enter the model. We argue that in applied research there is also uncertainty as to which method one should deploy, prompting model averaging over user-defined choices. Specifically, we propose, and detail, a nonparametric regression estimator averaged over choice of kernel, bandwidth selection mechanism and local-polynomial order. Simulations and an empirical application are provided to highlight the potential benefits of the method.

For kernel-based estimators, smoothness conditions ensure that the asymptotic rate at which the bias goes to zero is determined by the kernel order. In a finite sample, the leading term in the expansion of the bias may provide a poor approximation. We explore the relation between smoothness and bias and provide estimators for the degree of the smoothness and the bias. We demonstrate the existence of a linear combination of estimators whose trace of the asymptotic mean-squared error is reduced relative to the individual estimator at the optimal bandwidth. We examine the finite-sample performance of a combined estimator that minimizes the trace of the MSE of a linear combination of individual kernel estimators for a multimodal density. The combined estimator provides a robust alternative to individual estimators that protects against uncertainty about the degree of smoothness.

Estimators for derivatives associated with a density function can be useful in identifying its modes and inflection points. In addition, these estimators play an important role in plug-in methods associated with bandwidth selection in nonparametric kernel density estimation. In this paper, we extend the nonparametric class of density estimators proposed by Mynbaev and Martins-Filho (2010) to the estimation of m-order density derivatives. Contrary to some existing derivative estimators, the estimators in our proposed class have a full asymptotic characterization, including uniform consistency and asymptotic normality. An expression for the bandwidth that minimizes an asymptotic approximation for the estimators’ integrated squared error is provided. A Monte Carlo study sheds light on the finite sample performance of our estimators and contrasts it with that of density derivative estimators based on the classical Rosenblatt–Parzen approach.

Local polynomial regression is extremely popular in applied settings. Recent developments in shape-constrained nonparametric regression allow practitioners to impose constraints on local polynomial estimators thereby ensuring that the resulting estimates are consistent with underlying theory. However, it turns out that local polynomial derivative estimates may fail to coincide with the analytic derivative of the local polynomial regression estimate which can be problematic, particularly in the context of shape-constrained estimation. In such cases, practitioners might prefer to instead use analytic derivatives along the lines of those proposed in the local constant setting by Rilstone and Ullah (1989). Demonstrations and applications are considered.

We propose a nonparametric estimator of the Lorenz curve that satisfies its theoretical properties, including monotonicity and convexity. We adopt a transformation approach that transforms a constrained estimation problem into an unconstrained one, which is estimated nonparametrically. We utilize the splines to facilitate the numerical implementation of our estimator and to provide a parametric representation of the constructed Lorenz curve. We conduct Monte Carlo simulations to demonstrate the superior performance of the proposed estimator. We apply our method to estimate the Lorenz curve of the U.S. household income distribution and calculate the Gini index based on the estimated Lorenz curve.