This study uses Monte Carlo experiments to produce new evidence on the performance of a wide range of panel data estimators. It focuses on estimators that are readily available in statistical software packages such as Stata and Eviews, and for which the number of cross-sectional units (N) and time periods (T) are small to moderate in size. The goal is to develop practical guidelines that will enable researchers to select the best estimator for a given type of data. It extends a previous study on the subject (Reed and Ye, Which panel data estimator should I use? 2011), and modifies their recommendations. The new recommendations provide a (virtually) complete decision tree: When it comes to choosing an estimator for efficiency, it uses the size of the panel dataset (N and T) to guide the researcher to the best estimator. When it comes to choosing an estimator for hypothesis testing, it identifies one estimator as superior across all the data scenarios included in the study. An unusual finding is that researchers should use different estimators for estimating coefficients and testing hypotheses. The authors present evidence that bootstrapping allows one to use the same estimator for both.(Replication Study) JEL C23 C33

Ever since H. E. Hurst brought the concept of long memory time series to prominence in his study of river flows the origins of the so-called Hurst phenomena have remained elusive. Two sets of competing models have been proposed. The fractional Gaussian noises and their discrete time counter-part, the fractionally integrated processes, possess genuine long memory in the sense that the present state of a system has a temporal dependence on all past states. The alternative to these genuine long memory models are models which are non-stationary in the mean but for physical reasons are constrained to lie in a bounded range, hence on visual inspection appear to be stationary. In these models the long memory is merely an artifact of the method of analysis. There are now a growing number of millenial scale temperature reconstructions available. In this paper we present a new way of looking at long memory in these reconstructions and proxies, which gives support to them being described by the non-stationary models. The implications for climatic change are that the temperature time series are not mean reverting. There is no evidence to support the idea that the observed rise in global temperatures are a natural fluctuation which will reverse in the near future.

No abstract

No abstract

In the literature many papers state that long-memory time series models such as Fractional Gaussian Noises (FGN) or Fractionally Integrated series (FI(d)) are empirically indistinguishable from models with a non-stationary mean, but which are mean reverting. We present an analysis of the statistical cost of model mis-specification when simulated long memory series are analysed by Atheoretical Regression Trees (ART), a structural break location method. We also analysed three real data sets, one of which is regarded as a standard example of the long memory type. We find that FGN and FI(d) processes do not account for many features of the real data. In particular, we find that the data sets are not H-self-similar. We believe the data sets are better characterized by non-stationary mean models.