Dienstag, 17. April 2012, 12-13 Uhr - Raum: W9-109

Dipl.-Psych.  Kristian Kleinke (BSocSc Hons)
Universität Bielefeld

Multiple Imputation of Multilevel Count Data

Throughout the last couple of years multiple imputation has become a popular and widely accepted technique to handle missing data properly. Although various multiple imputation procedures have been implemented in all major statistical packages, currently available software is still highly limited regarding the imputation of incomplete count data. As count data analysis typically makes it necessary to fit statistical models that are suited for count data like Poisson or negative binomial models, also imputation procedures should be specially tailored to the statistical specialities of count data. We present flexible and easy to use software to create multiple imputations of incomplete ordinary and overdispersed multilevel count data, based on a generalized linear mixed model with multivariate normal random effects, using penalized quasi-likelihood. Our procedure works as an add-on for the popular and powerful mice software (van Buuren & Groothuis-Oudshoorn, 2011). The advantage is that users simply work in mice and call these functions directly from mice and do not have to familiarize themselves with yet another statistical software.  

 

Dienstag, 15. Mai 2012, 12-13 Uhr - Raum: W9-109

Prof. Dr. Martin Wagner
Karl-Franzens-Universität Graz

On the econometric analysis of the environmental Kuznets curve

Recent years have seen a growing literature on the environmental Kuznets curve (EKC) that resorts in a large part to unit root and cointegration techniques. The empirical EKC literature has failed to acknowledge that such regressions involve unit root nonstationary regressors and their integer powers (e.g. GDP and GDP squared), which behave differently from linear cointegrating regressions. In this talk we discuss the necessary tools for cointegration analysis of EKC type relationships, including testing for cointegration and estimation of nonlinear cointegrating relationships.
The estimation theory allows to include pre-determined stationary regressors, deterministic regressors, unit root nonstationary regressors and their integer powers. We consider fully modified OLS estimation and specification tests based on augmented and auxiliary regressions. Cointegration testing is based on extending the KPSS test as well as on an extension of a variance ratio test of Phillips and Ouliaris (1990). We present some simulation results illustrating the performance of the estimators and tests. In the empirical application for CO2 and SO2 emissions for 19 early industrialized countries over the period 1870-2000 we only find limited evidence for an EKC, and less evidence than is ‘found’ by applying the typically used inadequate methods developed for linear cointegrating relationships.

 

Dienstag, 29. Mai 2012, 12-13 Uhr - Raum: W9-109

Prof. Regina T. Riphahn, Ph.D.
Friedrich-Alexander-Universität Erlangen-Nürnberg

Wage Mobility in East and West Germany

This article studies long run patterns and explanations of wage mobility as a characteristic of regional labor markets. Using German administrative data we describe wage mobility since 1975 in West and since 1992 in East Germany. Wage mobility declined substantially in East Germany in the 1990s and moderately in East and West Germany since the late 1990s. Therefore, wage mobility does not balance recent increases in cross-sectional wage inequality. We apply RIF regression based decompositions to measure the role of factors associated with these mobility changes. Increasing job stability plays an important role in the East German mobility decline.

 

Dienstag, 19. Juni 2012, 12-13 Uhr - Raum: W9-109

Dipl.-Vw. Christian Heinze
Universität Bielefeld

Consumer Price Level Predictions: A State Space Approach

Based on price level indices at the county level for the year 1993 and inflation figures at state and national level for subsequent years I predict regional consumer price levels in Germany for the years 1993-2005. Current interest in regional cost of living has been fueled by the agglomeration wage differential debate, e.g. Wheeler (2006), Yankow (2006), Blien et al. (2009). However, price level indices for German counties are currently not available for the years following 1993. My predictions are derived from a state space model similar to Mardia et al. (1998), which allows to combine kriging (optimal prediction in space) and Kalman filtering and smoothing techniques (optimal prediction in time). The latter yield prediction standard errors which incorporate prediction uncertainty but ignore estimation uncertainty. The bootstrap of Pfeffermann and Tiller (2005) allows to account for the latter. My state space model involves unknown variances, which may be zero. For this case Andrews (2000) was able to show that off-the-shelf parametric and non-parametric bootstrap procedures are inconsistent. Furthermore, any modification necessarily shows poor performance for small but nonzero variances. I quantify the problem by means of a Monte Carlo experiment reflecting the above setting.

 

Dipl.-Vw. Oliver Jones
Universität Bielefeld

Forecasting Regional Levels of Employment by Occupation

We forecast regional labour demand in Germany on an aggregation level that brings us to roughly 180 regions and 50 to 350 occupations. The database consists of the social security information of the workforce employed on the 31st of June each year from 1984 to 2008. A highly automatic procedure is needed since it is impossible to manually generate more than 9000 forecasts. We start with univariate time series methods, like exponential smoothing, since they have proven to be successful in mass forecasts. Then panel data methods are evaluated against them. The occupational classification that is currently used when the social security information is recorded is regrouped. The aim is to achieve better forecasts through more homogeneity within the occupations regarding the educational level. Furthermore the resulting time series should be more stable because we combine occupations for which we observe a high interchange of workforce. Because of data privacy protection only aggregated results are shown.

 

Dr. Joachim Schnurbus
Universität Bielefeld

Forecasting in Nonlinear Panel Data Regression by Stepwise Updating of Product Kernel Weights

Forecasting of Y_iT+p using a general class of nonlinear panel data models Y_i,t = g(X_it, Z_it) + U_it with error process {U_it} and unknown smooth regression function g(.), cross-section i=1,...,N and time dimension t=1,...,T, continuous and categorical predictor variables X_it and Z_it, where the former may contain exogenous and lagged endogenous variables and the latter typically contains deterministic trend and seasonal components t and s, respectively, is considered. In the framework of multiple nonparametric mixed kernel regression, continuous and discrete predictor variables X and Z are weighted by product kernel functions W(h) based on simultaneously estimated bandwidths h. The aim of the paper is to demonstrate how the initially estimated bandwidths vector h_T=(h_X, h_Z) ---or parts thereof--- can be updated for p-step predictions, avoiding the computationally burdensome simultaneous estimation of h for each new cross-section of observations T+p, p=1,2,.... The updated bandwidth vector h_T+p is derived from the changes in the structure of the smoothing (hat) matrix H, where \hat{Y}=HY, after a new observation of X and Z is available. Besides an extensive Monte Carlo study we present a forecasting exercise for a well-known scanner panel data set from consumer good marketing.

 

Dienstag, 26. Juni 2012, 12-13 Uhr - Raum: W9-109

Andrew Pua
University of Amsterdam

Applying Projected Score Methods to Panel Data Models

Panel data models allow for the control of time-constant unobserved heterogeneity of cross-sectional units. Typically, time-constant heterogeneity is not of main interest and is treated as a high-dimensional nuisance parameter in large N panels. I show that the idea of projecting the score equations for the parameters of interest on an appropriate space of functions, capturing the effects of the presence of high-dimensional nuisance parameters, can be helpful in estimating the parameters of interest in nonlinear panel data models. I also show that explicitly calculating these projections can be used to trace the impact of misspecifications of the model. Furthermore, exogeneity requirements play a crucial role in these calculations. As a consequence, projected scores are somewhat hard to calculate but very transparent. I apply the method to a class of linear and nonlinear panel data models and show that projected scores can give the same answers as profile score adjustments for varying intercept models found in the literature.