Dependence in macroeconomic variables: Assessing instantaneous and persistent relations between and within time series

Abstract

The present thesis comprises two rather independent chapters. In general, the diagnosis and quantification of dependence is a major aim of econometric studies. Along these lines, the concept of dependence serves as an encompassing framework to analyze time series with two very different techniques. First, we consider a single macroeconomic time series. A series which incorporates only temporary deviations from deterministic terms provides a different starting point for economic interpretations than an ‘unpredictable’ (random) series. In the context of dependency, we are interested if a time series is stationary or if it exhibits persistent dependence on larger horizons. We consider univariate tests for stationarity under distinct model settings. More recently, panel unit root tests have been developed to overcome power deficiencies and provide a more general economic statement. The standard panel unit root tests are not robust under time-varying variances and trending data. Against this background, we introduce a new test procedure which performs well in this setting. Departing from the framework of a single time series the diagnosis of dependencies between several variables provides evidence on relations in a macroeconomic system. As- suming stationarity of the series, we analyze instantaneous and persistent effects between macroeconomic indices by means of vector autoregressive models. Dependencies can, be- yond the standard linear setting, be present in diverse forms. We first refer to the variety of dependencies. Subsequently, we review and compare nonparametric measures which are developed to robustly diagnose various dependence types. Drawing on these dependence diagnostics helps to conduct a preliminary analysis of the data. In macroeconomics, the analyst might be further interested in causalities. We analyze causalities by means of struc- tural vector autoregressive models tracing the variables back to unanticipated independent shocks. Hereby, we are mainly interested in the identification of instantaneous response ma- trices relying on non-Gaussianity of the structural shocks. We compare such independence based identification procedures relying on beforehand selected nonparametric dependence measures. Furthermore, we highlight their performance by means of a simulation study. The assumption of at most one Gaussian (structural) shocks is essential for unique identi- fication of the instantaneous response to independent shocks. However, in a system with multiple Gaussian shocks the non-Gaussian ones can still be uniquely identified. We prove this result and, thereby, enable a more general definition of identifiability.