Estimating long memory time-series-cross-section data

This paper considers the estimation and inference problems of a general class of time-series-cross-section (TSCS) models consisting of stationary or nonstationary long memory regressors and errors, while allowing for cross-correlations and serial correlations in cross-section and time dimensions, respectively. Although the applicability of this class of TSCS models is far-reaching, we show that each regression coefficient of these models can be easily tested with the critical values from the standard normal distribution based on the approach proposed in this paper. Furthermore, our approach is built on Robinson's [Robinson, P.M., 1998. Inference-without-smoothing in the presence of nonparametric autocorrelation. Econometrica 66, 1163-1182] long-run variance estimator and thus does not involve the difficult problems of choosing a kernel function or a bandwidth parameter. We also demonstrate that, under various combinations of long memory processes and cross-section dimensions, the finite sample performance of our method for this class of long memory TSCS models is promising even though the time span is only 20. We then apply this method to re-examine the welfare spending studies of Hicks and Swank [Hicks, A., Swank, D., 1992. Politics, institutions and welfare spending in industrialized democraticies, 1960-1982. American Political Science Review 86, 658-674]. The testing results are different from the findings in Hicks and Swank (1992) and those in Beck and Katz [Beck N., Katz, J.N., 1995. What to do (and not to do) with time-series cross-section data. American Political Science Review 89, 634-647], because we find a weak but significant positive voter turnout effects when the number of differencing is equal to 1.