Research That Matters (January 17 - 20, 2008)
Methods Traditionally, to analyze time series data combined with cross-sectional data like the PSID, diverse pooled cross-sectional time series models with somewhat different assumptions on error structures have widely been employed by social science researchers. Recently, among behavioral science researchers, multilevel analysis model (HLM) has been used to analyze pooled cross-sectional time series data. This paper raises the following methodological question: what is more appropriate between pooled cross-sectional time series models and multilevel model in analyzing such data? In this paper, first, two sets of pooled cross-sectional time series models are applied. The first set uses GEE estimation and allows different descriptions of the correlation matrix within cross-sections subject to the constraint that the same correlation matrix applies to all cross-sections. The following four models are used: (1) independent structure modeling that all observations are pooled into one file; (2) autoregressive structure modeling an exponentially decaying correlation in time within a cross-section; (3) stationary structure permitting a correlation between consecutive time points; and (4) non-stationary structure permitting correlations for all observations separated by time points. The second set of models uses GLS estimation. Compared with the GEE models, this set relaxes the restrictions of sameness within cross-section correlation matrices; hence, it permits the estimation of heteroscedastic variances across sections, and the estimation of cross-sectional specific autoregression. In addition, in this paper, multilevel models are used. The main difference between multilevel models and pooled cross-sectional time series models is that the association between the outcome (giving) and the independent variables is not fixed to be constant across sections. Thus, intercept and slopes are assumed to vary randomly across sections. In a two-level analysis, this variation can be predicted by variables at the level of the cross-sections.
Results In data analysis, this paper finds that the above mentioned models have similar results. The following socioeconomic variables of households have statistically significant effects on overall giving: household income, household wealth excluding home equity, itemization status, age of family head, marital status of family head, health condition of family head, the presence of an adult member volunteering in the household, the level of education of family head, religious affiliation of family head, and employment status of family head.
Iimplications This paper makes discussions on the methodological selection in analyzing panel data and some directions for future study.