Modeling the correct level of analysis in non-aggregated household panel data: A simulation approach

Multilevel models can deal with nested structures in household panel data to derive unbiased regression coefficients and standard errors for predictors from multiple hierarchical levels, e.g., households, products, or stores. Within the framework of multilevel modeling, researchers can apply purely nested models or cross-classified random effects models (CCREM). This paper explains the partially cross-classified structure in household panel data. Simulation study 1 demonstrates that standard errors for level-two predictors are severely downward biased when applying a nested three-level model to partially crossed data. Furthermore, the hierarchical location of interactions between predictors associated with two crossed levels is explained. Simulation study 2 demonstrates that with unbalanced real-world data, both standard errors and regression coefficients for interaction-level predictors can be biased when the "artificial" random interaction level is omitted from a CCREM. The simulation studies are followed by a discussion of implications for the application of multilevel models to household panel data.