When “contexts” are geographical areas, is multilevel model still a good choice to model hierarchical data, or a new approach is needed?
29 September 2015
Speaker: Dr Guanpeng Dong, Sheffield Methods Institute, University of Sheffield
It is very common that our research uses hierarchical data where the higher level units or “contexts” are defined as geographical areas—for example, individuals nest into census units or houses into districts. In such situations, we need to think about three questions:
- Are lower-level units correlated with each other if they are in the same “context” or group?
- Are the interactions or correlations among lower-level units strictly bounded within “contexts” or groups?
- Are contexts themselves independent of each other? The first effect is referred to as a vertical group dependence effect. The latter two can be considered as horizontal dependence effects at each level of the data hierarchy.
If the last two dependence effects were suspected, standard multilevel models would not be a good modelling choice. Instead, an integrated spatial and multilevel model could be used to deal with the vertical and horizontal dependence simultaneously.