University of Washington - Department of Statistics
It is well known that bias can result from making inference about individuals based on group level (ecological) data. However, ecological data are readily available for many applications, and we show that ecological data can provide information when used in combination with individual level data. We also show that information can be increased by using the ecological data to inform the sampling design for individual data.
Additionally, the interplay of group level and individual level effects in these multilevel models will complicate the interpretation of results. Correct causal inference will therefore depend on both the level of planned intervention and the true causal model. By representing the multilevel models with Directed Acyclic Graphs, we show that we can test between some of the possible causal models.