6 Multilevel modelling
Relatively unexplored area of research.
Li, Zaslavsky, and Landrum (2013) showed that “exploiting the multilevel structure in at least one stage can greatly reduce the bias”. They emphasize that propensity score methods offer a more robust alternative to regression adjustment, especially in complex multilevel observational data where correctly specifying the outcome model may be challenging.
However, to properly estimate standard error for the treatment effect estimate from a multi-level or complex survey (where clustering and or stratification are present), it is necessary to address clustering options through outcome modelling (Austin, Jembere, and Chiu 2018).