Table 1.

Summary of the contrasting perspectives on accounting for non-independence described here.

	Model-based	Design-based
Goal of accounting for clustering	Better approximating the probability model for the data generating process	Accounting for sampling strategy to allow inference to a finite population of interest
Implications for analysis	A tendency toward more complexity, potentially including cross-classified models to avoid misspecification	A tendency toward less complexity, focused attention on accounting for sampling may be seen as sufficient
Cluster definition source	Relatively more emphasis on the a priori structure of the data generating process, or empirical analysis suggestive of residual clustering	Relatively more emphasis on the investigator-controlled and empirically-informed model relating cluster membership to sampling probabilities
Key analytic technique(s)	Multi-level models, generalized estimating equations or cluster robust standard errors	Models incorporating complex sampling weights, which may include multi-level models or generalized estimating equations

Note: While we emphasize for clarity the divergent implications of the model-based and design-based perspectives, both perspectives are flexible and there is much potential for overlap and integration