br Conclusions br Ethics approval

Conclusions
Ethics approval
Acknowledgements
Background Human experience takes place in multiple overlapping contexts, including geographic contexts such as neighborhoods and cities, organizational contexts such as schools and clinics, and social contexts such as SR 3576 and friendship networks. Though the variability of health-relevant exposures and outcomes within and between these contexts has long been a focus of study (Mooney, Knox, & Morabia, 2014; Morabia, 2014; Pincus & Stern, 1937), in recent years, research teams have increasingly had opportunities to link measures from more than one type of context within the same study population (Box 1). The integration of multiple context types into our research reflects the multiplicity of overlapping contexts that shape our social experience and related health risks. Health-relevant sorting into neighborhoods (Bischoff & Reardon, 2013), schools (Reardon & Owens, 2014), clinics (Sarrazin, Campbell, Richardson, & Rosenthal, 2009), and workplaces (Goh, Pfeffer, & Zenios, 2015) has been well-documented in the literature, complicating our ability to study the implication of changing such contexts for our health. Beyond physical contexts there are social networks and affinity groups that affect the health of individuals. The numerous overlapping contexts in which individuals are embedded result in correlations within “clusters” (a term that we will use for brevity to indicate the spatial units, institutional settings, or other macro-units to which individuals in a study population are indexed via a cluster identifier). The availability of repeated measures over time in longitudinal studies bring further complexity as well as value (Leckie, 2009). One or more of the clusters may take on particular salience because of the study design or context characteristics available for linkage (Fig. 1). Doing so makes salient the often implicit tensions between two inferential perspectives labeled as model-based and design-based. This paper identifies two common perspectives and their implications when considering a clustering-based analytic approach (e.g. by using random effects or cluster robust standard errors) for studies linking context to health. As such analytic approaches have become easier to implement in standard statistical software (Diez Roux, 2000; Singer, 1998), how specifically to analyze clustered data, and whether hierarchical or cross-classified techniques are truly necessary, should be considered carefully (Mitchell, 2001). Attention to what have been called “model-based” and “design-based” inference goals (Snijders & Bosker, 2012c; Sterba, 2009), and the tensions between perspectives and within each perspective, can elucidate how we decide on which cluster definition (or definitions) to account for, a decision that in turn affects all subsequent analytic and inferential steps. We aim to provide insight into these perspectives and their implications for an applied population health research audience. We first discuss distinguishing features of each perspective, and then turn to how they offer divergent guidance under the increasingly common circumstance of having more than one type of context available to account for non-independence (Fig. 1). Consider, for example, an investigation of swimming skills (Hulteen et al., 2015) among children in a given city, with relevance to both physical activity (Fisher et al., 2005) and drowning risk (Brenner et al., 2009). The investigative team systematically samples schools within the city, and then children within those schools, such that sampling probabilities are known. Suppose also that residents of some neighborhoods have received frequent marketing of private swimming lessons at their local swimming pool (for the sake of illustration, we suppose this is unmeasured, as would often be the case for local social norms or other behaviorally-relevant characteristics of context). Empirically, it might be that residual clustering in the outcome is greater based on neighborhood than by school. Exposures of interest addressed by the investigative team across several empirical manuscripts are defined at the individual (e.g., gender), school (e.g., physical education hours/week), and neighborhood level (e.g., area-based socioeconomic indicators). A team that adopts a design-based perspective would be attentive to sampling weights and inference to the city population, but might not require adding a random effect to account for within-neighborhood clustering because that clustering is a reflection of the clustering truly present in the city (rather than being investigator-imposed). By contrast, the model-based perspective would primarily be focused on specification of the model, accounting for neighborhood clustering if the processes shaping the skills of two children within the same neighborhood are not considered independent; the model-based team might consider an unweighted analysis using adjustment as a possibly more efficient alternative to a weighted analysis. Both perspectives are flexible, and ideally the advantages of each will be considered, but we posit that being able to name and distinguish them will help to avoid confusion.