Analysts may empirically check if the outcome variable

Analysts may empirically check if the outcome variable has an ICC that lipid metabolism is distinguishable from zero, as introduced briefly above. Even those favoring a priori specification of how the model will account for clustering, coauthor or reviewer suggestions may warrant an empirically-informed response. The ICC can be approximated even if the outcome variable is dichotomous (Merlo et al., 2006; Ridout, Demetrio, & Firth, 1999). While often ICC is evaluated for the outcome variable, examining the ICC for model residuals is common as well, and is arguably even more closely aligned with the question of whether there is unexplained non-independence in our data. If the ICC for our outcome is estimated as zero, there is no variation between groups beyond what would be expected by chance. We may then have little to gain by explicitly accounting for clustering. However, what ICC we consider as substantially greater than zero is dependent on the research context and the dataset; small but statistically significant ICCs may easily result from the use of ‘Big Data’ (Mooney, Westreich, & El-Sayed, 2015). Point estimates and confidence intervals (Snijders & Bosker, 2012b) thus may be more informative than p-values alone, and researchers may find comparisons to an r statistic helpful when interpreting the magnitude of an ICC. Both the ICC and r are describing a portion of variance explained (by the cluster-level identifier or by covariates in a regression model, respectively), and for both of these we would expect to occasionally note values as low as 2% (ICC or r of 0.02) as statistically distinguishable from zero, and explaining 5% or more of the outcome variance (an ICC or r ≥ 0.05) is considered important (Subramanian & O’Malley, 2010) and occasionally much higher ICCs are noted in health research (J Merlo, Wagner, Ghith, & Leckie, 2016; Rodriguez & Goldman, 1995, 2001).
Conclusions Some caution is clearly warranted to make sure we are appropriately accounting for clustering with awareness of our perspectives and with attention to what is needed to address our research questions. Table 1 summarizes some key contrasts between the model-based and design-based perspectives. Whereas a model-based perspective emphasizes the probability model generating the data, a design-based perspective emphasizes the need to account for how the data were sampled. However, these perspectives only occasionally surface as a clear difference of opinion about how to proceed. Indeed, given that both our modeling strategies and our sampling are always imperfect aspects of each perspective are often co-mingled in teaching and in practice (Gelman & Hill, 2007; Snijders & Bosker, 2012d; Sterba, 2009). However, since investigative and mentorship teams may span multiple perspectives, attention to each is warranted, and may be particularly important as we select a cluster definition.
Ethics approval
Acknowledgements This work was supported by grants from the National Institute of Child Health and Human Development (G.S.L., grant K01HD067390, S.J.M., grant 5T32HD057822) and National Institute of Drug Abuse (D.S.F., grant number T32DA031099). The funding source had no role in the writing of this article nor the decision to submit it for publication. The authors would like to acknowledge Dr. Jeff Goldsmith for his critical review of previous draft.
Introduction Heart disease has remained the leading cause of death in the U.S. since 1950 (Heron and Anderson, 2016), despite substantial declines over 50 years (Cooper et al., 2000). This reduction is due, in part, to public health prevention and improved clinical treatment (Ford et al., 2007). Declines in heart disease mortality are evident across racial, gender, and geographic lines. For example, the average U.S. county experienced a 62% reduction in heart disease mortality rates between 1973 and 2010, but this varied dramatically by region, with many southern counties experiencing slower declines substantially less than 50%, and northern and western counties experiencing faster declines from 65–82%. (Casper et al., 2016). Furthermore, county-specific declines were faster for Whites (63% on average) than for Blacks (54% on average) nationwide (Vaughan, Quick, Pathak, Kramer, & Casper, 2015). In this context of notable but uneven progress in addressing a complex chronic disease, heart disease represents both a success story for public health efforts, as well as an opportunity to learn lessons about the challenges of promoting equitable improvements in population health.