| Summary: | Thomas P Ahern1, Jaclyn LF Bosco2, Rebecca A Silliman2, Marianne Ulcickas Yood3, Terry S Field4, Feifei Wei5, Timothy L Lash1, On behalf of the BOW Investigators1Department of Epidemiology, Boston University School of Public Health, Boston, MA, USA; 2Department of Medicine, Section of Geriatrics, Boston University School of Medicine, Boston, MA, USA; 3Department of Epidemiology and Public Health, Yale University School of Medicine, New Haven, CT, USA; 4Meyers Primary Care Institute, University of Massachusetts Medical School, Worcester, MA, USA; 5HealthPartners Research Foundation, Minneapolis, MN, USABackground: Comorbidity indices summarize complex medical histories into concise ordinal scales, facilitating stratification and regression in epidemiologic analyses. Low subject prevalence in the highest strata of a comorbidity index often prompts combination of upper categories into a single stratum (‘collapsing’).Objective: We use data from a breast cancer cohort to illustrate potential inferential errors resulting from collapsing a comorbidity index. Methods: Starting from a full index (0, 1, 2, 3, and ≥4 comorbidities), we sequentially collapsed upper categories to yield three collapsed categorizations. The full and collapsed categorizations were applied to analyses of (1) the association between comorbidity and all-cause mortality, wherein comorbidity was the exposure; (2) the association between older age and all-cause mortality, wherein comorbidity was a candidate confounder or effect modifier. Results: Collapsing the index attenuated the association between comorbidity and mortality (risk ratio, full versus dichotomized categorization: 4.6 vs 2.1), reduced the apparent magnitude of confounding by comorbidity of the age/mortality association (relative risk due to confounding, full versus dichotomized categorization: 1.14 vs 1.09), and obscured modification of the association between age and mortality on both the absolute and relative scales.Conclusions: Collapsing categories of a comorbidity index can alter inferences concerning comorbidity as an exposure, confounder and effect modifier.Keywords: epidemiology, breast neoplasms, comorbidity, confounding factors (epidemiologic), bias (epidemiologic), statistical models
|
|---|
