Donald A Redelmeier1, Robert J Tibshirani2. 1. Department of Medicine, University of Toronto, Toronto, Ontario M5S 1A1, Canada; Evaluative Clinical Sciences Program, Sunnybrook Research Institute, 2075 Bayview Avenue, Toronto, Ontario M4N 3M5, Canada; Institute for Clinical Evaluative Sciences, G1 06, 2075 Bayview Avenue, Toronto, Ontario M4N 3M5, Canada; Division of General Internal Medicine, Sunnybrook Health Science Centre, 2075 Bayview Avenue, Toronto, ON M4N 3M5, Canada; Center for Leading Injury Prevention Practice Education & Research, Toronto, Ontario M4N 3M5, Canada. Electronic address: dar@ices.on.ca. 2. Department of Biomedical Data Sciences, Stanford University, 450 Serra Mall, Stanford, CA 94305, USA; Department of Statistics, Stanford University, 450 Serra Mall, Stanford, CA 94305, USA.
Abstract
OBJECTIVES: To demonstrate analytic approaches for matched studies where two controls are linked to each case and events are accumulating counts rather than binary outcomes. A secondary intent is to clarify the distinction between total risk and excess risk (unmatched vs. matched perspectives). STUDY DESIGN AND SETTING: We review past research testing whether elections can lead to increased traffic risks. The results are reinterpreted by analyzing both the total count of individuals in fatal crashes and the excess count of individuals in fatal crashes, each time accounting for the matched double controls. RESULTS: Overall, 1,546 individuals were in fatal crashes on the 10 election days (average = 155/d), and 2,593 individuals were in fatal crashes on the 20 control days (average = 130/d). Poisson regression of total counts yielded a relative risk of 1.19 (95% confidence interval: 1.12-1.27). Poisson regression of excess counts yielded a relative risk of 3.22 (95% confidence interval: 2.72-3.80). The discrepancy between analyses of total counts and excess counts replicated with alternative statistical models and was visualized in graphical displays. CONCLUSION: Available approaches provide methods for analyzing count data in matched designs with double controls and help clarify the distinction between increases in total risk and increases in excess risk.
OBJECTIVES: To demonstrate analytic approaches for matched studies where two controls are linked to each case and events are accumulating counts rather than binary outcomes. A secondary intent is to clarify the distinction between total risk and excess risk (unmatched vs. matched perspectives). STUDY DESIGN AND SETTING: We review past research testing whether elections can lead to increased traffic risks. The results are reinterpreted by analyzing both the total count of individuals in fatal crashes and the excess count of individuals in fatal crashes, each time accounting for the matched double controls. RESULTS: Overall, 1,546 individuals were in fatal crashes on the 10 election days (average = 155/d), and 2,593 individuals were in fatal crashes on the 20 control days (average = 130/d). Poisson regression of total counts yielded a relative risk of 1.19 (95% confidence interval: 1.12-1.27). Poisson regression of excess counts yielded a relative risk of 3.22 (95% confidence interval: 2.72-3.80). The discrepancy between analyses of total counts and excess counts replicated with alternative statistical models and was visualized in graphical displays. CONCLUSION: Available approaches provide methods for analyzing count data in matched designs with double controls and help clarify the distinction between increases in total risk and increases in excess risk.