Jonathan J Shuster1, Almut G Winterstein. 1. Department of Epidemiology and Health Policy Research, College of Medicine, University of Florida, 1600 W. Archer Raod, #100212, Gainesville, FL 32610-0212, USA. jshuster@biostat.ufl.edu
Abstract
BACKGROUND AND OBJECTIVE: To provide researchers with an efficient yet methodologically robust method to design and analyze studies of the effectiveness of quality improvement interventions targeted at medical errors. METHODS: An interrupted time-series design was chosen. Error rates of a preintervention observational period are compared to those of an intervention period (interrupted by a brief transitional period not used in the analysis). Potential errors are flagged by computerized analysis. The positive predictive value of this automated method is established by targeted expert review of random samples of monthly admissions with flagged errors. The lengths of the preintervention and intervention observational periods and number of audits per month are determined via power and type I error analysis. The setting was a future study of electronic alert systems for errors related to nephrotoxic agents in a medium-sized hospital. RESULTS: Based on these methods, a study was deemed feasible to be conducted over a 38-month period, auditing 40 potential errors per month. CONCLUSION: A logistic monthly error process model with independent variables (1) time in months (same slope pre vs. postintervention), and (2) a discrete jump postintervention to assess the effect size, offers a flexible, easy-to-interpret way to attack this problem.
BACKGROUND AND OBJECTIVE: To provide researchers with an efficient yet methodologically robust method to design and analyze studies of the effectiveness of quality improvement interventions targeted at medical errors. METHODS: An interrupted time-series design was chosen. Error rates of a preintervention observational period are compared to those of an intervention period (interrupted by a brief transitional period not used in the analysis). Potential errors are flagged by computerized analysis. The positive predictive value of this automated method is established by targeted expert review of random samples of monthly admissions with flagged errors. The lengths of the preintervention and intervention observational periods and number of audits per month are determined via power and type I error analysis. The setting was a future study of electronic alert systems for errors related to nephrotoxic agents in a medium-sized hospital. RESULTS: Based on these methods, a study was deemed feasible to be conducted over a 38-month period, auditing 40 potential errors per month. CONCLUSION: A logistic monthly error process model with independent variables (1) time in months (same slope pre vs. postintervention), and (2) a discrete jump postintervention to assess the effect size, offers a flexible, easy-to-interpret way to attack this problem.