Seminar - Testing for weekly and seasonal cycles in medical errors via Poisson regression

27 September 2012

Full title: Testing for weekly and seasonal cycles in medical errors via Poisson regression with autocorrelation

Presenter:Dr David Bulger, Senior Lecturer, Statistics Department, Macquarie University

The medical errors considered here are errors made by medical staff in health care delivery, ranging in severity from the fatal to the trivial. Medical error data were collected over a five-year period by voluntary reporting within a large rural NSW public health service. We wished to test for weekly and seasonal cycles in the frequency of errors, so Poisson regression seemed appropriate, but a fit model showed residual autocorrelation. Incorporating autocorrelation into a Poisson regression model is not straightforward: the obvious approach is to model the logarithm of the daily expected error count linearly, incorporating autoregression terms and other predictors, but this is essentially a latent variable method, and can be very demanding computationally. We applied a GLARMA model, an alternative approach due to Davis, Dunsmuir and Streett, which uses yesterday's observed error count as a predictor for today's expected error count. Controlling for the autocorrelation reduced the model variance, and we established weekly and annual cycles, showing that errors (especially less severe errors) were more common in spring and on weekends.

About the speaker: Dr Bulger is a Senior Lecturer in Macquarie University's Statistics Department. His current research projects involve statistical modeling for applications in healthcare, political science and image processing, queue theory and statistics education. Much of his past work relates to stochastic and quantum algorithms for global optimisation.

Hosted by The George Institute for International Health

Time: 11am

Location: Level 10, King George V Building, Royal Prince Alfred Hospital

Contact: Serigne Lo

Phone: 02 9657 0329

Email: 213f3b29292125365434201a3e475c194135101f1b0a00632f4d