Pharmacoepidemiology & Drug Safety (2006), 15, 477-484.
John A. Mandryk1, Judith M. Mackson1, Fiona E. Horn1, Sonia E. Wutzke1, Caro-Anne Badcock2, Rob J. Hyndman3 and Lynn M. Weekes1
- National Prescribing Service Ltd, Surry Hills, Sydney, Australia.
- Covance Pty Ltd, Covance Biostatistics Asia-Pacific, North Ryde, Sydney, Australia.
- Department of Econometrics and Business Statistics, Monash University, Clayton VIC 3800, Australia.
Purpose: The National Prescribing Service Ltd (NPS) aims to improve prescribing and use of medicines consistent with evidence-based best practice. This study compares two statistical methods used to determine whether multiple educational interventions influenced antibiotic prescribing in Australia.
Methods Monthly data (July 1996 to June 2003) were obtained from a national administrative claims database. The outcome measures were the median number of antibiotic prescriptions per 1,000 consultations for each general practitioner (GP) each month, and the mean proportion (across GPs) of each subgroup of antibiotics (e.g. roxithromycin) out of the nine antibiotics having primary use for upper respiratory tract infection. Two methods were used to investigate shifts in prescribing: augmented regression, which included seasonality, autocorrelation and one intervention; and seasonally adjusted piecewise linear dynamic regression, which removed seasonality prior to modelling, included several interventions, a term for GP participation in NPS activities, and autocorrelated errors.
Results Both methods described a similar decrease in rates, with a non-significant increase after the first intervention in April 1999 – the inclusion of more interventions and the GP participation term made no difference. Using roxithromycin as an example of the analyses of proportions, both methods implied that after April 1999 the proportion significantly decreased. The statistical significance of this intervention disappears when all interventions and the GP term are included.
Conclusions The two analyses provide similar conclusions but raise questions about what is the best way to model drug utilization data, particularly regarding whether to include additional intervention terms when they are part of an extended roll-out of related interventions.
Keywords: claims data; longitudinal study; antibiotics; time series; seasonality; regression methods; evaluation.