Mini Oral Australian Epidemiology Association ASM 2018

Modelling antibiotics prescription: comparison of different statistical models in analysis of count data from otitis media clinical trial (#22)

Victor Oguoma 1 , Jemima Beissbarth 1 , Michael Binks 1 , Nicole Wilson 1 , Peter Morris 1 , Amanda Leach 1
  1. Division of Child Health , Menzies School of Health Research, Darwin, NT, Australia

Background:

Otitis media is one of the leading causes of antibiotic prescriptions in children. In our trial comparing pneumococcal conjugate vaccines, antibiotics prescribing is a problem. Poisson regression is a commonly applied method in the analysis of count outcome data. However, many real-life data violate the Poisson assumption. We examined different statistical models for count data using a preliminary dataset of antibiotic prescriptions for acute otitis media (AOM) and chronic suppurative otitis media (CSOM) to ascertain model fit of Poisson-based models in relation to others.

Methods:

Antibiotic prescriptions in children attending the health clinics for parent-driven presentation and active research examinations with a diagnosis of any AOM and CSOM was compared using different Poisson-based and negative binomial-based models for count data. Empirical model selection was assessed using log-likelihood test (LL), Akaike information criterion (AIC), Bayesian Information criterion (BIC) and the difference between observed and predicted probabilities of each model.

Results:

Over first 2 years of life ~ 1300 antibiotic prescriptions were made following ~ 1410 diagnosis of any AOM and CSOM. The variance was greater than the mean indicating potential over dispersion. The goodness-of-fit statistics based on LL, AIC and BIC gave varying results for different models, but the difference between the observed and predicted probabilities best fit the data for NB-based models.

Conclusion:

Comprehensive study specific model selection is essential to ensure appropriate reporting and inferences that are a true representation of sample data are made. However, in RCTs, its important to specify the model prior to data analysis.