| dc.description.abstract | Adverse events are commonly reported among patients using the highly active antiretroviral therapy (HAART). These adverse events present as count outcomes which have non-normal distribution. The regression models appropriate for counted data have seen little use in adverse events analysis and in HAART use specifically. Generalised linear models offer suitable alternatives to the analysis of count data.
The simplest, the Poisson regression model is the starting point for count data analysis. It is however inadequate and likely to be misleading unless restrictive assumptions are met because individual counts are usually more variable i.e. overdispersed than is implied by the model. In the negative binomial regression model, a random term reflecting unexplained between-subject differences is included in the regression model to account for overdispersion.
This study assesses two methods of dealing with deviations from normal assumptions, with a focus on the application of generalized linear models (GLMs), specifically the poisson and negative binomial models. The utility of these models is examined by assessing which factors are likely to predict adverse events to HAART. The likelihood ratios and the Bayesian Information Criteria (BIC) are used to adjudicate on the most parsimonious model.
Results demonstrated that the negative binomial model was more optimal in the assessment of adverse events compared to the poisson model. The data were found to have overdispersion and this was efficiently addressed by the negative binomial model. Months on treatment were associated with adverse events (RR 0.88; p-value 0.000, CI 0.82-0.94). Patients who were on treatment longer experienced less adverse events. | en_US |
