The Use of Poisson-Mixture Models for Evaluating the Risk of Typhoid Fever in Oyo State, Nigeria

Poisson and negative binomial regressions are popularly used to model count data. However, they have limitations of not accounting for excess zeros and over-dispersion. Hence, this study compared the performance of four Poisson-mixture models in identifying factors influencing the number of typhoid fever cases(TFC) in Oyo State. Surveillance data on TFC, extracted from the Oyo State Integrated Disease Surveillance and Response database (2011 to 2014), were used. Presence of over dispersion (variance exceeds mean) in the data was investigated. The zero-inflated Poisson, zero-inflated negative binomial (ZINB), zero inflated generalized Poisson, and zero-altered Poisson models were fitted to the data, and the best model was selected based on the least AIC value. There were 38,342 reported cases of TFC within the year 2011 to 2014 and 8,118 (73.2%) zero cases. The mean was 3.46, while the variance was141.39 which indicated over-dispersion. There was a decline (89.0%) in TFC between 2012 and 2014.There was a 74% lower risk of typhoid fever in 2014 (IRR=0.742, 95%CI: 0.647, 0.852) compared to the risk in 2011. The highest risk was recorded in Lagelu (IRR=4.072, 95% CI: 2.847, 5.823), while the lowest was in Ibadan North. (IRR = 0.599, 95%CI: 0.425, 0.846). The zero inflated negative binomial regression was the best model to estimate factors associated with typhoid fever cases in the presence of over-dispersion.