Majallah-i Dānishgāh-i ̒Ulūm-i Pizishkī-i Bābul (Jun 2006)
Length of hospital stay (LOS) modeling with mixture Poisson distribution
Abstract
BACKGROUND & OBJECTIVE: Modeling is one of the most fundamental methods of denoting statistical variables which by using it, we can found distribution of noted response variable. For analysis data same as length of hospital stay (LOS), we have not data normality and error variances homogeneity, so, we can use nonparametric methods or distribution correction likes logarithmic transferring for using parametric methods. According to experiences in this situation, considering mixture distributions approximately can improve goodness- of- fit distribution. The goal of this study was to introduce mixture Poisson modeling and using mixture Poisson regression models for explaining duration of patient hospitalization in hospital and gaining effective factors on this time of duration and also comparing these models with common regression models in these data.METHODS: After interdicting mixture Poisson modeling and its regression, we applied these models for modeling LOS in two wards in Arak Vali-e-Asr hospital. Variables of age, marriage status, birth location and living location as independent variables and duration of hospitalization in hospital as countable response variable were considered and LOS was considered as response variable for application these models.FINDINGS: The findings have shown that in base Log-likelihood value and more dispersion LOS data in the surgical ward mixture Poisson model was a suitable for explain LOS with the other variables and in internal ward the variation of hospitalization time is not great, so this model cannot describe this variable explanation.CONCLUSION: By consideration Log-likelihood value and variation of LOS in surgical ward, the Poisson mixture model is a good model for describing this variable. By using general models, the Log-likelihood value is more than mixture Poisson modeling and there are less significant factors in models. Application of these models in cases which the countable response variable has great variation, is recommended.
