Bayesian multinomial ordered categorical response model for the analysis of length of hospital stay

Length of hospital stay (LOS) is of primary importance in health services research because it is directly related to health care management and cost of health care. In some epidemiological settings the actual length of stay is not directly observed but it is known to have happened in a particular interval or for simple epidemiological interpretation time is categorized into ordered categorical responses. In this paper, we focus our attention on cumulative regression models for ordinal responses to analyze length of hospital stay for children admitted to a paediatric ward for malaria. Such models exploit the ordered scale of the outcomes. We approach our analysis using a Bayesian probit model. Our model incorporated random effects for hospital specific heterogeneity, while simultaneously investigating nonlinear effects in covariates within the general framework of semi-parametric regression models. Findings indicate children who died had relatively shorter LOS, which suggest worse prognosis at admission. Calendar time effects indicated changing seasonal effects with high peaks in wet season and low peak in dry season, largely explained by malaria transmission patterns. Age showed deviation from linearity, and early discharge was associated with much older children than infants.