Mixed effect quantile and M-quantile regression for spatial data

Observed data are frequently characterized by a spatial dependence; that is the observed values can be influenced by the "geographical" position. In such a context it is possible to assume that the values observed in a given area are similar to those recorded in neighboring areas. Such data is frequently referred to as spatial data and they are frequently met in epidemiological, environmental and social studies, for a discussion see Haining, (1990). Spatial data can be multilevel, with samples being composed of lower level units (population, buildings) nested within higher level units (census tracts, municipalities, regions) in a geographical area. Green and Richardson (2002) proposed a general approach to modelling spatial data based on finite mixtures with spatial constraints, where the prior probabilities are modelled through a Markov Random Field (MRF) via a Potts representation (Kindermann and Snell, 1999, Strauss, 1977). This model was defined in a Bayesian context, assuming that the interaction parameter for the Potts model is fixed over the entire analyzed region. Geman and Geman (1984) have shown that this class process can be modelled by a Markov Random Field (MRF). As proved by the Hammersley-Clifford theorem, modelling the process through a MRF is equivalent to using a Gibbs distribution for the membership vector. In other words, the spatial dependence between component indicators is captured by a Gibbs distribution, using a representation similar to the Potts model discussed by Strauss (1977). In this work, a Gibbs distribution, with a component specific intercept and a constant interaction parameter, as in Green and Richardson (2002), is proposed to model effect of neighboring areas. This formulation allows to have a parameter specific to each component and a constant spatial dependence in the whole area, extending to quantile and m-quantile regression the proposed by Alfò et al. (2009) who suggested to have both intercept and interaction parameters depending on the mixture component, allowing for different prior probability and varying strength of spatial dependence. We propose, in the current dissertation to adopt this prior distribution to define a Finite mixture of quantile regression model (FMQRSP) and a Finite mixture of M-quantile regression model (FMMQSP), for spatial data.