Modelos de regressão ajustados a dados espaciais de áreas com sementes melhoradas de milho em Moçambique

Abstract

A regionalized variable represented by area data is one in which the observations have a geographic reference and come from regions such as villages, localities, municipalities, districts, provinces or some delimited area in space. For each of these regions, these data are generally presented in the form of averages,rates, proportions, among others. These types of spatial data are often referred to simply as area data. In studies with this type of data, if the interest is to adjust regression models or another type of model, the existence of spatial dependence between the observations must be taken into account. In this case, classical linear regression (OLS) models may not be appropriate. In such cases, the option has been to use models indicated for area data, such as spatial lag autoregressive (SAR) or correlated spatial error (SEM) models. In this article, the objective was to evaluate, in a practical way, the quality of fit of these three models: SAR, SEM and OLS. In addition, the effect of the spatial weighting matrix W on the goodness of fit, an essential component in the first two models, was evaluated. As for the data, they come from an agricultural survey, referring to the use of improved maize seeds in Mozambique. The contribution of some covariates of interest to farmers using such seeds was also evaluated. The main result is that the SAR model was the one that best fitted the data, followed by SEM, and finally OLS. In addition, it was observed that the specification of the W matrix can influence the quality of the model's fit.