dc.description.abstract | This study reviews theory on logistic and extends to auto-logistic regression, which is
appropriate for binary data. This auto-logistic regression allows for local spatial autocorrelation
by the inclusion in a logistic model of a covariate deprived from neighbouring values
of response variable.
The statistical methodology for taking account of spatial autocorrelation is not as developed
as for these type of models as it is for least squares regression analysis. After factoring
in the auto-covariate factor, the model was compared with logistic regression model. Models
having smaller Akaike's Information Criterion (AIC) was considered the better model
in the sense of model fitting balanced with model parsimony. The results indicated that
auto-logistic model outperforms logistic model.
Auto-logistic model was then fitted to Somalia settlement census data to identify the key
possible covariates influencing poverty in Bari region. Parameter estimates were carried out
using maximum pseudo-likelihood method which is an alternative of maximum likelihood.
Finally, Receiver Operating Characteristics (ROC) curve was used to summarize the predictive
accuracy of the model in terms of true positive/false positive pairs given different cut-off
values for model prediction of spread events.
Key words: Logistic model, Newton-Raphson method, spatial autocorrelation, auto-logistic model and
maximum pseudo-likelihood. | en |