In their examinations of the causes of violence, researchers are often faced with applying spatial econometric methods to models with discrete outcomes. There are no appropriate methods for doing this when the outcomes are measured at intra-city areal units. This research aimed to fill this gap. The framework developed was applied to a real-world empirical problem. It examined the socioeconomic and demographic determinants of disaggregate homicide rates at two intra-city levels of areal aggregation and compared inferences derived from several sets of models. The analysis was performed on disaggregated homicide counts (1989-91) recorded in Chicago's census tracts and neighborhood clusters through the use of explanatory factors obtained from census sources. An extension of the Generalized Cross Entropy (GCE) methods was applied to the data in an effort to use their flexibility in allowing error structures across space. An information-based measure was developed and used in selecting the hypothesized error structure that best approximated the true underlying structure. The findings confirmed that ignoring spatial structures in the regression residuals often leads to severely biased inferences and, hence, a poor foundation on which to base policy. Also, evidence was found of homicide type-specific and areal units-specific models, thus exposing the need to disaggregate violence into distinct types. Resource deprivation was a reliable predictor of all types of violence analyzed and at both levels of areal aggregation. In addition, there was apparently a spill-over effect of resource deprivation on the amount of violence expected in neighboring areas. This indicates the importance of taking into account the spatial structure in a sample when planning and implementing policy measures. The GCE approach used in this project provides several paths for future research, particularly in the analysis of rare crimes.