This paper describes the development and an assessment of a spatio-temporal model for people-caused forest fires in a portion of boreal forest in northeastern Ontario, a central province in Canada. Space and time along with location-specific weather-based fire danger rating indices and anthropogenic effects are included in the modelling we present, which parallels the structure of recent methodology for assessing fire risk using logistic generalized additive models (GAMs) introduced in Brillinger et al. (Institute of Mathematical Statistics Lecture Notes, 2003). In these models, the data consist of observations on a very fine set of space-time cells, where fires are rare and the complete data set is too large to analyze. Consequently, the non-fire observations are sampled. This induces an offset in the additive structure, which we connect to the analysis of case-control studies. The model's fit and estimated partial effects are shown to be sensitive to large reductions in this inclusion probability. We also make comparisons between a model with an additive decomposition of spatial and temporal effects to one with a spatio-temporal interaction, and we investigate the impact of restricting fire-weather and anthropogenic effects to be linear. Our results suggest that, when using logistic GAMs to model our wildland fire occurrence data on this scale, there is no advantage to including space and time interaction effects, and that models with linear terms, which have dominated the fire risk literature, are inadequate.