© 2013 John Wiley & Sons, Ltd. Frequently, models are required to combine information obtained from different data sources and on different scales. In this work, we are interested in estimating the risk of wildfire ignition in the USA for a particular time and location by merging two levels of data, namely, individual points and aggregate count of points into areas. The data for federal lands consist of the point location and time of each fire. Nonfederal fires are aggregated by county for a particular year. The probability model is based on the wildfire point process. Assuming a smooth intensity function, a locally weighted likelihood fit is used, which incorporates the group effect. A logit model is used under the assumption of the existence of a latent process, and fuel conditions are included as a covariate. The model assessment is based on a residual analysis, while the False Discovery Rate detects spatial patterns. A benefit of the proposed model is that there is no need of arbitrary aggregation of individual fires into counts. A map of predicted probability of ignition for the Midwest US in 1990 is included. The predicted ignition probabilities and the estimated total number of expected fires are required for the allocation of resources.