Currently in its third phase, the North American Forest Dynamics (NAFD) project has launched nationwide processing of historic Landsat data to provide a comprehensive annual, wall-to-wall analysis of U.S. disturbance history over the last 30+ years. Because understanding the cause of disturbance is important to quantifying carbon dynamics, work is underway to attribute causal agents to these nationwide change maps. Developing empirical models of the diverse causal agents in this country involves many decisions. Alternative response designs (such as varying size, shape, quantity, and level of detail in training data) are being evaluated in terms of their costs and benefits for national mapping applications. Many classes of predictor variables such as spectral signatures, textural metrics, extant geospatial disturbance libraries, and bioclimatic information, are being tested for their contribution to classification models. Flexible modeling techniques, such as the Random Forests models used here, are powerful predictive tools but must be coupled with simple rule-based models reflecting expert knowledge. And decisions about appropriate modeling subpopulations are being made in light of available training data, diversity of ecological zones, and computational efficiency. We will be synthesizing results from our initial exploratory work as well as from pilot analyses conducted over 10 Landsat TM scenes representing diverse causal agents, forest types, and forest prevalence levels. We also discuss how these causal disturbance models will enable extensive analyses of temporal and spatial patterns in causal agents across the United States.