Although forest conservation activities particularly in the tropics offer significant potential for mitigating carbon emissions, these types of activities have faced obstacles in the policy arena caused by the difficulty in determining key elements of the project cycle, particularly the baseline. A baseline for forest conservation has two main components: the projected land-use change and the corresponding carbon stocks in the applicable pools such as vegetation, detritus, products and soil, with land-use change being the most difficult to address analytically. In this paper we focus on developing and comparing three models, ranging from relatively simple extrapolations of past trends in land use based on simple drivers such as population growth to more complex extrapolations of past trends using spatially explicit models of land-use change driven by biophysical and socioeconomic factors. The three models of the latter category used in the analysis at regional scale are The Forest Area Change (FAC) model, the Land Use and Carbon Sequestration (LUCS) model, and the Geographical Modeling (GEOMOD) model. The models were used to project deforestation in six tropical regions that featured different ecological and socioeconomic conditions, population dynamics, and uses of the land: (1) northern Belize; (2) Santa Cruz State, Bolivia; (3) Parana State in Brazil; (4) Campeche, Mexico; (5) Chiapas, Mexico; and (6) Michoacan, Mexico. A comparison of all model outputs across all six regions shows that each model produced quite different deforestation baseline. In general, the simplest FAC model, applied at the national administrative-unit scale, projected the highest amount of forest loss (four out of six) and the LUCS model the least amount of loss (four out of five). Based on simulations of GEOMOD, we found that readily observable physical and biological factors as well as distance to areas of past disturbance were each about twice as important as either sociological/demographic or economic/infrastructure factors (less observable) in explaining empirical land-use patterns. We propose from the lessons learned, a methodology comprised of three main steps and six tasks can be used to begin developing credible baselines. We also propose that the baselines be projected over a 10-year period because, although projections beyond 10 years are feasible, they are likely to be unrealistic for policy purposes. In the first step, an historic land-use change and deforestation estimate is made by determining the analytic domain (size of the region relative to the size of proposed project), obtaining historic data, analyzing candidate historic baseline drivers, and identifying three to four major drivers. In the second step, a baseline of where deforestation is likely to occur --a potential land-use change (PLUC) map is produced using a spatial model such as GEOMOD that uses the key drivers from step one. Then rates of deforestation are projected over a 10-year baseline period using any of the three models. Using the PLUC maps, projected rates of deforestation, and carbon stock estimates, baselineprojections are developed that can be used for project GHG accounting and crediting purposes: The final step proposes that, at agreed interval (eg, +10 years), the baseline assumptions about baseline drivers be re-assessed. This step reviews the viability of the 10-year baseline in light of changes in one or more key baseline drivers (e.g., new roads, new communities, new protected area, etc.). The potential land-use change map and estimates of rates of deforestation could be redone at the agreed interval, allowing the rates and changes in spatial drivers to be incorporated into a defense of the existing baseline, or derivation of a new baseline projection.