A physically-based computer model is developed for estimating soil temperatures at 0.05, 0.1, 0.2, and 0.5 m depths in sandy, forested soils, using daily minimum and maximum air temperatures and precipitation as input data. The algorithm is evaluated by comparing its output to 490 soil temperature observations for 1990-93 at stations in southern Michigan, USA. The mean errors of the soil temperature estimates are less than 0.1 C at the 0.5 m depth, thus comparing favorably to models requiring more detailed input. The model is applied to 14 stations across lower Michigan, using 1951-91 climatic data, in order to examine regional trends in soil temperature and freezing and to compare these trends to patterns of snow thickness and air temperature. Simulations reveal that soils seldom freeze within the deep snow area (lake-effect snow belt) of southern Michigan and along the coast of Lake Michigan. Soil freezing is more common at non-snowbelt, interior locations.