The different daylength calculation procedures that can be employed to estimate solar radiation using an Angstrom-Prescott regression have not been sufficiently evaluated. In this study, daily global solar radiation data measured in Toledo (Spain) during the period 1986-1995 were used to test five daylength estimation models using different definitions of sunrise/sunset and twilight. Models were compared using the root mean square error (RMSE), the mean bias error (MBE) and the t-statistic. In the first two cases, the differences found between the results from the different models were small. Analysis using the t-statistic, on the other hand, showed that the use of civil twilight to calculate daylength produced the best estimates of global solar radiation for Toledo. (C) 2004 Elsevier Ltd. All rights reserved.