The aim of this study is to develop a model which computes exit dose values from transmission dose data obtained during patient treatment with an electronic portal imaging device (EPID). The proposed model convolves the primary dose distribution, derived from transmission dose distributions at large air gaps, with a scatter kernel to obtain the exit dose. The influence of inhomogeneities on the scatter contribution is taken into account by using a radiological path length model. To determine the parameters of the model, an extensive set of transmission dose measurements was performed behind various phantoms in an 8 MV beam using a liquid-filled EPID. The influence on the transmission dose of field size, phantom thickness, air gap between phantom and detector, and source-phantom distance was investigated. At air gaps larger than 50 cm the distribution of scattered dose is almost flat and its contribution to the total dose is relatively small, thus allowing an accurate separation of the primary and scattered dose by subtraction. Scattered dose distributions for air gaps smaller than 50 cm were obtained by subtracting the primary dose (corrected for divergence) from the measured total transmission dose. The resulting scattered dose distribution behind homogeneous phantoms has a Gaussian shaped profile, which becomes wider with increasing air gap. The relative contribution of scattered dose depends on the phantom thickness and is maximal for a thickness of about 10 cm. Using these results, the parameters of the convolution model (i.e., the shape of the scatter kernel) were determined. With the model the absolute exit dose is predicted with an accuracy of about 2% (1 s.d.) within the entire radiation field for homogeneous phantoms. Inhomogeneities are taken into account by calculating the radiological path length from the measured primary dose, i.e., without using CT data. By using the measured radiological path length the exit dose can be determined for inhomogeneous phantoms with an accuracy of 2.5%. It is concluded that, using our convolution model, EPID measurements at large air gaps can be used to estimate absolute exit doses in an 8 MV beam with an accuracy of 2.5%.