The variation of dose and image quality parameters in the voltage range 25...33 kV have been studied on 3 mammographic phantoms simulating thin, average and thick breasts respectively. The global image quality has been objectively evaluated by an image quality index. The average glandular dose has been chosen as relevant indicator of risk. The study shows that the surface dose and the midplane dose are not suitable indicators of risk if the breast thickness is variable. For all phantoms, the deterioration of image quality with voltage increase was found to be relatively less important than the decrease of the average glandular dose. Consequently, the following optimum voltages are proposed: 28-29 kV for thin breasts, 29-30 kV for average breasts and 33-35 kV for thick ones. The study showed that breast thickness variation is not suitably compensated by the automatic exposure devices currently available.