The focus of this paper is the size of the wage penalty due to maternal leave incurred by working mothers in Germany. Existing estimates suggest large penalties with little rebound over time. We apply recent panel data methods designed to address problems of sample selectivity, unobserved heterogeneity and endogeneity to German Socio-Economic Panel (GSOEP) data. Heckman's classic treatment of selectivity is extended to deal with both heterogeneity and simultaneity. In order to exploit the actual working hours data available in GSOEP, we develop the case of a censored tobit participation model. We also investigate the sensitivity of the results to the choice of method. Our estimates imply a maternal leave wage penalty of 10 to 14% which is less persistent over time than other studies suggest. Five years after the return to work maternal wages have caught up.