Unemployment duration data derived from retrospective surveys often show an abnormal concentration of responses at certain durations. This common kind of measurement error is known as "heaping" in the statistical literature. Although heaping effects may lead to severe biases in estimated coefficients of duration models, in applied work researchers have either neglected them altogether or tried to account for them in an ad hoc way. This is also the case for recent microeconometric research based on unemployment duration data derived from the retrospective calendar information in the German Socio-Economic Panel, where a very high proportion of all unemployment spells beginning in January or end in December of each year. We show how this kind of heaping can be modelled within a maximum likelihood framework using external validation information and demonstrate for this particular data set how parameter estimates in discrete-time proportional hazard models of unemployment duration are affected by alternative specifications of the heaping mechanism. Durmain result is that parameter estimates are generally rather insensitive to whether or not heaping is explicitly taken into account and to different assumptions about the heaping mechanism, but may be substantially affected by ad hoc procedures to contral for heaping which tend to pick up selectivity effects and censoring.