The problem of estimating the effect of hormone replacement therapy (HRT) on breast cancer has attracted much attention in the causal inference literature. It is known that the HRT effect on diagnosis (observed) is potentially larger than the effect on true cancer status (of interest), and that the standard approach (stratification) to deal with this detection bias is not valid when screening pattern and cancer status share common causes. We propose a sensitivity analysis to deal with detection bias, and demonstrate that the earlier proposed approach of data restriction is not valid.