Extreme precipitation events are a major cause of economic damage and disruption, and need to be addressed for increasing resilience to a changing climate, particularly at the local scale. Practitioners typically want to understand local changes at spatial scales much smaller than the native resolution of most Global Climate Models, for which down scaling techniques are used to translate planetary-to-regional scale change information to local scales. However, users of statistically downscaled outputs should be aware that how the observational data used to train the statistical models is constructed determines key properties of the downscaled solutions. Specifically for one such downscaling approach, when considering seasonal return values of extreme daily precipitation, we find that the Localized Constructed Analogs (LOCA) method produces a significant low bias in return values due to choices made in building the observational data set used to train LOCA. The LOCA low biases in daily extremes are consistent across event extremity, but do not degrade the over all performance of LOCA-derived changes in extreme daily precipitation. We show that the low bias in daily extremes is a function of a time-of-day adjustment applied to the training data and the manner of gridding daily precipitation data. The effects of these choices are likely to affect other downscaling methods trained with observations made in the same way. The results developed here show that efforts to improve resilience at the local level using extreme precipitation projections can benefit from using products specifically created to properly capture the statistics of extreme daily precipitation events.