In attempting to contact households for a survey, it is necessary to determine the timing ofeach call. Often, average “best” times to call are used in order to determine when to place the first call(s). The timing of subsequent calls is then governed by very general rules. This paper tests an experimental method that uses multi-level models to predict the times that have the highest probability of contact for each household and uses the predictions from these models to prioritize cases for calling. The predictions are updated each day in real-time as additional data are gathered. The method is evaluated through a series of experiments on a telephone survey that used automated call scheduling and an experiment on a face-to-face survey, where a recommended calling time was delivered to interviewers. The results of these experiments are used to suggest directions for future research.