This paper sets out to explain modeling and correcting sample bias in Call Detail Records (CDRs). A proper understanding of sample bias is key to producing useful estimates derived from CDRs: such calculations rely heavily on a good understanding of how the sample (cell-phone users) relates to the larger populations it is drawn from. It could have major applications in crisis monitoring and response, as in the case of flood vulnerability predictions. Data-Pop Alliance uses both statistical and machine learning approaches, relying on data from Orange’s D4D challenges, official censuses and Demographic and Health Survey (DHS) program data.