Aspects of the invention relate to yield analysis techniques for generating root cause distribution information. Suspect information for a plurality of failing dies is first generated using a layout-aware diagnosis method. Based on the suspect information, potential root causes for the plurality of failing dies, and suspect feature weights and total feature weights for each of the potential root causes may then be determined. Next, the probability information of observing a particular suspect that is related to a particular root cause may be extracted. Finally, an expectation-maximization analysis may be conducted for generating the root cause distribution information based on the probability information and the suspect information. Heuristic information may be used to prevent the analysis from over-fitting.