It is well-known that the selection criteria make it possible to determine the order of a statistical model associated with the observed data. But in practice, the problem of missing values requires a modification of these criteria. For Akaike Information criterion, this problem of incomplete data was studied by Cavanaugh and Shumway (1998), they demonstrated an extension of Akaike’s criterion to take account of missing values. But this criterion does not always lead to correct model selection. In this paper, we propose a new information criterion of Schwarz. This criterion is based on the motivation provided for the posterior probability of the candidate model and the EM algorithm. We have validated the theoretical results on simulated data. The new criterion converges to the correct order of the candidate model for both small and large samples, even if the percentage of missing data increases.