The assumption that the sub-set of persons for whom data are not measured (i.e., incomplete or missing), data can be indicated by scores on variables that are measured (i.e., complete). The MAR assumption is different than the traditional assumption of ‘missing completely at random’ (MCAR), and it permits a correction for the bias due to selection using maximum-likelihood estimation (MLE). For complex data sets, there are software available for dealing with missing data in a principled way.
See Multilevel modeling