Descriptions of the Model Specification Options describes the options in the red triangle menu next to Model Specification.
The Informative Missing option constructs a coding system that allows estimation of a predictive model despite the presence of missing values. It codes both continuous and categorical model effects.
When a continuous main effect has missing values, a new design matrix column is created. This column is an indicator variable, with values of one if the main effect column is missing and zero if it is not missing. In addition, missing values for the continuous main effect are replaced with the mean of the non-missing values for rows included in the analysis. The mean is a neutral value that maintains the interpretability of parameter estimates.
The parameter associated with the indicator variable estimates the difference between the response predicted by the missing value grouping and the predicted response if the covariate is set at its mean.
For a higher-order effect, missing values in the covariates are replaced by the covariate means. This makes the higher-order effect zero for rows with missing values, assuming that Center Polynomials is checked (the default setting). This is because Center Polynomials centers the individual terms involved in a polynomial by their means.
In the Effect Tests report, each continuous main effect with missing values will have Nparm = 2. In the Parameter Estimates report, the parameter for a continuous main effect with missing values is labeled <colname> Or Mean if Missing and the indicator parameter is labeled <colname> Is Missing. Prediction formulas that you save to the data table are given in terms of expressions corresponding to these model parameters.
When a nominal or ordinal main effect has missing values, the missing values are coded as a separate level of that effect. As such, in the Effect Tests report, each categorical main effect with missing values will have one additional parameter.
In the Parameter Estimates report, the parameter for a nominal effect is labeled <colname>[ ]. For an ordinal effect, the parameter is labeled <colname>[-x], where x denotes the level with highest value ordering.
As with continuous effects, prediction formulas that you save to the data table are given in terms of expressions corresponding to the model parameters.