Variable Lists (MULTIPLE IMPUTATION command)

Analysis variables are specified after the command name. Two or more variables must be specified. The TO and ALL keywords can be used to refer to multiple variables. If any variable is specified more than once, the last instance of that variable is honored.

The variable list specifies variables to impute and to include in analyses of missingness. By default, analyis variables are also used as predictors to in imputation models of other analysis variables. The lists of variables to impute and predictors can be restricted via the CONSTRAINTS subcommand.

Variable Order

When imputing using the FCS and MONOTONE methods, variables are imputed sequentially in the order in which they are listed in the variables list. The order of the variables is ignored by the AUTO method except to break ties.

Predictors

The set of predictors that is used when a particular variable is imputed depends on the imputation method:

  • For the FCS method, when a variable is imputed all other analysis variables are used as predictors in the imputation model.
  • When the MONOTONE method is used, only variables that precede the variable to be imputed in the variable list are used as predictors.
  • When the AUTO method is used the set of predictors depends on the pattern of missingness in the data. If the data have a nonmonotone pattern of missingness, FCS is used and all other analysis variables are used as predictors. If the data have a monotone pattern, MONOTONE is used and all variables that precede the variable to be imputed are used predictors. Note that AUTO sorts the analysis variables to detect monotone pattern, so the actual order of variables may not correspond to their order in which they are specified in the variables list.

Measurement Level

Measurement level recorded in the data dictionary is honored for each analysis variable. Measurement level determines the following:

  • The default type of imputation model for variables whose values are imputed (linear regression or logistic regression).
  • Whether a variable is treated as a factor (categorical) or covariate (scale) when used as a predictor in imputation models.
  • Whether the variable is treated as scale or categorical in summaries of missing values.

The procedure treats ordinal and nominal variables equivalently as categorical variables.