Method (Multiple Imputation)
The Method tab specifies how missing values will be imputed, including the types of models used. Categorical predictors are indicator (dummy) coded.
Imputation Method. The Automatic method scans the data and uses the monotone method if the data show a monotone pattern of missing values; otherwise, fully conditional specification is used. If you are certain of which method you want to use, you can specify it as a Custom method.
- Fully conditional specification. This is an iterative
Markov chain Monte Carlo (MCMC) method that can be used when the pattern
of missing data is arbitrary (monotone or nonmonotone).
For each iteration and for each variable in the order specified in the variable list, the fully conditional specification (FCS) method fits a univariate (single dependent variable) model using all other available variables in the model as predictors, then imputes missing values for the variable being fit. The method continues until the maximum number of iterations is reached, and the imputed values at the maximum iteration are saved to the imputed dataset.
Maximum iterations. This specifies the number of iterations, or "steps", taken by the Markov chain used by the FCS method. If the FCS method was chosen automatically, it uses the default number of 10 iterations. When you explicitly choose FCS, you can specify a custom number of iterations. You may need to increase the number of iterations if the Markov chain hasn't converged. On the Output tab, you can save FCS iteration history data and plot it to assess convergence.
- Monotone. This is a noniterative method that can be used
only when the data have a monotone pattern of missing values. A monotone
pattern exists when you can order the variables such that, if a variable
has a nonmissing value, all preceding variables also have nonmissing
values. When specifying this as a Custom method,
be sure to specify the variables in the list in an order that shows
a monotone pattern.
For each variable in the monotone order, the monotone method fits a univariate (single dependent variable) model using all preceding variables in the model as predictors, then imputes missing values for the variable being fit. These imputed values are saved to the imputed dataset.
Include two-way interactions. When the imputation method is chosen automatically, the imputation model for each variable includes a constant term and main effects for predictor variables. When choosing a specific method, you can optionally include all possible two-way interactions among categorical predictor variables.
Model type for scale variables. When the imputation method is chosen automatically, linear regression is used as the univariate model for scale variables. When choosing a specific method, you can alternatively choose predictive mean matching (PMM) as the model for scale variables. PMM is a variant of linear regression that matches imputed values computed by the regression model to the closest observed value.
Logistic regression is always used as the univariate model for categorical variables. Regardless of the model type, categorical predictors are handled using indicator (dummy) coding.
Singularity tolerance. Singular (or non-invertible) matrices have linearly dependent columns, which can cause serious problems for the estimation algorithm. Even near-singular matrices can lead to poor results, so the procedure will treat a matrix whose determinant is less than the tolerance as singular. Specify a positive value.
How To Specify an Imputation Method
This feature requires the Missing Values option.
- From the
menus choose:
- In the Impute Missing Data Values dialog box, click the Method tab.