EM Imputation in SPSS: Procedure, Validation, and Recovery Paths

You have a dataset with missing continuous values and need to impute them using a method that preserves distribution shape and predictor relationships. This guide walks through SPSS Expectation-Maximization imputation, from prerequisites to convergence checks and recovery paths when assumptions fall short.

\n

Desired Outcome: Preserve Distribution Shape While Filling Gaps

\n

EM imputation is available in SPSS Statistics 21 and later through the GUI and syntax. Your dataset should have at least three complete cases after listwise deletion, and continuous variables should not exhibit systematic missingness that violates the Missing At Random (MAR) assumption. If missingness depends on the unobserved value itself (Missing Not At Random), EM may produce biased results without substantive justification.

\n

Procedure: GUI and Syntax Paths

\n

Using the GUI: Choose Analyze > Missing Value Analysis > Impute. In the dialog, select the continuous variables you want to impute, ensure Expectation-Maximization is selected as the method, and click OK. SPSS will generate output that includes an EM Imputation table and a new dataset with imputed values.

\n

Using syntax (recommended for reproducibility): Open the Syntax Editor and run

\n

MISSING VALUES v1 TO v5 (EM).

\n

Replace v1 TO v5 with your variable names. This command can be run in the Syntax Editor with File > Open > Syntax or by pasting into the command line. Required permission: edit access to the active dataset.

\n

Expected Checks: Verifying the Imputation

\n

After running EM, open the SPSS Output window and locate the EM Imputation table. Confirm that Converged appears in the status row and note the Number of iterations field; non-convergence after the default iteration limit suggests the model may not stabilize for your data pattern.

\n

Compare model coefficients (e.g., regression weights) using the original complete-case dataset and the imputed dataset. Substantial shifts may signal that the MAR assumption is tenuous. Additionally, inspect histograms of the imputed values against the distribution of original non-missing values to ensure the imputed range sits plausibly within the observed spread.

\n

Recovery Options: When EM Falls Short

\n

\n
• Switch to listwise or pairwise deletion for a sensitivity analysis.
\n
• Apply a logarithmic or other transformation to reduce skew before re-running EM.
\n
• Use the MIANALYSIS extension for multiple imputation, which generates several complete datasets and combines estimates to reflect imputation uncertainty.
\n

\n

Each option trades off bias against variance, so document your choice in the analysis log.

\n

Limitations and the MAR Boundary

\n

EM imputation assumes data are Missing At Random: the probability of a value being missing may depend on observed variables but not on the unobserved value itself. Applying EM to Missing Not At Random patterns without substantive justification can yield biased estimates that look deceptively precise. SPSS version also matters: output format and convergence diagnostics evolved in Statistics 27+, so verify that your session's tables match the version-specific expectations.