SPSS EM Imputation: Keep Cases Without Pretending Certainty

Your survey has 1,200 respondents, income is missing for 28% of cases, age for 12%, and education is complete. Dropping incomplete rows loses power and can skew the sample. Expectation-Maximization imputation in SPSS lets you retain those cases under a Missing at Random assumption, but only if you treat the output as a model-based estimate, not as real data.

The useful takeaway: EM is a single-imputation method that maximizes likelihood iteratively. It is appropriate for exploratory work and for generating a complete dataset for methods that cannot handle missingness, but inference requires multiple imputations and pooling with MIANALYZE to reflect imputation uncertainty.

What EM does in SPSS and when it fits

EM is an iterative algorithm. The E-step computes the expected value of missing data given observed data and current parameter estimates. The M-step re-estimates parameters as if the expected values were observed. Repeating E and M steps increases the likelihood until change falls below a tolerance.

SPSS implements EM in the Missing Value Analysis procedure. It handles numeric and categorical variables together under a multivariate model. The PROXIMITATE subcommand adds predictive mean matching, which draws imputed values from observed donors with close predicted means. That reduces distortion when normality is questionable.

EM assumes data are Missing at Random, meaning missingness can be predicted from observed variables but not from unobserved values after conditioning on observed data. If data are Missing Not at Random, for example income missing because high earners refuse to answer, EM will be biased and sensitivity analyses are needed.

Syntax, settings and practical checks

Run syntax in the SPSS Syntax Editor with a licensed SPSS Statistics installation and write access for any saved output files. Work on a copy of the source data.

MVA VARIABLES=Income Age Education
  /EM
    TOLERANCE=0.001
    ITERATIONS=50
    CONVERGENCE=0.001
    PROXIMITATE
  /SAVE OUTFILE='C:\temp\data_em_imputed.sav'

Meaningful placeholders: replace variable names with your numeric and categorical set, and set OUTFILE to a path you control. Default settings are 50 iterations and 0.001 convergence tolerance. You can raise ITERATIONS if convergence is slow, but more iterations increase runtime and do not fix model misspecification.

Expected checks after running:

• Convergence status in the output. A status code of 0 indicates termination within the iteration limit. Non-zero flags suggest non-convergence.
• Iteration history showing log-likelihood change decreasing toward tolerance.
• Imputed value statistics for means and standard deviations compared to observed distributions.

Risks: EM can fail to converge with extensive missingness or small samples. Check patterns with MVA descriptives first. Imputed datasets saved by EM are single imputations. Analyzing them directly understates standard errors because imputation uncertainty is ignored.

Verification and proper inference

After EM, do not run a single analysis on the imputed file and report p-values as if data were fully observed.

For proper inference, generate multiple imputations with the MULTIPLE IMPUTATION command, then analyze each imputed dataset with the same procedure, and pool results with MIANALYZE. Compare pooled standard errors with complete-case standard errors. Meaningful inflation confirms that missingness was impacting precision.

A practical diagnostic is to compare complete-case estimates with EM-based estimates on the same model. Large shifts suggest the missingness mechanism matters and warrants investigation of MAR assumptions.

Trade-off and limitation

EM retains sample size and avoids listwise deletion bias under MAR, which is valuable for descriptive summaries and for feeding models that require complete data. The cost is a single point estimate of missing values and reliance on multivariate normality for numeric variables. Predictive mean matching mitigates some distortion but does not remove model dependence.

Convergence can be unstable with high proportions of missing data per case. Running EM with varied starting seeds and inspecting stability of final parameter estimates helps assess robustness.

Actionable closing: use EM in SPSS to understand the impact of missingness and to create a complete dataset for preliminary exploration, always check convergence status and MAR plausibility, and move to multiple imputation with MIANALYZE pooling for any final inference.