SPSS Missing Data: Listwise vs Pairwise vs EM Imputation

A decision guide for SPSS missing data handling: when listwise deletion is fine, when pairwise deletion helps, when EM imputation is worth it, and how to validate the result with reproducible syntax.

The decision you actually face

Your dataset has gaps, and every SPSS procedure you run will quietly make a choice about them for you. By default, most SPSS procedures use listwise deletion: any case with a missing value on any variable in the analysis is dropped entirely. That default is rarely the best option, and it is almost never the one you chose deliberately. This guide compares the three built-in options — listwise deletion, pairwise deletion, and Expectation-Maximization (EM) imputation — against the two constraints that matter most: how much data you have, and how the values are distributed.

Comparing the supported options

Method	Where it applies in SPSS	Best when	Main risk
Listwise deletion	Default in most procedures (REGRESSION, GLM, FACTOR)	Missingness is small (<5%) and plausibly random	Biased estimates and lost power if data are not missing completely at random (MCAR)
Pairwise deletion	CORRELATIONS and procedures that accept a correlation matrix	You need a correlation/covariance matrix and want to keep every available pair	Each correlation uses a different N; matrices can be non-positive-definite, breaking downstream analyses
EM imputation	Missing Value Analysis (MVA) procedure	Moderate missingness, roughly multivariate normal data, and you need a single complete dataset	Slow convergence with high missingness; understates standard errors if treated as real observed data

Trade-offs in practice

Sample size. With a large sample and a few percent missing, listwise deletion costs little and keeps your analysis simple. With a small sample, dropping 15% of cases because one variable out of twelve is missing can destroy your power. That is when pairwise deletion or imputation earns its complexity.

Distribution. EM imputation assumes the data are approximately multivariate normal. It iterates between estimating parameters from current data and filling in expected values, repeating until the estimates stop changing (convergence). With skewed variables or heavy missingness, convergence can be slow and the imputed values can be pulled toward an unrealistic joint distribution. Check skewness before trusting it.

Missingness mechanism. All three methods assume data are at least missing at random (MAR) — the probability of a value being missing depends only on observed data. If missingness depends on the missing value itself (for example, high earners skipping the income question), none of these methods fully fixes the bias. Diagnose the pattern first with Missing Value Analysis rather than picking a method blindly.

Concrete implementation with reproducible syntax

Menu-driven work in SPSS is hard to reproduce. Use syntax, and keep it in a separate file you load with the INSERT command so the handling logic is versioned and repeatable:

* File: missing_handling.sps — run from SPSS Syntax Editor.
* Requires: SPSS Statistics Base + Missing Values option for MVA.

* Step 1: Quantify missingness before deciding.
FREQUENCIES VARIABLES=age income score
  /FORMAT=NOTABLE
  /STATISTICS=MEAN STDDEV SKEWNESS.

* Step 2: EM imputation via Missing Value Analysis.
MVA VARIABLES=age income score
  /EM(TOLERANCE=0.001 CONVERGENCE=0.0001 ITERATIONS=50)
     OUTFILE='C:\analysis\em_imputed.sav'.

Run this in the SPSS Syntax Editor (File → Open → Syntax, then Run → All). No special OS permissions are needed beyond write access to the output path; you do need the Missing Values add-on licensed for MVA. Replace the variable names and the output path with your own. TOLERANCE and CONVERGENCE control when the EM loop stops; if the output warns that convergence was not reached within 50 iterations, raise ITERATIONS and re-run.

To keep this reproducible across projects, call the file from a master syntax file:

INSERT FILE='C:\analysis\missing_handling.sps'.

One caution: a typo in the path or a syntax error inside the inserted file can leave you running analyses on unprocessed data without an obvious failure. Always check the SPSS Output Viewer for error notes after an INSERT, and confirm the output file exists and has the expected case count before proceeding.

Validating the choice

Do not trust the imputation until you have compared outputs. A practical check:

Run your target analysis (for example, a regression) three ways: listwise on raw data, pairwise correlations feeding the model, and the EM-imputed dataset.
Compare coefficients and sample sizes across the three. Large swings between listwise and EM results are a red flag that data are not MCAR and listwise deletion is biasing you.
Inspect the imputed values: in the EM output dataset, compare means and standard deviations of imputed variables against the observed cases. Imputed variables that are dramatically less variable than observed ones indicate the model is shrinking values toward the mean.

Limitations

SPSS's built-in EM produces a single imputed dataset, so standard errors in later analyses treat imputed values as if they were observed — they will be optimistically small. If inference matters, multiple imputation (available in SPSS via the Multiple Imputation procedure in recent versions) is the stronger choice. Version note: the MVA and INSERT syntax shown applies to IBM SPSS Statistics 25 and later; menu locations differ slightly across versions. Verify convergence messages and case counts in your own environment before relying on the output.

Choosing a Missing Data Method in SPSS: Listwise vs Pairwise vs EM Imputation