How to Handle Missing Data in SPSS: Multiple Imputation vs Listwise Deletion

Introduction: Why Missing Data Deserves a Plan, Not a Panic

Almost every real dataset has gaps. Respondents skip a question, a sensor fails, a participant drops out before the final survey wave, or a coding error leaves a column half empty. In SPSS, what you do with those gaps can change your sample size, your standard errors, and sometimes your conclusions — so it deserves a deliberate decision, not a default click.

This tutorial walks you through the two approaches most students and researchers meet first: listwise deletion (analyse only complete cases) and multiple imputation (fill in plausible values several times, analyse each filled dataset, then pool the results). You will learn when each one is defensible, how to run both in SPSS click by click, how to read the output, and how to report what you did in APA style. No advanced statistics background is assumed.

First, Understand Why Data Are Missing

Before choosing a method, ask why values are missing. Statisticians describe three patterns, and the pattern matters more than the percentage missing.

Missing completely at random (MCAR)

Data are MCAR when the chance of a value being missing has nothing to do with any variable, observed or unobserved. A page of a questionnaire lost in printing is a classic example. Under MCAR, complete cases are still a random subset of your sample, so listwise deletion does not bias your estimates — it mainly costs you statistical power.

Missing at random (MAR)

Data are MAR when missingness can be explained by variables you did observe. For example, older participants might be more likely to skip an income question, but once you account for age, the missingness itself carries no extra information. Multiple imputation works well under MAR, because it can use those observed variables to generate plausible replacement values.

Missing not at random (MNAR)

Data are MNAR when missingness depends on the missing value itself — for instance, people with very high incomes being less likely to report income. Neither listwise deletion nor standard multiple imputation fully fixes MNAR. If you suspect MNAR, say so openly in your write-up, consider a sensitivity analysis, and seek advice rather than pretending the problem away.

You cannot prove which pattern holds from the data alone, but you can gather evidence. A useful first step in SPSS is to look at how much data are missing, and in which variables, before you decide anything.

Step 1: Diagnose the Missing Data in SPSS

Do not jump straight to a fix. Spend five minutes describing the problem first.

Check the amount and pattern of missingness

Go to Analyze > Missing Value Analysis. Move the variables you plan to analyse into the Quantitative Variables or Categorical Variables box, depending on their measurement level. Under Patterns, ask for a tabulated pattern and a chart if you want a visual overview. Click OK.

The output shows, for each variable, how many values are missing and what percentage that represents, plus a table of missing-data patterns — which combinations of variables tend to be missing together. A variable missing for 2% of cases is a different problem from a variable missing for 40%, and a pattern in which the same participants miss a whole block of questions may point to dropout rather than random skipping.

Run Little’s MCAR test

In the same Missing Value Analysis dialog, under the EM tab, tick the option for Little’s MCAR test. A non-significant result is consistent with MCAR, while a significant result suggests the data are not MCAR — which pushes you toward multiple imputation rather than simply deleting incomplete cases. Treat this test as supporting evidence, not a verdict: with very large samples it can be significant even for trivial departures, and with small samples it may lack power to detect a real problem.

Decide on a threshold for action

There is no universal rule that a certain percentage forces imputation, but a practical working approach is this: if only a tiny fraction of cases have any missing value and the pattern looks random, listwise deletion may be acceptable and is easy to defend. If a meaningful share of cases would be dropped, or if missingness relates to observed variables, multiple imputation usually preserves more information and gives more honest standard errors. Whatever you choose, report the amount of missing data and your reasoning — reviewers notice silence on this point far more than they penalise a sensible choice.

Step 2: Option A — Listwise Deletion in SPSS

Listwise deletion, also called complete-case analysis, is SPSS’s default in most procedures: any case missing a value on any variable in that analysis is excluded from that analysis.

How to run it

You usually do not need a special menu. For example, to run a regression on complete cases only, go to Analyze > Regression > Linear, move your dependent and independent variables into their boxes, click Statistics if you want extra output, and click OK. SPSS will report the number of cases actually used, which will be smaller than your total sample if any cases have missing values.

If you want to be explicit, you can first create a complete-cases flag. Go to Transform > Compute Variable, name the new variable CompleteCase, and use a function that counts missing values across your key variables, setting the flag to 1 when the count is zero and 0 otherwise. You can then filter on that flag with Data > Select Cases if you need a visibly complete dataset for a particular table.

When listwise deletion is defensible

Listwise deletion is simplest to explain and keeps your analysis on real, observed data only. It is most defensible when the amount of missing data is small, Little’s test does not reject MCAR, and dropping incomplete cases still leaves you with enough power for your analysis. Its main cost is lost cases: if each of ten variables is missing for a different 5% of participants, listwise deletion can easily discard a much larger share of your sample than any single percentage suggests.

Its limits

When data are MAR rather than MCAR, complete cases are no longer representative, and listwise deletion can bias estimates. It also wastes information from participants who answered most questions. If your complete-case sample drops sharply, or if the people you lose differ systematically from those you keep, do not settle for the default — move to multiple imputation.

Step 3: Option B — Multiple Imputation in SPSS

Multiple imputation replaces each missing value with several plausible values, creating several complete datasets. You analyse each dataset in the usual way, then SPSS pools the results into a single set of estimates whose standard errors reflect the uncertainty introduced by the missing data. That last point is crucial: a single imputation (filling each gap once, for example with the mean) pretends you know the missing values, and so understates uncertainty. Multiple imputation does not make that pretence.

How to run it click by click

Go to Analyze > Multiple Imputation > Impute Missing Data Values. On the Data tab, move all variables that will appear in your analysis — plus any auxiliary variables that predict missingness or the missing values themselves — into the Variables in Model box. Including good predictors here is what makes the imputations plausible, so do not restrict the list to only the variables in your final model if other measured variables carry useful information.

On the Method tab, the automatic method is a sensible starting point: SPSS chooses a suitable imputation model based on your variable types. Set the number of imputations. Five was once the standard recommendation, but with modern computing there is little reason to stop there if a substantial share of data is missing; ten to twenty imputations are commonly used, and more imputations mainly cost computing time. Set the maximum percentage of missing values if you want to exclude variables that are mostly empty, and decide whether to cap imputed values within observed ranges for scale variables.

On the Output tab, choose to create a new dataset containing the imputed datasets, and ask for an imputation model summary if you want to inspect how well the model fits. Click OK. SPSS creates a new dataset in which the original data appear as Imputation 0, followed by Imputation 1, Imputation 2, and so on, with a variable marking which imputation each row belongs to.

Paste-ready SPSS syntax

If you prefer syntax, or want a reproducible record of exactly what you did, the following template runs a basic multiple imputation. Replace the variable names with your own, adjust the number of imputations, and run it from a Syntax window:

MULTIPLE IMPUTATION Age Income Score1 Score2
  /IMPUTE METHOD=AUTO NIMPUTATIONS=20 MAXPCTMISSING=NONE
  /CONSTRAINTS Age(MIN=18 MAX=90) Income(MIN=0)
  /MISSINGSUMMARIES NONE
  /IMPUTATIONSUMMARIES MODELS
  /OUTFILE IMPUTATIONS=ImputedData.

The CONSTRAINTS line is optional but useful: it stops the procedure from imputing impossible values, such as a negative income or an age below your sample’s minimum. After running the syntax, check the new dataset, then run your analysis as normal — SPSS will automatically repeat it across imputations and pool the results, provided you analyse the imputed dataset rather than the original.

How to interpret the pooled output

After imputation, your output contains two layers: results for each individual imputation, and a pooled result. Report the pooled result. In a regression, for example, you will see pooled coefficients, standard errors, t values, and significance levels, plus the fraction of missing information, which tells you how much of the uncertainty in each estimate comes from the missing data. A high fraction of missing information is a warning that your conclusions lean heavily on the imputation model — worth noting, not hiding.

Also inspect the imputed values themselves before trusting the pooled numbers. Compare the distributions of observed and imputed values for key variables: imputed values should look plausible and broadly similar to observed values, not wildly out of range or piled up at a single point. If they look wrong, revisit your imputation model — usually by adding predictors or fixing constraints — rather than proceeding anyway.

Multiple Imputation vs Listwise Deletion: Which Should You Choose?

Use this comparison when you write your methods section and when a supervisor asks you to justify your choice.

Choose listwise deletion when missing data are minimal, the pattern looks consistent with MCAR, and dropping incomplete cases leaves your sample comfortably large enough. It is transparent, easy to replicate, and uses only observed data.

Choose multiple imputation when a substantial share of cases would otherwise be lost, missingness relates to observed variables (MAR), or you want standard errors that honestly reflect the uncertainty of filling in gaps. It preserves sample size and reduces bias under MAR, at the cost of a more complex workflow and a methods paragraph you must write carefully.

Avoid single imputation shortcuts such as replacing every missing value with the variable mean or the last observed value. These methods distort distributions, shrink variances, and understate standard errors. They may look tidy in a dataset, but they make your results look more precise than they are.

Reporting in APA Style (7th ed.)

APA style expects transparency about missing data, not a particular method. In your Method or Results section, state how much data were missing, what you assumed about the pattern, what you did, and how many cases the final analysis used. A template you can adapt — replacing the bracketed parts with your own numbers — is:

“Missing data ranged from [x]% to [y]% across variables. Little’s MCAR test was [non-significant / significant], χ²(df) = [value], p = [value]. [Listwise deletion was used, leaving N = [value] complete cases for analysis. / Missing values were addressed with multiple imputation (m = [number] imputations) using [variables included in the imputation model]; pooled estimates are reported.]”

For a pooled regression, report the pooled coefficients with their pooled standard errors and confidence intervals, and note that estimates are pooled across imputations. Do not report a single imputation’s results as if they were the final analysis, and do not omit the original sample size: readers should be able to see both how many people you recruited and how many contributed to each analysis.

Key Takeaways

  • Diagnose first: report how much data are missing and in which variables, and run Little’s MCAR test as supporting evidence before choosing a method.
  • Listwise deletion is simple and defensible when missing data are few and plausibly MCAR, but it discards whole cases and can bias results under MAR.
  • Multiple imputation creates several plausible complete datasets, analyses each one, and pools the results, so standard errors honestly reflect the uncertainty of filling in gaps.
  • Never use single mean substitution as a quick fix: it distorts distributions and makes results look more precise than they are.
  • Always report what you did: the amount missing, the method chosen, the number of imputations if applicable, and the final analysis sample size.

Frequently Asked Questions

How much missing data is too much for listwise deletion?

There is no fixed cutoff. The decision depends on the pattern as well as the percentage: a small, random amount is usually fine to delete, while a larger amount, or missingness linked to observed variables, favours multiple imputation. If listwise deletion would remove a large share of your sample, imputation is usually the safer choice.

Does multiple imputation invent data?

No — and this is a common misunderstanding. Multiple imputation does not claim to recover the true missing values. It generates several plausible values consistent with the observed data, analyses each version, and pools the results so that the final uncertainty reflects the fact that the true values are unknown.

How many imputations should I use?

Five was the traditional minimum, but ten to twenty are commonly used today, especially when a substantial share of data is missing. More imputations improve the stability of pooled estimates and cost little extra computing time, so erring toward more is reasonable.

Should I impute the dependent variable?

Imputing predictor variables is standard practice. Cases with a missing outcome are often handled differently: many analysts impute all variables for the imputation model but then exclude cases with imputed outcomes from the final analysis model, a strategy sometimes called impute-then-delete. Whichever approach you take, state it clearly in your methods section.

What if my data are missing not at random (MNAR)?

Standard multiple imputation assumes MAR, so it does not fully solve MNAR. If you have good reason to suspect MNAR, acknowledge it, run sensitivity analyses under different assumptions if you can, and discuss the limitation honestly rather than presenting imputed results as if the problem were solved.

Need this done on your data? Get a free quote → /services/