How to Handle Missing Data in Your Dataset

You open your SPSS dataset, run a regression with 240 respondents, and the output says "N = 178." Sixty-two cases vanished. No warning, no error message. SPSS silently dropped every row that had at least one blank cell across your variables, and your analysis now rests on 74% of the data you actually collected. That silent deletion is listwise exclusion, and it is the default behavior in almost every statistical procedure. Whether those 62 missing cases matter depends entirely on why they are missing.

The Three Types of Missing Data

Rubin's 1976 classification divides missing data into three categories based on the relationship between the missingness and the data values themselves. The distinction is not academic. It determines which correction methods are valid and which will introduce bias.

MCAR (Missing Completely at Random) means the probability of a value being missing has no relationship to any variable in your dataset, observed or unobserved. A survey respondent skips a question because they accidentally turned two pages at once. A lab technician spills a blood sample. The missingness is pure accident. MCAR is the easiest type to handle because the remaining data is still a random sample of the full dataset. Simple deletion methods work here without introducing bias. In practice, true MCAR is rare.

MAR (Missing at Random) is more common and more confusing than its name suggests. "At random" does not mean the missingness is random. It means the probability of missing data can be explained by other observed variables in your dataset, but not by the missing value itself. For example, younger respondents in a health survey are more likely to skip the income question. Age predicts the missingness, but once you account for age, the actual income value does not predict whether income is missing. MAR data requires methods that use the observed variables to fill in the gaps, such as multiple imputation.

MNAR (Missing Not at Random) is the hardest case. Here the missingness depends on the unobserved value itself. People with very high incomes refuse to report their income. Patients with severe depression drop out of a treatment study precisely because their symptoms are worst. No amount of observed data can fully explain the pattern, and standard imputation methods will produce biased estimates. MNAR requires specialized models such as selection models or pattern-mixture models, and it often demands sensitivity analyses to test how different assumptions about the missing mechanism affect your conclusions.

How to Diagnose the Missing Data Pattern

Before choosing a treatment method, you need to know how much data is missing and whether the pattern looks random. Start with the raw numbers. In SPSS, go to Analyze > Missing Value Analysis. The summary table shows the percentage of missing values for each variable. A variable with 2% missing is a different problem than one with 25% missing.

Next, look at the pattern. SPSS produces a missing data pattern chart that shows which combinations of variables tend to be missing together. If the same 30 respondents skipped both the income question and the savings question but answered everything else, that is a pattern worth investigating. It might indicate that financially sensitive questions trigger non-response as a block.

Little's MCAR test, available in the same SPSS dialog, tests whether the missing data pattern is consistent with MCAR. The null hypothesis is that the data is MCAR. A non-significant result (p > 0.05) means you cannot reject MCAR, so simple deletion may be acceptable. A significant result (say, chi-square = 47.3, df = 22, p = 0.001) tells you the missingness is related to other variables in the dataset, pointing toward MAR or MNAR. The test cannot distinguish between MAR and MNAR because that distinction depends on the unobserved values themselves.

We have seen students skip this diagnostic step entirely and jump straight to deletion or imputation. That is the equivalent of choosing a statistical test without checking your assumptions. The most common statistics mistakes in theses include exactly this kind of shortcut.

Treatment Methods: From Simple to Sophisticated

Listwise Deletion

Listwise deletion removes any case with at least one missing value on any variable in the analysis. It is the SPSS default for regression, ANOVA, and most other procedures. The advantage is simplicity: every analysis runs on the same complete set of cases. The disadvantage is data loss. In a regression with 12 predictor variables, even 3% missingness per variable can eliminate 30% or more of your cases if the missing values are spread across different respondents. Listwise deletion is only safe when the data is MCAR and the total proportion of incomplete cases is small, typically under 5%. Beyond that threshold, you are throwing away too much information and reducing statistical power unnecessarily.

Pairwise Deletion

Pairwise deletion keeps each case for every analysis where that case has complete data on the specific variables being tested. A respondent who answered items 1 through 15 but skipped item 16 would be included in all correlations that do not involve item 16. This preserves more data than listwise deletion, but it creates a problem: different analyses within the same study use different subsets of respondents with different sample sizes. A correlation matrix built with pairwise deletion might use N = 230 for one pair and N = 195 for another. In some cases, the resulting matrix is not positive definite, meaning it contains mathematical contradictions that make factor analysis or structural equation modeling impossible.

Mean Imputation

Mean imputation replaces every missing value with the variable's mean. It preserves the sample size and keeps the overall mean unchanged. That is where the advantages end. By replacing real variation with a single constant, mean imputation reduces the variable's standard deviation and compresses its distribution. Correlations between the imputed variable and all other variables are artificially weakened. In a dataset of 200 cases with 20 missing values on a stress scale (10% missing), mean imputation would reduce the standard deviation of that variable by roughly 5% and bias every correlation involving stress toward zero. Most methodologists now consider mean imputation unacceptable for published research. Avoid it.

Multiple Imputation

Multiple imputation is the current gold standard for handling MAR data. The procedure creates multiple complete versions of your dataset (typically 5 to 20), each with different plausible values substituted for the missing entries. Those values are drawn from a predictive model that uses the observed data to estimate what the missing values might have been, with random variation added to reflect the uncertainty. You run your planned analysis on each imputed dataset separately, then pool the results using Rubin's rules, which combine the estimates and adjust the standard errors to account for the imputation uncertainty.

In SPSS, go to Analyze > Multiple Imputation > Impute Missing Data Values. Set the number of imputations to at least 20 for less than 10% missing data, or 40 or more for higher rates. SPSS creates a stacked dataset with an Imputation_ variable that identifies each copy. After imputation, run your analysis as usual. SPSS pools the results automatically when you analyze the imputed dataset. Academic Stats Agent can also walk you through the imputation setup and help interpret the pooled output.

Expectation-Maximization (EM)

The EM algorithm estimates the parameters (means, variances, covariances) of the complete data by iterating between two steps. The E-step estimates the expected values of the missing data given the current parameter estimates. The M-step re-estimates the parameters using both the observed and estimated data. The algorithm repeats until the estimates converge. EM produces a single set of parameter estimates rather than multiple datasets, which makes it faster and simpler than multiple imputation. The tradeoff is that EM does not account for imputation uncertainty in the standard errors, so confidence intervals and p-values may be slightly too narrow. It works well as a diagnostic tool and for exploratory analysis, but multiple imputation is preferred for final reported results in a thesis or publication.

Decision Table: Choosing the Right Method

The right approach depends on two factors: how much data is missing and what type of missingness you are dealing with. The table below summarizes the recommended method for each combination.

% Missing	Type	Recommended Method
< 5%	MCAR	Listwise deletion (safe, simple)
< 5%	MAR	Multiple imputation or listwise deletion
5 – 20%	MCAR	Multiple imputation (preferred) or pairwise deletion
5 – 20%	MAR	Multiple imputation (required)
> 20%	MCAR	Multiple imputation with 40+ imputations
> 20%	MAR	Multiple imputation with 40+ imputations
Any %	MNAR	Sensitivity analysis + selection or pattern-mixture models

Mean imputation is absent from the table because it should not be used regardless of the percentage or type of missingness. If your supervisor suggests it, point them to Schafer and Graham (2002), which showed that mean imputation produces biased parameter estimates even under MCAR conditions with as little as 10% missing data.

The Most Common Mistake: Doing Nothing

The single worst approach to missing data is ignoring it. And it is by far the most common one we see. A student collects 300 survey responses, imports the file into SPSS, and runs an independent-samples t-test. SPSS reports N = 261 in the output. The student copies the result into the thesis, reports t(259) = 2.14, p = 0.033, and moves on. Nowhere in the methodology or results chapter do they mention that 39 cases were excluded, explain why, or test whether the missingness was random.

This is a problem for two reasons. First, if the data is MAR or MNAR, the t-test result is biased. The 261 remaining cases are not representative of the original 300. Second, thesis committees increasingly expect a missing data section in the methodology chapter. Reporting the percentage of missing values, running Little's MCAR test, and describing how you handled the gaps takes one paragraph. Failing to do so suggests either unawareness or carelessness, and neither reflects well during a defense.

In every project we complete, the missing data diagnostic is one of the first things we run. We check the pattern, run Little's test, and choose the method before any substantive analysis begins. If you are writing up your own results, make sure you address missing data before reporting test statistics. Our guide on common thesis statistics mistakes covers this and 9 other errors that trip up students regularly.