Missing data are an inevitable reality in psychological research. Participants who do not complete all sessions of a longitudinal study, items left blank on a questionnaire, technical failures in response recording, dropouts in clinical trials. The question is not whether you will have missing data, but how you are going to handle them. And the way you do so can profoundly affect your study's conclusions, from the estimation of effects to the validity of statistical inferences.
The first thing to understand is that not all missing data are equal. Rubin (1976) proposed a classification into three mechanisms that remains the foundation of the entire field. Data that are Missing Completely At Random (MCAR) are those whose absence is unrelated to any variable in the study, whether observed or unobserved. It is as if you had flipped a coin to decide which data to remove. Data that are Missing At Random (MAR) are those whose absence can be explained by other observed variables in the study, but not by the value of the missing datum itself. And data that are Missing Not At Random (MNAR) are those whose absence is related to the very value that is missing: for example, when the most depressed participants are the ones who drop out of a depression study precisely because they are too depressed to continue.
Why listwise deletion is problematic
The most common strategy for handling missing data: and also the worst in most situations: is listwise deletion: you simply remove all participants who have any missing datum and analyze only the complete cases. This strategy has two serious problems. The first is the loss of statistical power. If you have 200 participants and each has a 5% probability of being missing on any given variable, with 10 variables you can easily lose 40% of your sample. The second problem is more subtle but potentially more harmful: if the data are not Missing Completely At Random (i.e., if the mechanism is MAR or MNAR), listwise deletion produces biased estimates. The participants remaining in the analysis are not representative of the original sample, and the conclusions you draw may not generalize to the population of interest.
Despite these problems, listwise deletion remains the default method in most statistical software. SPSS applies it automatically unless you specify otherwise, and many researchers are not even aware they are using it. This is especially concerning because the simplest and most recommended alternative by methodologists has been available for decades: multiple imputation.
Multiple imputation: the current standard
Multiple imputation (MI) works as follows. Instead of deleting missing data or filling them in with a single value (such as the mean, which is another discouraged practice because it underestimates variability), the procedure generates multiple copies of the dataset, each with the missing values replaced by plausible values generated from the distribution of the observed data. Typically, between 20 and 50 imputed datasets are generated. The statistical analysis of interest is then performed on each of the imputed datasets, and finally the results are combined using Rubin's rules, which account for both the variability within each imputation and the variability between imputations.
The beauty of multiple imputation is that it produces unbiased estimates under the MAR assumption (one of the statistical assumptions worth checking) (which is the most realistic in most psychology research situations) and that the standard errors adequately reflect the additional uncertainty caused by the missing data. A confidence interval obtained after multiple imputation is correctly wider than one obtained with listwise deletion, because it incorporates the uncertainty about what the actual values of the missing data would have been.
In R, the mice package (Multivariate Imputation by Chained Equations) is the reference for multiple imputation. Its operation is relatively straightforward: the mice() function generates the imputed datasets, with() applies the statistical model to each one, and pool() combines the results. In SPSS, the multiple imputation module has been available since version 17, although its flexibility is less than that of mice. JASP does not offer multiple imputation natively, which is an important limitation for studies with missing data.
Practical considerations and recommendations
Before applying any method for handling missing data, it is essential to describe the pattern and extent of missing data in your sample. How many participants have complete data, how many have partial data, which variables have the most missing data, and whether there are systematic patterns in the missingness. The naniar package in R offers excellent visualization tools for exploring missing data patterns. It is also important to assess whether the missing data mechanism is MCAR (using Little's test) or whether the missingness is associated with observed variables (which suggests MAR). The MNAR mechanism cannot be tested with the available data, but it can be evaluated through sensitivity analyses.
My general recommendation is to use multiple imputation as the default method whenever you have more than 5% missing data. Below that threshold, listwise deletion usually produces results very similar to multiple imputation, as long as the data are missing at random. But if you have 10%, 20%, or more missing data (a frequent situation in longitudinal and clinical studies, where mixed models are often preferred), multiple imputation is not only advisable but necessary to obtain valid conclusions. And in any case, always report the proportion of missing data and the method you used to handle them, because this transparency allows the reader to evaluate the robustness of your results. If you need guidance with your missing data, check our thesis analysis service.
Before you submit: writing "cases with missing data were excluded" without saying how many they were or why they were missing is one of the fastest ways to earn a methods comment in the first review round. If your manuscript is already written, run it through the Q1 Reviewer: a free Reviewer 2 style pre-review that tells you, in under a minute, what a reviewer would question about your sample and your analyses while you can still fix it.