A topic in the Open Knowledge Graph — a free, open map of 15,290 topics and the order to learn them in.
Missing data is ubiquitous in psychological research and can bias results if not properly addressed. Mechanisms of missingness—missing completely at random (MCAR), missing at random (MAR), and missing not at random (MNAR)—determine appropriate handling strategies. Deletion methods (listwise, pairwise) are simple but can bias results and reduce statistical power when data are not MCAR. Multiple imputation and maximum likelihood estimation are sophisticated methods that can provide unbiased estimates when data are MCAR or MAR. Understanding the mechanism and pattern of missing data is essential for choosing analytical strategies.
Examine a dataset with missing data and determine the likely mechanism (MCAR, MAR, MNAR) by exploring patterns and relationships between missing status and observed variables.
Missing data can be ignored if the sample size is large enough (actually, bias from missing data depends on the mechanism, not sample size). Listwise deletion is appropriate because it uses only complete cases (actually, listwise deletion can introduce bias and reduces power unless data are MCAR).
Missing data is not just an inconvenience — it is a measurement and inference problem that, if handled naively, can systematically distort your conclusions. From your work on inferential statistics, you know that valid inference requires your observed sample to represent the target population. When data are missing, you no longer have a clean random sample; you have a sample shaped by a process that determined who or what is missing. Understanding that process — the missingness mechanism — is the essential first step, because the right remedy depends entirely on why data are absent.
The three mechanisms form a hierarchy of seriousness. Missing Completely At Random (MCAR) means the probability of a value being missing is unrelated to anything — not to the variable itself, not to any other measured variable. A participant's questionnaire page getting coffee spilled on it is MCAR. Under MCAR, your complete cases are a random subset of your intended sample, and simple deletion methods (listwise, pairwise) produce unbiased estimates — just with reduced power. Missing At Random (MAR) is more subtle: missingness is related to other *observed* variables in the dataset, but not to the unobserved missing values themselves. Women in a survey might be less likely to report income, but if you can model who is missing income based on other observed variables (gender, education, age), the missingness is "explainable" by things you've measured. Under MAR, sophisticated methods can recover unbiased estimates. Missing Not At Random (MNAR) is the hardest case: missingness is related to the missing value itself. Depressed individuals are less likely to complete depression measures precisely because of their depression. No statistical method can fully correct for MNAR without additional assumptions or external data.
Listwise deletion — dropping any case with any missing value — is the default in most software and the most commonly misused approach. Under MCAR it gives unbiased (but underpowered) results. Under MAR or MNAR it introduces selection bias: your "complete case" sample is systematically different from the intended sample in ways that distort your estimates. Imagine a longitudinal study where participants with worsening symptoms are most likely to drop out. Your remaining sample of "completers" will look healthier than the true population, biasing outcome estimates downward. This isn't a statistical technicality — it's a substantive distortion of your research conclusions.
Multiple imputation (MI) addresses this by replacing each missing value not with a single number but with a set of plausible values drawn from a distribution estimated from observed data. Running analyses on multiple completed datasets and combining results using Rubin's rules propagates the uncertainty from the imputation into your final estimates, producing correct standard errors. Full information maximum likelihood (FIML) takes a different approach: instead of filling in missing values, it uses all observed information to estimate model parameters directly, including cases with partial data. Under MAR, both MI and FIML produce valid inferences. Under MNAR, both are biased — and so is any other method — but MI and FIML typically produce *less* biased estimates than listwise deletion, making them the preferred default.
The practical workflow starts with diagnosing the mechanism: examine whether missingness correlates with observed variables (test MCAR formally with Little's test, explore MAR patterns by regressing missingness indicators on observed covariates). Then choose your method accordingly — and always report how you handled missing data so readers can evaluate the validity threat. The key mindset shift is treating missing data as a data quality issue to be modeled, not a nuisance to be removed. A dataset with 30% missing data handled thoughtfully via MI can yield more valid conclusions than a "complete" dataset where missingness was ignored.