A Missing-Data Decision Tree: From Mechanisms to Sensitivity Analysis
Assess missingness mechanisms before deleting observations and choose complete-case analysis, multiple imputation and sensitivity analyses appropriately.
Why missing data matter
Missing observations reduce information and can bias estimates depending on why data are missing. There is no universal percentage below which missingness is harmless.
MCAR, MAR and MNAR
MCAR: missingness is independent of observed and unobserved values.
MAR: conditional on observed information, missingness no longer depends on the missing value.
MNAR: missingness still depends on unobserved values after conditioning.
These mechanisms cannot be identified conclusively from one statistical test; process knowledge and reasons for missingness are essential.
Methods
Complete-case analysis can lose efficiency and produce bias. Single imputation methods such as mean substitution or last observation carried forward treat imputed values as known and underestimate uncertainty.
Multiple imputation creates several plausible datasets, analyses each and combines estimates using Rubin's rules. The imputation model should include outcomes, exposures, analysis covariates and useful auxiliary predictors of missingness.
Sensitivity analysis
MAR cannot be fully verified from observed data. Pattern-mixture, delta-adjustment or tipping-point analyses can assess robustness to plausible MNAR departures.
References
Accessed 20 June 2026.