Missing data

Missing data are values absent from a dataset where observations would otherwise be expected. Their patterns and causes affect analysis, interpretation, and statistical validity.

Missing completely at random: A missingness mechanism in which absence is independent of both observed and unobserved data. This strongest randomness condition allows complete-case analysis to remain unbiased under suitable assumptions.
Multiple imputation: A statistical procedure that replaces missing values with multiple draws, analyzes each completed dataset, and combines results. It represents uncertainty about absent values instead of treating one replacement as certain.
Null value: A marker indicating that a value is absent, unknown, or not applicable in a data system. Null markers encode absence but can have different meanings across systems and analyses.
Nonresponse bias: Bias arising when people who do not respond differ systematically from those who do. Survey nonresponse creates absent answers that can distort population estimates.
Missing at random: A missingness mechanism in which absence depends on observed data but not on unobserved values, conditional on those observations. This condition justifies many likelihood-based and imputation methods when the relevant predictors are included.
Complete-case analysis: An analysis restricted to observations with no missing values on the variables being used. Its simplicity is attractive, but its validity depends on the missingness process and analysis target.
NaN: A special floating-point value representing an undefined or unrepresentable numeric result. NaN is a common computational marker for numeric missingness, though its behavior differs from null.
Attrition bias: Bias caused when participants leave a study in ways related to outcomes or exposures. Longitudinal studies can lose measurements as participants drop out selectively.
Missing not at random: A missingness mechanism in which absence depends on the unobserved value itself, even after conditioning on observed data. When missing values depend on their own unseen levels, ordinary correction methods may remain biased.
Inverse probability weighting: A method that weights observations by the inverse of their estimated probability of being observed or assigned. It can correct selection from observed predictors when observation probabilities are modeled adequately.

Linked from 14 pages