What We Miss About Missing Values

Missing data should be understood as a byproduct of the data-generating process, not merely a defect to be fixed. The reasons for missingness are categorized into three mechanisms: Missing Completely at Random (MCAR), Missing at Random (MAR), and Missing Not at Random (MNAR). The distinction between these, particularly MAR and MNAR, is an untestable assumption that relies on domain knowledge rather than statistical tests. Various methods exist to handle missing values, such as Heckman's selection models, inverse probability weighting, multiple imputation, and sensitivity analysis, each with its own underlying assumptions. Ultimately, the appropriate method depends on the specific question being asked and the assumed reason for the data's absence.