10th U6+ Consortium of African Universities International Conference

Cook's distance remains one of the most widely used case-deletion diagnostics for locating influential observations in regression analysis, yet its ordinary least squares formulation is not well suited to panel data in which observations are correlated within units and across time. This article develops and evaluates a generalized Cook's distance for two-way random effects panel data models estimated by generalized least squares, extending the classical case-deletion argument to accommodate hierarchical dependence, serial correlation, and heteroscedastic disturbances. The generalized statistic is derived analytically from the block-deletion normal equations using the Sherman-Morrison-Woodbury identity and is expressed in closed form through a generalized hat matrix and whitened residuals. Its performance is assessed through an extensive Monte Carlo experiment spanning balanced and unbalanced panels of varying dimension, AR(1) and ARMA(1,1) error processes, and unit-, time-, and observation-level contamination at 5%, 10%, and 15%. Across every design point examined, the generalized Cook's distance achieved higher true positive rates, negligible false positive rates, and superior F1-scores relative to the classical statistic. The findings support the routine use of covariance-adjusted Cook's distance whenever regression diagnostics are applied to correlated panel data.