Nonignorable Attrition in Multi-period Panels with Refreshment Samples


In panel surveys, some observation units drop out before the end of the observation period. This panel attrition cannot be ignored if it is related to the variables of interest. Hirano, Imbens, Ridder and Rubin (Econometrica, 2001) propose the Additively Nonignorable (AN) attrition model to correct for the potential selectivity of the attrition in panels with two periods. If a refreshment sample is available in the second period, their model nonparametrically just-identifies the population distribution and the observation probability. We propose the Sequential Additively Nonignorable attrition model, that just-identifies the population distribution and the sequence of observation hazards for panels with more than two periods.

Journal of Business and Economic Statistics, 37(3)