Abstract
An argument in favour of projecting the score function for models involving incomplete data is presented. Projection is then applied to the aggregate Markov‐chain model resulting in weighted least‐squares estimators. The limit theory and efficiency of these estimators are studied using martingale limit theory.