The effect of missing data on latin square designs

Abstract

Three potential patterns of missing observations in Latin squares are considered (i) the loss of replicates of one or more treatments, (ii) removal of complete blocks, and (iii) plots missing at random throughout the design. For (iii) the loss of 1, 2, or 3 observations is studied extensively and the best and worst cases are examined in general. The robustness of the designs to missing observations is assessed in terms of average and maximum variances of treatment differences and the efficiency relative to the complete design.