Possible Advantages of Matching Concepts in Agricultural Experiments with Blocks
In many agricultural studies the experimental units are grouped into blocks based on covariates available when the experiment is planned. The treatments are randomly assigned to experimental units within the blocks, and the blocks are included in the model used to analyze the study. In some studies, additional covariates are measured while the experiment is ongoing. A question then becomes, can the information from the additional covariates be used for “post blocking” such that the treatment effect is more effectively measured (similar to matching in observational studies)? Issues that could impact this answer include: 1) how unbalanced are the new covariates among the treatments; 2) how much do the new covariates impact the response and what is the nature of the covariate impact on the response; 3) how much do the treatments influence the new covariates; 4) how much information is used or lost if the new covariates are used to improve the blocks; 5) exactly how is the “post blocking” done (some sort of matching, or some sort of ANCOVA); and so forth.
To obtain a very preliminary answer to this question, a study of sheep grazing endophyte infected fescue was considered. There were four treatments in the study and eight blocks of sheep (based on covariates related to body condition score). The sheep were on the treatments for approximately 140 days, and some additional covariates were measured during this period. Using these covariates to “post block” did in fact modify the estimated treatment effect, but we are still investigating properties of the new estimated treatment effect. Some simulation studies to address the question are underway and will be discussed.