Design and analysis of pen studies in the animal sciences.

Increasingly, research is being performed where animals subjected to a common treatment are also housed in a common pen. Issues have been raised regarding the proper planning of experiments and conduct of statistical analyses in these instances. This presentation reviews the problems associated with ignoring animal grouping during data analyses, and gives appropriate methods to use when animals are grouped in pens. Using animals as the error term when treatments are applied to pens can cause two types of biases. The first one is one of location, which biases point (parameter) estimates of the treatment effects. The pen effect includes unrecognized, systematic non-random effects other than that of the treatment, which is why pens must be replicated and randomized if treatment effects are to be estimated without biases. Pens also result in non-systematic random effects. These affect the variance of the sub-pen units (cows). That is, cows within a pen have more in common than cows across pens. In essence, pen studies have an implicit split-plot design where the main plots (pens) receive the treatments of interest, while the sub-plots (cows) receive all the same sub-plot treatment. Using the sub-plot error to test the effect of main plot treatment effects causes a second type of bias by creating artificial degrees of freedom, and hence biasing severely the test of significance for the treatment effects. Behind all statistical analyses is a mathematical model with its associated assumptions. The assumption with pen-based treatment is that pens have a random effect. Thus pens, or the interaction of pens with other model elements is/are the correct error term(s). The same statistical designs used with cows as experimental units can be used with pens. The number of experimental units to achieve a given power is considerably less with pens because the variance associated with pens is substantially less than the variance of cows within pens. Pens must be replicated, randomized, and included in the statistical model if treatment effects and their significances are to be estimated without biases. [ABSTRACT FROM AUTHOR]