Encyclopaedia of DesignTheory: Latin squares
|Experimental||Other designs||Math properties||Stat properties|
Here is a Latin square of order 5. We will consider a number of different experiments where it could provide the basic structure for the design. The discussion is due to R. A. Bailey: see for example "Orthogonal partitions for designed experiments", Designs, Codes and Cryptography 8 (1996), 45-77.
The structure of the experimental units is a rectangular grid (which happens to be square in this case); there is no structure on the set of treatments.
We can use the Latin square to allocate treatments. If the rows of the square represent patients and the columns are weeks, then for example the second patient, in the third week of the trial, will be given drug D. Now each patient receives all five drugs, and in each week all five drugs are tested.
Again the Latin square gives an allocation of varieties to plots in the same way as before.
Again, we can assign treatments to units using the Latin square. Let us number the cells of the square from 1 to 25 as follows.
We assume that the rows of the table represent days, so that units 1, ..., 5 are the timeslots on the first day. Now we use the five columns of the square to assign the five temperatures, and the letters to assign the recipes A, B, C, D, E.
Label the rows of the square by temperatures; the columns by mixing times; and let the letters A, ..., E stand for recipes as before. Each of the 25 cells of the square represents the treatment to be applied to one experimental unit.
We have 25 different cheeses to be tasted. Fifteen tasters are available; each taster can only taste and compare five of the cheeses. Thus, not all pairs of cheeses can be directly compared by the same taster; we require that any pair should be compared by at most one taster. Also, though this is not a formal requirement, it would be nice if two cheeses which are not compared with one another are both compared with a constant number of other cheeses.
Suppose further that only five tasting booths are available, so that three tasting sessions must be run.
We number the cheeses from 1 to 25, and associate each cheese with a cell of the square array according to the earlier figure. In the first tasting session, we assign to the five tasters the cheeses in the five rows of the array: that is, the first taster gets cheeses 1, 2, 3, 4 and 5; the second gets 6, 7, 8, 9, 10; and so on. In the second session, we use the columns, so that the first taster gets 1, 6, 11, 16, 21 and so on. In the final session we use the letters of the Latin square, so that the first taster gets 1, 7, 13, 19, 25 and so on.
Note that the sets of cheeses assigned to the fifteen tasters are the blocks of the net associated to the Latin square, while the three tasting sessions correspond to the parallel classes in the resolution of the net. Properties of nets ensure that our optional condition is indeed satisfied.
Table of contents | Glossary | Topics | Bibliography | History
R. A. Bailey,
Peter J. Cameron
9 October 2002