Using the Factor Relationship Diagram

to Identify the Split-plot Factorial Design

Research Branch
B.C. Ministry of Forests
P. O. Box 9519 STN Provo Govt.
Victoria, B.C.

Introduction

Split-plot experimental designs can be hard to identify and understand. This paper will
show how to correctly identify split-plot designs and will contrast them to simple factorial
designs. The process of developing a factor relationship diagram (FRO) can help
elucidate the correct design. An example will be used to show how the FRO works and
the appropriate SAS code to obtain the correct ANOV A F-tests.

Factors and their relationships

A treatment factor is an independent variable that may affect the response variable. It is
controlled by the researcher and is often of specific interest to the researcher. Its levels
can be randomly assigned to the experimental material. A classification factor is similar
to a treatment factor except that its levels are not assignable because it is an inherent
property of the experimental material. Examples include gender, occupation, and tree
species.
The experimental material or subject of the experiment is usually arranged in a hierarchy
of differently sized nested units (the unit hierarchical structure). For instance, trees may
be arranged into rows or groups of, say, 20 trees. These rows may then be grouped into
blocks. Measurements may be taken by subsampling the trees. For example, the number
of cones may be counted on only two of the many branches on each tree. . Another
example would be a study of school children in which the children are arranged into
classrooms, classrooms into schools and schools into school districts. This is a nested
hierarchical structure, with each level a different unit/actor.
While unit factors are generally nested within each other. treatment and classification
factors are often crossed. When two factors are crossed this means that the levels of one
factor occur in the experiment with all the levels of the other factor. The treatment and
classification factors, together with their relationships, form the treatment structure of an
experiment. We develop the full description of the experimental design when we assign
the different treatment and classification factors to appropriate unit factors in the unit
hierarchical structure.

126
Factor Relationship Diagram Construction

The factor relationship diagram is used to display the relationships between the various
treatment, classification and unit factors of an experimental study. It can be constructed
using the following three steps.

Step 1: Identify all the treatment, classification and unit factors in the study. Assign
unique numbers to each unique level for each factor. This will help clarify whether
factors are crossed or nested.

Step 2: Draw the unit structure with the largest grouping of experimental material placed
on the top row and successively smaller groupings placed below.

Step 3: Determine which unit factor is the experimental unit for each treatment and
classification factor. Then place the treatment and classification factors above their
experimental unit. If two or more factors have the same experimental units then they can
be placed in any order. The actual random assignment of factor levels to their
experimental units is omitted to assist in the identification of the factor relationships.

Factorial Designs

The general factorial design has two or more treatment or classification factors which are
crossed with each other. In the simple factorial, all the combinations of factor levels are
randomly assigned to the elements of just one unit factor. This design is known as the
Completely Randomized Design (CRD).
Suppose a forestry experiment is to be done with three different types of fertilizer
(treatment factor F) and two different amounts of a boron supplement (treatment factor A)
to be applied to orchard trees. If these treatments are crossed then there will be six
different treatments in all. Suppose that the experimental material are young trees
arranged into rows of two trees each. The unit structure has two factors: individual trees
which are nested within the second factor of rows. A simple, completely randomized
design (CRD) results if all six treatment levels are randomly assigned to the rows of 2
trees each. The rows are the experimental units. Trees are the subsamples, since
measurements for most response variables must be taken on each tree individually.
The FRD for this design is shown in Figure I, and the SAS program is shown below.
Note that the test statement is necessary to get the correct F-tests. With subsampling, the
default error term is not the correct error term for the treatment factors.

127
ANDVA Table:
Source of Variation df Test
Fertilizer F 2 R(FA)
Boron A 1 R(FA)
FxA 2 R(FA)
Rows R(FA) 6 E(RFA)
Trees E(RFA) 12
Total 23

SAS Program:
* variable names: f for fertilizer types a for boron supplement
row for row of trees y for the response variable;
proc glm;
class f a row;
model y= f I a row(f*a);
test h = f I a e=row(f*a); *<== required because of the sub-sampling;
run;

Fertilizer F 1 2 3

/~ /~ /~
Boron A 2 1 2 2

Rows of Trees
/\ /\ /\ /\ /\ /\
I .2 3 4 5 6 7 8 9 IO II 12

Trees
/\ /\ /\ /\ /\ /\ /\ /\ /\ /\ /\ /\
2 3 4 S 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20' 21 22 23 24

Figure 1. Factor Relationship Diagram for a two-way factorial design with subsampling

128
Split-plot Designs

Split-plot designs are a specific form of a factorial design, where different treatment or
classification factors are assigned to different levels in the unit hierarchy. To show how
this works, lets change the example. Suppose that the three different fertilizers are
randomly assigned to the rows of two trees as before, but that the boron supplement is
now assigned to individual trees within each row. This is shown in Figure 2. Notice that
there is no subsampling in this design. Thus the default error is the split-plot error term
and does not need to be specifically mentioned in the SAS program.

ANOV A Table:
Source of Variation df Test
Fertilizer F 2 R(F)
Rows R(F) 9
Boron A 1 AxR(F)
FxA 2 A x R(F)
AxRows R(FA) 9 E(RFA)
Trees E(RFA) 0
Total 23

SAS program.:
* variable names: f for fertilizer types a for boron supplement
row for row of trees y for the response variable;
proc glm;
class f a row;
model y = f I a row(f);
test h = f e=row(f);
run;
Suppose that tree response was determined by counting the number of cones produced on
two branches of each tree. In this case, these branch responses are subsamples at the
split-plot level. The SAS program must then include the subsampling and appropriate
test statements.

129
SAS program.:
* variable names: f for fertilizer types a for boron supplement
row for row of trees y for the response variable;
proc glm;
class fa row
model y = f I a row(f) a*(row*f);
test h = f e=row(f);
test h = a f*a e = a*(row*f); *<== required because of the sub-sampling
run;

Fertilizer F 1 2 3

Row of Trees
/!~ 2 3 4
/!~
5 6 7 8 9
/!~ 10 11 12

Boron A
/\ /\ /\ /\ /\ /\. 1\ /\ /\ /\ /\ /\
121212121212121212121212

Trees
II I \ I I \ \ I I I I \ I \ I \ \ I \ I I I I
1 2 . 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 i9 20 21 22 23 24

Figure 2. Factor Relationship Diagram for a split-plot design without subsampling

This example split-plot design was derived from the completely randomized factorial
design whose experimental units or mainplots (rows of 2 trees) were split into individual
trees for the split-plot unit factor. Another common split-plot design arises when a
randomized block factorial design (RBD) is the starting point instead. In this case, the
experimental units or mainplots which are assigned at least one treatment or classification
factor are grouped into blocks. The split-plot design arises when these experimental units·
are split into smaller units. These smaller units or split-plots are then either randomly
assigned the levels of another treatment factor or their level of another classification
factor is observed.

Bibliography
Bergerud, W.A. 1996. Displaying factor relationships in experiments. The American
Statistician, 50(3): 228-233.
Milliken, G.A., and Johnson, D.E; 1984. Analysis of messy data, Volume 1: Designed
experiments Lifetime Learning Publications, Belmont, CA.

130