Graphics Reference
In-Depth Information
Example - A Scatterplot Matrix with Conditioning
6.3.3
he goal of a chiropractic research project (Harrison et al.,
) was to model pain
indices constructed from patient pain questionnaires as functions of skeletal mea-
surements summarizing patient posture:
SBA:
sacral base angle
API:
angle of pelvic incidence
PTPIA:
posterior tangent pelvic incidence angle
Sex:
female, male
Group:
pain category: normal, chronic, acute
Associations between these two classes of variables could suggest chiropractic skele-
tal adjustments to address pain. We illustrate with a subset of the data set that in-
cludes three continuous skeletal measurements for both sexes and three pain cate-
gories from a sample of
subjects. Figure
.
was prepared as an initial look at
these data, not as a presentation of an interim or final analysis.
Figure
.
exemplifies what is meant by a structured set of graphs. It systemati-
cally unifies
interrelated graphs defined by the Cartesian product of several sets.
he figure consists of a
arrangement of scatterplot matrices (sploms). he two
columns are defined bysex of the patient and the three rowsbypain category. Within
each of the sploms, the nine panels are defined by the Cartesian product of the set of
three continuous variables (SBA,API, PTPIA)crossed with itself. heuppertriangle
of each splom contains the mirror image of the set of plots in the lower triangle.
Evidently these skeletal measurements donot differ bysex but dodifferaccording
to pain category. he measurements are more tightly clustered for subjects classified
aspain-free(
normal
)thanforthosehaving
acute
or
chronic
pain.Inaddition,
we see that measurements
API
and
SBA
are more highly correlated for pain subjects
than those without pain. Toease the reader'stask in seeing both the tightness and the
correlation, we collect in Fig.
.
all the
SBA ~ API
panels from Fig.
.
and also
show the marginal distributions for the Sex, Group, and Total.
Coordinating Sets of Related Graphs
6.3.4
he graphical issues that needed attention in Fig.
.
are
Positioning: the panels containing marginal displays need to be clearly delineated
as distinct from the panels containing data from just a single set of levels of the
factors. We do this by placing extra space between the set of panels for the indi-
vidual factor values and the panels containing marginal displays.
Scaling: allpanelsneedtobeonexactly thesamescaletoenhancethereader'sability
to compare the panels visually. We use the automatic scaling feature of trellis
plots toscale simultaneously both the individual panels and the marginal panels.
Labeling: we indicate the marginal panels by use of the strip labels. Each panel label
is constructed from the levels of the two factors defining it. he marginal pan-
els contain a level name from just one factor. he Total panel is named without
reference to the factor level names.
