Académique Documents
Professionnel Documents
Culture Documents
Fisher (3) has shown the advan- sible to obtain measurements on the
tages of the factorial experiment over dependent variable, so that statisti-
the classical method of "one vari- cal analysis of the outcomes is by way
able." The following gains accrue of the "analysis of variance." In
from consideration of the effects of many research areas, however, phe-
independent variables (treatments) nomena are not yet amenable to scal-
upon the dependent variable in the ing so that one has counts or fre-
context of other independent vari- quencies within given categories ra-
ables: (a) With a sample of size n, ther than measures, e.g., male versus
and k treatment classifications which female rather than degrees of sexu-
do not interact, "hidden replication" ality. There is no logical hindrance
enables estimation of all k main ef- to the use of factorial experimenta-
fects with the same precision as tion with these phenomena, and such
would be achieved for one in a single is to be recommended in light of the
factor experiment of the same size. advantages to be gained. The prob-
The economy of the factorial design lem is to find a method of statistical
is indicated by the fact that to obtain analysis appropriate to this type of
the same amount of information by data, x2 methods are available for
the "rule of one variable," one would the comparison of sampled frequen-
need k sets of n replicates, (b) If cies and for assessing association in
there is interaction among the treat- simple contingency tables. These
ments, the factorial arrangement en- cases are in effect instances of single
ables its isolation and evaluation and and double classification designs, and
thereby sets the limits of generaliza- if contingency association is the
tion. One can specify the effect of analogue of interaction in analysis of
the independent upon the dependent variance, then a method of assessing
variable in a variety of contexts; and multiple contingency is needed for
conversely, if interaction is zero one the analysis of frequency data from
may conclude that the relationship is higher order designs. Pearson (6)
constant through all contexts con- described a procedure for assessing
sidered, (c) A further virtue of the multiple contingency but failed to
factorial design lies in the informa- consider the question of additivity
tion it may provide about the rela- of x2 components. Bartlett (1) of-
tive efficacy of different combina- fered a method for the 2 X 2 X 2 case
tions of conditions for the production which involves the solution of a cubic
of given effects. Most use has been equation and is difficult to apply in
made of this in agriculture and indus- practice. Recently, Lancaster (5)
try, but it has its scientific as well as following proofs by Irwin (2) and
its technological applications, such Lancaster (4) has devised a general
as in sorting out necessary and suf- method of partitioning a total %2 and
ficient conditions. degrees of freedom into independent
In practice the factorial design has additive components due to given
most often been used where it is pos- sources of variation. This completes
134
ANALYSIS OF FREQUENCY DATA 135
the parallel with the analysis of vari- ment. The method will be developed
ance in which the total sum of squares through a notation which, while
and degrees of freedom are parti- perfectly general, will for simplicity
tioned into sums of squares and df be set out for an A XB X C design.
for all main effects, interactions and Let the classifications be symbo-
error. This paper presents for psy- lized as A, B, C, , L. Let A be
chologists a general form of multiple subdivided into a categories repre-
contingency analysis based on Lan- sented generally by the subscript i
caster's work, provides an illustra- which thus takes values from 1 to a.
tion, and comments on the generality Similarly B is represented by (j
of application of the method. = 1, , b), C by (ft = l, , c),
etc. Let pijk the probability of an
MULTIPLE CONTINGENCY observation falling in the ijkth cell;
ANALYSIS Oijk = the observed frequency in the
ijkth cell; and ,-;& = the expected
Complex contingency tables of frequency in the ijkth cell. Let a dot
frequency data from multiple classi- in place of a subscript represent sum-
fication designs may be of several mation across the values represented
forms according as the sampling of by the subscript, e.g.,
main effects is random or restricted,
(a) The random case imposes no
sampling restrictions. For example,
after a random sample has been
taken, it may be classified in various Let the total sample size o ... N; and
ways and the frequencies within finally ^.,. = 1.0. On the hypothesis1
classes will be due only (within of zero interaction, Pak pi..Xp.}.
sampling limits) to the population Xp..k, Pa.=pi..Xp.j., etc. These
proportions, (b) The mixed case in- parameters are used to find the ex-
volves restrictions upon the propor- pected frequencies, e.g., <# = *
tions within categories of given classi- X N, and hence x2 may be calculated
fications and freedom with respect to as ^(oeY/e. Now some or all of
others, e.g., arranging in advance that the values of the parameters pi...,
a total sample will involve equal p.j., p.,h may be (a) known from the
proportions of the sexes. Parameters population; or (b) estimated from
for a classification are denned by its the sample, e.g., pi.. = Oi../N. These
restriction. Whichever case is in- situations taken with the random
volved, for each observed frequency in and mixed designs provide four cases
the table there will be an expected each requiring separate consideration.
value, and hence divergence of the Case (la) will be presented in full
total table from expectation may be for the A XB X C design.
tested through x2- Within a total (la) Random sampling, known pa-
table, however, there will be a num- rameters
ber of sources of variation comprising
main effects and interactions, and to The partition of total x2 ar>d df
isolate them one would need to par- into component values for this case,
tition the total x2 and df into inde- 1
Ordinarily one works with the null hy-
pendent additive components due pothesis, but population hypotheses of non-
zero interactions maybe entertained, e.g., in a
to such sources. The problem is test of goodness of fit with case la, and again
to specify the expected frequen- in determining the power of the test of signifi-
cies which will meet this require- cance for a given situation.
136 J. P. SUTCLIPPR
TABLE 1
PARTITION OF x2 AND DF FOR CASE (la)
Number Source