Newsom

Psy 525/625 Categorical Data Analysis, Spring 2021 1

Three-Way Contingency Tables

Three-way contingency tables involve three binary or categorical variables. I will stick mostly to the
binary case to keep things simple, but we can have three-way tables with any number of categories with
each variable. Of course, higher dimensions are also possible, but they are uncommon in practice and
there are few commonly available statistical tests for them.

Notation
We described two-way contingency tables earlier as having I rows and J columns, forming an I × J
matrix. If a third variable is included, the three-way contingency table is described as I × J × K, where I
is the number of rows, J is the number of columns, and K is the number of superordinate columns each
level containing a 2 × 2 matrix. The indexes i, j, and k are used to enumerate the rows or columns for
each dimension. Each within-cell count or joint proportion has three subscripts now, nijk and pijk. There
are two general types of marginal counts (or proportions), with three specific marginals for each type.
The first type of marginal count combines one of the three dimensions. Using the + notation from two-
way tables, the marginals combining counts across either I, J, or K dimensions are n+jk, ni+k, or nij+,
respectively. The second general type of marginal collapses two dimensions and is given as ni++, n+j+, or
n++k. The total count (sample size) is n+++ (or sometimes just n). The simplest case is a 2 × 2 × 2 shown
below to illustrate the notation.

k=1 k=2
n111 n121 n1+1 n112 n122 n1+2
n211 n221 n2+1 n212 n222 n2+2
n+11 n+21 n++1 n+12 n+22 n++2
Note: marginals collapsing across levels of K are not shown.

In the Azen and Walker text as well as elsewhere, the three variables are referred to by letter names X,
Y, and Z. 1 Although we do not need to assign explanatory (independent) and response (dependent)
roles to the variables for all hypotheses that might be tested, it can be easier to put the table into more
familiar terms by imagining one response variable Y (with its J number of levels) and two explanatory
variables, X and Z (with I and K number of levels, respectively). Thinking in this way, we now have the
conceptual equivalent of a ANOVA design, where Y is the (binary) dependent variable and X and Z are
two independent variables. One common hypothesis of interest is to examine the X effect on Y at
different levels of Z, which is the same question asked in the test of the two-way factorial ANOVA (e.g., is
the X group difference on Y the same for Z1 and Z2?). It is important to note that this is the real analogy to
ANOVA for at least one possible or common hypothesis that can be tested with three-way contingency
table. Azen and Walker use the three-way ANOVA as an analogy throughout the chapter for a different
reason. Their intention is to use the three-way ANOVA analogy for the three-way contingency table to
describe the process by which the researcher should proceed in analysis steps (e.g., examining the
three-way table first, and then depending on the results pursuing two-dimensional or one-dimensional
comparisons). So, be careful not to confuse their process analogy to ANOVA with the statistical test
analogy to ANOVA.

Partial Tables
There are two 2 × 2 tables above that make up the three-way table—an X × Y table within Z1 and an X ×
Y table within Z2. These are known as partial tables. Although partial tables can be constructed for levels
of the X or levels the Y variable also, it is most common to set up three-way tables in this manner and
discuss the X × Y partial tables within levels (or strata) of Z. Marginals for the two-way partial tables can
be expressed as conditional proportions, similar to the simple conditional proportion in the two-way case.
The notation pij|k denotes one type of conditional proportion, which is the within-cell count relative to the

1
Many authors use A, B, and C, but this is easily confused with the fourway table notation.
 Newsom
Psy 525/625 Categorical Data Analysis, Spring 2021 2

number of cases within one level of Z, pij|k = nijk/n++k. Of course, we could also compute other conditional
proportions, such as pi|jk or pjk|i and so on.

If the two I × J contingency tables are considered separately, we could compute odds ratios as before,
adding to the notation to indicate that the odds ratio is specific to one level of the Z variable. Following
the notation in your text, we use X, Y, and Z and the Greek letter θ (“theta”) for the odds ratio.

= X 1|=
YZ1 X 1|YZ1 odds n /n n n p
XY | Z1
211 221
θ
=111 221
= = =
= X 0|=
YZ1 111 121 211 121 X 0|YZ1 odds n /n n n p

If we were to calculate these odds ratios for each level of Z and then average them taking into account
the different number of cases in each stratum of Z, we get a common odds ratio called the Mantel-
Haenszel (MH) odds ratio.

∑ (n n22 k / n++ k )
K
11k
θ MH = k =1

∑ (n n21k / n++ k )
K
k =1 12 k

The Mantel-Haenszel is the common odds ratio and assumes that the odds ratio does not differ across
the levels of Z, so it may not be of interest itself in many circumstances. It can be used, however, if the
test of whether the X and Y are associated overall by assessing the weighted average of this relationship
over the levels of Z is of interest and the odds do not differ substantially across the levels.

Breslow-Day Test
Perhaps the most useful test is a test of the hypothesis that the X-Y association differs across levels of Z,
known as the Brelow-Day test (Breslow & Day, 1980). Later modified by Tarone (1985), it is sometimes
referred to as the Breslow-Day-Tarone test. If there is no difference across levels of Z, the X-Y
relationship is homogeneous. Another way to state this hypothesis test is as an evaluation of whether
the partial table odds ratios are equal across the levels of Z, that θ XY |Z = θ XY |Z . 1 2

2
K OY − E (Y1k | θ MH ) 
χ 2
= ∑  1k
V (Y1k | θ MH )
BD
k =1

In the numerator, we have an alteration of the usual chi-squared statistics in which of the expected
values are subtracted from the observed values, except that the observed and the expected values are
for partial tables for each level of Z (indexed by subscript k for each stratum of Z), with the expected
values taking into account the Mantel-Haenszel odds ratio, θMH. The expected values, E (Y1k | θ MH ) , and
variance estimate, V (Y1k | θ MH ) , in the denominator are considerably more complicated than for the two-
way contingency table. 2 The Breslow-Day test is only for 2× 2× k tables, so X and Y must be binary. The
χ BD
2
can be evaluated against the chi-squared distribution with df = K -1 if the Tarone modification is
used. Significance implies θ XY |Z ≠ θ XY |Z but also that θ XZ |Y ≠ θ XZ |Y and θYZ | X ≠ θYZ | X for the 2× 2× 2 case. If
1 2 1 2 1 2

there are many strata for Z (i.e., K is large) and data are sparse, then an exponential weighted average
of the common odds ratio (Woolf, 1955) is more efficient.

2
Since you are not afraid to look, here is the computations for these two quantities (Hosmer & Lemeshow, 2000):
 1  
{ } {θ } − {4 (θ − 1) θ MH ( n11k + n12 k )( n11k + n21k )} 
2
E (Y1k | θ MH ) 
=   θ MH ( n11k + n12 k ) + ( n11k + n21k )  + ( n21k + n22 k ) − ( n11k + n21k )  ± ( n11k + n12 k ) + ( n11k + n21k )  + ( n21k + n22 k ) − ( n11k + n21k ) 
 2 (θ MH − 1)  
MH MH

−1
 1 1 1 1 
and V  E (Y1k | θ MH ) = + + +
   E Y |θ 
 ( 1k MH ) ( n11k + n12k ) − E (Y1k | θ MH ) ( n11k + n21k ) − E (Y1k | θ MH ) ( n21k + n22 k ) − ( n11k + n21k ) + E (Y1k | θ MH ) 
Newsom
Psy 525/625 Categorical Data Analysis, Spring 2021 3

I suspect that most social scientist are interested in testing the hypothesis for which the Breslow-Day test
is designed, that the odds ratio at each level of Z is equal. If this is the question of interest, here is an
overview of the process you might use:

sig Interpret, test cell differences

with levels of Z if desired
Test partial tables
sig within levels of Z ns Test marginals

χ 2
BD
sig Interpret, test cell differences
ns Collapse across levels (collapsing across level of Z) if desired
of Z and test 2 x 2
(i.e., test θMH) ns Test marginals

Cochran-Mantel-Haenszel
A special case of the homogeneity of association is the Cochran-Mantel-Haenszel (CMH) test that the
odds ratios for all partial tables are equal to 1, which is referred to as a test of conditional independence.
For the 2× 2× 2 case, it is a test of the null hypothesis that θ=
XY | Z θ=
XY | Z 1 , but also whether θ=
XZ |Y θ=
XZ |Y 1 1 2 1 2

and θ=
YZ | X
1
θ=
YZ | X 1 . If the test is significant, it may be due to any one of the partial tables with an odds
2

ratio above or below 1.0. The computation of the CMH follows the general Pearson chi-squared equation
and can be stated as a likelihood ratio, G2, as well. The computation of the expected values is
considerably simpler than for the Breslow-Day test.

(O − Eijk )
2
I J K I J K  Eijk 
= ∑∑∑ = −2∑∑∑ Oijk ln 
ijk
χ 2
or GCMH
2

CMH
Eijk 
=i 1 =j 1 =
k 1 =i 1 =j 1 =
k 1  Oijk 

ni ++ n+ j + n++ k
where Eijk = 2
n+++

Both CMH tests are evaluated against the chi-squared distribution with df = IJK – I – J – K + 2.

Logistic Regression
Logistic regression can be used to test the same hypotheses as those investigated with the Mantel-
Haenszel odds ratio and the Breslow-Day tests, and the logistic regression approach is probably the
more commonly employed. As you may have guessed, the corresponding hypothesis in logistic
regression is investigated by testing the X*Z interaction in predicting Y. Such tests are not exactly equal
even though they address the same hypothesis (see Hosmer & Lemeshow, 2000, for a discussion). The
association tests discussed above do not require assigning explanatory (independent) and response
(dependent) roles to the variables per se. Logistic regression models assume that one of the three
variables is a response (usually Y in our examples above) and the other two (X and Z in our examples
above) are explanatory. The maximum likelihood estimation as well as the more general framework of
the logistic regression modeling set the logistic approach apart somewhat from the Mantel-Haenszel
odds ratio and Breslow-Day tests. We will return to how to test these hypotheses with logistic regression
in a later section of the course.

Other Three-Way Table Hypotheses

The Breslow-Day and the Cochran-Mantel-Haenszel are the most commonly employed hypothesis tests,
but, with a three-way contingency table, there are many possibilities. With the properly derived expected
frequencies, a variety of tests can investigate whether variables X, Y, and Z are independent from one
another. Wickens (1989) discusses association tests can be formulated with loglinear models. Using
bracket notation, different tests can be distinguished by which variables are independent from one
another. For example, [X][Y][X] indicates that each of the three variables are independent from one
Newsom
Psy 525/625 Categorical Data Analysis, Spring 2021 4

another, whereas [XY][Z] indicates that Z is unrelated to X and Y. A nodal diagram can also be used to
depict these relations, with the diagram below reflecting the independence implied by [XY][Z].
Z

X Y

Loglinear modeling is a general and very flexible framework which will be introduced in the next section
of the course.

Software Examples
To illustrate statistical tests of three-way contingency tables, I used data from the 2020 Quinnipiac
presidential poll conducted in Ohio and Georgia. 3 The Breslow-Day test examines whether
independents were more likely to favor Biden or Trump, comparing this association across the two
states.

SPSS
get file='c:\jason\spsswin\cdaclass\quinnipiac 2020 3way.sav'.

output comment text="OH".

temporary.
select if state eq 0.
crosstabs /tables=ind by response
/statistics=none
/cells=count row.

output comment text="GA".

temporary.
select if state eq 1.
crosstabs /tables=ind by response
/statistics=none
/cells=count row.

crosstabs /tables=ind by response by state

/statistics=cmh
/cells=none.

OH

GA

3
Data source: https://poll.qu.edu/georgia/release-detail?ReleaseID=3679. Note that the data extrapolated cell sample sizes and used some
rounding, so the results should be taken as only approximate.
Newsom
Psy 525/625 Categorical Data Analysis, Spring 2021 5

R
> library(haven)
> d = read_sav("c:/jason/spsswin/cdaclass/quinnipiac 2020 3way.sav")
>
> library(lessR)
> #if need to convert data types in order to compute a correlation in R--lessR function
> #d <-Transform(response = as.numeric(response))
>
> d1 <- Subset(state=0)

>
> #this lessR BarChart function produces a chi-square test by default
> BarChart (response, by=ind, horiz = FALSE, stat = "proportion", beside = TRUE, xlab="Response: Trump=0,
Biden=1", data=d1)

Joint and Marginal Frequencies

------------------------------

response
ind 0 1 Sum
0 631 775 1406
1 235 319 554
Sum 866 1094 1960

Cramer's V (phi): 0.022

Chi-square Test: Chisq = 0.975, df = 1, p-value = 0.323

Cell Proportions within Each Column

Newsom
Psy 525/625 Categorical Data Analysis, Spring 2021 6

-----------------------------------

response
ind 0 1
0 0.729 0.708
1 0.271 0.292
Sum 1.000 1.000

> d2 <- Subset(state=1)

> #this lessR BarChart function prosduces a chi-square test by default

> BarChart (response, by=ind, horiz = FALSE, stat = "proportion", beside = TRUE, xlab="Response: Trump=0,
Biden=1", data=d1)

Joint and Marginal Frequencies

------------------------------

response
ind 0 1 Sum
0 631 775 1406
1 235 319 554
Sum 866 1094 1960

Cramer's V (phi): 0.022

Chi-square Test: Chisq = 0.975, df = 1, p-value = 0.323

Cell Proportions within Each Column

-----------------------------------

response
ind 0 1
0 0.729 0.708
1 0.271 0.292
Sum 1.000 1.000

> counts <-array(

+ c(308,401,132,191,323,374,103,128),
+ dim=c(2, 2, 2),
+ dimnames=list(state=c("OH", "GA"),
+ ind =c("party aff", "independent"),
+ response =c("Biden", "trump"))
+ )

> library(DescTools)
> BreslowDayTest(counts,correct=FALSE)

Breslow-Day test on Homogeneity of Odds Ratios

data: counts
X-squared = 0.02913, df = 1, p-value = 0.8645

> mantelhaen.test(counts)

Mantel-Haenszel chi-squared test with continuity correction

data: counts
Mantel-Haenszel X-squared = 0.69909, df = 1, p-value = 0.4031
alternative hypothesis: true common odds ratio is not equal to 1
95 percent confidence interval:
0.896715 1.335499
sample estimates:
common odds ratio
1.094332
Newsom
Psy 525/625 Categorical Data Analysis, Spring 2021 7

SAS
proc import datafile="c:\jason\spsswin\cdaclass\quinnipiac 2020 3way.sav" out=one
dbms = sav replace;
run;

proc freq data=one;

tables state*ind*response /cmh ;
title 'Test of 3-Way Table';
run;
The FREQ Procedure

Table 1 of ind by response

Controlling for state=OH

ind(ind) response(response)

Frequency ‚
Percent ‚
Row Pct ‚
Col Pct ‚Trump ‚Biden ‚ Total
ƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒˆƒƒƒƒƒƒƒƒˆƒƒƒƒƒƒƒƒˆ
party affiliate ‚ 308 ‚ 401 ‚ 709
‚ 29.84 ‚ 38.86 ‚ 68.70
‚ 43.44 ‚ 56.56 ‚
‚ 70.00 ‚ 67.74 ‚
ƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒˆƒƒƒƒƒƒƒƒˆƒƒƒƒƒƒƒƒˆ
independent ‚ 132 ‚ 191 ‚ 323
‚ 12.79 ‚ 18.51 ‚ 31.30
‚ 40.87 ‚ 59.13 ‚
‚ 30.00 ‚ 32.26 ‚
ƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒˆƒƒƒƒƒƒƒƒˆƒƒƒƒƒƒƒƒˆ
Total 440 592 1032
42.64 57.36 100.00

Table 2 of ind by response

Controlling for state=GA

ind(ind) response(response)

Frequency ‚
Percent ‚
Row Pct ‚
Col Pct ‚Trump ‚Biden ‚ Total
ƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒˆƒƒƒƒƒƒƒƒˆƒƒƒƒƒƒƒƒˆ
party affiliate ‚ 323 ‚ 374 ‚ 697
‚ 34.81 ‚ 40.30 ‚ 75.11
‚ 46.34 ‚ 53.66 ‚
‚ 75.82 ‚ 74.50 ‚
ƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒˆƒƒƒƒƒƒƒƒˆƒƒƒƒƒƒƒƒˆ
independent ‚ 103 ‚ 128 ‚ 231
‚ 11.10 ‚ 13.79 ‚ 24.89
‚ 44.59 ‚ 55.41 ‚
‚ 24.18 ‚ 25.50 ‚
ƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒˆƒƒƒƒƒƒƒƒˆƒƒƒƒƒƒƒƒˆ
Total 426 502 928
45.91 54.09 100.00
Newsom
Psy 525/625 Categorical Data Analysis, Spring 2021 8

Test of 3-Way Table

13:38 Wednesday, April 7, 2021 4

The FREQ Procedure

Summary Statistics for ind by response

Controlling for state

Cochran-Mantel-Haenszel Statistics (Based on Table Scores)

Statistic Alternative Hypothesis DF Value Prob

ƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒ
1 Nonzero Correlation 1 0.7864 0.3752
2 Row Mean Scores Differ 1 0.7864 0.3752
3 General Association 1 0.7864 0.3752

Common Odds Ratio and Relative Risks

Statistic Method Value 95% Confidence Limits

ƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒ
Odds Ratio Mantel-Haenszel 1.0943 0.8967 1.3355
Logit 1.0943 0.8967 1.3355

Relative Risk (Column 1) Mantel-Haenszel 1.0521 0.9395 1.1781

Logit 1.0517 0.9393 1.1777

Relative Risk (Column 2) Mantel-Haenszel 0.9615 0.8824 1.0477

Logit 0.9613 0.8823 1.0474

Breslow-Day Test for

Homogeneity of Odds Ratios
ƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒƒ
Chi-Square 0.0291
DF 1
Pr > ChiSq 0.8645

Total Sample Size = 1960

Sample Write-Up
A Breslow-Day test was used to investigate whether the association between identification as an
independent (identified with a major party vs. independent) and candidate preference (Biden vs. Trump)
was the same in Ohio and Georgia. Among Ohioans, 56.56% of respondents identifying with a major
party favored Biden, whereas 59.13% of respondents identifying as independent favored Biden. Among
Georgians, 53.66% of respondents identifying with a major party favored Biden, whereas 55.41% of
respondents identifying with a major party favored Biden. The Breslow-Day test indicated that the
association between party identification and the candidate preference did not differ significantly between
Ohio and Georgia, χ BD2
(1) = .029 , p = .86. The common odds ratio indicated no overall difference in
candidate preference by party identification, ORMH = 1.09, 95% CI = .90, 1.34. Follow-up analyses [not
shown above] indicated that there was no association between independent identification and the
candidate favored within state, Pearson χ2(1) = .60, p = .44, for Ohio, and Pearson χ2(1) = .21, p = .64, for
Georgia. 4

4
It might be desirable to go one step further and test the marginals to determine whether Biden or Trump was favored overall.
Newsom
Psy 525/625 Categorical Data Analysis, Spring 2021 9

References and Further Readings

Breslow, N. E., & Day, N. E. (1980). Statistical methods in cancer research. Vol. 1. The analysis of case-control studies (Vol. 1, No. 32).
Distributed for IARC by WHO, Geneva, Switzerland.
Tarone, R. E. (1985). On heterogeneity tests based on efficient scores. Biometrika, 72(1), 91-95.
Wickens, T. D. (1989). Multiway contingency tables analysis for the social sciences. New York: Erlbaum.
Woolf, B. (1955). On estimating the relation between blood group and disease. Annals of Human Genetics, 19(4), 251-253.