McNemar Tests of Marginal Homogeneity


Back to Agreement Statistics page


The McNemar test

The McNemar test (McNemar, 1947; Sheskin, 2000, pp. 491-508; Somes, 1983) is an extremely simple way to test marginal homogeneity in K×K tables.

The basic McNemar test applies to 2×2 tables. Consider such a table that summarizes agreement between two raters on a dichotomous trait:

Table 1. Summary of binary ratings by Rater 1 (rows) and Rater 2 (columns)
  - +  
- a b a + b
+ c d c + d
  a + c b + d total
Marginal homogeneity implies that row totals are equal to the corresponding column totals, or
(a + b) = (a + c)
(c + d) = (b + d)

Since the a and the d on both sides of the equations cancel, this implies b = c; this is the basis of the McNemar test.

The McNemar statistic is calculated as

X2 = (b - c)2/(b + c).     (1)

The value X2 can be viewed as a chi-squared statistic with 1 df.

Some authors recommend a version of the McNemar test with a correction for discontinuity, calculated as:

X2 = (|b - c| - 1)2/(b + c).     (2)

but this is controversial.

Statistical significance is determined by evaluating the probability of X2 with reference to a table of cumulative probabilities of the chi-squared distribution or a comparable computer function. A significant result implies that marginal frequencies (or proportions) are not homogeneous. The test is inherently two-tailed. For a one-tailed test, one could divide the obtained p value by two.

When b and/or c are small, the McNemar test X2 is not well approximated by the chi-squared distribution. When, say, (b + c) < 10 a two-tailed exact test, based on the cumulative binomial distribution with p = q = .5, can be used instead.

Example

Let the cells of a 2×2 table be as follows:

Table 2. Example data
40 10
20 50

By Eq. 1, the McNemar test X2 = (10 - 20)2/(10 + 20) = 100/30 = 3.33 (1 df, p = .068).

Using the continuity correction (Eq. 2), X2 = 2.70 (1 df, p = .100).

With the exact test, p = 0.099.
 

Test of marginal homogeneity for a single category

Given ratings on a K-level categorical variable, agreement between two raters is summarized by a K×K crossclassification table. Table 3 below is an example with three rating categories of 1 = low, 2 = moderate, and 3 = high.

Table 3. Summarization of ratings by Rater 1 (rows) and Rater 2 (columns)
  low   mod. high  row total
low n11 n12 n13 n1.
moderate n21 n22 n23 n2.
high n31 n32 n33 n3.
column total n.1 n.2 n.3 n..
with, nij being the number of cases assigned category i by Rater 1 and category j by Rater 2.

To test marginal homogeneity for a single category, one collapses the full table into a 2×2 table. Specifically, to test row/column marginal homogeneity for category k, one collapses all rows and columns corresponding to the other categories. For example, to test marginal homogeneity for the category "low," one would collapse the table above to produce:

Table 4. Table 3 collapsed to test row/column homogeneity for the "low" category




Rater 1
             Rater 2
    low moderate
or high
  low n11 n12+n13
  moderate
or high
n21+
n31
n22+n23+
n32+n33

and then apply the basic McNemar test to this table. The test has 1 df. A significant X2 value would imply that the Rater 1 and Rater 2 marginals for this category differ.

Similarly, to test the raters' marginal rates for the "moderate" category, one would collapse rows/columns 1 and 3 to produce the 2×2 table:

Table 5. Table 3 collapsed to test row/column homogeneity for the "moderate" category
  moderate low or high
moderate n22 n21+n23
low or high n12+
n32
n11+n13+
n31+n33

and perform the basic McNemar test on this table.

In this way marginal homogeneity with respect to each category can be tested. Because there are multiple tests, one may wish to adjust the overall alpha. For example, a simple Bonferroni adjustment can be applied. With K categories, there are K - 1 independent tests. For an "experiment-wise" alpha of .05, the Bonferroni method would make .05/(K - 1) the significance criterion for each test.
 

Stuart-Maxwell test

Whereas the method above tests row/column homogeneity with respect to each individual category, the Stuart-Maxwell test (Stuart, 1955; Maxwell, 1970; Everitt, 1977) tests marginal homogeneity for all categories simultaneously.

The test is calculated in the following way. Consider a K × K frequency table of the same form as Table 3. Let column vector d contain any K - 1 of the values,

d1, d2, ..., dK

where

di = ni. - n.i  (i = 1, ..., K)

Let S denote the (K - 1) × (K - 1) matrix of the variances and covariances of the elements of d. The elements of S are equal to:

sii = ni. + n.i - 2nii
sij = -(nij + nji)

The Stuart-Maxwell statistic is calculated as:

X2 = dS-1 d.

where d' is the transpose of d and matrix S-1 is the inverse of S

X2 is interpreted as a chi-squared value with df equal to K - 1. In the case of K = 2, the Stuart-Maxwell statistic and the McNemar statistic (Eq. 1) are identically equal.

If there is perfect agreement for any category k, that category must be omitted in order to invert matrix S. (Note that if there is perfect agreement on a category, the corresponding row and column marginal frequencies are equal.) Such categories should be ignored in calculations and the Stuart-Maxwell test performed with respect to the remaining categories. The df in this case can still be considered K - 1, where K is the number of original categories; this treats omitted categories as if they were included but contributed 0 to the value of X2--a reasonable view since such categories have equal row and column marginals.

Example

Consider the hypothetical data in Table 6.

Table 6. Hypothetical summary of ratings by Rater 1 (rows) and Rater 2 (columns)
  low mod. high row total
low 20 10 5 35
moderate 3 30 15 48
high 0 5 40 45
column total 23 45 60 128

We first calculate any K - 1 of the (row sum - column sum) differences; we arbitrarily choose those for rows/columns 1 and 2. This produces d as follows:

     12
      3

The corresponding variance/covariance matrix S is:

      18      -13
     -13       33

The inverse, S-1, is:

     0.0776   0.0306
     0.0306   0.0424

The value of d' S-1 d = X2 = 13.76. With 2 df, p = 0.001.
 

Bhapkar test

The Bhapkar (1966) test is a more powerful alternative to the Stuart-Maxwell test. It is calculated in a way similar to the Stuart-Maxwell test (see above), but the formulas for the elements of S are different. See Agresti (2002) for details.

The Bhapkar and Stuart-Maxwell tests are asymptotically equivalent (Keefe, 1982). With a large N, both will produce the same chi-squared value. As the Bhapkar test is more powerful, it is preferred.

The MH program, free software available on this website, will perform both the Bhapkar test and the Stuart-Maxwell test.
 

Test of equal category thresholds

The Concept of Rater Thresholds

With ordered-category ratings, it is often theoretically reasonable and intuitively appealing to consider the idea of rater thresholds. By this view, raters begin with a subjective continuous impression of how much trait a case has. Then they apply subjective thresholds or cutpoints which map that impression into a particular rating category. For example, if the trait is "mobility," a rater first perceives a given patient's level as falling somewhere on a continuum. The rater then applies thresholds to assign a specific rating category of, say, low, moderate, or high, as illustrated below.


   low      moderate        high
<--------|------------|---------------->
         t2           t3

Actual Trait Level (continuous)

In the example above, a case whose judged trait level is below threshold t2 would be assigned the rating category "low." A case whose judged trait level is above threshold t3 would be assigned the rating category "high." A case whose judged trait level is between the two thresholds would be assigned the rating category "moderate."

Threshold tk (k = 2, ..., K) is the minimum trait level a case must display to be assigned rating level k or higher. There is no threshold t1; a case is assigned rating level 1 if the case's trait level does not exceed threshold t2.

Threshold locations potentially differ between raters. The locations of a rater's thresholds determine how often the rater uses each rating category. For example in the situation below,


<--------|------------|------------>   Rater 1
         t2          t3

<---------------|-----|------------> Rater 2 t2 t3
Rater 2 has a higher threshold t2. This corresponds to a wider definition of the lowest rating category and a narrower definition of the middle rating category. Rater 2, then, would tend to use the lowest rating category more often, and the middle category less often, than Rater 1.

We now return to the 3×3 crossclassification in Table 3. Suppose one wishes to test whether the lowest threshold (t2) is the same for both raters. To do this one would first collapse all rows after Row 1 and all columns after Column 1. Then one would perform the McNemar test on the resulting 2×2 table. A significant result would imply that threshold t2 differs between the two raters. (Note that here the 2×2 table and associated McNemar test is the same as with Table 4.)

To test equality of threshold t3 between raters, one would collapse Rows 1 and 2, and Columns 1 and 2 to produce the following 2×2 table.

Table 7. Table 3 collapsed to test equality of row and column thresholds for "high" rating level
  low or
moderate
high
low or
moderate
n11+n12+
n21+n22
n13+
n23
high n31+n32 n33
and perform a McNemar test on this table.

In general, with a K × K table, one can test equality of a given threshold k (k = 2, ..., K) by collapsing rows/columns 1 to k-1 and collapsing rows/columns k to K, and performing the basic McNemar test on the resulting 2×2 table.

The tests for thresholds t2 and tK are identical to the tests of marginal homogeneity for categories 1 and K (although the results are interpreted differently). However, the tests for thresholds t3, ..., tK-1 are unique.
 

Test of overall bias

With ordered-category ratings, the McNemar test can also be used to assess overall bias of raters--defined as a tendency of one rater to make ratings generally higher or lower than the other rater.

This simple test is described by Bishop, Fienberg and Holland (1975; pp. 284-285). For a K × K table, let b = the sum of frequencies in cells above the main diagonal, and let c = the sum of frequencies in cells below the main diagonal. For example, with reference to Table 3,

b = n12 + n13 + n23
c = n21 + n31 + n32
One then uses these values of b and c in Eq. 1. The test has 1 df. A significant X2 value implies that one raters' ratings are generally higher or lower than those of the other rater.
 

Software

The MH program will perform all the tests described on this page for a K × K crossclassification table, where K can be as large as 50.

SAS will perform a McNemar test for 2×2 tables. It is possible SPSS has similar features. Other specialized biostatistics and epidemiological software, such as Epistat, perform the McNemar test. For additional suggestions, one might search the web using the key words "McNemar test" and "software".
 

References


Go to Agreement Statistics
Go to Latent Structure Analysis
Go to My papers and programs page

Last updated: 30 August 2006 (Bhapkar test)


(c) 2006 - 2009 John Uebersax PhD    email