Reading time 3 minutes

McNemar’s test for paired proportions

• What is McNemar’s Test?
• The Null Hypothesis
• How Does the Test Work?
• Test Assumptions
• What if Assumptions Aren’t Met?
• Performing McNemar’s Test
  • Table 1 a,b: Matched Case-Control Study of Asthma Death and Short-Acting B2 Agonist Use [3] (Two Presentations)
• References

This material is reproduced based on the examples from: Oxford handbook of medical statistics (Janet Peacock Philip J Peacock) [1].

What is McNemar’s Test?

• A statistical method to assess the association between two paired proportions
• Applicable for matched case-control studies or ‘before and after’ studies
• Based on the chi-squared distribution with 1 degree of freedom
• Generates a P value, estimates, and a confidence interval

See Bland, Chapter 13 [2].

The Null Hypothesis

• The prevalence in the population remains consistent under both conditions

How Does the Test Work?

• The test focuses on discordant pairs (yes/no, no/yes) where exposure differs
• Concordant pairs (yes/yes, no/no) are disregarded since they don’t provide information on differences within pairs
• Expected frequencies are calculated assuming no association (null hypothesis is true), meaning the frequencies are equal in both discordant pairs (yes/no, no/yes)
• Observed frequencies are compared to expected values
• If observed frequencies deviate significantly from expected values, this implies a real association
• The test employs a formula based on the chi-squared distribution to calculate a P value

Test Assumptions

• Requires a large sample
• For the test to be valid, each expected frequency should be greater than 5

What if Assumptions Aren’t Met?

• The P value might be too small, resulting in potentially false significant outcomes
• If numbers are small but the rule of thumb is satisfied, apply the version of the test with a continuity correction (see Bland, Chapter 13) [2]

Performing McNemar’s Test

• Always use frequencies, not percentages, for calculations
• The test is typically carried out using a computer program – the calculations following Table 1 demonstrate how the test operates.

Table 1 a,b: Matched Case-Control Study of Asthma Death and Short-Acting B2 Agonist Use [3] (Two Presentations)

(a)

Died (case)	Survived (control)	No. of pairs	Notation
No	No	411	a
Yes	No	69	b
No	Yes	45	c
Yes	Yes	7	d

(b) Results arranged as a 2x2 table

Died (case)	Used β2 agonist	Yes	No	Total
Survived	Yes	411	45	456
(control)	No	69	7	76
Total		480	52	532

• Expected frequency = (b+c)/2 = (69+45)/2 = 57
• Test statistic is: \(\frac{\sum_{\text{discordant cells}}(O - E)^2}{E} = \frac{(69 - 57)^2 + (45 - 57)^2}{57} = 5.05\)

This follows a chi-squared distribution with 1 degree of freedom and has P=0.031, indicating a relationship between the use of short-acting $\beta$2 agonist and death from asthma.

References

1. Oxford handbook of medical statistics (Janet Peacock Philip J Peacock).
2. Bland M. An introduction to medical statistics. 3rd ed. Oxford: Oxford University Press, 2000.
3. Anderson HR, Ayres JG, Sturdy PM, Bland JM, Butland BK, Peckitt C et al. Bronchodilator treatment and deaths from asthma: case-control study. BMJ 2005; 330(7483):117.