Paired Proportion Plot Explained: How to Make, Read and Report It
A paired proportion plot shows how the proportion of "yes" outcomes changes between two measurements on the same units, for example the proportion passing a screening before and after training. The matching statistical test for paired binary data is McNemar's test.
Why pairing matters
If the same 60 people are classified twice, the two proportions are dependent. An ordinary test for two independent proportions ignores this and gives wrong standard errors. Only the discordant pairs, those who changed, carry information about the change.
The 2 × 2 paired table
| After: Pass | After: Fail | |
|---|---|---|
| Before: Pass | a | b |
| Before: Fail | c | d |
pbefore = (a + b)/n pafter = (a + c)/n Δ = (c − b)/n χ² = (b − c)² / (b + c)
Original example (illustrative)
Sixty students take a screening quiz before and after a revision course: a = 30, b = 5 (pass to fail), c = 15 (fail to pass), d = 10. The proportion passing rises from 35/60 = .583 to 45/60 = .750, so Δ = 10/60 = .167. McNemar's statistic is (5 − 15)²/20 = 5.00, p = .025 (asymptotic); with continuity correction, (|5 − 15| − 1)²/20 = 4.05, p = .044. The exact test treats b as Binomial(20, .5): p = 2 × .0207 = .041. An approximate 95% CI for Δ is .167 ± 1.96 × √(20 − 100/60)/60 = .167 ± .140, i.e. [.03, .31]. The conditional odds ratio is c/b = 3.0.
The interval is wide, so the data support an improvement but leave its size uncertain. Without a control group, the change cannot be attributed to the course.
How to make it in StatFAS
Open the Paired Proportion Plot Generator. StatFAS has an existing McNemar's test post; link it after verifying the URL. For an exact test on b and c alone, see the binomial test calculator.
Reporting
The proportion passing increased from .58 to .75 (Δ = .17, 95% CI [.03, .31]); an exact McNemar test indicated this change was significant, p = .041, n = 60.
Common mistakes and alternatives
Using a chi-square test of independence on paired data, using the asymptotic test when b + c is small (below about 25, prefer the exact test), reporting only p, and ignoring that the plot cannot show who switched in which direction unless lines are drawn per pair. For three or more paired binary measurements use Cochran's Q.
Conclusion
Show paired proportions with their uncertainty, test the discordant pairs, and keep the design (paired, uncontrolled) in mind when interpreting.
1.
statfas-paired-proportion-plot-interface.png, alt: "StatFAS Paired Proportion Plot Generator interface"2.
statfas-paired-proportion-plot-example-data.png, alt: "Example data entered in the StatFAS Paired Proportion Plot Generator"3.
statfas-paired-proportion-plot-settings.png, alt: "Paired Proportion Plot Generator settings in StatFAS"4.
statfas-paired-proportion-plot-result.png, alt: "Paired Proportion Plot Generator result generated in StatFAS"References
- McNemar, Q. (1947). Note on the sampling error of the difference between correlated proportions or percentages. Psychometrika, 12(2), 153–157.