Use the McNemar Test for Paired Yes/No Data

Ask the right paired question

McNemar applies to paired binary outcomes: the same people answer Yes or No before and after an event, for instance. Each person's two observations are linked, while different people should be independent. Two unrelated groups need a different method. The test asks whether No-to-Yes and Yes-to-No changes occur at similar rates.

Example: Sixty students report whether they use a library database before and after a workshop. The people whose response changes are central. Those answering Yes twice or No twice describe stability, but do not form the contrast between the two change directions.

Build a correctly labelled table

Put the before response in rows and the after response in columns, or vice versa, and label both clearly. Verify that every person appears once and an absent follow-up answer was not silently coded No. The four cells must sum to the number of complete pairs.

Report percentages at each time point and both discordant counts before testing. Similar overall proportions can hide many individual switches in opposite directions. Conversely, a modest marginal shift may involve a meaningful subgroup. Paired data contain information that separate frequency tables lose.

Check sparse counts and interpretation

The common asymptotic approximation may be unreliable with very few discordant pairs. Look for an appropriate exact procedure in your software and state which version you used. Continuity corrections can also change the statistic, so a p-value without its method is hard to assess.

Describe direction and size using the observed changes and, where useful, an uncertainty interval. A significant result indicates asymmetric switching, not necessarily that the workshop caused it. In an uncontrolled before–after design, outside events may also influence responses.

Write a result readers can check

Give the number of complete pairs, both change counts, the selected test variant and its result. Show or clearly describe the 2×2 table. Saying “18 switched to Yes and six switched to No” before the inference lets readers verify the claimed direction.

Consider measurement consistency too. Were the question and response choices identical at both times? If “use” meant something different in the follow-up, the test may compare different constructs. Record that limit in methods rather than discovering it only after a surprising p-value.

Frequently asked questions