Chi-Square Test for Trend (Cochran-Armitage): Ordered Categorical Exposures

Why order is information the standard test throws away

A standard chi-square test of independence is built to detect any deviation from independence, in any direction, across any pairing of categories. That generality has a cost: the test has no concept of "higher" or "lower" among the exposure categories. Relabel small, medium, and large as, say, medium, large, and small, and the Pearson chi-square statistic does not change, because the test never used the order to begin with.¹

The Cochran-Armitage trend test fixes this by assigning a numeric score to each ordered level, equally spaced by default, and testing whether the outcome proportion moves linearly across those scores.² When the real relationship is a genuine dose-response pattern, concentrating the test on that one specific pattern, instead of spreading it across every possible kind of association, gives it more power to detect it.

A real result: does body frame predict a diabetes diagnosis?

In a cross-sectional screening cohort of 403 adults from Buckingham and Louisa counties, Virginia, each participant's body frame was recorded as small, medium, or large, a natural three-level ordered exposure.⁴ Using the American Diabetes Association's diagnostic threshold (glycosylated haemoglobin ≥ 6.5%), 379 participants have both frame and diabetes status recorded.

Example data: Vanderbilt Diabetes Study, Buckingham and Louisa counties, Virginia (Willems, Saunders, Hunt, and Schorling, 1997, courtesy of the Vanderbilt University Department of Biostatistics, hbiostat.org/data); see references. Computed with scipy, not generated.

The proof: shuffle the labels and the standard chi-square doesn't move

Run a standard, order-blind chi-square test of independence on this same table and it returns χ²(2) = 11.22, P = .004. Reorder the three rows into any other sequence, large first, then small, then medium, and recompute: the statistic is exactly χ²(2) = 11.22 again, to every decimal place. The test cannot tell the difference between a clean dose-response trend and a scrambled one, because nothing in its arithmetic ever looks at row order.

The trend test does not have that property. Its whole calculation is built from the scores assigned to each ordered level; scramble the order without also scrambling the scores and the linear contrast breaks. That dependence on order is not a limitation of the trend test. It is the entire reason it exists: to capture the one piece of information, "this goes up as that goes up", that the standard test is mathematically incapable of seeing.

This isn't yet a one-click preset

StatsPlease's Category Patterns preset runs the standard chi-square test of independence on two uploaded categorical variables today, the same test used for the shuffle demonstration above. The Cochran-Armitage trend statistic shown in the card above was computed directly from this repository's own implementation to illustrate the concept accurately; it is not yet exposed as its own preset button in the product. If your analysis specifically calls for a trend test on an ordered exposure, note that in your methods section alongside how you computed it.

Writing the AMA sentence

Name which test you actually ran and why the exposure's order mattered to that choice. "A chi-square test of independence was used" invites the question of whether the exposure was ordered and whether that order was worth testing directly. "A Cochran-Armitage trend test was used because body frame is an ordered exposure with an expected dose-response relationship to diabetes risk" answers the question before a reviewer has to ask it.

Frequently asked questions

What does a chi-square test for trend do that a regular chi-square doesn't?

A regular chi-square test of independence treats category labels as unordered; it would return the identical statistic if the categories were shuffled. The Cochran-Armitage trend test assigns an ordered score to each exposure level and specifically tests whether the outcome proportion rises or falls monotonically across that order, concentrating power into the dose-response pattern clinicians usually expect.

When should I use a trend test instead of a standard chi-square?

When the exposure has a genuine order, such as disease stage, dose tier, or a graded risk category, and you expect the outcome to change monotonically across that order. If the exposure is unordered (nominal), or you expect a non-monotonic pattern, a standard chi-square is the correct choice instead.

Does StatsPlease compute the Cochran-Armitage trend test automatically?

Not yet as a dedicated one-click preset. The Category Patterns preset computes the standard chi-square test of independence on your uploaded categorical variables today; the trend-specific statistic shown in this post was computed directly from the same public dataset to illustrate the concept.