Ordinal Trend Tests for Contingency Tables (Linear-by-Linear and Cochran-Armitage)

Description

One-degree-of-freedom tests and a matching model for whether an ordered outcome rises or falls across ordered groups, far more powerful than a nominal chi-square.

A chi-square test of independence ignores category order and spreads its power across many degrees of freedom. When rows, columns or both are ordinal, scores are assigned to categories and the correlation r between row and column scores is tested: M squared = (n - 1) r squared, referred to chi-square on 1 df. With a binary column this becomes the Cochran-Armitage test for a linear trend in a proportion across ordered groups; with a binary row it compares mean scores, and midrank scores give rank tests such as Wilcoxon-Mann-Whitney. The linear-by-linear association loglinear model gives the same test plus an estimated local odds ratio. Scores should reflect plausible distances, and because conclusions can shift with unbalanced data, a sensitivity check with two or three score sets is advised.

Attributes

Loss type
Unmigrated
Effort
Hours

Steps

  1. Assign monotone scores to row and column categories (midpoints, equal spacing or midranks).
  2. Compute the Pearson correlation r between row and column scores over all n observations.
  3. Compute M squared = (n - 1) r squared and refer it to chi-square on 1 df, or M to the standard normal for a one-sided trend.
  4. If one variable is binary, read the result as a trend in the proportion across ordered groups (Cochran-Armitage) or a difference in mean scores.
  5. If one variable is nominal with more than two categories, use the row-means score statistic on I-1 df instead.
  6. Optionally fit the linear-by-linear association model to estimate the local odds ratio and check fit with G squared.
  7. Repeat with alternative sensible scores and report whether the conclusion holds.

Sources

  • An Introduction to Categorical Data Analysis (Agresti, 1996)
  • Bayesian Statistical Modelling (Congdon, 2006)

Also known as

  • Linear-by-linear association test
  • M squared trend test
  • Cochran-Armitage trend test
  • Uniform association model
  • Test for linear trend in proportions

Categories

  • Analytics & Behaviour
  • Statistics & Modelling