Square-Table Analysis of Paired Categorical Responses (Symmetry, Quasi-Symmetry, Marginal Homogeneity)

Description

Tests whether the same people's multi-category responses on two occasions or two items have the same distribution, and in which direction they shift.

When each respondent is classified twice on the same I categories, such as preferred plan before and after a redesign, the I x I table has a heavy diagonal and independent-sample chi-square tests are wrong. Symmetry (mirror cells equal) implies marginal homogeneity. Quasi-symmetry allows different margins with a symmetric association; under it marginal homogeneity equals symmetry, so the likelihood ratio difference between the two fits tests marginal homogeneity on I-1 df. When quasi-symmetry fits poorly, the Stuart-Maxwell statistic on marginal differences is used. For ordered categories an ordinal quasi-symmetry model gives a 1-df test of directional shift, and a paired score test compares mean scores. Significant results are followed by category-versus-rest McNemar tests with intervals to locate the change. It generalises McNemar's test beyond two categories.

Attributes

Loss type
Unmigrated
Effort
Hours

Steps

  1. Cross-classify each respondent's first and second response in an I x I table.
  2. Fit the symmetry model (df I(I-1)/2) and the quasi-symmetry model (df (I-1)(I-2)/2).
  3. Test marginal homogeneity by G squared(symmetry) minus G squared(quasi-symmetry) on I-1 df, or by the Stuart-Maxwell statistic if quasi-symmetry fits badly.
  4. For ordered categories fit ordinal quasi-symmetry and test its slope on 1 df, or compare marginal mean scores with the paired score test.
  5. Locate the change by collapsing to each category versus the rest and applying McNemar's test with a confidence interval.
  6. Consider quasi-independence (independence off the diagonal) as a simpler description of who switches where.

Sources

  • An Introduction to Categorical Data Analysis (Agresti, 1996)

Also known as

  • Marginal homogeneity test
  • Symmetry model
  • Quasi-symmetry model
  • Ordinal quasi-symmetry
  • Bowker symmetry test
  • Stuart-Maxwell test
  • Generalised McNemar test

Categories

  • Analytics & Behaviour
  • Statistics & Modelling