Median Polish of Two-Way Tables

Description

Fit common, row and column effects to a two-way table by repeatedly removing medians, then inspect residuals and a diagnostic plot suggesting re-expression.

Median polish is a resistant alternative to two-way analysis of variance for exploring a table such as task by design or segment by month. Row medians are subtracted, then column medians of the remainder, and the cycle repeats until changes are negligible, leaving a common term, row effects, column effects and residuals. Because medians are used, a few extreme cells do not distort the effects; instead they stand out as large residuals. Tukey then draws a two-way plot of the fit, codes the residuals and computes comparison values (row effect times column effect divided by the common term). Plotting residuals against comparison values gives a diagnostic: a slope near 1 suggests a multiplicative structure, so logs should be taken; other slopes give a PLUS-one fit. The method extends to three-way tables.

Attributes

Loss type
Unmigrated
Effort
Hours

Steps

  1. Arrange the response (for example median task time) in a rows-by-columns table.
  2. Subtract each row's median, then each column's median of what remains.
  3. Repeat until the changes are negligible, typically two to four cycles.
  4. Record the common value, row effects, column effects and residuals.
  5. Draw a two-way plot of the fit and a coded display that highlights large residuals.
  6. Compute comparison values and plot residuals against them; fit a slope.
  7. If the slope is near 1, take logs and refit; otherwise consider a PLUS-one fit, and investigate cells with large residuals.

Sources

  • Exploratory Data Analysis (Tukey, 1977)
  • Design and Analysis of Experiments, Vol. 1 (Hinkelmann & Kempthorne, 2008)

Also known as

  • Median polish of two-way tables, with two-way plots, coded residuals and diagnostic plot for re-expression
  • Median polish
  • Row-PLUS-column fit
  • Resistant two-way analysis
  • Tukey's diagnostic plot
  • PLUS-one, extended and three-way fits
  • One degree of freedom for non-additivity

Categories

  • Analytics & Behaviour
  • Statistics & Modelling