Re-expression with the Ladder of Powers

Description

Choose a power, root, log or folded expression for amounts, counts or fractions so batches become symmetric with similar spreads and curved relationships straighten.

Tukey's re-expression family replaces fixed habits (always log task times) with a data-driven choice. Expressions form a ladder: squares, raw, square roots, logs (the zeroth power), negative reciprocal roots, negative reciprocals. For several batches, plotting log spread against log median gives a tilt: near 1/2 points to roots, near 1 to logs, and each step down the ladder lowers the tilt by about 1/2; aim for no tilt. For a curved scatter, three well-spaced points and the bulging rule ('move on the ladder as the bulge points') choose expressions of y and x that straighten it. Counted fractions get their own folded scales: start the counts to avoid 0 and 1, then use froots or flogs, which are zero at 50 percent and treat 'are' and 'aren't' symmetrically; differences of flogs in a 2 x 2 table express disproportion. Comparison is then made by differences on a well-chosen scale.

Attributes

Loss type
Unmigrated
Effort
Hours

Steps

  1. Classify the data as amounts, counts, balances or counted fractions; start counts with zeros (add a constant) before taking logs, and start fractions as (count + 1/6) / (total + 1/3).
  2. For several batches, compute the median and hinges of each, then plot log H-spread against log median and read the tilt.
  3. Move down the ladder (raw, roots, logs, reciprocals) until the tilt is close to zero, recomputing for typical points at each end.
  4. For a curved relationship, pick three representative points, compare the two segment slopes, and move y and/or x along the ladder in the direction of the bulge until the slopes agree.
  5. For proportions, re-express on a folded scale (froot or flog) and compare groups by differences on that scale.
  6. Check the result with midsummaries, displays and residual plots, and use parallel data sets to guide the choice where available.

Sources

  • Exploratory Data Analysis (Tukey, 1977)
  • Quantitative Psychological Research (Clark-Carter, 2004)
  • Design and Analysis of Experiments, Vol. 1 (Hinkelmann & Kempthorne, 2008)
  • Nonlinear Regression Analysis and Its Applications (Bates & Watts, 1988)
  • Linear Statistical Inference and Its Applications (Rao, 1973)
  • Foundations of Measurement, Vol. 1: Additive and Polynomial Representations (Krantz et al., 1971)

Also known as

  • Ladder of powers
  • Spread-level plot
  • Spread-versus-level plot
  • Variance-stabilising re-expression
  • Bulging rule
  • Tukey's ladder for straightening
  • Three-point method
  • Folded re-expression of counted fractions
  • Folded logs (flogs)
  • Froots
  • Started fractions

Categories

  • Analytics & Behaviour
  • Statistics & Modelling