Transform-Both-Sides and Power-of-the-Mean Weighting for Nonlinear Models

Description

Keep a meaningful curve while fixing variance that grows with the mean, by transforming response and model with the same power or weighting by a power of the fitted mean.

Spread often grows with level: task times, revenue and counts vary more where they are larger. Transforming only the response to fix this destroys the curve's meaning. Carroll and Ruppert's transform-both-sides applies the same Box-Cox power to the observed response and to the model function, so the mechanistic relationship is kept while the errors are stabilised; the power is estimated by maximising a concentrated likelihood over a grid. The alternative leaves the scale alone and weights observations by a power of the fitted mean, estimated with the curve in Box and Hill's staged procedure or jointly by maximum likelihood. Treating the estimated power as known costs little in the transform-both-sides model, unlike a Box-Cox transformation to linearity. Back-transformed predictions estimate the median and need a correction for the mean.

Attributes

Loss type
Unmigrated
Effort
Days

Steps

  1. Plot residuals against fitted values, or replicate standard deviations against means, to confirm spread grows with level.
  2. Transform-both-sides: for a grid of powers, transform response and model function with the same Box-Cox power, fit by nonlinear least squares, and choose the power that maximises the concentrated log likelihood.
  3. Weighting alternative: fit unweighted, set weights to the fitted values raised to minus twice the variance power, refit, re-estimate the power and iterate to convergence (or estimate all jointly).
  4. Check weighted residuals against weighted fitted values for constant spread, and run a weighted lack-of-fit F test if replicates exist.
  5. Report curve parameters on their original scale and apply a retransformation correction when predicting means after a log.

Sources

  • Nonlinear Regression (Seber & Wild, 1989)

Also known as

  • Transform both sides (TBS)
  • Power-transform both sides (PTBS)
  • Power-transformed weighted least squares (PTWLS)
  • Variance proportional to a power of the mean
  • Box-Hill weighting
  • Carroll-Ruppert transformation

Categories

  • Analytics & Behaviour
  • Statistics & Modelling