Compositional Data Regression via Log-Ratio Transform

Description

Turns each observation's shares into log-ratios against a reference part, regresses them jointly, then maps predictions back to shares that respect the sum-to-one constraint.

Many outcomes are compositions: non-negative parts that sum to one, such as how a user's time splits across features, how a budget splits across categories, or vote shares across parties. Ordinary regression on each share separately ignores the constraint, can predict impossible values and produces spurious negative correlations. Aitchison's log-ratio approach transforms each composition into J - 1 log-ratios against a reference part (additive log-ratio) or against the geometric mean (centred log-ratio), which are unconstrained and can be modelled with multivariate regression. Katz and King found electoral log-ratios heavier tailed than a normal distribution even after covariates, so they used a multivariate t regression with estimated degrees of freedom; Jackman fits it by Bayesian MCMC. Coefficients describe effects on log-ratios, so results are interpreted by simulating predicted log-ratios and back-transforming them into shares with uncertainty. Zero parts must be handled before transforming, and results are relative to the reference part.

Attributes

Loss type
Unmigrated
Effort
Days

Steps

  1. Confirm the outcome is a composition (non-negative parts summing to one) and choose a reference part or use centred log-ratios.
  2. Handle zero parts (replacement, amalgamation of categories, or a separate model for presence) and compute the J - 1 log-ratios.
  3. Specify a multivariate regression for the log-ratios, using a multivariate t error if residuals are heavy tailed.
  4. Fit by maximum likelihood or MCMC (several chains from dispersed starting values) and check convergence.
  5. Interpret coefficients by simulating predicted log-ratios for chosen profiles and back-transforming them to shares with intervals.
  6. Note cases excluded because a part was absent and consider how that might bias the results.

Sources

  • Bayesian Analysis for the Social Sciences (Jackman, 2009)

Also known as

  • Log-ratio regression for shares
  • Additive log-ratio (alr) regression
  • Aitchison log-ratio analysis
  • Katz-King multiparty model
  • Compositional data regression via log-ratio transform (multivariate t)

Categories

  • Analytics & Behaviour
  • Statistics & Modelling