Bootstrap Likelihood Ratio Test for the Number of Mixture Components

Description

Decide whether data need g+1 rather than g segments by comparing the observed likelihood ratio with its distribution in samples simulated from the fitted g-segment model.

The ordinary likelihood ratio test fails for the number of mixture components: under the null the extra component sits on the boundary of the parameter space and its parameters are not identifiable, so the statistic has no chi-squared reference distribution and approximations such as Wolfe's are unreliable. McLachlan's remedy is a parametric bootstrap: simulate data sets from the fitted g-component model, fit g and g+1 components to each with exactly the same procedure as for the real data, and use the replicated statistics as the null distribution. Rejecting when the observed value exceeds the j-th smallest of B replicates gives a test of size about 1 - j/(B+1), so 19 replicates suffice for a 5% test, though more give power. Applied sequentially (1 against 2, 2 against 3) it picks the smallest number of segments the data support.

Attributes

Loss type
Siloed
Effort
Days

Steps

  1. Fit the mixture with g and g+1 components from many random starts and compute twice the gain in log likelihood.
  2. Simulate B samples of the same size from the fitted g-component model (B of 99 or more is advisable; 19 is the minimum for a 5% test).
  3. Fit g and g+1 components to each simulated sample with exactly the same starts, stopping rule and handling of spurious solutions as for the real data.
  4. Estimate the p-value from the rank of the observed statistic among the replicates.
  5. If significant, increase g by one and repeat; stop at the first non-significant step but keep an eye on later jumps in log likelihood.
  6. Report the sequence of tests alongside BIC or ICL rather than as the sole criterion.

Sources

  • Finite Mixture Models (McLachlan & Peel, 2000)

Also known as

  • BLRT
  • Bootstrapped likelihood ratio test
  • Parametric bootstrap LRT for the number of clusters
  • Resampling test of g versus g+1 components
  • Monte Carlo test for the number of components

Categories

  • Analytics & Behaviour
  • Statistics & Modelling