Evidential model ranking without likelihoods

We present a probabilistic formulation of the Approximate Bayesian Computation scheme that allows evidential ranking of alternative models without direct use of a likelihood function. This approach is particularly important when ranking of several sophisticated stochastic models is desired, and the likelihood is either too complex or impossible to define. We suggest a modification of a Sequential Monte-Carlo sampler that uses ideas of Path Sampling to estimate an approximation to marginal likelihoods as a measure of evidence support. We demonstrate applications of this method on a problem of ranking alternative models of cancerous tumour growth using unique data from three cancerous spheroid lines.