Goodness of fit for models with intractable likelihood

Routine goodness-of-fit analyses of complex models with intractable likelihoods are hampered by a lack of computationally tractable diagnostic measures with well-understood frequency properties, that is, with a known sampling distribution. This frustrates the ability to assess the extremity of the data relative to fitted simulation models in terms of pre-specified test statistics, an essential requirement for model improvement. Given an Approximate Bayesian Computation setting for a posited model with an intractable likelihood for which it is possible to simulate from them, we present a general and computationally inexpensive Monte Carlo framework for obtaining $$p$$ p -valuesthat are asymptotically uniformly distributed in [0, 1] under the posited model when assumptions about the asymptotic equivalence between the conditional statistic and the maximum likelihood estimator hold. The proposed framework follows almost directly from the conditional predictive p-value proposed in the Bayesian literature. Numerical investigations demonstrate favorable power properties in detecting actual model discrepancies relative to other diagnostic approaches. We illustrate the technique on analytically tractable examples and on a complex tuberculosis transmission model.