Taking error into account when fitting models using Approximate Bayesian Computation

Authored by Richard M Sibly, der Vaart Elske van, Dennis Prangle

Date Published: 2018

DOI: 10.1002/eap.1656

Sponsors: United Kingdom Natural Environment Research Council (NERC)

Platforms: R NetLogo RNetlogo

Model Documentation: Other Narrative Pseudocode Mathematical description

Model Code URLs: Model code not found

Abstract

Stochastic computer simulations are often the only practical way of answering questions relating to ecological management. However, due to their complexity, such models are difficult to calibrate and evaluate. Approximate Bayesian Computation (ABC) offers an increasingly popular approach to this problem, widely applied across a variety of fields. However, ensuring the accuracy of ABC's estimates has been difficult. Here, we obtain more accurate estimates by incorporating estimation of error into the ABC protocol. We show how this can be done where the data consist of repeated measures of the same quantity and errors may be assumed to be normally distributed and independent. We then derive the correct acceptance probabilities for a probabilistic ABC algorithm, and update the coverage test with which accuracy is assessed. We apply this method, which we call error-calibrated ABC, to a toy example and a realistic 14-parameter simulation model of earthworms that is used in environmental risk assessment. A comparison with exact methods and the diagnostic coverage test show that our approach improves estimation of parameter values and their credible intervals for both models.

Tags

Individual-based model individual-based models Dynamics Parameter estimation IBM calibration population growth Inference Monte-carlo Eisenia-foetida savigny Fecundity Approximate bayesian computation (abc) Model fitting Stochastic computer simulation Likelihoods