Stellenbosch University
Welcome to Stellenbosch University
Seminar: Department of Logistics
Start: 05/10/2018, 13:00
End: 05/10/2018, 14:00
Contact:Linke Potgieter -
Location: Van der Sterr 3022

​​​​Presenter: Prof Wim Delva (Deputy-director for research: SACEMA)

Co-authors: Marijn Hazelbag, Jonathan Dushoff

Title of talk: Recasting model calibration as a missing data problem: Approximate Bayesian Computation with MICE

Introduction

Approximate Bayesian Computation (ABC) is a simulation-based inference method used to calibrate individual-based models (IBMs) to summary statistics of empirical data (target features). However, the basic rejection ABC scheme requires a prohibitively large number of simulations for IBMs with many parameters; several alternatives have therefore been proposed. We present a new adaptive sequential ABC scheme (MABC), based on the idea that Multivariate Imputation by Chained Equations (MICE) can be used to connect model inputs and outputs.

Methods

A first wave of model input parameters is generated through sobol sequences. Parameter combinations are ranked according to distance from the target features. A fraction of best-matching model features with corresponding parameter combinations are selected as complete observations. Next, the vector of target features is appended with missing values for the model parameters and MICE is used to impute plausible values. The imputed parameter combinations are then sampled with inverse probability density weights, and used for a new wave of model simulations, and the process of MICE-based parameter guesses is repeated until a sufficiently high quality of calibration is achieved. We calibrated an IBM with 14 unknown parameters to 37 target features, using MABC and Lenormand's adaptive population sequential ABC, and computed the root mean squared error (RMSE) of the 100 best-fitting parameter combinations for both methods.

Results

The RMSEs for MABC ranged between 0.14 and 0.26, while those for Lenormand's method ranged between 0.18 and 1.06.

Discussion

MABC shows great potential for calibrating IBMs with many unknown parameters and a large array of target features. MABC also has the advantage of greater user control over the duration of the calibration process and the quality of the final calibration output, as the end point of the MABC procedure can be the starting point for additional waves of simulations.