Sensitivity analysis of agent-based models: a new protocol

Agent-based models are widely used in management research but are often criticised as hard to interpret, and most contain non-parametric elements that standard sensitivity methods do not handle. We set out a systematic protocol with four goals: testing robustness, finding the most influential elements, examining how elements interact, and establishing the direction of change.

Randomised finite change indices combined with a factorial design address the first three goals; adapted individual conditional expectation (ICE) plots address the fourth. We demonstrate the protocol on the Garbage Can Model of organisational decision-making.