Abstract
At the outset of an analysis, there is a need to interact with the problem-owners to understand their perspectives on the issues. This understanding leads to the construction of one or more models to reflect their views, their values and their uncertainties. Some models are qualitative; others, quantitative. Quantitative models need populating with numbers, either from data or from further judgements elicited from the problem-owners, their stakeholders or their experts. The model(s) may then be analysed to provide feedback to the problem-owners on the possible resolution of the problem. In practice, the process may iterate, cycling through more sophisticated models, which require further inputs from the problem-owners. This paper discusses the elicitation processes involved, arguing that the current literature has developed if not in silos, then in pockets of activity that do not reflect the more joined up processes that often take place in practice. Furthermore, it is suggested that potential psychological and behavioural biases that may occur in quantitative elicitation are reasonably well understood and guarded against, whereas less attention has been paid to similar biases that may affect more qualitative model building.
Acknowledgements
I am grateful to many colleagues for discussions on modelling, soft and hard elicitation, though we did not always use that terminology: in particular, Martine Barons, Mark Burgman, Anca Hanea, John Maule, Nadia Papamichail, Larry Phillips, Tudor Rickards, Duncan Shaw, Jim Smith, Bartel van der Walle and Mike Yearworth. My interest in the topic was stimulated by Doug White, one of my earliest mentors, sadly no longer with us.
Disclosure statement
No potential conflict of interest was reported by the author(s).
Notes
1 Ackermann (Citation2012) and Checkland (Citation2019) also reflect on the long debate between soft and hard OR.
2 Pidd (Citation1996) used the terms interpretive and mathematical/logical for qualitative and quantitative, respectively.
5 In many complex cases there will be several models either networked together or run in parallel. But that does not change the fundamental question here.