A discriminant model for classifying contractor performance on public works projects

Abstract

Contractor selection practices in the UK construction industry have long been criticised and presently a divergent range of methods and preferences exists. Albeit, many of the practices adopted comply with good guidance practices and recommendations from construction reports and commentators. This research focused on UK construction clients' contractor selection preferences i.e. prequalification criteria (PC) and project-specific criteria (PSC). The main aim was to develop a contractor classification framework to assist construction clients' decisionmaking during tender evaluation. Investigating client selection preferences and behaviours are the main focus of this research. However, attention was also given to the contractors' views upon selection, from prequalification to invitation-to-tender. Factors affecting clients' non-use of standard prequalification practices were found to be a perceived: lack of flexibility and tolerance to clients' specific needs; and a long term confidence with 'in-house' selection practices. With regard to the use of PC and PSC, there appears to be much concordance among clients and contractors, but levels of importance assigned by public clients and clients' representatives were found to be significantly different to some extent in building and civil engineering works. Based on data from 68 small to medium size UK minor works (below £50 million), a contractor classification model (i.e. Z2 model) was developed. Multivariate discriminant analysis is used to classify contractors' past performance into good and poor groups. The classification model is made up of 5 variables: (i) contractors' plant and equipment resources; (ii) past performance in time on similar projects; (iii) past performance in cost of similar projects; (iv) reputation and image; and (v) relationship with local authority. The developed model has a 90% accuracy in classifying contractors into 'good' and 'poor' groups and a 70% accuracy when tested against independent data.