Log-linear modelling of paired comparison data from consumer tests

Paired comparison (PC) is a method which requires no assessor training; on the other hand, more statistical work and calculations must be performed before PC data can be converted to a product rating on a linear scale. Often Bradley¯Terry or Thurstone¯Mosteller models are used to position the products on a linear scale. This paper shows how the parameters of the Bradley¯Terry model can be estimated using a log-linear model which is accessible in most statistical packages. Relevant hypotheses, such as product equality and fit of the Bradley¯Terry model, are tested. Extensions of the Bradley¯Terry model with assessor segmentation and product composition are shown in the general framework of log-linear modelling. Modelling and hypothesis testing are demonstrated on orange soft drink data.