Hedonic regressions: mis-specification and neural networks
There is an emerging literature in the field of hedonic regressions which successfully applies flexible functional forms. In contrast, numerous authors report hedonic regressions with relatively simple functions, often with little or no attempt to consider interaction terms and higher level powers. Such simple functions perform relatively well in spite of good a priori grounds for quite complex interaction effects. Indeed, the studies are characterised by a general indifference to testing these more flexible models, even when degrees of freedom permit. Using detailed scanner data, hedonic functions for the UK TV market are estimated and tested for interaction effects. For a case study of the market, higher level interaction effects are tested for and compare the results with those from a neural network, with its property of 'universal approximation'. The use of neural networks is particularly novel in this context and some general findings on their suitability emerge.