Duration Models and Neural Networks to Analyse Unemployment Spells

A set of completed spells of unemployment was obtained from the periodic labour force survey with a rotating panel design, carried out by ISTAT in Emilia Romagna, Italy, from the first quarter of 1993 to the first quarter of 1995. The data set was analysed through Cox's proportional hazards model with a Weibull specification of spell length distribution. The empirical distributions of durations showed pronounced spikes which suggests the presence of a considerable telescopic effect. This effect was modelled by an exponential heaping function. Heterogeneity was controlled by adding a suitable set of covariates and unobserved heterogeneity was introduced on the scale parameter of the baseline hazard function. The results were compared with those obtained by a neural network model, but the performance of the latter did not show strong improvement over the former because presumably strong non-linearities and interactions did not underlie the relationships between the spell lengths and the covariates.