Nearly 400 composite railway bridge decks of a new kind belonging to the trough type with U-shaped cross section have been constructed in Belgium over the last fifteen years. The construction of these bridge decks is rather complex with the preflexion of precambered steel girders, the prestressing of a concrete slab and the addition of a 2nd phase concrete. Until now, they have been designed with a classical computation method using a pseudo-elastic analysis with modular ratios. Globally, they perform according to the expectations but variability has been observed between the measured and the computed camber of these bridge decks just after the transfer of prestressing and also at long-term. A statistical analysis of the variability of the relative difference between the measured camber and the computed camber is made for a sample of 36 bridge decks using no less than 10 variables. The most significant variables to explain this variability at prestressing are the ratio between the maximum tensile stress reached in the steel girders during the preflexion and the yield strength and the type of steel girder. For the same sample, the long-term camber under permanent loading is computed by two methods and compared with measurements taken one or two years after the construction. The camber computed by the step-by-step method shows a better agreement with the measured camber than the camber computed by the classical method. The purpose of the paper is to report on the statistical analysis which was used to determine the most significant parameters to consider in the modeling in order to improve the prediction of the behaviour of these composite railway bridge decks.