The actions on road safety concerning the speed limits had been the subject of more than one hundred efficiency evaluations in term of victim and accident numbers reduction since 1966. These country scale evaluations led Elvik and al. to make méta-analyzes allowing them to establish a certain adequacy with "power" models in which the degree vary from 4 to 8 for the number of the victims. Thus, we are seeking through times series of victims number in Switzerland, France and Hungary, to assess the impact of actions on speed limits. Indeed, Switzerland in 1984 and 1985 has launched a campaign action on speed by decreasing the authorized speed limits on its roads, by 10km/h for the urban zones and motorways, by 20km/h for the rural roads; in 1990 France has lowered the speed limits in town to 50 km/h; on the contrary, Hungary has increased the speed limit by 10 km/h outside built-up areas. The most widespread methodology for efficiency evaluation uses Poissonnian models of the killed number in before/after plans. For monthly observations, this approach reduces the variability of the data, major concept in statistics and no take account the dynamic of the series. We propose in our article to tackle the question of evaluation through the modeling of times series, thus taking into account the possible deformation in time. For that we use two families of models. The first family of model based on the ARMA processes theory, well-known in the field of the road safety, belongs to the class of the parametric models. Within the framework of the parametric estimate, one postulates a model and considers a number finished of parameters. In addition to the strong a priori brought by the model, it is also necessary to check assumptions on the distribution of the observations. The second family of model based on the functional estimate is built from the nonparametric predictor of Nadaraya Watson. Compared to the ARMA models, non-parametric methods have the advantage compared to nearly use no a priori information. At the matter of fact, the question is now to estimate a function from a limited number of observations; to some extent the model let the data express themselves. For the last decades, non-parametric methods have been experiencing a significant development in the theoretical field, enhancing the scientific rigour as for their implementation. Two parametric models, autoprojectif and with transfer function, and a nonparametric model are proposed. The model with transfer function allows a direct effect modelling by the use of a dummy variable to characterize the event. These models then allow us to assess the impact of the measure on speed by comparing the observed values to observed ones. We also give the confidence intervals for the effects and compare the results obtained by the various modelling approaches.