Ordre : Université De Paris Sud U . F . R . Scientifique D '

In this chapter we rst review results on a famous method to obtain estimators in a semiparametric setup, the pro le likelihood method, which is a two-step maximization method based on the likelihood. Then we discuss a generalization called penalized pro le likelihood method, which is introduced in Murphy & van der Vaart (1999) and in Dalalyan et al. (2006). It allows among other properties to deal with second order features in some semiparametric contexts. We obtain new families of semiparametric estimators by this method in a white noise model, the period model, and in the famous survival analysis Cox model. Moreover, the classical semiparametric estimators in those models are derived as particular cases of the preceding families. The study of the properties of the rst family are the object of chapters 2 and 3 of this thesis, in particular second order properties. The existence of analog results for the Cox model remains an open question.