1: \begin{abstract} 

2:  We consider the problem of estimating the structural function in  nonparametric instrumental regression,

3:  where  in the presence of an instrument $W$ a response~$Y$ is modeled in dependence of an endogenous explanatory variable  $Z$. 

4: 

5:  The proposed estimator is based on dimension reduction and additional

6:  thresholding. The minimax optimal rate of convergence of the

7:  estimator is derived assuming that the structural function belongs to

8:  some ellipsoids which are in a certain sense linked to the

9:  conditional expectation operator of $Z$ given $W$.  We illustrate

10:  these results by considering classical smoothness assumptions.

11:  However, the proposed estimator requires an optimal choice of a

12:  dimension parameter depending on certain characteristics of the

13:  unknown structural function and the conditional expectation operator

14:  of $Z$ given $W$, which are not known in practice.  The main issue

15:  addressed in our work is a fully adaptive choice of this dimension

16:  parameter using a model selection approach under the restriction that

17:  the conditional expectation operator of $Z$ given $W$ is smoothing in

18:  a certain sense.  In this situation we develop a penalized minimum

19:  contrast estimator with randomized penalty and collection of models.

20:  We show that this data-driven estimator can attain the lower risk

21:  bound up to a constant over a wide range of smoothness classes for

22:  the structural function.

23: \end{abstract}

24: