1: \begin{abstract}

2: In this paper we study nonparametric estimation

3: in a binary treatment model where the outcome

4: equation is of unrestricted form, and the selection equation

5: contains multiple unobservables that enter through a nonparametric

6: random coefficients specification.  This specification is flexible

7: because it allows for complex unobserved heterogeneity of economic

8: agents and non-monotone selection into treatment.  We

9: obtain conditions under which both the

10: conditional distributions of $Y_{0}$ and $Y_{1}$, the outcome for

11: the untreated, respectively treated, given first stage unobserved random

12: coefficients, are identified.  We can thus identify an average

13: treatment effect, conditional on first stage unobservables called

14: UCATE, which yields most treatment effects parameters that depend on

15: averages, like ATE and TT.  We provide sharp bounds on the

16: variance, the joint distribution of $(Y_{0},Y_{1})$ and the

17: distribution of treatment effects.  In the particular case where

18: the outcomes are continuously distributed, we provide novel and weak

19: conditions that allow to point identify the joint conditional distribution

20: of $Y_{0},Y_{1},$ given the unobservables.  This allows to derive

21: every treatment effect parameter, {\it e.g.} the distribution of treatment

22: effects and the proportion of individuals who benefit from

23: treatment.  We present estimators for the marginals, average and distribution of treatment effects,

24: both conditional on unobservables and unconditional, as well as

25: total population effects.  The estimators use all

26: the data and discard tail values of the instruments when they

27: are too unlikely.

28: We provide their rates of convergence, and analyze their finite sample behavior in

29: a simulation study. Finally, we also discuss the situation where some of the instruments

30: are discrete.

31: \end{abstract}