1: \begin{abstract}

2:   Mixture of regression models are useful for regression analysis in

3:   heterogeneous populations where a single regression model may not be

4:   appropriate for the entire population. We study the nonparametric

5:   maximum likelihood estimator (NPMLE) for fitting these models. The

6:   NPMLE is based on convex optimization and does not require prior

7:   specification of the number of mixture components. We establish

8:   existence of the NPMLE and prove finite-sample parametric (up to

9:   logarithmic multiplicative factors) Hellinger error bounds for the

10:   predicted density functions. We also provide an effective procedure

11:   for computing the NPMLE without ad-hoc discretization and prove a

12:   theoretical convergence rate under certain assumptions. Numerical

13:   experiments on simulated data for both discrete and non-discrete

14:   mixing distributions demonstrate the remarkable performances of our

15:   approach. We also illustrate the approach on two real datasets.\\ 

16: \keywords{conditional gradient method, empirical Bayes, finite-sample parametric

17:   rate, Hellinger distance, mixture of regression, nonparametric

18:   maximum likelihood estimator (NPMLE), random coefficient

19:   regression. } 

20: 

21: \end{abstract}

22: