1: \begin{abstract}

2:   In this paper a drift-randomized Milstein method is introduced 

3:   for the numerical solution of non-autonomous stochastic differential

4:   equations with non-differentiable drift coefficient functions. 

5:   Compared to standard Milstein-type methods

6:   we obtain higher order convergence rates in the $L^p(\Omega)$ and 

7:   almost sure sense. An important ingredient in the error analysis

8:   are randomized quadrature rules for H\"older continuous stochastic 

9:   processes.  By this we avoid the use of standard arguments based on the

10:   It\=o-Taylor expansion which are typically applied in

11:   error estimates of the classical Milstein method but

12:   require additional smoothness of the drift and diffusion coefficient

13:   functions. We also discuss the optimality of our convergence rates. Finally,

14:   the question of implementation is addressed in a numerical experiment.

15: \end{abstract}

16: