1: \begin{abstract}

2: A number of variable selection methods have been proposed involving

3: nonconvex penalty functions.  These methods, which include the smoothly

4: clipped absolute deviation (SCAD) penalty and the minimax concave

5: penalty (MCP), have been demonstrated to have attractive theoretical

6: properties, but model fitting is not a straightforward task, and the

7: resulting solutions may be unstable.  Here, we demonstrate the

8: potential of coordinate descent algorithms for fitting these models,

9: establishing theoretical convergence properties and demonstrating that

10: they are significantly faster than competing approaches.  In addition,

11: we demonstrate the utility of convexity diagnostics to determine

12: regions of the parameter space in which the objective function is

13: locally convex, even though the penalty is not.  Our simulation study

14: and data examples indicate that nonconvex penalties like MCP and SCAD

15: are worthwhile alternatives to the lasso in many applications.  In

16: particular, our numerical results suggest that MCP is the preferred

17: approach among the three methods.

18: \end{abstract}