[R-sig-ME] nlmer: gradient for random change point model

[Ken Beath]

On 13/12/2007, at 3:32 PM, Ken Beath wrote:

>
> The zero derivatives are due to the interval around the breakpoint not
> containing any data, I suggest placing the initial changepoint in the
> centre of the data and then increasing eps until it works.
>
>

This is not correct, reason is that starting values for both slopes  
must be different, otherwise changing the breakpoint location doesn't  
change anything. It also helps to get the starting values close to the  
true values.
Following works fine

test.nls <- nls(dv ~ hockey(weeks, alpha1, beta1, beta2, brk),
						start= list(alpha1=30,beta1=-6,beta2=0,brk=5),
						trace=TRUE,
                        data = small.df)

Ken