[R] nls - convergence problem

[Spencer Graves]

Hi, Doug: 

      How would you diagnose something like this?  For example, might 
the following (from ?nlsModel) help: 

     DNase1 <- DNase[ DNase$Run == 1, ]
     mod <-
      nlsModel(density ~ SSlogis( log(conc), Asym, xmid, scal ),
               DNase1, start=list( Asym = 3, xmid = 0, scal = 1 ))
     mod$Rmat()        # R matrix from the QR decomposition of the gradient

      Wouldn't the first zero or nearly diagonal element of Rmat 
identify a variable to which the model was insensitive and could 
therefore be fixed or removed from the model? 
      Thanks,
      Spencer Graves    

Douglas Bates wrote:

> peter.schlattmann at t-online.de wrote:
>
>> Dear list,
>>
>> I do have a problem with nls. I use the following data:
>>
>>
>>> test
>>
>>
>>     time  conc dose
>>    0.50  5.40    1
>>    0.75 11.10    1
>>    1.00  8.40    1
>>    1.25 13.80    1
>>    1.50 15.50    1
>>    1.75 18.00    1
>>    2.00 17.00    1
>>    2.50 13.90    1
>>    3.00 11.20    1
>>   3.50  9.90    1
>>   4.00  4.70    1
>>   5.00  5.00    1
>>   6.00  1.90    1
>>   7.00  1.90    1
>>   9.00  1.10    1
>> 12.00  0.95    1
>> 14.00  0.46    1
>> 24.00    NA    1
>> 30.00    NA    1
>> 36.00    NA    1
>>
>> I use the self-starting function SSfol:
>>
>> nls(conc~SSfol(dose,time,lKe,lKa,lCl),data=test,trace=T,control=nls.control(maxiter=13,tol=0.001,minFactor=1.E-500),na.action=na.omit) 
>>
>>
>> This gives the following output:
>>
>> 99.15824 :  -1.2061792  0.1296157 -4.3020997 86.07567 :  -0.7053265 
>> -0.3873204 -4.1278009 85.19743 :  -0.5548499 -0.5333776 -4.1173627 
>> 85.19246 :  -0.5466376 -0.5415731 -4.1173247 85.1922 :  -0.5444637 
>> -0.5437461 -4.1173223 85.1922 :  -0.5442240 -0.5439857 -4.1173223 
>> 85.1922 :  -0.5441337 -0.5440760 -4.1173223 85.1922 :  -0.5441104 
>> -0.5440993 -4.1173223 85.1922 :  -0.5440984 -0.5441113 -4.1173223 
>> 85.1922 :  -0.5441089 -0.5441008 -4.1173223 85.1922 :  -0.5441006 
>> -0.5441091 -4.1173223 85.1922 :  -0.5441051 -0.5441046 -4.1173223 
>> 85.1922 :  -0.5441051 -0.5441046 -4.1173223 85.1922 :  -0.5441051 
>> -0.5441046 -4.1173223
>> Error in nls(conc ~ SSfol(dose, time, lKe, lKa, lCl), data = test, trace
>> = T,  :         number of iterations exceeded maximum of 13
>>
>> I do not understand this error message for two reasons:
>>
>> 1. Apparently the algorithm has converged, at the end of the output the
>> change of parameter estimates seems to be less than 0.001
>>
>>
>> 2. If the maximum number of iterations is reached I would expect a
>> warning in the case that the algortihm did not converge. However, here
>> the algorithm apparently has converged. Thus I do not understand what
>> happened.
>
>
> But it hasn't converged.  The nls function uses a relative offset 
> convergence criterion that is not being satisfied here.  By setting 
> the minimum step factor to a very small number you are more small 
> steps than you would otherwise but you are not converging.  It appears 
> that the parameter estimates are not defined for this model/data set 
> combination.
>
> ______________________________________________
> R-help at stat.math.ethz.ch mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide! 
> http://www.R-project.org/posting-guide.html