[R] Errors with maxNR: NA in gradient

[Emma MONTARSOLO]

Hi !

I use maxNR function (from maxLik package) to estimate parameters via maximum likelihood and Newton–Raphson method.

Context :
I have a variable Y which follows a Gumbel distribution of parameters Beta1X1 + Beta2X2 (location parameter) and sigma (scale parameter).
I have to estimate sigma, beta1 and beta2

The following errors appear when I run maxNR :

> algo=maxNR(loglikelihood, start=c(sigma:8.686603,beta1=-4.976215,beta2=7.313875))
Iteration 1
Parameter:
                  beta1       beta2
   4.251275 -249.985063    5.376905
     beta1 beta2
NaN   NaN   NaN
Error in maxNRCompute(fn = function (theta, fnOrig, gradOrig = NULL, hessOrig = NULL,  :
  NA in gradient
In addition: There were 15 warnings (use warnings() to see them)

> warnings()
Messages d'avis :
1: In log(sigma) : NaNs produced
2: In log(sigma) : NaNs produced
3: In log(sigma) : NaNs produced
4: In log(sigma) : NaNs produced
5: In log(sigma) : NaNs produced
6: In log(sigma) : NaNs produced
7: In log(sigma) : NaNs produced
8: In log(sigma) : NaNs produced
9: In log(sigma) : NaNs produced
10: In log(sigma) : NaNs produced
11: In log(sigma) : NaNs produced
12: In log(sigma) : NaNs produced
13: In log(sigma) : NaNs produced
14: In log(sigma) : NaNs produced
15: In log(sigma) : NaNs produced

My code is the following :

n=length(data_reg$y)

loglikelihood<-function(parameters)
{
  sigma<-parameters[1]
  beta1<-parameters[2]
  beta2<-parameters[3]

  Z=(data_reg$y-(beta1*data_reg$x1)-(beta2*data_reg$x2))/sigma
  ll=-(n*log(sigma))+sum(Z+exp(-Z))
  ll
}

algo=maxNR(loglikelihood, start=c(sigma:8.686603,beta1=-4.976215,beta2=7.313875))



I use results from a previous estimation method as starting values.

Thanks for your help. ��

EM

	[[alternative HTML version deleted]]