Dear Alfonso On 10 May 2013 12:51, <alfonso.carf...@uniparthenope.it> wrote: > we are computing maximum likelihood estimations using maxLik package and we > realized that the results of the estimation depend on the version of the > package installed > > for example if we try to estimate this function with an old version of > maxLik under R 2.13.1 (32 bit version installed 2 years ago): > > > L<-function (param) {b0t<-param[1] > p1t<-param[2] > p2t<-param[3] > p3t<-param[4] > p4t<-param[5] > > for(i in 17:T) {n[i,]<- b0t + p1t*a[i-1] + p2t*sum(a[(i-4):(i-1)]) + > p3t*(sum(a[(i-8):(i-1)])) + p4t*(sum(a[(i-16):(i-1)])) > m[i,]<-exp(n[i])/(1+exp(n[i])) > ll[i-16,]<-a[i]*log(m[i])+(1-a[i])*log(1-m[i]) } > sum(ll)} > b2<-maxLik(L, start=c(-2.8158,5,-1,0.3213,-0.3112)) > > > we obtain this results: > > summary(b2) > > Maximum Likelihood estimation > Newton-Raphson maximisation, 16 iterations > Return code 2: successive function values within tolerance limit > Log-Likelihood: -38.11285 > 5 free parameters > Estimates: > Estimate Std. error t value Pr(> t) > [1,] -2.81578 0.43548 -6.4660 1.007e-10 *** > [2,] 50.50024 13.17046 3.8344 0.0001259 *** > [3,] -11.53344 3.31075 -3.4836 0.0004947 *** > [4,] 0.32130 0.42978 0.7476 0.4547096 > [5,] -0.31121 0.23245 -1.3388 0.1806280 > --- > Signif. codes: 0 ?***? 0.001 ?**? 0.01 ?*? 0.05 ?.? 0.1 ? ? 1 > -------------------------------------------- > > NB: a is a binary time series > > if we try to estimate the same function using the last version of maxLik > under R.3.0 (64 bit latest version) the estimation do not converge and this > is the error message: > > Iteration 15 > Parameter: > [1] -2.8146429 51.3042554 -11.7373313 0.3245214 -0.3125767 > Gradient: > [1] NaN NaN NaN NaN NaN > Errore in maxNRCompute(fn = logLikAttr, fnOrig = fn, gradOrig = grad, > hessOrig = hess, : > NA in gradient > > What causes this?
It could be that the NaNs in the gradients are caused by rounding errors and/or approximation errors in the numerical (finite-difference) derivatives when using R.3.0 64 bit, because different hardware and different software versions (e.g. R, mathematical libraries, OS) could lead to different rounding errors. In this case, the specification of a function that returns analytical gradients could solve the problem. If this does not solve the problem and you cannot find out the reason for the NaNs in the analytical gradients yourself, please provide a reproducible example so that we could help you with this. Please note that you could also ask questions regarding the maxLik package via a forum at maxLik's R-Forge site: https://r-forge.r-project.org/projects/maxlik/ ... and please do not forget to cite maxLik in your publications :-) Best regards, Arne -- Arne Henningsen http://www.arne-henningsen.name ______________________________________________ R-help@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide http://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code.