R error - Gov2001 - lists.gking.harvard.edu

Sean Ingham

30 Mar 30 Mar

4:47 p.m.

I think you left out this argument, control=list(fnscale = -1) which tells optim to look for the minimum of the negative of your function zero.poiss, i.e., its maximum. If you omit it, optim will look for the minimum of your function, which presumably doesn't exist. Sean On Wed, Mar 30, 2011 at 4:38 PM, Adela Soliz <ars431(a)mail.harvard.edu>wrote;wrote:

...

Hi Class, I keep receiving this message when I try to optimize my function in problem 2.4: Error in optim(par = rep(0, 8), y = y, X = X, Z = Z, fn = zero.poiss, : non-finite finite-difference value [6] I've checked my function to make sure all the parts of it are working correctly, I've tried changing things in optim. I think we talked about this error is section once- can anyone tell me what's going on? -Adela _______________________________________________ gov2001-l mailing list gov2001-l(a)lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l

_______________________________________________ gov2001-l mailing list gov2001-l(a)lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l

Reply

Brandon Stewart

5:52 p.m.

Can you send us some code for the log-likelihood function? If I had to guess I would say that you left a factorial of y in there. This becomes an incredibly large number and overloads the program. That's why you have to simplify that stuff out. Brandon On Wed, Mar 30, 2011 at 5:50 PM, Adela Soliz <ars431(a)mail.harvard.edu>wrote;wrote:

...

Thanks, Sean, but I have that. This is my optim command: optim(par=c(rep(0,10)), y=y, X=X, Z=Z, fn=zero.poiss, control=list(fnscale=-1), method="BFGS", hessian=TRUE) On Wed, Mar 30, 2011 at 4:47 PM, Sean Ingham <stingham(a)gmail.com> wrote:

I think you left out this argument, control=list(fnscale = -1) which tells optim to look for the minimum of the negative of your function zero.poiss, i.e., its maximum. If you omit it, optim will look for the minimum of your function, which presumably doesn't exist. Sean On Wed, Mar 30, 2011 at 4:38 PM, Adela Soliz <ars431(a)mail.harvard.edu>wrote;wrote:

Hi Class, I keep receiving this message when I try to optimize my function in problem 2.4: Error in optim(par = rep(0, 8), y = y, X = X, Z = Z, fn = zero.poiss, : non-finite finite-difference value [6] I've checked my function to make sure all the parts of it are working correctly, I've tried changing things in optim. I think we talked about this error is section once- can anyone tell me what's going on? -Adela _______________________________________________ gov2001-l mailing list gov2001-l(a)lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l

_______________________________________________ gov2001-l mailing list gov2001-l(a)lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l

Reply

Brandon Stewart

6:01 p.m.

Unfortunately, I think your derivation is off. It's hard to tell why this particular function isn't working but I suppose it doesn't entirely matter. Two things to help push you in the right direction: (1) you need an indicator variable for where y=0 (2) my guess from looking at this is that you don't have the right PDF for the poisson (that or you have oversimplified it) Brandon On Wed, Mar 30, 2011 at 5:54 PM, Adela Soliz <ars431(a)mail.harvard.edu>wrote;wrote:

...

Here is my function: zero.poiss<-function(par, X, Z, y){ beta<-par[1:6] gamma<-par[7:8] phi<-1/(1+exp(-Z%*%gamma)) lambda<-(exp(X%*%beta)) sum((1-y)*log((1-phi)*lambda)+y*log(phi+(1-phi)*(1-lambda))) } On Wed, Mar 30, 2011 at 5:52 PM, Brandon Stewart < brandonmstewart(a)gmail.com> wrote:

Can you send us some code for the log-likelihood function? If I had to guess I would say that you left a factorial of y in there. This becomes an incredibly large number and overloads the program. That's why you have to simplify that stuff out. Brandon On Wed, Mar 30, 2011 at 5:50 PM, Adela Soliz <ars431(a)mail.harvard.edu>wrote;wrote:

Thanks, Sean, but I have that. This is my optim command: optim(par=c(rep(0,10)), y=y, X=X, Z=Z, fn=zero.poiss, control=list(fnscale=-1), method="BFGS", hessian=TRUE) On Wed, Mar 30, 2011 at 4:47 PM, Sean Ingham <stingham(a)gmail.com> wrote:

I think you left out this argument, control=list(fnscale = -1) which tells optim to look for the minimum of the negative of your function zero.poiss, i.e., its maximum. If you omit it, optim will look for the minimum of your function, which presumably doesn't exist. Sean On Wed, Mar 30, 2011 at 4:38 PM, Adela Soliz <ars431(a)mail.harvard.edu>wrote;wrote: > Hi Class, > > I keep receiving this message when I try to optimize my function in > problem 2.4: > > Error in optim(par = rep(0, 8), y = y, X = X, Z = Z, fn = zero.poiss, > : > non-finite finite-difference value [6] > > I've checked my function to make sure all the parts of it are working > correctly, I've tried changing things in optim. I think we talked about > this error is section once- can anyone tell me what's going on? > > -Adela > > _______________________________________________ > gov2001-l mailing list > gov2001-l(a)lists.fas.harvard.edu > http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l > >

_______________________________________________ gov2001-l mailing list gov2001-l(a)lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l

Reply