similar to: Optimization max likelihood problem

Hi R-help! I'm trying to understand how R's nlm function updates its estimate of the Hessian matrix. The Dennis/Schnabel book cited in the references presents a number of different ways to do this, and seems to conclude that the positive-definite secant method (BFGS) works best in practice (p201). However, when I run my code through the optim function with the method as "BFGS",

I've been working on a new package and I have a few questions regarding the behaviour of the nlm function. I've been (for better or worse) using the nlm function to fit a linear model without suppling the hessian or gradient attributes in the objective function. I'm curious as to why the nlm requires 31 iterations (for the linear model), and then it doesn't work when I try to add

This practical problem in maximum likelihood estimation must be encountered quite a bit. What do you do when a data point has a probability that comes out in numerical evaluation to zero? In calculating the log likelihood you then have a log(0) problem. Here is a simple example (probit) which illustrates the problem: x<-c(1,2,3,4,100) ntrials<-100 yes<-round(ntrials*pnorm((x-3)/1))

Dear All, I am running a simulation to obtain coverage probability of Wald type confidence intervals for my parameter d in a function of two parameters (mu,d). I am optimizing it using "optim" method "L-BFGS-B" to obtain MLE. As, I want to invert the Hessian matrix to get Standard errors of the two parameter estimates. However, my Hessian matrix at times becomes

> > On 9 Feb 2017, at 16:00, G?ran Brostr?m <goran.brostrom at umu.se> wrote: > > > > In my package 'glmmML' I'm using old C code and linpack in the optimizing procedure. Specifically, one part of the code looks like this: > > > > F77_CALL(dpoco)(*hessian, &bdim, &bdim, &rcond, work, info); > > if (*info == 0){ > >

In my package 'glmmML' I'm using old C code and linpack in the optimizing procedure. Specifically, one part of the code looks like this: F77_CALL(dpoco)(*hessian, &bdim, &bdim, &rcond, work, info); if (*info == 0){ F77_CALL(dpodi)(*hessian, &bdim, &bdim, det, &job); ........ This usually works OK, but with an ill-conditioned data

Hi I am trying to recover the hessian of a problem optimised with box-constraints. The problem is that in some cases, my estimates are very close to the boundary, which will make optim(..., hessian=TRUE) or optimHessian() fail, as they do not follow the box-constraints, and hence estimate the function in the unfeasible parameter space. As a simple example (my problem is more complex though,

Hi, I've been trying to use optim to minimise least squares for a function, and then get a guess at the error using the hessian matrix (calculated from numDeriv::hessian, which I read in some other r-help post was meant to be more accurate than the hessian given in optim). To get the standard error estimates, I'm calculating sqrt(diag(solve(x))), hope that's correct. I've found

Hi , I am trying to get some estimator based on lognormal distribution when we have left,interval, and right censored data. Since, there is now avalible pakage in R can help me in this, I had to write my own code using Newton Raphson method which requires first and second derivative of log likelihood but my problem after runing the code is the estimators were too high. with this email ,I provide

Hey, R Users my windows is 64 bit windows 7.?I am trying to install the package ucminf into my 64 bit version R but cannot.??the package I downloaded is from http://cran.r-project.org/web/packages/ucminf/index.html?and I installed it with the "install from local zip files", due to I did not connect my computer to internet. did anyone meet this problem and is there a version of

> On 10 Feb 2017, at 14:53, G?ran Brostr?m <goran.brostrom at umu.se> wrote: > > Thanks to all who answered my third question. I learned something, but: > > On 2017-02-09 17:44, Martin Maechler wrote: >> >>>> On 9 Feb 2017, at 16:00, G?ran Brostr?m <goran.brostrom at umu.se> wrote: >>>> >>>> In my package 'glmmML'

Hi, my name is Marcel R. Lopes. My problem is, I made a code to calculate the estimates of a Cox model with random effects. Used to optimize the R command for this. The estimates were calculated correctly, but the Hessian matrix does not have good values. The same thing was done in SAS and gave good results for the Hessian Matrix. Where is the problem in R? As the Hessian is calculated?. How

GENERAL REFERENCE ON NONLINEAR OPTIMIZATION? What are your favorite references on nonlinear optimization? I like Bates and Watts (1988) Nonlinear Regression Analysis and Its Applications (Wiley), especially for its key insights regarding parameter effects vs. intrinsic curvature. Before I spent time and money on several of the refences cited on the help pages for "optim",

Dear all R users, Is there any way to calculate hessian matrix of a given function at any given point? Regards [[alternative HTML version deleted]]

Dear UseRs, I wrote the following function to use MLE. --------------------------------------------- mlog <- function(theta, nx = 1, nz = 1, dt){ beta <- matrix(theta[1:(nx+1)], ncol = 1) delta <- matrix(theta[(nx+2):(nx+nz+1)], ncol = 1) sigma2 <- theta[nx+nz+2] gamma <- theta[nx+nz+3] y <- as.matrix(dt[, 1], ncol = 1) x <- as.matrix(data.frame(1,

Hello: Below is a toy logistic regression problem. When I wrote my own code, Newton-Raphson converged in three iterations using both the gradient and the Hessian and the starting values given below. But I can't get nlm() to work! I would much appreciate any help. > x [1] 10.2 7.7 5.1 3.8 2.6 > y [1] 9 8 3 2 1 > n [1] 10 9 6 8 10 derfs4=function(b,x,y,n) {

Does anybody know of a function that implements the derivative (gradient) of the multivariate normal density with respect to the *parameters*? It?s easy enough to implement myself, but I?d like to avoid reinventing the wheel (with some bugs) if possible. Here?s a simple example of the result I?d like, using numerical differentiation: library(mvtnorm) library(numDeriv) f=function(pars, xx, yy)

I'm trying to obtain numerical derivative of a probability computed with mvtnorm with respect to its parameters using grad() and jacobian() from NumDeriv. To simplify the matter, here is an example: PP1 <- function(p){ thetac <- p thetae <- 0.323340333 thetab <- -0.280970036 thetao <- 0.770768082 ssigma <- diag(4) ssigma[1,2] <- 0.229502120

I have a question about passing arguments to the function f that nlm minimizes. I have no problems if I do this: x<-seq(0,1,.1) y<-1.1*x + (1-1.1) + rnorm(length(x),0,.1) fn<-function(p) { yhat<-p*x+(1-p) sum((y-yhat)^2) } out<-nlm(fn,p=1.5,hessian=TRUE) But I would like to define fn<-function(x,y,p) { yhat<-p*x+(1-p) sum((y-yhat)^2) } so

I'd like to be able to test linear hypotheses after setting up and running a model using optim or perhaps nlm. One hypothesis I need to test are that the average of several coefficients is less than zero, so I don't believe I can use the likelihood ratio test. I can't seem to find a provision anywhere for testing linear combinations of coefficients after max. likelihood. Cheers