Forum: help

RE: maxLik: Return code 100: Initial value out of range. [ Reply ] By: Nancy Chou on 2019-05-16 15:19	[forum:46735]
Hi Arne, I actually solve the problem. It turned out my main variable contains missing data. Once I removed the missing data, I was able to run. Thank you for responding to my e-mail Joanne

RE: maxLik: Return code 100: Initial value out of range. [ Reply ]
By: Arne Henningsen on 2019-05-16 14:36

[forum:46733]

If I run your code, I get an error message:

R> S <- sum(cases) # total number of cases
Error: object 'cases' not found

Please provide a minimal reproducible example.

I guess that you get the error message "Initial value out of range", because the log-likelihood value at the initial parameters is infinite, NA, or NaN. You can check this by:

R> c(as.numeric(glmfit$coefficients), .1)

Please also check your "constraints".

maxLik: Return code 100: Initial value out of range. [ Reply ]
By: Nancy Chou on 2019-05-15 16:37

[forum:46725]

To whom it May Concern,

I'm trying to fit a Poisson excess relative risk models using a paper I came across (https://www.researchgate.net/publication/329865390_Poisson_excess_relative_risk_models_new_implementations_and_software)

In there, the author used the maxLik function to fit the Poisson excess relative risk model.

When I ran the program using my own data, I came across the error term that the initial value out of range. I'm puzzled by the error message and not sure how to problem solve. Below is my code.

S <- sum(cases) # total number of cases
bcvr <- cbind(rep(1, length(cases)), age) # covariance matrix
M <- ncol(bcvr) # number of covariates at the background risk term
Sk <- as.numeric(cases%*%bcvr) # inner product of cases with covariates
doses <- cbind(dose) # RR term covariate
Nd <- ncol(doses) # number of RR term covariates
#log-likelihood
llk <- function(b) sum(log(dpois(cases, py*exp(bcvr%*%b[1:M])*(1 + doses%*%b[(M + 1):(M + Nd)]))))
# gradient of the log-likelihood with respect the model parameters
grd <- function(t){
pyEi <- as.numeric(py*exp(bcvr%*%t[1:M]))
ERRi <- as.numeric(1 + doses%*%t[(M + 1):(M + Nd)])
c( Sk - ((ERRi*pyEi)%*%bcvr), (cases/ERRi)%*%doses - (pyEi)%*%doses )
}
# definition of constrains for ERR parameters: beta1 + beta2 > -1/max(D) and beta1 > -1/max(D)
A <- matrix(c(rep(0, M), 1), ncol=M + 1)
B <- 1/max(dose)
constraints <- list(ineqA = A, ineqB = B)
# model fitting
fitERR <- maxLik(logLik = llk, grad = grd, start = c(as.numeric(glmfit$coefficients), .1), constraints = constraints)

Thanks
Nancy