Estimate parameters by the method of maximum likelihood.
mle2(minuslogl, start, method, optimizer, fixed = NULL, data=NULL, subset=NULL, default.start=TRUE, eval.only = FALSE, vecpar=FALSE, parameters=NULL, parnames=NULL, skip.hessian=FALSE, hessian.opts=NULL, use.ginv=TRUE, trace=FALSE, browse_obj=FALSE, transform=NULL, gr, optimfun,...) calc_mle2_function(formula,parameters, links, start, parnames, use.deriv=FALSE, data=NULL,trace=FALSE)
- Function to calculate negative log-likelihood, or a formula
- Named list. Initial values for optimizer
- Optimization method to use. See
- Optimization function to use. Currently available choices are "optim" (the default), "nlm", "nlminb", "constrOptim", "optimx", and "optimize". If "optimx" is used, (1) the
optimxpackage must be explicitly loaded with
require(Warning: Options other than the default may be poorly tested, use with caution.)
- Named list. Parameter values to keep fixed during optimization.
- list of data to pass to negative log-likelihood function: must be specified if
minusloglis specified as a formula
- logical vector for subsetting data (STUB)
- Logical: allow default values of
minusloglas starting values?
- Logical: return value of
minuslogl(start)rather than optimizing
- Logical: is first argument a vector of all parameters? (For compatibility with
TRUE, then you should use
parnamesto define the parameter names for the negative log-likelihood function.
- List of linear models for parameters. MUST BE SPECIFIED IN THE SAME ORDER as the start vector (this is a bug/restriction that I hope to fix soon, but in the meantime beware)
- (unimplemented) specify transformations of parameters
- List (or vector?) of parameter names
- gradient function
- Further arguments to pass to optimizer
- a formula for the likelihood (see Details)
- Logical: print parameter values tested?
- Logical: drop into browser() within the objective function?
- (stub) list of link functions/parameter transformations ("log"=log/exp, "logit"=plogis/qlogis, etc.)
- Bypass Hessian calculation?
- Options for Hessian calculation, passed through to the
- Use generalized inverse (
ginv) to compute approximate variance-covariance
- user-supplied optimization function. Must take exactly the same arguments and return exactly the same structure as
- (experimental, not yet implemented): construct symbolic derivatives based on formula?
optim optimizer is used to find the minimum of the negative log-likelihood. An approximate covariance matrix for the parameters is obtained by inverting the Hessian matrix at the optimum.
minuslogl argument can also specify a formula, rather than an objective function, of the form
x~ddistn(param1,...,paramn). In this case
ddistn is taken to be a probability or density function, which must have (literally)
x as its first argument (although this argument may be interpreted as a matrix of multivariate responses) and which must have a
log argument that can be used to specify the log-probability or log-probability-density is required. If a formula is specified, then
parameters can contain a list of linear models for the parameters.
- as a list containing lists, with each list corresponding to the starting values for a particular parameter;
- just for the higher-level parameters, in which case all of the additional parameters generated by
model.matrixwill be given starting values of zero (unless a no-intercept formula with
-1is specified, in which case all the starting values for that parameter will be set equal)
- [to be implemented!] as an exhaustive (flat) list of starting values (in the order given by
trace argument applies only when a formula is specified. If you specify a function, you can build in your own
cat() statement to trace its progress. (You can also specify a value for
trace as part of a
control list for
skip.hessian argument is useful if the function is crashing with a "non-finite finite difference value" error when trying to evaluate the Hessian, but will preclude many subsequent confidence interval calculations. (You will know the Hessian is failing if you use
method="Nelder-Mead" and still get a finite-difference error.) If convergence fails, see the manual page of the relevant optimizer (
optim by default, but possibly
constrOptim if you have set the value of
optimizer) for the meanings of the error codes/messages.
An object of class
Do not use a higher-level variable named
parameters -- this is reserved for internal use.
Note that the
minuslogl function should return the negative log-likelihood, -log L (not the log-likelihood, log L, nor the deviance, -2 log L). It is the user's responsibility to ensure that the likelihood is correct, and that asymptotic likelihood inference is valid (e.g. that there are "enough" data and that the estimated parameter values do not lie on the boundary of the feasible parameter space).
The requirement that
data be specified when using the formula interface is relatively new: it saves many headaches on the programming side when evaluating the likelihood function later on (e.g. for profiling or constructing predictions). Since
data.frame uses the names of its arguments as column names by default, it is probably the easiest way to package objects that are lying around in the global workspace for use in
mle2 (provided they are all of the same length).
optimizer is set to "optimx" and multiple optimization methods are used (i.e. the
methods argument has more than one element, or
all.methods=TRUE is set in the control options), the best (minimum negative log-likelihood) solution will be saved, regardless of reported convergence status (and future operations such as profiling on the fit will only use the method that found the best result).
x <- 0:10 y <- c(26, 17, 13, 12, 20, 5, 9, 8, 5, 4, 8) d <- data.frame(x,y) ## in general it is best practice to use the `data' argument, ## but variables can also be drawn from the global environment LL <- function(ymax=15, xhalf=6) -sum(stats::dpois(y, lambda=ymax/(1+x/xhalf), log=TRUE)) ## uses default parameters of LL (fit <- mle2(LL)) fit1F <- mle2(LL, fixed=list(xhalf=6)) coef(fit1F) coef(fit1F,exclude.fixed=TRUE) (fit0 <- mle2(y~dpois(lambda=ymean),start=list(ymean=mean(y)),data=d)) anova(fit0,fit) summary(fit) logLik(fit) vcov(fit) p1 <- profile(fit) plot(p1, absVal=FALSE) confint(fit) ## use bounded optimization ## the lower bounds are really > 0, but we use >=0 to stress-test ## profiling; note lower must be named (fit1 <- mle2(LL, method="L-BFGS-B", lower=c(ymax=0, xhalf=0))) p1 <- profile(fit1) plot(p1, absVal=FALSE) ## a better parameterization: LL2 <- function(lymax=log(15), lxhalf=log(6)) -sum(stats::dpois(y, lambda=exp(lymax)/(1+x/exp(lxhalf)), log=TRUE)) (fit2 <- mle2(LL2)) plot(profile(fit2), absVal=FALSE) exp(confint(fit2)) vcov(fit2) cov2cor(vcov(fit2)) mle2(y~dpois(lambda=exp(lymax)/(1+x/exp(lhalf))), start=list(lymax=0,lhalf=0), data=d, parameters=list(lymax~1,lhalf~1)) ## try bounded optimization with nlminb and constrOptim (fit1B <- mle2(LL, optimizer="nlminb", lower=c(lymax=1e-7, lhalf=1e-7))) p1B <- profile(fit1B) confint(p1B) (fit1C <- mle2(LL, optimizer="constrOptim", ui = c(lymax=1,lhalf=1), ci=2, method="Nelder-Mead")) set.seed(1001) lymax <- c(0,2) lhalf <- 0 x <- sort(runif(200)) g <- factor(sample(c("a","b"),200,replace=TRUE)) y <- rnbinom(200,mu=exp(lymax[g])/(1+x/exp(lhalf)),size=2) d2 <- data.frame(x,g,y) fit3 <- mle2(y~dnbinom(mu=exp(lymax)/(1+x/exp(lhalf)),size=exp(logk)), parameters=list(lymax~g),data=d2, start=list(lymax=0,lhalf=0,logk=0))
Documentation reproduced from package bbmle, version 1.0.17. License: GPL