# Profile Likelihood Computation and Confidence Intervals for `twinstim`

objects

`twinstim_profile.Rd`

Function to compute estimated and profile likelihood based confidence
intervals for `twinstim`

objects. Computations might be cumbersome!

WARNING: the implementation is not well tested, simply uses
`optim`

(ignoring optimizer settings from the original fit),
and does not return the complete set of coefficients at each grid point.

## Arguments

- fitted
an object of class

`"twinstim"`

.- profile
a list with elements being numeric vectors of length 4. These vectors must have the form

`c(index, lower, upper, gridsize)`

.`index`

:index of the parameter to be profiled in the vector

`coef(fitted)`

.`lower, upper`

:lower/upper limit of the grid on which the profile log-likelihood is evaluated. Can also be

`NA`

in which case`lower/upper`

equals the lower/upper bound of the respective 0.3 % Wald confidence interval (+-3*se).`gridsize`

:grid size of the equally spaced grid between lower and upper. Can also be 0 in which case the profile log-likelihood for this parameter is not evaluated on a grid.

- alpha
\((1-\alpha)\%\) profile likelihood based confidence intervals are computed. If alpha <= 0, then no confidence intervals are computed. This is currently not implemented.

- control
control object to use in

`optim`

for the profile log-likelihood computations. It might be necessary to control`maxit`

or`reltol`

in order to obtain results in finite time.- do.ltildeprofile
If

`TRUE`

calculate profile likelihood as well. This might take a while, since an optimisation for all other parameters has to be performed. Useful for likelihood based confidence intervals. Default:`FALSE`

.- ...
unused (argument of the generic).

## Value

list with profile log-likelihood evaluations on the grid, and
– not implemented yet –
highest likelihood and Wald confidence intervals.
The argument `profile`

is also returned.

## Examples

```
# profiling takes a while
if (FALSE) { # \dontrun{
#Load the twinstim model fitted to the IMD data
data("imdepi", "imdepifit")
# for profiling we need the model environment
imdepifit <- update(imdepifit, model=TRUE)
#Generate profiling object for a list of parameters for the new model
names <- c("h.(Intercept)","e.typeC")
coefList <- lapply(names, function(name) {
c(pmatch(name,names(coef(imdepifit))),NA,NA,11)
})
#Profile object (necessary to specify a more loose convergence
#criterion). Speed things up by using do.ltildeprofile=FALSE (the default)
prof <- profile(imdepifit, coefList,
control=list(reltol=0.1, REPORT=1), do.ltildeprofile=TRUE)
#Plot result for one variable
par(mfrow=c(1,2))
for (name in names) {
with(as.data.frame(prof$lp[[name]]),
matplot(grid,cbind(profile,estimated,wald),
type="l",xlab=name,ylab="loglik"))
legend(x="bottomleft",c("profile","estimated","wald"),lty=1:3,col=1:3)
}
} # }
```