Penalty on Likelihood Function
Values of the parameter estimates can result in encounter history probabilities outside the range 0-1, or in values of derived parameters that are nonsense. Such a situation often arises with the identity link function, but can also arise with other link functions and the multi-strata model, where the psi values for a particular strata may sum to >= 1. When encounter history probabilities are outside the interval 0-1, or other constraints are not being met, the optimization procedure assesses a penalty to the likelihood in an attempt to find valid parameter estimates that provide valid encounter history probabilities.
Occasionally, this method does not work, so that the final parameter values have a penalty value assessed. The user is warned that the parameter values are not valid. To generate valid estimates, the initial estimates of the parameters should be specified such that valid encounter history probabilities are generated.
The simulated annealing optimization procedure is more robust to finding estimates that meet the contraints than the default optimization method.