An opportunity is offered to calculate likelihood profiles for
specified parameters.
Likelihood profiling proceeds by sequentially fixing the parameter
of interest at one of a vector of predefined values for that parameter,
then maximizing the likelihood of that value with respect to all
the other parameters.
A vector of likelihood values is created corresponding to each
fixed parameter value.
The vector of negative ln-likelihoods is transformed and scaled
to a vector of likelihoods from which a probability distribution
of parameter values can be created and plotted.
This so-called likelihood profile can be used to draw statistical
inferences concerning the parameter’s values.
In particular, likelihood profiles facilitate a likelihood ratio
test which permits an analyst to make a probabilistic statement
of confidence concerning a parameter’s value with respect
to a postulated value.
A likelihood ratio test is based on the premise that twice the
difference in negative ln-likelihood between two competing nested
models (let’s say A and B) is chi-square distributed with
the degrees of freedom determined by the difference in the number
of free parameters between the two competing models.
In the case of likelihood profiling, model A is any accepted fit
to the model with an associated value for the parameter of interest
and model likelihood, model B is any one of the values in the vector
of profiled model likelihoods corresponding to a fixed value of
the parameter of interest, and there is one (1) degree of freedom.
By the way, comparison of a negative ln-likelihood profile with
its corresponding probability distribution nicely illustrates the
relationship between the shape of the negative ln-likelihood profile
and the corresponding probability distribution.
A Gaussian (normal) probability distribution has an associated
quadratic negative ln-likelihood profile, and the second partial
derivative of the likelihood function with respect to the parameter
is constant, as expected for a quadratic function.
An asymmetric negative ln-likelihood profile results in a skewed
associated probability distribution, which when calculating a covariance
matrix numerically will result in inconsistent estimates of
parameter covariances across the three grids of numerically calculated
second partial derivatives.
Smartstats© will guide you through the procedure for calculating
likelihood profiles.
The numerical output of the procedure to calculate the likelihood
profiles can be saved in a text file.
|