Re: calcuating goodness of fit in the statistical tool box

Peter Perkins <Peter.PerkinsRemoveThis@xxxxxxxxxxxxx> wrote
in message <fdu91v$g9g$1@xxxxxxxxxxxxxxxxxx>...
Jing Cao wrote:

the fitting, can I use the loglikelihood as an indicator for
comparing my models? (I used normal, lognormal, and Weibull
for the same dataset - the results are more or less
similar, but I still want to rank the results with some kind
of indicator of the fit).

Those are alltwo-parameter models, but none is a special
case of the
others. So while technically, you should not be using a
likelihood
ratio test, you could compute the AIC using the
log-likelihoods.

However, I don't recommend doing (only) that. A single
number cannot
possibly tell you why/how a model doesn't fit as well as
another. If
you're already using DFITTOOL, take the opportunity to
make plots of the
PDF against the data, or better yet, CDFs or probability
plots. Look
and see _how_ the various model do or don't fit the data
-- skewness?
heavy tails? and so on. Decide whether any discrepancies
are important
or not.

Hope this helps.

Thank you for the confirmation, Peter! I have already done
the things you suggested in the toolbox and decided Weibull
to be the best fit, but am tempted to produce a single
number which can sum up the conclusion. I really appreciate
your confirmation and suggestions. I will use the AIC as
one of the indicators, and make my conclusion based on more
aspects of the fitting results. Thanks again!


.

Relevant Pages

  • Re: Distribution fitting
    ... > Peter Perkins wrote: ... >> Jan Van den Bulcke wrote: ... >>> like a distribution. ... >> which, if you plotted them against 1:n, look like a Weibull (or ...
    (comp.soft-sys.matlab)
  • Re: Weibull Distribution but with a varied cycle time
    ... line on a Weibull plot can be used to tell us whether the failure rate ... Weibull paper provides ... an easy way of finding the slope of the line along with printed ... I use Kaleidagraph for making plots of data, ...
    (sci.stat.math)