Back to the Top
The following message was posted to: PharmPK
Hello,
can somebody comment on what should happen to SSEs when you "fix" a
parameter in a PK model, and why?
thanks
Ravi Kuppuraj
Back to the Top
The following message was posted to: PharmPK
Ravi Kuppuraj writes:
can somebody comment on what should happen to SSEs when you "fix" a
parameter in a PK model, and why?
That depends on what you fix it to. If you fix a value to its maximum
likelihood estimate, SSE should be unaffected. If you fix it to
something other than SSE than then SSE should increase. Now, MSE, on
the other hand, will change. The denominator of MSE is (n-p) where p
is the number of estimable parameters. Because you have fixed MSE, p
is smaller and so MSE will increase. Hope this helps,
Pete Bonate
Back to the Top
The following message was posted to: PharmPK
Dear Ravi Kuppuraj,
With respect to your question:
> can somebody comment on what should happen to SSEs when you "fix" a
> parameter in a PK model, and why?
SSE (Sum of squared errors) will always increase if a parameter is
fixed during fitting of a model. This is simply so because fixing a
parameter does not allow to explore the entire parameter space to find
the minimum SSE. So there will be always a different value for the
fixed parameter resulting in a lower SSE (unless you fixed the value at
the minimum, but this makes no sense; see below).
Please note that the SSE is only one possible 'objective function' to
be minimized during fitting; e.g. the objective function may be 'minus
two log-likelihood' or any expressed related to it (including SSE). The
numerical value of the objective function is usually not relevant; it
is used only to find the minimum (and AIC, see below).
In case of ordinary least=squares (OLS) or weighted least-squares
(WLS), 'minus two log-likelihood' can be replaced by log(SSE) and
log(WSSE), respectively (log refers to natural logarithm with base e),
leaving out various constant terms.
Whether or not the change of the objective function by fixing or
varying a particular parameter is 'relevant', is usually judged from
Akaike's Information Criterion (AIC):
AIC = -2 log(likelihood) + 2 P
where P is the number of estimated parameters. Comparing two models,
the one with lowest AIC is the 'best' model.
So, fixing a parameter decreases P and increased -2 log(likelihood).
Besides, fixing a parameter implies that the number of assumptions
about the model increases; one investigates only a restricted model.
This can be done only if there is some plausible value for that
parameter, e.g. a value obtained from an independent source. Fixing a
parameter simply because the fitting procedure does not provide a
plausible estimate (e.g. a negative or zero value) is definitely not
good practice! (although reporting a negative PK parameter is also not
recommended). Here we are in the darker side of modeling ...
Best regards,
Hans Proost
Johannes H. Proost
Dept. of Pharmacokinetics and Drug Delivery
University Centre for Pharmacy
Antonius Deusinglaan 1
9713 AV Groningen, The Netherlands
tel. 31-50 363 3292
fax 31-50 363 3247
Email: j.h.proost.aaa.farm.rug.nl
PharmPK Discussion List Archive Index page
Copyright 1995-2010 David W. A. Bourne (david@boomer.org)