NONMEM Users Network Archive

Hosted by Cognigen

Re: Bootstrap analysis

From: Leonid Gibiansky <LGibiansky>
Date: Tue, 21 Apr 2009 17:43:02 -0400

I do not think that bootstrap mean is a useful statistics. Median of the
bootstrap distribution could/should be compared with the final-model
point estimates. Precision of the parameter estimates can be evaluated
as a 95% confidence interval defined as an interval between 2.5 and 97.5
percentiles of the bootstrap distributions. Another useful application
of the bootstrap parameter estimates is the investigation of the
correlation between those. A scatter-plot matrix of parameters versus
parameters readily reveals existing correlations. If those are very
strong, the model could/should be improved to remove
over-parameterization. Histograms can be useful if you put the
final-model estimate on top of those distributions. It should be
somewhere close to the center of the bootstrap parameter distribution.
If you like some p-value, you can compute the one-sided or two-sided
probability of observing the value as extreme as the final-model
parameter estimates based on the bootstrap distribution (using a percent
of bootstrap estimates that are below or above the final-model. estimate).

Thanks
Leonid



--------------------------------------
Leonid Gibiansky, Ph.D.
President, QuantPharm LLC
web: www.quantpharm.com
e-mail: LGibiansky at quantpharm.com
tel: (301) 767 5566




Nick Holford wrote:
> Varsha,
>
> Congratulations on discovering how to use a bootstrap to evaluate the
> distribution of your model parameter estimates.
>
> The bootstrap mean is probably a more robust estimate of the true value
> of the parameter than the value estimated from the original data. I
> prefer to report the bootstrap mean for this reason.
>
> The uncertainty, e.g. 95% confidence interval, can sometimes be useful
> for model evaluation but more commonly is is best used to keep journal
> reviewers 'happy'. There are very few other real applications of knowing
> the uncertainty of a single parameter but it might be used to try to
> demonstrate that a PD parameter (e.g. Emax) is different from zero and
> thus indicate that the drug does something useful.
>
> The good news is that you don't have to worry about using bootstraps "to
> confirm the fact that the model I have is the best fit for the data".
> The bootstrap can never confirm this for you. You need to buy a
> subscription to 'Talk to God' in order to get that kind of information.
>
> Nick
>
>
> Varsha Mehta wrote:
>> Group:
>>
>> I have bootstrap analysis (my first) parameter estimates and model
>> parameters. The PDxPOP/NONMEM manual I have does not provide
>> any guidance as to how I can statistically compare these two (or do I
>> need to?). I also have histograms for the thetas in bootstrap analysis.
>> I can make some visual judgements but is there a way to statistically
>> compare the two results (bootstrap v model) built in to the NONMEM
>> that I can use to quickly get some statistical comparison results?
>>
>> How else can I use the bootstrap results to confirm the fact that the
>> model I have is the best fit for the data?
>>
>>
>> Thanks in advance.
>>
>> Varsha Mehta, MS(CRDSA), Pharm.D., FCCP
>> Clinical Associate Professor
>> Pharmacy, Pediatrics and Communicable Diseases
>> Clinical Pharmacist Neonatal Critical Care
>> University of Michigan
>> (O) 734-936-8985
>> (F) 734-936-6946
>> varsham
>>
>> **********************************************************
>> Electronic Mail is not secure, may not be read every day, and should
>> not be used for urgent or sensitive issues
>>
>
Received on Tue Apr 21 2009 - 17:43:02 EDT

The NONMEM Users Network is maintained by ICON plc. Requests to subscribe to the network should be sent to: nmusers-request@iconplc.com.

Once subscribed, you may contribute to the discussion by emailing: nmusers@globomaxnm.com.