From: Gastonguay, Marc <*marcg*>

Date: Sun, 29 Jan 2012 18:02:56 -0500

Dear Charles,

When using $PRIOR, the df specifies the degrees of freedom for the

Inverse-Wishart prior distribution on the covariance matrix of the

individual random effects (OMEGA). In practical terms, the larger the df,

the more informative the prior will be. It's not surprising, then, that

your parameter estimates approach the results of the previous model when df

is large.

The choice to use an informative prior distribution should not be made

based on objective function changes. Instead, you might want to consider

the rationale for including the prior information in the first place. One

strategy would be to use informative priors only where necessary to support

components of a previously defined or otherwise known model. If the new

data alone do not support estimation of parameters required in this model

structure, then you might want to include informative prior distributions

on those specific components. Sensitivity to the "informativeness" of the

prior could be explored by varying the df (or prior variance for fixed

effects), and conclusions from your analysis should probably be viewed in

the context of this prior sensitivity.

As you indicate, for this particular example, you may not need to use

$PRIOR at all, and could simply pool the data in a single analysis.

Hope this is useful.

Marc

On Sun, Jan 29, 2012 at 5:02 PM, Charles Steven Ernest II <

ernest_charles_steven_ii

*> I have previously conducted a meta-analysis of PK data that contained
*

*> extensive and sparse sampling from 330 patients with a run time of ~ 4
*

*> days. I know have data from another 200 patients with sparse data. I have
*

*> created the median and 95th PI from the previous model and overlaid the
*

*> current data. The results demonstrate that the new data is well described
*

*> by that model. When the new data is fit with that model, the data does not
*

*> support using the model as some parameters were unidentifiable. I could
*

*> conducted an analysis of all the data simultaneously but was interested in
*

*> another method. Therefore, I have implemented $PRIOR into the model and
*

*> noticed that with each successive increase of the df, the objective
*

*> function significantly decreases. However, the THETA values do not change
*

*> much and are different from the prior estimates used. The only other
*

*> things that changed besides the objective function were the estimates and
*

*> SE of the covariance terms and the BSV estimate of the peripheral volume of
*

*> distribution. These values become more in line with the those observed
*

*> previously, and the correlation values between them becomes stronger. My
*

*> question is it justifiable to use such a high df (df=330) based on these
*

*> significant decreases of objective function and covariance as the
*

*> information from this meta-analysi would be highly informative.
*

*> Thanks
*

*>
*

--

Marc R. Gastonguay, Ph.D. <marcg

Scientific Director

Metrum Institute <http://metruminstitute.org>

*Metrum Institute is a 501(c)3 non-profit organization.*

Received on Sun Jan 29 2012 - 18:02:56 EST

Date: Sun, 29 Jan 2012 18:02:56 -0500

Dear Charles,

When using $PRIOR, the df specifies the degrees of freedom for the

Inverse-Wishart prior distribution on the covariance matrix of the

individual random effects (OMEGA). In practical terms, the larger the df,

the more informative the prior will be. It's not surprising, then, that

your parameter estimates approach the results of the previous model when df

is large.

The choice to use an informative prior distribution should not be made

based on objective function changes. Instead, you might want to consider

the rationale for including the prior information in the first place. One

strategy would be to use informative priors only where necessary to support

components of a previously defined or otherwise known model. If the new

data alone do not support estimation of parameters required in this model

structure, then you might want to include informative prior distributions

on those specific components. Sensitivity to the "informativeness" of the

prior could be explored by varying the df (or prior variance for fixed

effects), and conclusions from your analysis should probably be viewed in

the context of this prior sensitivity.

As you indicate, for this particular example, you may not need to use

$PRIOR at all, and could simply pool the data in a single analysis.

Hope this is useful.

Marc

On Sun, Jan 29, 2012 at 5:02 PM, Charles Steven Ernest II <

ernest_charles_steven_ii

--

Marc R. Gastonguay, Ph.D. <marcg

Scientific Director

Metrum Institute <http://metruminstitute.org>

*Metrum Institute is a 501(c)3 non-profit organization.*

Received on Sun Jan 29 2012 - 18:02:56 EST