|
|
 Up |
|
|
  |
Author: AnjaAnja Date: Mar 21, 2008 14:45
Hello everyone,
I have used a Gibbs Sampler to generate a posterior distribution. The
three parameters I had was mean, variance and standard deviation.
Is there a way to tell if there is a correlation between the
parameters from looking at the posterior distribution.
Thanks,
Anja
|
| |
|
| | 5 Comments |
|
|
Author: FrancogrexFrancogrex Date: Mar 21, 2008 16:07
On Mar 21, 10:45Â pm, Anja googlemail.com> wrote:> I have used a Gibbs Sampler to generate a posterior distribution. The
> three parameters I had was mean, variance and standard deviation.
> Is there a way to tell if there is a correlation between the
> parameters from looking at the posterior distribution.
If you mean by looking at the chains, then I don't think so, you'll
have to test for correlation. you'll also have to plot an
autocorrelation plots to detect if there is high autocorrelation. if
so consider thinning the chains.
|
| |
|
| | no comments |
|
|
Author: AnjaAnja Date: Mar 21, 2008 18:11
On Mar 21, 11:07Â pm, Francogrex grex.org> wrote:> On Mar 21, 10:45Â pm, Anja googlemail.com> wrote:
>>> I have used a Gibbs Sampler to generate a posterior distribution. The
>> three parameters I had was mean, variance and standard deviation.
>> Is there a way to tell if there is a correlation between the
>> parameters from looking at the posterior distribution.
>
> If you mean by looking at the chains, then I don't think so, you'll
> have to test for correlation. you'll also have to plot an
> autocorrelation plots to detect if there is high autocorrelation. if
> so consider thinning the chains.
So are you suggesting that the only way is to sample the distribution
and then do the correlation test?
Thanks,
Anja
|
| |
| no comments |
|
|
Author: FrancogrexFrancogrex Date: Mar 22, 2008 03:39
On Mar 22, 2:11Â am, Anja googlemail.com> wrote:> So are you suggesting that the only way is to sample the distribution
> and then do the correlation test?
yes, but something is not clear. You know that the variance and the
standard deviation of a distribution are related and also that the
variance and the mean are related... I do not understand why you want
to do a correlation test for those three parameters coming from one
distribution (I am supposing it's a univariate normal distribution).
|
| |
| no comments |
|
|
Author: AnjaAnja Date: Mar 22, 2008 06:21
On Mar 22, 10:39Â am, Francogrex grex.org> wrote:> On Mar 22, 2:11Â am, Anja googlemail.com> wrote:
>>> So are you suggesting that the only way is to sample the distribution
>> and then do the correlation test?
>
> yes, but something is not clear. You know that the variance and the
> standard deviation of a distribution are related and also that the
> variance and the mean are related... I do not understand why you want
> to do a correlation test for those three parameters coming from one
> distribution (I am supposing it's a univariate normal distribution).
Actually I made a mistake in my original post. I was trying out some
Bayesian linear regression and the parameters were the gradient,
intercept and the variance. Now, I am trying to see if I can find some
correlation among these.
Thanks,
Anja
|
| |
| no comments |
|
|
|
|
|
Author: Herman RubinHerman Rubin Date: Mar 22, 2008 18:43
>On Mar 21, 10:45=A0pm, Anja googlemail.com> wrote:>> I have used a Gibbs Sampler to generate a posterior distribution. The
>> three parameters I had was mean, variance and standard deviation.
>> Is there a way to tell if there is a correlation between the
>> parameters from looking at the posterior distribution.
>If you mean by looking at the chains, then I don't think so, you'll
>have to test for correlation. you'll also have to plot an
>autocorrelation plots to detect if there is high autocorrelation. if
>so consider thinning the chains.
When interested in many items when sampling, use the
same random variables to approximate all nof them.
Then one can answer such questions.
|
| Show full article (1.31Kb) |
| no comments |
|
|
