# Non-centred parameterisation for Dirichlet distribution

**URL:** <https://discourse.mc-stan.org/t/non-centred-parameterisation-for-dirichlet-distribution/1843>\
**Category:** Modeling\
**Created:** [September 10, 2017, 4:38pm UTC](https://discourse.mc-stan.org/t/non-centred-parameterisation-for-dirichlet-distribution/1843 "2017-09-10T16:38:46Z")\
**Posts on this page:** 1\
**Showing post:** 14

<div class="post-metadata">

**Author:** ![stemangiola](https://yyz2.discourse-cdn.com/flex030/user_avatar/discourse.mc-stan.org/stemangiola/32/260_2.png) [@stemangiola](https://discourse.mc-stan.org/u/stemangiola)\
**Post date:** [June 16, 2018, 4:24am UTC](https://discourse.mc-stan.org/t/non-centred-parameterisation-for-dirichlet-distribution/1843/14 "2018-06-16T04:24:30Z")

</div>

> [@Bob\_Carpenter](#):
>
> It’s basically to deal with unmodeled overdispersion. The idea is that you might be able to make it multinomial with the right information, but without that, you’re left expanding the posterior intervals to get calibrated coverage.

With this are you saying that is better off starting from multinomial and build up a more complex model, than using NB framework for coping with the complex overdispersion. Am I understanding well?

(I am trying to implement a gamma-multinomial model [More flexible than Dirichlet-multinomial: gamma-multinomial (?) - #2 by stemangiola](http://discourse.mc-stan.org/t/more-flexible-of-dirichlet-multinomial-gamma-multinomial/4519/2))

> [@Bob\_Carpenter](#):
>
> The other problem with the Dirichlet is lack of covariance structure. But it’s hard to imagine estimating covariance over 10s of thousands of genes’ expression. It would require some kind of low-rank approximation.

For the amount of patients analised in the majority of studies at this stage I am more worried about solving the overdispersion modelling, rather than covariance structure.

---

_[View the full topic](https://discourse.mc-stan.org/t/non-centred-parameterisation-for-dirichlet-distribution/1843)._