Uncertainty of estimated proportions

[kvittingseerup]

I was just wondering if it is possible to use the posterior distribution to estimate a confidence/credible interval on the estimated cell type proportions?

Cheers
Kristoffer

[tinyi]

Dear Kristoffer, Sorry for the delay. This is a great question. Confidence interval is not possible, as it is frequentist. Credible interval is possible, but the current version does not output the full posterior samples. I have given some thought about whether to output the full posterior when developing the package, and decided not to do so. My thoughts are as follows. 1) To save the memory so that the memory overhead does not scale up as a function of MCMC chain length. 2) Often the posterior is heavily peaked due to the high sequencing depth (>=1E8) for a typical RNA-seq dataset. 3) There would be very few applications involving the full posterior. Two potential applications that I would imagine would be 1) to test if the fraction of some cell type > 0, or 2) to test if the fraction of cell type A is greater in sample B than that in sample C. These, however, should be tested on biological replicates rather than on MCMC samples, if one thinks carefully about the actual null hypothesis being tested. Besides, due to the dirichlet distribution, all fractions are by nature strictly positive, so it does not make sense to set up test for the "cell type > 0" hypothesis. Let me know if this answers your questions. Also, feel free to share your thoughts on the application on the full posterior. We may also consider adding this feature for future development. Best, Tinyi

…

On Fri, Feb 7, 2020 at 3:06 AM Kristoffer Vitting-Seerup < ***@***.***> wrote: I was just wondering if it is possible to use the posterior distribution to estimate a confidence/credible interval on the estimated cell type proportions? Cheers Kristoffer — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#2?email_source=notifications&email_token=AB4NHS3WK5YJTVKAGCZ526TRBUJALA5CNFSM4KRKHB62YY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4ILXROZQ>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AB4NHS2YODF4CXDHEPNKLC3RBUJALANCNFSM4KRKHB6Q> .

[kvittingseerup]

Hi Tinyi

Thanks for the detailed response - I think your arguments are sound.

Guess the only real usecase would be cases where you have rare cell populations and the difference between existing and non-existing is crucial but hard to determine without the credibility intervals (as all fraction values per default >0).

If you wanted it could be an optional argument (defaulting to false) together with a workflow in the vignette to showcase how to do it.

Cheers
Kristoffer

[tinyi]

Hi Kristoffer, In your case, we should not test the credible interval on the posterior distribution, as dirichlet is strictly positive. We would have to test using the fraction inferred from multiple biological replicates, and test theta > epsilon, for some epsilon > 0. Best, Tinyi

…

On Wed, Feb 12, 2020 at 4:28 AM Kristoffer Vitting-Seerup < ***@***.***> wrote: Hi Tinyi Thanks for the detailed response - I think your arguments are sound. Guess the only real usecase would be cases where you have rare cell populations and the difference between existing and non-existing is crucial but hard to determine without the credibility intervals (as all fraction values per default >0). If you wanted it could be an optional argument (defaulting to false) together with a workflow in the vignette to showcase how to do it. Cheers Kristoffer — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#2?email_source=notifications&email_token=AB4NHSZSDXV5MGUOVWN65Z3RCO6KRA5CNFSM4KRKHB62YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOELQB2CY#issuecomment-585112843>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AB4NHS5LU76UKENGPEBA7NLRCO6KRANCNFSM4KRKHB6Q> .