search

emvolz-phylodynamics / phyland

Phylogeographic inference with structured coalescent models
Other
0 stars 1 forks source link

Allow constrained fits? #11

Open sdwfrost opened 7 years ago

sdwfrost commented 7 years ago

When sample sizes are small, estimates of effective population sizes can be extremely large. Do you think it would be a good idea to provide upper bounds for Ne?

emvolz commented 7 years ago

In this situation, I would suggest 1) optimising a penalized likelihood, eg by adding a weighted L2 norm of the Ne vector. Unfortunately, it would be difficult to choose a prinicipled value of the weight. Maybe simulation-based cross-val, but complicated 2) MAP, perhaps with default exponential prior for Ne, and user supplies rate parameter What do you think?

Erik M Volz

Department of Infectious Disease Epidemiology Imperial College London erik.volz@gmail.com http://www.erikvolz.info

On 12 September 2017 at 11:12, Simon Frost notifications@github.com wrote:

When sample sizes are small, estimates of effective population sizes can be extremely large. Do you think it would be a good idea to provide upper bounds for Ne?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/emvolz-phylodynamics/phyland/issues/11, or mute the thread https://github.com/notifications/unsubscribe-auth/AOeAn3wfN1o6o7LXitRke7RBIrKhYWuMks5shlkagaJpZM4PUakQ .

sdwfrost commented 7 years ago

When we had similar issues with model fitting (see here) we used a Jeffrey's prior. Might upset a few people, but requires little user input and provides an objective means of smoothing out the very high Ne.

emvolz commented 7 years ago

A potential problem with Jeffrey's is that the estimates we get will depend on the timescale of the tree, and we would prefer that estimates not depend on whether branches are in days, years etc.

What if we penalise the entropy of Ne, since what we really want is to avoid large outliers? \sum^m -log( Ne_i / N )

Erik M Volz

Department of Infectious Disease Epidemiology Imperial College London erik.volz@gmail.com http://www.erikvolz.info

On 12 September 2017 at 12:41, Simon Frost notifications@github.com wrote:

When we had similar issues with model fitting (see here https://github.com/johnros/chords) we used a Jeffrey's prior https://en.wikipedia.org/wiki/Jeffreys_prior. Might upset a few people, but requires little user input and provides an objective means of smoothing out the very high Ne.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/emvolz-phylodynamics/phyland/issues/11#issuecomment-328827149, or mute the thread https://github.com/notifications/unsubscribe-auth/AOeAnxXiO7PZFsrQp-LTVqs5TVuyoHpKks5shm3hgaJpZM4PUakQ .