Version 2

[seabbs]

We discussed the fact this still seems like a concept people are struggling with. This issue is for planning a version 2.

I thought a sensible approach would be:

• [ ] Simulate several plausible epidemics
• [ ] Simulate observed data with short, medium, and long delays
• [ ] Implement backwards sampling
• [ ] Implement deterministic deconvolution
• [ ] Implement a fixed delay
• [ ] Implement GP powered deconvolution with fixed delays
• [ ] Visual each one against the simulated data
• [ ] explore impact of misspecification with each method, both in terms of getting mean/variance of the delay distribution wrong as well as what happens if the delay distribution is e.g. bimodal.
• [ ] score the estimates, including over time, using CRPS/similar
• [ ] Comment on additional complexities such as reporting periodicity, delay censoring when estimated from linelists, and use cases for deconvolution vs working directly from a generative process for your given analysis question.
• [ ] Comment on available methods

There is a question for how much this should be linked to reproduction number estimation and/or if it should be treated as its own topic. As it comes up in many applications I would perhaps argue for the general case and link to applications (like CFR, Rt etc) but can see it being easier to show bias if using an application directly.

Resources:

• This work
• Gist: https://gist.github.com/seabbs/fb1bc9c79c3dd7117f9314cb97e71615
• Gostic et al.: https://pubmed.ncbi.nlm.nih.gov/33301457/
• https://smw.ch/article/doi/smw.2020.20307

Work to do

• [ ] What is the scope/length of this piece
• [ ] Decide on scenarios and generate (perhaps starting from gist code)
• [ ] Implement a simplified GP stan model for deconvolution (borrow from inc2prev and/or epinowcast)
• [ ] Implement backsampling (gist code helps here)
• [ ] Run all methods and plot
• [ ] Write draft
• [ ] Discuss extensions

[sbfnk]

Yes, I think that covers pretty much all we need to cover.

I thought a sensible approach would be:

Also (trivially):

• [ ] Implement fixed delay

This I'd leave as a discussion point / follow-up paper as it adds quite a lot of complexity:

• [ ] Scope more complex versions with time-varying or uncertain delays

There is a question for how much this should be linked to reproduction number estimation and/or if it should be treated as its own topic. As it comes up in many applications I would perhaps argue for the general case and link to applications (like CFR, Rt etc) but can see it being easier to show bias if using an application directly.

I'd tend the same way

The other things that I think would be potentially interesting to explore (but might explode scope) are:

• [ ] explore impact of misspecification with each method, both in terms of getting mean/variance of the delay distribution wrong as well as what happens if the delay distribution is e.g. bimodal.
• [ ] score the estimates, including over time, using CRPS/similar

[seabbs]

Yes agree,

I like the idea of doing a bit of scoring but might save until other things are in place to see what it adds. I think looking at misspecification makes sense but would want to try and keep it as simple as possible I think.

I might look at this once the surrogate draft is done as getting at least the bones in place should be quite fast. I thought this might be quite a nice project to bring someone else in on, for example Hannah.