So here is the announcement for our next joint meeting on 17 November 2022 at 1pm (UK time).
Sam is currently trying to set up a permanent zoom link which we can then always use for our meetings in the future.
Anyone who is interested is welcome to attend this public meeting and also to suggest topics for discussion here in the thread. For example, we are very happy to discuss ideas for future developments or problems you have stumbled across while using or developing Epinowcast.
Of course, we can also make excellent use of the meetings to present newly developed functionality, for example, I plan to briefly present new plotting functionality of epinowcast at the next meeting!
I will prepare a small agenda and post more details here a few days before the meeting!
Sadly Sam has failed to organise a meeting link or finish the mission statement draft .
Unfortunately I also have an unavoidable and out of my control clash at 2.30 so will have to dash unless we can push to tomorrow (which I assume would not be good for others?). My bad!
My excuse is that I’ve been very distracted working with @sangwoopark on estimating censored delay distributions truncated by an underlying outbreak process. It is pretty fascinating stuff and I have lots of questions about what it means for the kind of analysis we often do. Maybe someone else has a better handle on this!
This paper by Shaun Seaman has some fascinating stuff in it (including the explicit connection to nowcasting) but also nicely room for more to be done. I was thinking of asking him to come and present on it at our monthly seminar (that does mean I need to organise it!).
@sbfnk suggested we use Ebola lineliest data for the case study in this work and it turns out to be the perfect target for nowcasting (including with confirmed/probable case complexity). The repo has code to process the data, visualise etc if wanting to play.
Exciting work @samabbott and @sangwoopark! I also like your way of plotting the observations by secondary event as stacked bars with different observation time points.
Do you think your simulation code is something that we could export into a separate package for epinowcast? I am still planning on doing this with my simulation code for the nowcasting, but maybe yours is better by now, or we can combine both.
Ah I don’t. Would it work for everyone if we pushed to the same time tomorrow? I know that is very bad but perhaps just this once.
Thank you @adrianlison - I was very pleased with myself! Yes, I think it would be great to pull this all together somewhere. We were also thinking of making our actual delay distribution estimation tooling stuff more robust and so having that as a package as well.
Methods for estimation of reporting delays under right truncation
Basically see @samabbott post above. Together with @sangwoopark, they have been working on estimation of delay distributions from right-truncated and possibly censored data. The methods used are closely related to nowcasting, and there may be some hidden potential to avoid a joint model of the epidemic curve and the reporting delay in certain situations, and use much more efficient models instead. Currently, Sam is still in a, quote, “state of confusion” about this, so this hunch has to be developed further in the future.
There may also be something going on with insights about the individual-level relationship between forward and backward delays that we could use to improve our renewal models in the future (e.g. modeling changes in the generation time distribution).
Renewal models & growth rate models
We had a short but lively discussion on the pros and cons of renewal models vs. growth rate models. Main points:
If you want to be negative, you can view (daily) growth rate models as renewal models with per default misspecified generation interval distribution (fixed, one day generation interval). For any given R trajectory, they require a more complex time series smoothing prior to account for this misspecification (which may also have distributional assumptions). In fact, if you use an AR process to model the growth rate, this becomes a renewal model in which the generation time distribution is also estimated.
By being “misspecified per default” and trying to account for this through an appropriate time series prior, growth rate models can in principle be more robust compared to simple renewal models which assume a known and fixed generation interval distribution.
It can be harder to identify very sharp changes in transmission dynamics from a growth rate model. Especially if you have some additional assumptions about transmission (e.g. further covariates, changepoints etc.) renewal models are better suited.
Growth rate models are probably faster to fit, however there may be issues if you have outlier days with unusually high or low infections. Renewal models have some natural smoothing over such outliers, growth rate models may get stuck more easily. This may be relaxed in both cases with stochastic modeling of infections.
Independent of the modeling, growth rates may be useful to compare between models using different generation time assumptions.
An epinowcast instance for mastodon
Sam is planning to host a mastodon instance and dedicate it to the epinowcast platform. This would likely be an invite-only space without moderation. If you are interested, reply to this thread or get in touch directly.