Case studies: Main paper structure example

[sbfnk]

The example in Section 3 discusses an AR/ARMA model as a type of AbstractLatentProcess. It's not clear why this was chosen but I think it's a problem that this is not an infectious disease epidemiological / biological example but a generic statistical model that isn't even that common in Epi. I think the section should have an example for an AbstractEpiModel, an AbstractLatentModel and an AbstractObservationModel - and if it can only have one it should be an AbstractEpiModel.

[seabbs]

This was discussed here for context: #50 (comment)

The original intent of this example is that it is simple, is a clear case of using the composability, and is then built on by the majority of the other case studies. I agree that having moved this into the SI it makes it less clear that it was an obvious choice (see the linked discussion).

Do you have any minimal change suggestions to make the current organisation work better?

Do you also have any more concrete suggestions for a more major shift as hinted at already? In particular, we need something that expresses the core ideas without being overly complicated and without duplicating the case studies.

My personal preference is to nuance what we have but open to bigger change suggestions. Something to reflect on is that if this is changed out then something will need to be done to the case studies as they are all built around this at the moment.

[seabbs]

@sbfnk please let me know if you have a suggested action before the first preprint and then anything else for the future.

[sbfnk]

I realise I failed to express an opinion when this was discussed earlier but I for one would bring the case studies back to the main text as opposed to just showing the ARIMA. I worry that infectious disease modellers will see this and consider it not relevant to their work.

[seabbs]

Can you expand what you mean here? Do you mean bring them back as they are to the main text in full? Or expand on how they are discussed in the main text at the moment? Or replace the current ARIMA case study with a different case study or other?

is there a way to address your concerns without major structure changes i.e. using better is signposting/flavour text?

[sbfnk]

I don't think we can address the concern that we're focusing on a statistical model in a paper that is supposed to be for epidemiologists without major structure change so let me try to suggest an alternative improvement instead.

In
https://github.com/EpiAware/ComposableProbabilisticIDModels/blob/5ed057a8cef53b5a02695ae07c9b43a61d30fe30/index.qmd#L235C1-L236C1
I know it's mentioned before but perhaps you can go then say: "This model will appear in the epidemiological case studies to describe the trajectory of the reproduction number."

And then later perhaps before

ComposableProbabilisticIDModels/index.qmd

Line 316 in 5ed057a

Other latent models extend modelling options through combining models additively, multiplicative scaling, and piecewise processes.

add:
"We will later use this latent process to generates trajectories of the time-varying reproduction numbers $R_t$. This can then linked with a model for incidence of new infections, where the model of $R_t$ is composed with a model of infection dynamics such as Renewal(), which implements the discrete-time renewal equation $I_t = R_t \sum_s g_s I_{t-s}$ to model incidence $I_t$."

[seabbs]

Yes, agree. To be clear though you think it would be better to restructure. Do you have a recommendation for that?

I don't think we can address the concern that we're focusing on a statistical model in a paper that is supposed to be for epidemiologists without major structure change so let me try to suggest an alternative improvement instead.

This seems quite negative so it would be useful to clarify.

[seabbs]

f116ab8 adds additional linking structure and moves the figure as this now needs to be referenced in the ARIMA workflow. The slight issue with this is the case studies aren't discussed until later but I think that is fine.

[sbfnk]

I don't think we can address the concern that we're focusing on a statistical model in a paper that is supposed to be for epidemiologists without major structure change so let me try to suggest an alternative improvement instead.>
This seems quite negative so it would be useful to clarify.

It was intended as clarification of

is there a way to address your concerns without major structure changes
in relation to the specific concern.

There's a trade-off here between reducing additional amount of time/work on this and the potential unknown benefit of substantial restructures, and therefore good reasons to keep what we have in place now. I also haven't engaged deeply with previous versions so am likely missing at least some of the rationale for the choice of showing the ARIMA here. But I do worry that an infectious disease modeller (our main audience) will see the ARIMA and decide it's not for them before they get to the supplement which contains the epi parts.

In response to

Do you have a recommendation for that?

Could we elaborate the renewal and an observation model in a similar style to the way AR is elaborated now, move the AR content to the supplement, and show and example of how the three form an EpiProblem? Or would some nuance get lost?

[seabbs]

Did the minor changes linked here go anyway to addressing those concerns? They are similar to the ones flagged in that discussion issue.

Yes, we could do that. The tradeoffs remain the same as in the original issue. What do you think of those?

• The ARIMA is a composition of several submodels that interact and can then be reused. This is not the case of composing a EpiProblem as they are currently written. If we had a more flexible Renewal this could also be done but we don't at the moment.
• It is nice as an example due to its simplicity, its reuse as a submodel across all the other case studies, and it doesn't require much/any prior knowledge or narrow us into a form of model i.e. a renewal. These were the main points that motiovated choosing it as a demonstrator for the our approach text. I think a particular worry is making people overly focus on the renewal.

What do you mean by "get to the SI"? The case studies are discussed in the main text so do you mean that they have to wait for it or that there isn't enough in those for a modeller to engage? Is there something more that a modeller would need to engage with them? Perhaps the equations? Or is the major concern that regardless of changes they won't have the patience to wait so long?

Something to perhaps note is that their is a schematic composition example in the our approach summary section which gives a birds eye/no code view of what this looks like.

To be honest I think having any code at all in the paper may lose people as well but I don't see that we can avoid it really and have any chance of communicating what we are doing.

Another option we have here is to bring in the block building schematic from the Wellcome slides. Would that help people stay engaged? I personally found it a bit simplified to be that interesting but perhaps good for others?

[sbfnk]

Did the minor changes linked here go anyway to addressing those concerns? They are similar to the ones flagged in that discussion issue.

Yes.

Yes, we could do that. The tradeoffs remain the same as in the original issue. What do you think of those?

Yes, I agree it's a nice example. Perhaps it shouldn't be the only one.

The case studies are discussed in the main text so do you mean that they have to wait for it.

It's people skimming / reading and thinking that this isn't for their kind of modelling and thus doesn't apply.

Is there any reason not to have one example for latent nonmechanistic and one for latent mechanistic and one for observation model? We could perhaps shorten the ARIMA to enable/balance that.

[seabbs]

Is there any reason not to have one example for latent nonmechanistic and one for latent mechanistic and one for observation model? We could perhaps shorten the ARIMA to enable/balance that.

This would essentially be bringing in the case studies wouldn't it? Do you think it could be done in a shorter way than that? The answer to why not those was losing people via length.