Change "back-calculation" to "non-mechanistic" infection model throughout

[jamesmbaazam]

At some point we should change back-calculation to nonparametric infection model throughout as that's what we're calling it in the model definition vignette.

Originally posted by @sbfnk in #521 (comment)

[sbfnk]

To be revisited after #345

[jamesmbaazam]

To confirm, this is just a simple re-phrasing exercise right?

[sbfnk]

Yes I think so (only additional thing being gradual deprecation of backcalc_opts()).

[seabbs]

Are we sure non-mechanistic is the right name?

EpiNow2/inst/stan/functions/infections.stan

Line 69 in dbb6536

infections = infections + shifted_cases .* exp_noise;

The model is really a few things right:

1. A pretty standard deconvolution approach but in the Bayesian context
2. An exponential growth model
3. A log link model (this is closest to being a "non-mechanistic model" in work with @SamuelBrand1 we call this a direct infections model)

One option could be to break those things up either with or without an overarching naming wrapper

[seabbs]

Looking into this in the using the estimate_infections vignette we call this non-parametric and in the model definition vignette we call it non-mechanistic

[sbfnk]

The way I see it the key distinction from the default model is that it doesn't use the renewal equation so if we call that the mechanism (for generating new infections) it makes it non-mechanistic. We could call it "non-renewal" instead to be more explicit?

[jamesmbaazam]

Non-renewal

Nice!

[seabbs]

One option would be a larger breaking change to make the interface:

For renewal

estimate_infections(...
  model = rt_opts()
...
)

For deconvolution

estimate_infections(
	model = deconvolution_opt()

For growth rate

estimate_infections(
	model = growth_opts()
)

[sbfnk]

I really like this idea - for names perhaps we could use

estimate_infections(...
  model = renewal_opts()
...
)

instead of rt_opts()?

For growth rate

estimate_infections(
	model = growth_opts()
)

We never implemented this as it's the same as renewal with gt = Fixed(1).

[sbfnk]

Noting that the original issue was about wording in the docs so whilst obviously related the change of options might be better suited to a separate issue/PR combo.

[sbfnk]

non-renewal

On reflection I think this is the most explicit but it's not what people will search for or be familiar with when choosing this model. That would be an argument of calling it "deconvolution".

[seabbs]

My point was really I would be favour of closing this as a not do and update the interface instead.

We never implemented this as it's the same as renewal with gt = Fixed(1).

also for another issue but this would open up having syntax sugar to make this more visible.

[sbfnk]

My point was really I would be favour of closing this as a not do

I think we still want to have consistent wording in the doc (which is what the original issue was) - if we change the interface to use deconvolution_opts() and call it backcalculation in the docs I think it'll just be confusing.

[seabbs]

yes but changing the interface changes the naming as there wont be two binary categories anymore

[sbfnk]

sorry, I'm not following - isn't it still two categories of renewal_model() vs. deconvolution_model() so we should update the language in the vignettes etc. accordingly?

[sbfnk]

pushing this to the next release as it's turning into a breaking change and we already have enough things to deprecate.

[seabbs]

L> so we should update the language in the vignettes etc. accordingly?

because there is more than one model in the decovolution approach. I think this is all getting a bit lost. Essentially we need to make the docs match the code and as a follow up issue consider a different interface where we nest models less and each has its own opts function