How-to case studies

[kaitejohnson]

The first case study will be led by Laura and focused on syndromic surveillance data (I think NSSP ED visits due to a specific syndrome definition). Here are some options for the second case study, these are highly motivated (i.e. biased) by some of the things we have discussed in the context of baselinenowcast:

• nowcasting NHSN data (counts of confirmed hospital admissions by pathogen, with potential for positive and/or negative backfill)
• nowcasting a non-count based indicator e.g. a composite metric such as % of ED visits due to flu (would also require positive/negative back fill)
• nowcasting across multiple strata with sparse data (e.g. using baselinenowcast and epinowcast/other methods)
• other common challenges I am missing?

[seabbs]

We want a non-count example here I think. We have thrown around using a composite ED metric or the variant data

[seabbs]

@lauraajones2 for awareness

[kaitejohnson]

Agreed that a non-count example would be good here. % ED visit might work nicely and could apply to the data in the local level forecasting Hub or FluSight as both of these accept % ED visits due to flu as a target.

I think variant data will be quite hard as an example for using an existing tool...

[seabbs]

I just spoke to Laura and we thought we would touch base on thoughts for this at the end of the meeting on Wednesday.

[kaitejohnson]

Update:
Could break into three total case studies with

1. NSSP and data hierarchies (facility level counts, hsa level counts, % ED vists/sparse data for different strata)
2. outbreak example (need vintages to do nowcasting for an outbreak!)
3. variant nowcasting (related to 1 in that the proportions are made up of counts which have high degree of back fill at the lab level)

[seabbs]

Here is the workflow paper that has excamples of "schematic" case studies: https://samabbott.co.uk/a-workflow-for-infectious-disease-modelling/

I think for all of them making sure we discuss different hierarchies of data and what that means for modelling is useful. All of them have this to some degree i.e case counts on a website with wayback through to a linelist for outbreak data with a shadow date for all dates etc. For variants proportions through to counts per sequence through to ... lower level stuff

[seabbs]

I changed the title to be more general but perhaps we want issues for each. The idea is that the leads on this coordinate/lead they don't need to necessarily do everything!

[kaitejohnson]

@lauraajones2 Perhaps we make three issues for the three case studies? I think I'd be less knowledgeable about 2 since I haven't seen/worked with that data from an internal perspective, but I would be happy to draft something for 3 and contribute to something for 1

[lauraajones2]

@kaitejohnson Makes sense to me! I too haven't done much with 2 besides high-level exploration but I think between the two of us can probably scaffold it out a bit and pull in others as needed

[kaitejohnson]

Adding notes here for how to do this:
Generally want to know for each case study data set, what are the question you would ask and what are the possible answers?

• What are the next questions?
  example: What data do you have? aggregate, static, individual level
• Going to include all the data hierarchies for each case study dataset e.g. line list, case counts, aggregate data for

[seabbs]

in order i.e question 1 leads to question 2a or 2b depending on answer etc etc