I am interested in “correcting” the current cumulative incidence (normalized by population) to the “stage” of the incidence curve

This is a little off-topic, but there are a lot of experts here….
I am interested in “correcting” the current cumulative incidence (normalized by population) to the “stage” of the incidence curve. In other words, I’d like to compare incidence while accounting for the fact that some counties have still yet to “peak” while others are already on slow decline. Anyone have any metrics of interest that inform the normalization of cases by local “epidemic stage”?

Hi Sean, very intriguing question, and I see that Ryan has already amplified your message in the general channel, so I hope folks much better informed than me would chime in. A few questions/clarifications/comments:

a. what do we mean by local “epidemic stage”?
Are we trying to figure out where are we on the incidence distribution/curve for a particular county? If we can figure out where we are on the curve for that county (which hasn’t peaked), then can we use the curves of ‘similar counties’ to normalize future incidence rates?

b. similar counties(?): perhaps we can start by looking at the regional/density of PoIs from SafeGraph (SG) as an indicator/proxy for economic activity, and correlate it to the census data that SG has packaged into the more user friendly Open Census Data. That is, maybe we can first identify the factors
that might make counties ‘similar’ - demographics/age/economic activity from PoIs. I think this is partly your question, I am just thinking aloud.

c. some metrics to watch out for: apart from population, we have to adjust for testing numbers per 1000; instead of just looking at the slope of the curve (rate of change of incidence/time, just for representation’s sake let’s call it di/dt), perhaps look at the rate of change of the slope itself (i.e. d^2t/dt^2).

I might be asking trivial questions here, and again, I am not an epidemiologist, so I hope I am not throwing us off track! In the meantime, I’ll start noodling around with the data to see if I can offer more than just suggestions/comments.

I thought this application of Shewhart control charts to categorizing epi phase was interesting: COVID-19 Data Tracker | IHI - Institute for Healthcare Improvement

Hi @Sean_Davis_NIH - did you ever solve the problem? @Yaniv_Carleton_College and I were toying with the idea of fitting Variational Auto Encoders (VAEs) of the timeline of each county (timeline = percent change in log cases, for example - but it should be bounded). We propose to use a rolling window of time (county X at weeks 0-10, 0-11, etc… where week 0 is the time of, say, 100’th case). Then, fit a VAE, and use the resulted fitting, so that if county X has a similar trend to the trend county Y had 5 weeks ago, we can use county’s Y trend to predict how county X will look like in the next 5 weeks. The divergences from these predictions are what you’re looking for.

However, you should also “throw in” change in number of tests (the actual number of tests doesn’t matter, just the change), the introduction of various NPIs (Non Pharmaceutical Interventions such as stay at home orders), and other time-specific changes. For example, you can throw in a measure of change in mobility of people, using the social distancing data.

I might explain it in hand-waiving without showing my hands, but thankfully @Yaniv_Carleton_College already wrote a notebook: Google Colab

And is working on a blog post on this.

Thanks, @Dana_Turjeman_UMich_Ross and @Joe_Wasserman_RTI_International. Both look promising. And yes, both address my question very nicely.