Submitted (finally) revisions to this analysis of using the test-negative study design in the presence of public health measures. Some new thoughts, in light of pandemic!
Work w/ @rozeggo, @TJHladish, & John Edmunds.
When criteria for testing vary systematically, the design can be biased. In outbreak & pandemic response, testing happens passively (e.g. you're sick => seek care => get tested) & actively (e.g. co-worker is test+ => contact-tracing => you get tested).
NB: NOT PEER REVIEWED
These represent diff risks and thus differ in thresholds for testing. e.g. in passive route, generic risk => testing conditional on symptoms, but for active scenario => high exposure risk => unconditional testing (e.g. COVID19 TTI protocols).
NB: NOT PEER REVIEWED
So: test-negative designs under such conditions are potentially biased. Not particularly surprising in hindsight!
What to do?
We proposed a hybrid design that deals w/ the baked-in bias, but didn't analyze its performance under other biases.
NB: NOT PEER REVIEWED
Key to that design: recording why test performed. e.g. for COVID19 control, there's testing given symptoms, testing contacts of test+, testing for travel or work, etc.
To potentially use those data in a study, however, we need *why* test was performed.
NB: NOT PEER REVIEWED
Keeping that info doesn't happen automatically - has to be planned for.
We developed this analysis (for a totally different disease, Ebola) in recognition that keeping that data wasn't a top priority. During a public health emergency, lots going on!
NB: NOT PEER REVIEWED
But we hope we've made the case for integrating this requirement into planning efforts for response activities - e.g. in data system choices, in integrated study protocols, for analytical cells in public health operations.
We forecast dates that African countries would *report* 1K and 10K cases. We used very sparse data (first 25 or fewer cases prior to 23 March), w/ deliberately low-detail method (branching processes), assuming global epi estimates & discounting any potential interventions.
That very simple approach worked well for countries that couldn't (or didn't) respond to early warnings like ours. My previous tweets about this report are from 27 March; since then 12 countries have reported >1K cases: who.int/docs/default-s…