Admissions & inpatient test rate trends
Squeezing the juice out of NHS test & trace data
In my last newsletter I finished up with the promise…
I shall be taking a look at nose poking rates in my next newsletter. More noses means more cases!
Although Her Majesty’s Government publish some rather useful reports and figures for the test & trace programme, together with some rather useful figures for testing by pillar, anybody who has tried to reconcile the two datasets will have likely banged their head on the table. Let me illustrate.
More people tested than tests!
In the week beginning 27th January ‘22 (week 105 of the test & trace programme) we observe 807,307 people tested under pillar 11 of the national scheme and 4,368,158 people tested under pillar 22 according to the latest test & trace datafile. If we flip to the coronavirus dashboard data we find 516,054 viral tests undertaken under pillar 1 of the scheme and 7,197,674 viral tests undertaken under pillar 2 of the scheme for the same week.
Combining the two data sources we find 7,197,674 pillar 2 tests undertaken on 4,368,158 people, this offering a sensible rate of 1.65 tests per person for the week in question. However, we also find 516,054 pillar 1 tests undertaken on 807,307 people, this offering a nonsensical rate of 0.64 tests per person. This defies logic for 807,307 tested people will require at least 807,307 tests!
Duplication and cherry fudge
A rummage through the data definitions sections for both sources reveals historic attempts by HMG number crunchers to avoid duplication that started with one method (a running count of unique cases) and ended with quite another (weekly adjustment for duplication). Duplication may arise with the same person testing themselves more than once within the space of a week and duplication may also arise from a person being tested both under pillar 1 and pillar 2 for the same week.
Oh, and by the way, positive results take precedence over negative, so we’re looking at the ripest cherries we can imagine when we look at alleged COVID case counts.
Duplication is not an easy problem to solve and I’m glad I’m not in the hot seat. Except that this is a problem we should not be solving!
If we start down the path of removing subsequent test results then we’ll end up biasing data in favour of positive results and we’ll also loose track of just how many tests were actually conducted. A person testing themselves eight times in a week is a rather different beast to a person testing themselves just once. HMG number crunchers handle this situation by ignoring seven tests; instead of declaring the actual count of 9 tests they’ll declare a count of just 2. This introduces 100% best bias with extra cheese and a side order of fries.
A similar situation applies to those who have tested themselves silly under pillar 2, to be tested again under pillar 1 when they go to have a bunion removed. In this instance HMG’s rule of using the earliest (positive) test result will wipe out subsequent negative testing under pillar 2 and eradicate all results obtained under pillar 1. Since testing in the wider community is rife then this will adversely impact on pillar 1 counts, and is why we get a nonsensical rate of 0.64 tests per person.
Toward a solution
It is possible to sit down with a large spreadsheet and unpick the mess caused by HMG number cruncher attempts at de-duplicating data. This task is greatly assisted by the fact that we didn’t have this problem right up until March ‘21, when tests were tests and people were people and weekly ratios for both pillar 1 and pillar 2 schemes made sense. The task is also assisted by a little known variable available within the download section of the UK GOV coronavirus dashboard called…
uniquePeopleTestedBySpecimenDateRollingSum
…the definition of which may be found here. For those who want to fiddle I can report that the rolling sum is a 7-day rolling sum and that it is entirely possible to unpack the series to provide daily counts for unique people tested rather than a rolling sum. Bizarrely, back in early 2021 a data analyst argued with me that derivation of the daily count was not possible despite me sending him a spreadsheet revealing the solution. There’s nowt so queer as folk. What is also darn queer is what this series reveals when unpacked into daily counts but that is another story for another day! Suffice it to say that using a few basic tricks and a smidgen of modelling I was able to put back all the pillar 1 test results that were removed and derive virus tests per person for each pillar.
The pudding…
There’s our steaming fruit pudding right there! Note how the pillar 1 scheme starts in earnest with a burst in repeat testing of inpatients in May ‘20. Both schemes settle down to minimum repeat testing during the summer but pillar 1 clinical testing becomes positively schizophrenic, swinging between bouts of testing just once on admission and repeat testing of inpatients with increasing vigour. Note the sudden hike for both schemes after Dec ‘20 - this is when lateral flow kits started shifting off the shelves. That elevated period during the whole of ‘21 will have generated more COVID cases than earlier periods simply because people are rolling the dice more often. In this regard the additional hike for ‘22 is going to push us into La La Land.
Is HMG chasing COVID cases like dogs chase cats, ‘coz it sure feels that way! For us sober analysts this represents a right pain, being a substantial bias that needs to be accounted for before we turn the handle on things like disease prevalence and case counts over time.
Pillar 1 testing is for those with a clinical need (sick folk and frontline workers).
Pillar 2 testing is for the wider community (schools, colleges, places of work, self-testing etc).