The PCR Test As A Predictor Of Acute Respiratory Conditions (part 1)

I utilise data from an unknown NHS Trust to determine the real-world value of a COVID-19 diagnosis in the EPR of in-hospital deaths

In part 11 of ‘Do COVID Vaccines Work?’ I promised to have a think how best to proceed given the ambiguity inherent in COVID designation prior to death. That’s a fancy way of saying that the COVID-Dx indicator variable I have been relying on is a pile of tripe without the benefit of onion sauce. I first looked at this issue in relation to my sample of in-hospital deaths in Catastrophic Health Collapse (part 7) back in April 2023 where we discovered a peculiar propensity for asymptomatic COVID death during the pandemic first wave, these being people who tested positive but never developed an acute respiratory phase. I had thought I’d undertaken a bit of modelling to investigate matters more deeply but it transpired that I hadn’t… hence today’s article and the beginning of a new mini-series.

Essentially what I am about to do is flip the logistic regression modelling process on its head and try to predict onset of acute respiratory conditions prior to death using a few basic demographics to level the playing field followed by the COVID-Dx indicator variable. We’ll then have a better idea of whether this variable is a plate of pure tripe or tripe served with boiled potatoes and some healthy green veg. I realise that not all acute respiratory conditions are going to be the result of SARS-COV-2 infection and thus the data will be peppered with true negatives but let’s just see how the analysis pans out…

Data Sample

A total of 19,857 in-hospital deaths aged 18 years and over were present in the EPR data dump for the period 2020/w10 - 2021/w38 and the weekly frequency of these is present in the following slide:

The dashed grey lines mark three periods that I elected to serve in split file analysis mode, these being 2020/w10 – 2020/w40 (n=7,420), 2020/w41 – 2021/w20 (n=8,980) and 2021/w21 – 2021/w36 (3,252). Deaths tail-off dramatically after week 36 owing to administrative delays hence the cut-off at 2021/w36.

We ought to pile straight in with a simple crosstabulation of acute respiratory conditions against COVID-19 diagnosis for all periods combined:

A Pearson Chi-squared test of association confirmed that this distribution of deaths is very unlikely to have arisen by chance (p<0.001), so we’re looking at a genuine effect.  We observe 1,843/3,412 (54.0%) of cases with a positive test result but no associated acute respiratory diagnosis; these being the so-called ‘asymptomatic COVID’ cases that I am rather suspicious about since this label is a rather convenient cover for an inadequate viral test. I’d rather label these as false positives! We also observe 1,616/16,240 (10.0%) of cases with an acute respiratory diagnosis but a negative test result. By the same token we ought to label these ‘false negatives’ but, of course, we don’t really have a clue since this category may indicate non-COVID respiratory disease!

We get a bit of a clue when we run the analysis again in split-file mode. Try this:

The false positive/asymptomatic category is seen to fall from 69.2% to 48.6% to 34.6% with each passing pandemic phase and this is worth deep contemplation. Clearly the goalposts have been moving but exactly what constitutes the ‘posts’ is not certain for we’re looking at a raft of factors ranging from changing PCR test design and use to changing clinical protocol, changing patient population and a mutating virus.

The false negative/non-COVID respiratory disease category is seen to hover from 10.3% to 9.7% to 9.9% and I am inclined to think this represents a consistent level of genuine non-COVID respiratory disease. What we better do now is start off with a basic logistic regression model to see if this can shed any light.

Logistic Regression Ahoy!

Herewith the structure for the base logistic regression model comprising five main effects in the prediction of acute respiratory conditions prior to death. Diagnoses represents the total number of medical diagnoses made on the medical record, which may range from 1 to 10, this being a rough proxy for how sick people are. CDR (case detection rate) is derived from the UK GOV coronavirus dashboard being the rolling 7-day count of new COVID cases detected in England divided by the rolling 7-day count of viral tests undertaken, this figure being a reasonable proxy for disease prevalence across the nation and thus a crude proxy for individual risk of infection. Chronic Respiratory is an indicator for long-term diseases like asthma, COPD and emphysema and thus an indicator for individual risk of developing an acute respiratory condition.

The five main effects and their ten 2-way and ten 3-way interactions were submitted to the modelling process using a forward (conditional) selection procedure which built a predictive model from the constant upwards by incorporating only those terms that reached statistical significance at the 95% level of confidence. This approach ensures a parsimonious result; that is to say it yields the simplest model possible that explains as much as possible. The dependent variable – that which we are trying to predict/explain/model – is the indicator variable Acute Respiratory.

In a nutshell I’m trying to fathom which factors matter when it comes to the risk of somebody developing an acute respiratory condition prior to death.

Base Model Results

Ok, so these are the terms that matter most since they are all highly statistically significant predictors of acute respiratory diagnosis. Risk of such a diagnosis increases with Age (OR = 1.02, p<0.001), Diagnoses (OR = 1.74, p<0.001) and CDR (OR = 1.02, p<0.001). Females are more likely to exhibit an acute respiratory condition (OR = 3.50, p<0.001) as are folk with a pre-existing Chronic Respiratory condition (OR = 2.80, p<0.001). On top of these main effects there are four interactions that modify the odds ratios slightly. All this makes a great deal of sense so let’s run the model again but with the addition of COVID-19 Dx that is entered (forced) at stage 2.

Results with COVID Dx

Here’s pretty much the same table of model coefficients but with COVID-19 Dx now sitting at the bottom after being levered into place. The obvious point of interest is that COVID-19 Dx is a highly statistically significant predictor of Acute Respiratory Conditions prior to death with a sizeable odds ratio (OR = 6.92, p<0.001) indicating a six-fold increase in risk when associated with a positive test result. This is the sort of result we may expect to see if the PCR test was doing its thing and identifying genuine COVID cases. However, what we cannot rule out is the possibility of a largely ineffective test that was persistently applied to severe respiratory cases until a confirmatory result was obtained. This is proof of association and not efficacy!

The less obvious point of interest is that CDR and Age by Chronic Respiratory have both been made redundant by inclusion of COVID-19 Dx, with p-values now up at p=0.629 and p=0.137 respectively. This means there is a relationship between CDR and COVID-19 Dx as well as a relationship between Age by Chronic Respiratory and COVID-19 Dx. The former makes sense in that we would expect risk of COVID death to follow disease prevalence; the latter also makes sense in that it points to elderly folk with pre-existing chronic disease also being contenders for genuine infection. Taken together these results suggest a degree of efficacy to the PCR test.

Split File Results

So far so good but we can go one step further and use split file analysis. What we are looking for is invariance; that is to say, if these results represent real world relationships then those relationships should remain more or less constant across the three periods outlined above. Here, then, is exactly the same logistic model but run using the split file mode:

Now this is where it gets rather interesting! We have three completely different model structures, with the earlier period of 2020/w10 – 2020/w40 yielding the most complex predictive model. This tells us that the goalposts were shifting in more ways than one, and so assessing PCR test performance (as well as vaccine efficacy) is going to be mighty tricky if not impossible! Ideally the Office for Nobbled Statistics and others will take note but then again we are no longer living in an ideal world.

COVID-19 Dx, as a predictor of acute respiratory conditions, rises from an odds ratio of 3.30 (p<0.001) to 8.38 (p<0.001) to 14.26 (p<0.001) and we ought to ask how this is possible. To my way of thinking the test should be working equally well at detecting severe COVID infection across all three periods but, bizarrely, it seems to get better over time! How is this possible? Sure, we can argue that the tests were improved or test protocols (who to test) improved over time but this sure looks to me like something else was going on that casts a shadow over test performance for the virus didn’t mutate to become more deadly. What is impossible to ignore is that the relationship between acute respiratory illness and positive test result became stronger as vaccination rollout proceeded. Thus, we may be looking at evidence of vaccine harm more so than improved testing.

Unvaccinated As Control

If vaccine harm is the driver behind this curious result then we shouldn’t see progression in the odds ratio for COVID-19 Dx for the unvaccinated cohort. Here’s the exact same model again but with vaccinated folk excluded:

Again, we note the moving goalposts but, again, we observe progression in the odds ratio for COVID-19 Dx from 3.30 (p<0.001) to 8.51 (p<0.001) to 11.41 (p<0.001) so this cannot be a function of vaccine harm. How utterly fascinating! Thus, something is driving a closer relationship between acute respiratory conditions and incidence of positive test results over time such that test efficacy appears to improve.

At this stage I must conclude that the PCR test was doing something of value but that something changed over time. I am aware that they changed primers over this period and moved from triple to double to single sequences; could it be that the test that was initially very specific to COVID simply became a catch-all for both COVID and non-COVID viral infection?

Well, now, there’s yet another can of worms! I am inclined to re-label COVID-19 Dx as VIRAL Dx and go looking for other evidence of non-COVID viral respiratory infection in the EPR. Until then…

Summary

• Analysis of 19,652 adult in-hospital deaths for the period 2020/w10 - 2021/w36 revealed 1,843/3,412 (54.0%) of cases with a positive test result but no associated acute respiratory diagnosis.

• Split-period multivariate logistic regression modelling in the prediction of acute (life-threatening) respiratory conditions yielded three differing model structures indicative of relationships changing over time (goalpost shifting). Many studies indicative of vaccine benefit assume constancy in this respect and are thus likely to be invalid.

• The capability of the PCR test in identifying infection leading to acute respiratory conditions was also found to vary over time and it is suggested that the test became less specific to COVID. This casts doubt over the validity of test-dependent diagnoses within the EPR.

Kettle On!