FOI requests reveal NHS Trusts cranked up the max threshold to loony levels early on, so yes, it's not likely this. Not all cases are confirmed using PCR as you say, with CXR and bloods in the driving seat. But here's the thing... what may be written in the casenotes by the SHO on duty is not necessarily what gets coded into the Trust PAS system. I have had 3 coders confide they were required by management to code for COVID when casenotes had been clearly marked as not COVID. What we are looking at imho are signs of management interference.
- Regarding the vaccination misclassification: Is this a semi-retraction of everything regarding vaccine efficacy and safety you've ever done thus far? Though if anything, I would expect a lack of matching between data systems to only dilute prior findings, though I suppose that assumes mismatches are a random occurrence and do not target a certain type of patient.
- How does anyone come to know the above mismatching is even happening? And how could I know if something similar were happening in studies in the US?
Apparently, no casedemic happened in Queensland, even after vaccinating 80% of people. Their borders were closed up until that point and there was allegedly no covid. So either vaccines don't cause covid-like illnessess that lead to tests, or they were never testing to begin with. Though that's all outside of a hospital setting potentially.
- Wouldn't (UK covid hospitalization count) / (PillarOneTestsCount) yield a curve whose shape would add insight?
- Is it a sin to pick and choose which periods to model with one method or the other to make some hybrid?
More of a refinement than a complete retraction. The two key variables - COVID Dx and vaccine Tx - upon which my work hangs are dodgy though not completely misleading. Thanks for the heads-up on Vx not causing FP - we can rule that out! I've penned four articles that take this further - the next two are attempts at predicting Vx status and the two after that introduce what I am calling PROD (prior risk of death) - there's a few surprises in store as you'll see. In order to string all these models together I've settled on 2020/w11 - 2021/w36 as the base period. At some point (probably early October) I'll be penning part 13 of 'Do Vaccines Work?' - it will be interesting to see how this all pans out.
Can a Simpson’s paradox manifest in the time domain, whereby some key factors to which the model(s) are sensitive are not time invariant? Is there potential for additional insight if multivariate LR would be applied separately to more than one time period where modalities are suspected to differ?
Interesting you should ask this because I have a factor that permits split file analysis for both MLP and LR runs using the periods 2020/w11-2020/w40, 2020/w41-2021/w14 and 2021/w15-2021/w36 and I have been using this technique after the main modelling runs. This approach was first used for 'Do COVID Vaccines Work? (part 3)' and features regularly in that series. When I can find time I'll be writing up the results for there are some strange things among the sensible! This method will be employed for part 13 of that same series, which I hope to produce early October (four articles extending the current work are already in the pipeline).
From what I can tell it's not the PCR threshold.
I don't kow about your data, but generally speaking COVID diagnoses are not always PCR confirmed.
Most of these surplus cases fall into the non-PCR-confirmed category, judging by my familiarity of the US data.
Besides, how would that just happen? Nobody ordered labs to increase cycle thresholds and later reduce them again.
I'll write you an email with two exemplary scatter charts to illustrate the US situation.
FOI requests reveal NHS Trusts cranked up the max threshold to loony levels early on, so yes, it's not likely this. Not all cases are confirmed using PCR as you say, with CXR and bloods in the driving seat. But here's the thing... what may be written in the casenotes by the SHO on duty is not necessarily what gets coded into the Trust PAS system. I have had 3 coders confide they were required by management to code for COVID when casenotes had been clearly marked as not COVID. What we are looking at imho are signs of management interference.
Well that's interesting.
- Regarding the vaccination misclassification: Is this a semi-retraction of everything regarding vaccine efficacy and safety you've ever done thus far? Though if anything, I would expect a lack of matching between data systems to only dilute prior findings, though I suppose that assumes mismatches are a random occurrence and do not target a certain type of patient.
- How does anyone come to know the above mismatching is even happening? And how could I know if something similar were happening in studies in the US?
- Regarding vaccines causing false positives:
https://news.rebekahbarnett.com.au/p/first-covid-deaths-were-fully-vaccinated
Apparently, no casedemic happened in Queensland, even after vaccinating 80% of people. Their borders were closed up until that point and there was allegedly no covid. So either vaccines don't cause covid-like illnessess that lead to tests, or they were never testing to begin with. Though that's all outside of a hospital setting potentially.
- Wouldn't (UK covid hospitalization count) / (PillarOneTestsCount) yield a curve whose shape would add insight?
- Is it a sin to pick and choose which periods to model with one method or the other to make some hybrid?
More of a refinement than a complete retraction. The two key variables - COVID Dx and vaccine Tx - upon which my work hangs are dodgy though not completely misleading. Thanks for the heads-up on Vx not causing FP - we can rule that out! I've penned four articles that take this further - the next two are attempts at predicting Vx status and the two after that introduce what I am calling PROD (prior risk of death) - there's a few surprises in store as you'll see. In order to string all these models together I've settled on 2020/w11 - 2021/w36 as the base period. At some point (probably early October) I'll be penning part 13 of 'Do Vaccines Work?' - it will be interesting to see how this all pans out.
Can a Simpson’s paradox manifest in the time domain, whereby some key factors to which the model(s) are sensitive are not time invariant? Is there potential for additional insight if multivariate LR would be applied separately to more than one time period where modalities are suspected to differ?
Interesting you should ask this because I have a factor that permits split file analysis for both MLP and LR runs using the periods 2020/w11-2020/w40, 2020/w41-2021/w14 and 2021/w15-2021/w36 and I have been using this technique after the main modelling runs. This approach was first used for 'Do COVID Vaccines Work? (part 3)' and features regularly in that series. When I can find time I'll be writing up the results for there are some strange things among the sensible! This method will be employed for part 13 of that same series, which I hope to produce early October (four articles extending the current work are already in the pipeline).