> A particularly concerning finding was the doubling of the diagnosis of dementia – which is typically irreversible – three months after testing positive for Covid-19, versus the other health conditions.
Unless they controlled for age when calculating that ratio (they didn't, as far as I can see), this is questionable. Covid-19 is disproportionately more likely to be diagnosed in the elderly (because most of the severe cases are in the elderly). Dementia is disproportionately more likely to occur in the elderly.
According to the Lancet paper that Infinitesimus linked to, the researchers divided the patients into cohorts for statistical analysis based on a number of variables:
> We identified a set of established and suspected risk factors for COVID-19, as follows: age, sex, race, obesity, hypertension, diabetes, chronic kidney disease, asthma, chronic lower respiratory diseases, nicotine dependence, ischaemic heart disease, and other forms of heart disease... We also identified an additional set of established risk factors for death due to COVID-19 (which we take to be risk factors for severe forms of COVID-19 illness), as follows: cancer (particularly haematological cancer), chronic liver disease, stroke, dementia, organ transplantation, rheumatoid arthritis, lupus, psoriasis, and other immunosuppression.
Yes, I read the paper. They controlled for these factors in assembling the cohorts.
Beyond this, it is ambiguous. They say the following:
"For analysis of psychiatric sequelae, propensity score matching was directly applied to each cohort pair. For analysis of psychiatric antecedents, given their much larger sample sizes (which exceeded the maximum number of 1·5 million patients possible per matched cohort), cohorts were first stratified by sex and age (18–30 years, 31–45 years, 46–60 years, 61–75 years, and ≥76 years) and propensity score matching (including for age) was achieved within each stratum separately."
First, I'm not sure exactly what this means: did they assemble the marginals as you would to calculate p(psych_symptom|illness), then perform propensity score matching on those? Or did they perform propensity score matching on the subsets of the data that are "people who had covid" vs. "people who had influenza"? It makes a big difference.
Second, propensity score matching is, essentially, regression followed by clustering. The details of the regression therefore matter: they are defining cluster cutoffs in terms of standard deviations, which tells you nothing without knowing the size of a standard deviation in the context of the split.
In short, I don't know if this method of matching will control for age properly in the analysis in question. You would need to see a plot of the age distributions of the data for the groups in question to be sure.
EDIT: Also, Figure 3 quite clearly shows that the relative risk for all psychological illnesses increases with age. This would seem to rebut the idea that they have controlled for this factor.
EDIT 2: The supplementary materials show tables for characteristics of Covid vs. X for all of the control illnesses, which makes me believe that they have applied cohort matching on the marginals. However, they do not show age in these tables, making me doubt that they properly controlled by age (see supp. table 1):
EDIT 3: Not directly relevant to the question, but it hints at it -- per supp. table 8, Cholelithiasis (gallstones) show the strongest association with psychiatric illness (i.e. the weakest Hazard Ratio relative to Covid). In other words: of all of the control diseases, having Covid-19 only makes you 1.58x more likely to have psychiatric illness than having gallstones.
Gallstones are strongly associated with age and gender.
EDIT 4: definitely not related, but supp. fig. 8-9 shows that if you require a confirmed Covid-19 test, the differences between Covid and the controls decreases dramatically. THEY DIDN'T CONTROL FOR ACTUALLY HAVING THE DISEASE!!
Just for example, Figure 2 from the text shows a gap of ~10% between covid and flu for all psychiatric illness. The corresponding plot in supp. fig. 9 shows a gap of less than 2% when you require a confirmed Covid test! This paper is falling apart.
What is with that automatic assumption on HN that whatever scientific paper is under discussion, they surely did ignored <insert completely obvious possibility that they actually did dealt with>?
I'm being equivocal in my wording, because that's what you do when you're being polite, and you're giving the authors the benefit of the doubt. Maybe I missed something, but I didn't see them control for this in a way that satisfies me that the issue was addressed.
Unless they controlled for age when calculating that ratio (they didn't, as far as I can see), this is questionable. Covid-19 is disproportionately more likely to be diagnosed in the elderly (because most of the severe cases are in the elderly). Dementia is disproportionately more likely to occur in the elderly.