Presentation

HCQ cocktail in early
COVID-19 infection
n << p logistic regression: deep learning solution?
Richard D. Gill - Leiden - 30 July 2020

• Leila Schneps (Paris)
• Dipro Mondal (Leiden)
• Hoang Van Truan (Vietnam and Marseilles)
• Philippe Gautret, Didier Raoult and others in Marseilles
• Rob Elens and others in Meijel
• Also thanks to many friends and colleagues for help with R,
JAGS, JASP, lasso, splines, Bayesian Cox regression, ...
Collaborators
and Acknowledgments

• Corona: many are exposed; some become infected, some
become infectious, some have symptoms
• Not many get very sick; but many of those who do get sick die or
suffer permanent harm
• Tests are very unreliable
• Symptoms are very unspecific
• Hydroxychloroquine, with azithromycine and Zinc, might be a
useful prophylactic
• There exists a plausible biochemical explanation
HCQ + AZT + Zn as preventive
treatment
HCQ cocktail as prophylaxis

• Suppose 50% of those who fear they are exposed and who fear some symptoms do not actually have
Covid-19
• Suppose 75% of the remainder will have a mild episode of sickness and completely recover in a few weeks
• Thus: only 1 in 8 is going to get seriously ill
• Suppose the HCQ prophylactic treatment could halve that to 1 in 16
• Then we want to find out whether p1 = p2 = 1/8, or p1 = 1/8, p2 = 1/16
• Suppose n = 128 × 9 (approx 1100)
• Then in treatment group we expect 16 × 9 negative outcomes, in control group either 16 × 9 or 8 × 9
• The square roots of the numbers are approx N(12, 1/4) and N(12, 1/4), or N(12, 1/4) and N(6√2, 1/4)
• The difference of the square roots is very approx N(0, 1/2) or N(3.5, 1/2)
• Root 2 times the difference of square roots is very approx N(0, 1) or N(5, 1)
• We have a pretty good chance of getting the right answer. A 4 times smaller sample is some use but not
much. n = 40 is pretty useless (we have to decide if we have an observation from N(0, 1) or from N(1, 1)).
Power calculation
Why we still don’t know

• First publicised “trials” had n = 40
• Spectacular positive results
• Politics: Macron, Trump, the alt-right (Corona deniers and anti-
vaxers )
• Fish pond cleaning fluid and a witch doctor from the Cameroons,
conspiracy theorists and the interests of Big Pharma, distrust of
models and science, ...
Two early publicised observational studies
Marseilles; Meijel

• Age [effect is very nonlinear]
• Sex [certainly has interaction with most other variables]
• Various comorbidities (“existing conditions”)
• Severity and duration of various symptoms
• Ethnicity?
• Air pollution?
• Bloodgroup!
• Earlier exposure to similar virus infections
Confounders
Known confounders

• n = 40, p = 40++
• Age, sex, comorbidities, symptoms
• What we now know: your chance of death from Covid-19 this year
is very similar to your chance of death from natural causes
• Risk groups: the old, the sick, obese, diabetics, Alzheimer
patients, poor, financially insecure, ... , some ethnic minorities, ...
• [Those who come in close contact with many sick people]
Marseilles; Meijel
Two much too small but very influential early studies

• Marseilles and Nice: observational study of patients at two
different clinics
• “Treatment” confounded with “clinic”
• Noncompliance (3 patients “dropped out”)
• Published analysis (Fisher exact test; some comparison of some
covariates — sex, age [t-test],...) — did not control for covariates
and did not use ITT (intention to treat) principle
Marseilles (and Nice)
Gautret ... Hoang ... Raoult (2020)

• GP and “orthomolecular doctor” Rob Elens was early adopter of
HCQ. But forbidden from giving it at some point. So treatment
was determined exogenously!
• Still, his groups are unbalanced regarding age and sex, and
anyway, “time” may also be a confounder
Meijel
A small town between Eindhoven and Venlo in the South of the Netherlands

• With perfect matching we could have “identical twins” with either
both the same outcome [usually good], or a small number with
HCQ has good outcome, without HCQ has bad outcome.
• If HCQ works then: 1 in 16 twins could be like that. 7 in 8 twins,
both have a good outcome. 1 in 16 twins, both have a bad
outcome.
• If HCQ doesn’t work, then: 7 in 8 twins both good outcome, 1 in 8
both have bad outcome.
• 20 twins would not be enough but 200 twins would start to provide
pretty convincing evidence.
Why controlling for confounders can help
Perfect matching

• Fisher exact test gives highly significant group difference
[NB: I do not say “treatment effect”], p = 0.005
• Odds ratio is estimated to be about 10
• Bayesian analysis with slab and spike prior gives much milder
conclusion
• With prior of 50% “no difference”, 50% uninformative prior on log
odds ratio, the posterior probability of “no difference” is still about
5%. The remaining posterior probability says “HCQ group does
better”, but size of improvement could well be less
Results with Marseilles, Meijel data
Two studies with n = 40

• Crazy coefficients with huge standard errors or even breakdown because
“perfect fit”, ie log odds ratios diverging
• Lasso throws away all covariates except either sex or treatment!
• Spline curve for effect of age gives nice looking results - from age 20 to 55
about flat, sharply rising from 55 to 95; but cannot fit age and sex, let alone
age/sex interaction
• Some other end-points (e.g. Cox regression for duration of hospitalisation)
gave even worse/crazier results.
• Nicely, however, leaving out some extremely young and extremely old patients
did not alter Meijel results
• ITT analysis did not alter Marseilles results
Logistic regression with end-point “Covid-19 free” after 7 days
Covariates: age (linear), sex, #comorbidities, #symptoms

• Use standard life insurance tables to combine age, sex and
comorbidities to one “Corona effective age”
• Use results of other studies aimed at different treatments to give
one index of severity/progression of Corona infection
• Now we have n = 40, one treatment variable, and two continuous
covariates. Group them or use splines. We can bring p down to
ca. 10 before testing the effect of treatment
Idea
Cf: deep learning: a neural net shouldn’t have to be taught the laws of physics every time anew. Once is enough.

• To be written
Dipro Mondal’s results
To be written

Dr Fauci chooses his words very carefully:
On Wednesday Dr Anthony Fauci, a leading member of the White House coronavirus
task force, told the BBC that hydroxychloroquine was not effective against the virus.
"We know that every single good study - and by good study I mean randomised control
study in which the data are firm and believable - has shown that hydroxychloroquine is
not effective in the treatment of Covid-19," he said.
Last month, the US Food and Drug Administration (FDA) cautioned against using the
drug to treat coronavirus patients, following reports of "serious heart rhythm problems"
and other health issues.
The FDA also revoked its emergency-use authorisation for the drug to treat Covid-19.
The World Health Organization (WHO) says "there is currently no proof" that it is
effective as a treatment or prevents Covid-19.

Presentation

More Related Content

Similar to Presentation (20)

More from Richard Gill (20)

Recently uploaded (20)

Presentation