Mon 2022-Mar-07

Ivermectin Revenant

Tagged: COVID / JournalClub / PharmaAndBiotech / SomebodyAskedMe / Statistics

Somebody asked about a recently published abstract comparing ivermectin vs remdesivir in treating COVID-19. (Sigh.)

Ivermectin just won’t go away

Despite repeated debunkings, ivermectin is a mania repeatedly rebunked, and just won’t go away. I should know better than to play scientific whack-a-mole. But just a quick look can’t hurt, right? (Right?)

The latest paper

Efimenko, et al. @ IJID: ivermectin vs remdesivid The paper in question [1] is in fact an abstract a conference on emerging diseases last year. That means, among other things, that it is very, very short: just a single page. So we can read the whole thing in just a few minutes.

The general idea is they took a look through a large EMR (“electronic medical records”) system, and:

  1. Isolated adult (≥ 18yr) patients with (a) a solidly believable COVID-19 diagnosis, and (b) who received either ivermectin or remdesivier, but not both. Cutoff dates were 2020-Jan-01 to 2021-Jul-11, so it does not include Delta or Omicron.
  2. Controlled for the usual confounding factors: age, gender, race/ethnicity, nicotine, diabetes, obesity, lower respiratory disease, heart disease, anti-inflammatory drug use of a couple kinds, and being on a ventilator (precise control method unspecified in the abstract).
  3. Then they looked for death as the primary outcome, demanding statistical significance at the usual $p \lt 0.05$ (statistical method unspecified in the abstract).

I mean… it looks like a reasonable thing to try. And fighting your way through the general chaos of US healthcare records cannot have been fun or easy! So we have to start from a position of offering kudos to the authors for doing something that was probably difficult, for the usual bizarre US healthcare reasons.

Also, we should note that this is just an abstract from a conference; from the id reported it might even be an abstract of a poster. That means 2 things:

  • The usual peer review is somewhat lighter, since in short conference talks (and especially in posters), the idea is to include new ideas to think about rather than final, decisive studies.
  • It’s very, very brief. Thus we should expect it to be light on details, frustrating as that might be. Any lack of detail here is not the fault of the authors, but of the process.

I want to make that clear, because although I’m going to be critical, the authors did some rather uphill work in a system that probably made data access difficult and forced them into making a brief, sketchy report. It’s not a full paper, and we shouldn’t hold it to the standards for that.

But let’s not be rubes, either.

The result

Still, they dutifully report a result that should make people at least take a look:

  • Odds Ratio: The odds ratio for reduced mortality with ivermectin vs remdesivir of OR = 0.30 with a 95% confidence interval of 0.20 – 0.48. Since the confidence interval on the odds ratio is bounded away from 1 (equal odds of death), this is statistically significant (though no reported $p$-value.)
  • Risk Difference: The risk difference was -5.2%, with a confidence interval of -7.1% – -3.37%, and a significance of $p \lt 10^{-4}$. The statistical significance here is considerable, though the effect size may perhaps not be: 5% decrease in death rates? (As this is an abstract, there are not enough details for us to dig into their math. So I’m frustrated, but probably the authors are frustrated at not being able to say it before a full publication, too.)

Some caveats

Now, what should we think about that? I have numerous reservations:

  • Electronic Medical Records: EMR systems in the US are a patchwork quilt, unlike those in Europe or Israel.
    • Yes, it’s available data of generally ok-to-high quality. But there’s a severe selection bias toward people with medical insurance, often with a particular provider.
    • That’s only a moderate concern here, since TriNetX covers 68 million patients; though it’s probably intermittent coverage as patients change jobs and insurers.
  • Relevant SARS-CoV2 variants: The study dates, 2020-Jan-01 to 2021-Jul-11, do not include any time period where there was significant Delta or Omicron circulation in the US. Those are the only relevant strains (until the next one!). So the applicability is, while not irrelevant, at least not obviously extensible to Omicron.
  • Patient gating factors: The TriNetX EMR system contains intermittent records on 68 million patients, of whom 1,761,060 met the critiera for diagnosis with COVID-19, but of whom only 41,608 were eligible for this study. Was the criterion of receiving either remdesivir or ivermectin but not both that severe so as to restrict attention to only $ 41,608 / 1,761,060 = 2.4\%$ of the COVID-19 patient population? Shouldn’t that make us worry about the general applicability of the results?
  • Unbalanced sample: The sample was quite unbalanced, by an order of magnitude: 1,072 ivermectin patients vs 40,536 remdesivir patients. While this reflects current practice (ivermectin kind of “underground” vs remdesivir standard of care in the early pandemic), it worries me.
    Groenwald & van Smeden @ Epidemiology: Efficient sampling in case-control studies
    • Case/Control Sampling: It would have been more usual to do some sort of case-control sampling. [2] In that case, one would take all 1,072 ivermectin cases, and then sampled the remdesivir cases to get a similar sized cohort so the sampling is closer to 1:1. (For bonus points that make grizzled old statisticians smile at you, assess the stability of the result under repeated bootstrap resampling of the remdesivir patients.)
    • Effect on statistical significance: This would have balanced the sampling, but reduced the sample size from 41k to 2k. Elimination of the bias would have cost statistical significance, and I wonder if that’s why they didn’t use this well-known technique: maybe the effect goes away? (In a real paper, not just an abstract, perhaps they will discuss this.)
  • Weird choice of comparison: Why did they choose to go up against remdesivir?
    • Strengh of effect: Remdesivir is not a super-strong COVID-19 therapy. I mean, it helps a little, but it didn’t constitute a game-changing therapy. It was available from the beginning of the pandemic, so that extends the baseline of relevant patients. I suspect they chose remdesivir for that reason, to get more patients to get to statistical significance. But with paxlovid and molnupiravir available, it’s not so relevant going forward.
    • Mode of administration: Very few physicians will prescribe ivermectin for COVID-19. Consequently, we’re comparing a self-administered drug usually taken at home, early in infection against a late-stage drug always given in hospitals by infusion, usually at later stages. This is like comparing 2 populations for headache: one group takes ibuprofen and nobody dies; the other group goes to the hospital for surgery to remove a brain tumor and some of them die. They’re so unlike that it’s almost a Gilbert Ryle-style category error to compare them.
  • Age difference: The ivermectin cohort is a decade younger than the remdesivir cohort (52 vs 62, with of course some spread that the abstract does not explain). While they did control for age, they can’t take the space in an abstract to explain how. So I’m a little antsy to make sure that they accounted for this risk factor by a sensible method. [3]
  • Sex difference: 10% fewer of the ivermectin cohort were male vs the remdesivir cohort (43% vs 54%). Again, they claim to have controlled for this, but I want to see how, given that they’ve (commendably!) admitted up front this bias in the input data.
  • Credentials: Honestly, I want to soft-pedal this, because I hate the creeping credentialism that sometimes is used as an excuse not to listen to people. But… of the 4 authors, only 1 is an infectious diseases person. The other 3 are in plastic surgery and urology, which while fine disciplines in and of themselves, are of doubtful relevance here. Still… I’m a bit undecided about whether the author specialties should be an issue. It both bugs me, and also bugs me that it bugs me. So let’s be generous and spot them some benefit of the doubt on this one.

Finally, just because I want to bend over backward to give these folks credit for doing something difficult, they close by admitting a large randomized controlled trial is required to really decide the question.

That’s fair enough, but people only run large RCT’s when there’s at least some preliminary evidence of an effect that’s both large enough to be clinically meaningful, and statistically significant enough that it’s probably real. So far, neither of those is the case with ivermectin.

The Weekend Conclusion

It appears to be a brave effort, though we need to see the details to be sure. But it also appears to be an irrelevant effort, because:

  • The cohort of patients has some problems with age and gender biases, though this might have been corrected by their methods; we’ll see when there’s a full report. On the other hand, the use of a US EMR system biases you toward employed people with health insurance, and I seen no way around that.
  • The sample is unbalanced, and probably should have done case/control sampling to make it so. That would have reduced statistical significance dramatically, so I’d like to hear from the authors about that, too.
  • The SARS-CoV2 variants circulating at the time of the study are no longer relevant.
  • The reported effect size (risk differnce of about 5%) is not large.
  • The comparison with remdesivir is odd, since remdesivir is safe but at best only mildly effective, and administered only in the hospital at relatively late stages. Ivermectin likely self-administered at home in early stages.
  • With the advent of paxlovid, molnupiravir, and bebtelovimab, the comparison with remdesivir is irrelevant. The small effect size for ivermectin reported here has no hope of beating the new antivirals.
  • As this is an abstract of a conference short talk/poster, we have little idea of the specific methods they used for controlling for confounders or assessing association with mortality. That’s not their fault; there’s just no room in an abstract. Still, I’m not gonna just believe based on a few random assertions in an abstract! I want to see the math.

So, no: this does not make me change my mind. I’m not on the ivermectin bandwagon for anything other than treating parasite infestations.

(And somehow, I’ve written a “summary” that’s longer than the publication itself! Shame on me…)

Notes & References

1: I Efimenko, et al., “Treatment with Ivermectin Is Associated with Decreased Mortality in COVID-19 Patients: Analysis of a National Federated Database”, Intl Jnl Infect Dis 116: Suppl, S40, 2022-March. Publication of abstract PS05.04 (947) from the Eight International Meeting on Emerging Diseases And Surveillance, IMED 2021-Nov 4-6 virtual meeting.

2: RHH Groenwald & M van Smeden, “Efficient Sampling in Unmatched Case–Control Studies When the Total Number of Cases and Controls Is Fixed”, Epidemiology 28:6, 834-837, 2017-Nov.

3: Yes, I’m a fussy old statistician. But I came by that fussiness honestly.

Published Mon 2022-Mar-07

Gestae Commentaria

Comments for this post are closed pending repair of the comment system, but the Email/Twitter/Mastodon icons at page-top always work.