Student-written meta analysis of a very small subset of studies exhibiting very high bias and significant flaws. Some of the problems:
- As of the publication date, there are 35 studies, authors include only 4. (They list 5, but two are the same study, preprint and published version).
- From the 17 RCTs, authors include 0.
- Authors include only late treatment studies, excluding all 10 early treatment studies and all 10 prophylaxis studies.
- Authors did not locate 13 studies, despite this being trivial from existing meta analyses.
- There is no logic in the exclusion reasons. For example, they include the most biased study to date, Soto-Becerra, and assign the highest weight to it.
- Authors randomly exclude letters but include preprints (excluding letters to help avoid positive results, including preprints to include Soto-Becerra).
- Soto-Becerra has clear evidence of extreme bias. The study presents 30 day results and extended KM curves up to day 43 for ivermectin. At 30 days the result is negative but reverts (as do all treatments in the study) and becomes positive before day 43. Authors of this meta analysis ignore the extended followup. Soto-Becerra is a database analysis that includes anyone with ICD-10 COVID-19 codes which includes asymptomatic PCR+ patients, therefore many patients in the control group are likely asymptomatic with regards to SARS-CoV-2, but in the hospital for another reason. For those that had symptomatic COVID-19, there is also likely significant confounding by indication. In this study all medications show higher mortality at day 30, which is consistent with asymptomatic (for COVID-19) or mild condition patients being more common in the control group. For ivermectin they show 30 day mortality aHR = 1.39 [0.88 - 2.22]. KM curves show that the treatment groups were in more serious condition, and also that after about day 35 survival became better with ivermectin. More than the total excess mortality happened on the first day. This is consistent with treated patients being in more serious condition, and with many of the control group patients being in hospital for something unrelated to COVID-19. Authors use a machine learning based propensity scoring system that appears over-parameterized and likely to result in significant overfitting and inaccurate results. Essentially they test for all interactions between two and three covariates. The nature and large number of covariates means many random correlations may be found. COVID-19 severity is not used. In summary, this is the lowest quality ivermectin study to date. This study also does not compare treatments with a control group not receiving the treatment - authors put patients receiving treatments after 48 hours in the control group. Authors also state that outcomes within 24 hours were excluded, however KM curves show significant mortality at day 1 (only for the treatment groups).
- We checked the reported results for the mortality outcome and found they do not appear to match the actual papers.
- Rajter: authors list mortality as 13/85 (treatment), 24/74 (control), the paper shows (for the matched cohort) 13/98 (treatment), 24/98 (control). The adjusted result in the paper is OR 0.27 [0.09-0.80] (multivariate) or OR 0.47 [0.22-0.99] (PSM). These correspond to RR 0.33 and 0.54 respectively, or logRR -1.1 and -0.62. However authors here show logRR 0.54 and 0.85 - they include the study twice (preprint and published). The preprint and published papers have the same multivariate result, the PSM result was added in the published paper. Neither of the two results the authors use match the actual results.
- Khan: the paper shows RR 0.13, logRR -2.0. Authors show logRR 0.13.
- Soto-Becerra at day 30 shows wHR 1.39 [0.88-2.22], and day 43 weighted KM 0.82 [0.76-0.88]. These correspond to logRR 0.33 and -0.19. Authors show logRR 1.75.
- Gorial: there is zero mortality with treatment in this paper. Using the typical continuity correction, the paper shows RR 0.29 when accounting for the different group sizes, or 0.86 when using naive continuity correction that does not account for the very different group sizes. These correspond to logRR -1.24 or -0.15. Authors show logRR 0.60.
- Authors did not locate and reference the existing widely known meta-analyses from well-known researchers - Kory et al., Hill et al., Lawrie et al.
For more issues see: [1, 2]
Authors on Twitter: [3, 4, 5, 6]
Castaneda-Sabogal et al., 1/27/2021, preprint.