Study findings updated . AI-assisted evidence summary, not individual medical advice or a clinician endorsement.
THE SHORT VERSION
What’s worth knowing.
The key findings, what they mean, and the context that matters.
THE HEADLINE TAKEAWAY88% not confirmedOf 32 non-randomized studies finding survival benefit, share big trials didn't confirm
Most non-randomized studies showing COVID drug survival benefits were not confirmed.
Of 32 non-randomized studies (no random assignment) linking a COVID drug to better survival, large randomized trials failed to confirm 28. Only 2 of 29 randomized trials found a survival benefit, both for corticosteroids (steroid drugs). Just 8% of 104 non-randomized studies corrected for immortal time bias, a timing error that flatters treatments.
In a crisis, promising-looking studies spread fast. Most early 'this drug saves lives' claims collapsed under fair testing, which is why randomized trials, not early impressions, are the test a treatment must pass. A cross-sectional review of studies published during 2020 that reported deaths; it shows how often early findings held up, not why each drug failed.
See the evidence
A cross-sectional review of 117 studies (133 comparisons) of possible COVID drugs published online January to October 2020, most often hydroxychloroquine.
From the study: “The results of the majority (28/32, 88%) of non-randomized studies reporting survival benefit were not replicated by large-scale RCTs.”
88% not confirmed
Shepshelovich D et al. Concordance between the results of randomized and non-randomized interventional clinical trials assessing the efficacy of drugs for COVID-19: a cross-sectional study. The Journal of antimicrobial chemotherapy. 2021.
From the study: “Only two RCTs (2/29, 7%) reported significant survival benefit, both reporting treatment with corticosteroids”
88% not confirmed
Shepshelovich D et al. Concordance between the results of randomized and non-randomized interventional clinical trials assessing the efficacy of drugs for COVID-19: a cross-sectional study. The Journal of antimicrobial chemotherapy. 2021.
From the study: “Seventy-one percent (74/104) of non-randomized studies reported adjusted survival results and only 8% (8/104) adjusted for immortal time bias.”
88% not confirmed
Shepshelovich D et al. Concordance between the results of randomized and non-randomized interventional clinical trials assessing the efficacy of drugs for COVID-19: a cross-sectional study. The Journal of antimicrobial chemotherapy. 2021.
RECOVERY found within 100 days that dexamethasone improves survival in severe COVID.
Answer in 100 daysTime RECOVERY took to show dexamethasone improves survival in severe disease
RECOVERY is a streamlined randomized trial run within routine hospital care. Over 39,000 patients from 178 hospital sites in the UK enrolled. Within 100 days it showed that dexamethasone, a steroid, improves survival in severe disease, and that hydroxychloroquine and azithromycin have no meaningful benefit for hospitalised patients.
Randomly assigning large numbers of patients settled questions that many smaller studies could not, and did it quickly. It showed that fast, reliable answers are possible even in the middle of a pandemic. This is a descriptive account by the trial's own team rather than the formal results paper, and it covers UK hospital patients only.
See the evidence
An article by RECOVERY investigators describing how the UK randomized trial enrolled over 39,000 hospitalised patients at 178 sites by April 2021.
From the study: “Within 100 days of its initiation, RECOVERY demonstrated that dexamethasone improves survival for patients with severe disease”
Answer in 100 days
Pessoa-Amorim G et al. Making trials part of good clinical care: lessons from the RECOVERY trial. Future healthcare journal. 2021.
From the study: “Importantly, it also showed that other widely used treatments (such as hydroxychloroquine and azithromycin) have no meaningful benefit for hospitalised patients.”
Answer in 100 days
Pessoa-Amorim G et al. Making trials part of good clinical care: lessons from the RECOVERY trial. Future healthcare journal. 2021.
Ivermectin's apparent survival benefit vanished when low-quality and suspect trials were dropped.
Benefit fell 51% → 4%Apparent survival gain: all 12 trials vs low-bias trials only (4% not significant)
Pooling all 12 ivermectin trials showed an apparent 51% survival benefit. Keeping only low-risk-of-bias trials (the most carefully run) left a nonsignificant 4%: no reliable benefit. One potentially fraudulent trial showed about 12 times the effect of the low-risk trials. The authors had retracted their own earlier, positive analysis.
Public interest in ivermectin grew on pooled results that leaned on flawed and possibly fake trials. Checking raw patient data, not just published tables, is what exposed the problem. A re-analysis of 12 trials with 2,628 participants, many of them small; fraud checks were possible only where patient-level data were shared.
See the evidence
A meta-analysis that re-sorted 12 ivermectin treatment trials by risk of bias, after the team retracted its own July 2021 analysis.
From the study: “The analysis demonstrates that on including all 12 studies, ivermectin results in a significant 51% increase in survival”
Benefit fell 51% → 4%
Hill A et al. Ivermectin for COVID-19: Addressing Potential Bias and Medical Fraud. Open forum infectious diseases. 2022.
From the study: “According to the potentially fraudulent study (risk ratio [RR], 0.08; 95% CI, 0.02-0.35), ivermectin improved survival ~12 times more in comparison with low-risk studies”
Benefit fell 51% → 4%
Hill A et al. Ivermectin for COVID-19: Addressing Potential Bias and Medical Fraud. Open forum infectious diseases. 2022.
An estimated quarter of US COVID deaths were missing from death certificates.
1 in 4 deaths missedFrom 24% of estimated US COVID deaths missing from death certificates
A statistical model compared all deaths with those expected without the virus, using test data to track its spread. It estimated 766,611 COVID-19 deaths in the United States from March 8, 2020 to May 29, 2021; 184,477 (24%) were not documented on death certificates. Eighty-two percent of these unrecognized deaths were among persons aged ≥65 years.
The pandemic's true toll was larger than the official counts most people saw, and the missing deaths fell mostly on older adults. Counting all deaths, not only certified ones, gives a more honest picture. A model-based estimate built on all-cause deaths and virus testing data, not a count of confirmed individuals, and the most recent weeks may still be underestimated.
See the evidence
A modelling study of weekly all-cause deaths across 50 states, New York City and the District of Columbia, March 2020 to May 2021.
From the study: “We estimated that 766,611 deaths attributable to COVID-19 occurred in the United States from March 8, 2020-May 29, 2021.”
1 in 4 deaths missed
Iuliano AD et al. Estimating under-recognized COVID-19 deaths, United States, march 2020-may 2021 using an excess mortality modelling approach. Lancet regional health. Americas. 2021.
From the study: “Of these, 184,477 (24%) deaths were not documented on death certificates.”
1 in 4 deaths missed
Iuliano AD et al. Estimating under-recognized COVID-19 deaths, United States, march 2020-may 2021 using an excess mortality modelling approach. Lancet regional health. Americas. 2021.
From the study: “Eighty-two percent of unrecognized deaths were among persons aged ≥65 years”
1 in 4 deaths missed
Iuliano AD et al. Estimating under-recognized COVID-19 deaths, United States, march 2020-may 2021 using an excess mortality modelling approach. Lancet regional health. Americas. 2021.
BNT162b2's '95% protection' meant 8 COVID cases versus 162 on placebo.
8 vs 162 casesCOVID-19 cases, vaccine vs placebo group, from at least 7 days after the second dose
In this placebo-controlled trial, 43,548 people 16 years of age or older were randomly assigned to two doses of BNT162b2 (the Pfizer-BioNTech vaccine) or placebo. From at least 7 days after the second dose, there were 8 COVID-19 cases among vaccine recipients evaluated and 162 among placebo recipients. That gap in case rates is what '95% protection against Covid-19' means.
The figure compares case rates between two groups chosen by chance; it is not the chance that a vaccinated person gets sick. Random assignment is why the comparison can be trusted. An early report with a median of 2 months of safety follow-up, funded by BioNTech and Pfizer, so how long protection lasts was not yet known.
See the evidence
A randomized, observer-blinded trial at 152 sites, 130 of them in the United States, enrolling 43,548 people from July to November 2020.
From the study: “There were 8 cases of Covid-19 with onset at least 7 days after the second dose among participants assigned to receive BNT162b2 and 162 cases among those assigned to placebo”
8 vs 162 cases
Polack FP et al. Safety and Efficacy of the BNT162b2 mRNA Covid-19 Vaccine. The New England journal of medicine. 2020.
From the study: “A total of 43,548 participants underwent randomization, of whom 43,448 received injections: 21,720 with BNT162b2 and 21,728 with placebo.”
8 vs 162 cases
Polack FP et al. Safety and Efficacy of the BNT162b2 mRNA Covid-19 Vaccine. The New England journal of medicine. 2020.
After infection, 39 of 40 symptoms were likelier than in uninfected people.
39 of 40 symptomsSymptoms more common after infection than in uninfected controls, 50 pooled studies
Across 50 studies of 14,661,595 people, infected people had higher risk than uninfected controls for 39 of 40 symptoms; loss of smell was 4.31 times as likely. Fatigue and poor sleep, common in everyone, did not rank among the top risks once controls were counted. In never-hospitalised people, a few symptoms, such as sore throat, were not raised.
Many symptoms, like tiredness, are common anyway. Only a comparison with uninfected people separates problems tied to infection from everyday background, and here nearly every symptom stood out. Pooled observational studies that cannot prove cause; most identified uninfected comparison people by self-report rather than tests.
See the evidence
A systematic review and meta-analysis of 50 controlled studies from Europe, Asia, North America and elsewhere, searched January 2022 to August 2024.
From the study: “In all populations combined, there was an increased risk of a wide range of 39 out of 40 symptoms in those infected with SARS‑CoV‑2 compared to uninfected controls.”
39 of 40 symptoms
O'Mahoney LL et al. The risk of Long Covid symptoms: a systematic review and meta-analysis of controlled studies. Nature communications. 2025.
Adding unreviewed preprints rarely changed what COVID treatment reviews concluded.
96.7% same verdictComplement of the 3.3% (2 of 60) whose verdict changed with preprints added
Preprints are studies shared before peer review (expert checking). Across 60 comparisons of pooled COVID treatment results (meta-analyses) with and without preprints, the verdict of benefit, no effect or harm differed in only two (3.3%). Key methods and results showed few important differences from later journal versions, which took a median of about six months to appear.
In an emergency, waiting months for journals delays answers. COVID trial preprints were mostly trustworthy, though each study's quality still needs judging on its own. A retrospective review of 356 trials up to August 2021; certainty ratings still shifted in 9 of 60 comparisons, and many preprints had at least one reporting difference.
See the evidence
A review of 356 COVID treatment trials available as preprints, journal papers or both, with 120 meta-analyses run with and without preprints.
From the study: “Apart from two (3.3%) of 60 comparisons, point estimates were consistent between meta-analyses including preprints versus those excluding preprints”
96.7% same verdict
Zeraatkar D et al. Consistency of covid-19 trial preprints with published reports and impact for decision making: retrospective review. BMJ medicine. 2022.
From the study: “Key methods and results showed few important differences between trial preprints and their subsequent published reports.”
96.7% same verdict
Zeraatkar D et al. Consistency of covid-19 trial preprints with published reports and impact for decision making: retrospective review. BMJ medicine. 2022.
From the study: “The median time to publication of preprints was about six months.”
96.7% same verdict
Zeraatkar D et al. Consistency of covid-19 trial preprints with published reports and impact for decision making: retrospective review. BMJ medicine. 2022.
COVID-19 papers won journal acceptance far faster than flu papers in 2020.
8 vs 92 daysMedian days from submission to journal acceptance: COVID-19 vs influenza papers
Among COVID-19 research articles, the median time from submission to journal acceptance was 8 days, versus 92 days for human influenza papers. Nearly half (47.0%) were accepted within a week and 19.5% within one day, which the authors call a warning sign for quality control. There were 82 retractions and withdrawals for COVID-19 articles, versus 1 for influenza.
Speed helped share urgent findings, but very fast acceptance raises doubts about how closely papers were checked. A journal's stamp did not guarantee that a study was sound. A descriptive PubMed study of 31,319 COVID-19 and 4,287 influenza articles that measures speed, not each paper's quality, and COVID papers far outnumbered flu papers.
See the evidence
An analysis of editorial dates for research articles indexed in PubMed and published during 2020, comparing COVID-19 with human influenza.
From the study: “The median time to acceptance for COVID-19 was lower than that for human influenza (8 vs. 92 days).”
8 vs 92 days
Schonhaut L et al. Scientific publication speed and retractions of COVID-19 pandemic original articles. Revista panamericana de salud publica = Pan American journal of public health. 2022.
From the study: “47.0% of COVID-19 research articles were accepted within the first week of submission, and 19.5% within one day.”
8 vs 92 days
Schonhaut L et al. Scientific publication speed and retractions of COVID-19 pandemic original articles. Revista panamericana de salud publica = Pan American journal of public health. 2022.
From the study: “There were 82 retractions and withdrawals for COVID-19 articles, 1 for human influenza, and 5 for articles that contain both terms”
8 vs 92 days
Schonhaut L et al. Scientific publication speed and retractions of COVID-19 pandemic original articles. Revista panamericana de salud publica = Pan American journal of public health. 2022.
Long COVID guidance varied in practical applicability
A review found that guidance documents varied in methodological quality and practical applicability. Even among the prioritized, higher-scoring documents, applicability was the weakest assessed domain.
These were separate scores for guidance domains, not percentages of patients helped or a combined overall score. Diverse development methods and scarce guidance from lower-income settings limit comparison.
Study details & original results
Montoya Monsalve, Ricardo et al.. Methodological quality and clinical applicability of long COVID guidelines and scientific statements: a systematic review using AGREE II and AGREE-REX.. Frontiers in public health. 2026.
Separate rescaled AGREE II domain scores across the guidance documents; these are domain scores rather than proportions of documents and were not summed into an overall score.
Rigor of development 16.7-97.9%; applicability 0.0-95.8%; editorial independence median 100%
Applicability was the weakest domain among four prioritized documents contributing 99 recommendations, despite those documents scoring higher than the remaining corpus in all six domains.
68%
Reporting of a GRADE-compatible evidence-to-decision framework among the 20 appraised documents.
1 of 20 documents
Distribution across five NICE-derived care domains for recommendations extracted from the 19 documents identified in the initial search.
510 recommendations; 63.3% clinical management and 29.6% care planning
Model design changed the reported antiviral-treatment contrast
Antiviral-treatment estimates differed by model. Modified time-varying PyCox had higher reported scores for telling outcomes apart than time-fixed PyCox. Training was faster with modified PyCox than original PyCox or Cox.
This was methodological work in older Medicare beneficiaries. Confounding and diagnosis coding could affect results. The clinical outcome and treatment comparison group for the antiviral contrast are unavailable in this summary; the predicted health outcome is also unspecified. Neither prediction scores nor training speed establishes a patient benefit.
Study details & original results
Baik, Seo Hyon et al.. Deep learning survival analysis with time-varying covariates: extending PyCox to assess COVID-19 antiviral treatments and long-COVID associations.. JAMIA open. 2026.
Gradient in model-estimated antiviral treatment contrasts when comparing time-varying and time-fixed survival models
3.5-4.0 percentage points for time-varying models, 4.6 percentage points for time-fixed Cox, and 5.7 percentage points for time-fixed PyCox
Predictive discrimination of time-varying versus time-fixed PyCox models.
Time-dependent AUC 0.536-0.609 with modified PyCox versus 0.481-0.519 with time-fixed PyCox
Computational-efficiency comparison among survival-analysis approaches.
Training time: 1 min 14 s for modified PyCox, 3 min 18 s for original PyCox, and 5 min for traditional Cox
Fraud screening changed which survey responses were retained
One international web survey of adults living with Long COVID retained 17% of responses at Time 1 and 49% at Time 2 after fraud screening.
These are retention proportions in one survey, not the proportion of all Long COVID research that is fraudulent. Screening does not itself establish that every retained response was accurate.
Study details & original results
McDuff, Kiera et al.. Battling the Bots and Defending Against Fraudulent Responses in an International Community-Engaged Web-Based Survey With People Living With Long COVID: Methodological Study.. Journal of medical Internet research. 2026.
Responses retained after fraud screening in one international web-based survey of adults living with Long COVID
17% (798/4663) at Time 1; 49% (629/1281) at Time 2
Vagus-nerve-stimulation trials had substantial risk of bias
83%
A review of vagus nerve stimulation for hospitalized COVID-19 found high risk of bias in 83% of trials; the evidence base comprised 6 randomized trials and 221 patients.
This assesses trial methods, rather than establishing that stimulation helps or harms. The source identifies hospitalized COVID-19 patients; the findings do not establish results for a Long COVID treatment group.
Study details & original results
Balan, Adrian et al.. Efficacy and Safety of Vagus Nerve Stimulation for Hospitalized COVID-19 Patients: A Systematic Review and Methodological Evaluation of Randomized Controlled Trials.. Medicina (Kaunas, Lithuania). 2026.
Proportion of VNS for COVID-19 RCTs with high risk of bias
Dates show when findings were added here, not when papers were published. Study populations and comparisons differ.
WHEN YOU WANT TO GO DEEPERSee all 10 findingsOriginal results, comparisons and study details.
4 of 4 source groupsFindings stay together with their study.
SOURCE 054 findings
Montoya Monsalve, Ricardo et al.. Methodological quality and clinical applicability of long COVID guidelines and scientific statements: a systematic review using AGREE II and AGREE-REX.. Frontiers in public health. 2026.
Distribution across five NICE-derived care domains for recommendations extracted from the 19 documents identified in the initial search.
Added
510 recommendations; 63.3% clinical management and 29.6% care planningSource [5]
Separate rescaled AGREE II domain scores across the guidance documents; these are domain scores rather than proportions of documents and were not summed into an overall score.
Added
Rigor of development 16.7-97.9%; applicability 0.0-95.8%; editorial independence median 100%Source [5]
All 4 findings from this source
Reporting of a GRADE-compatible evidence-to-decision framework among the 20 appraised documents.
Applicability was the weakest domain among four prioritized documents contributing 99 recommendations, despite those documents scoring higher than the remaining corpus in all six domains.
Baik, Seo Hyon et al.. Deep learning survival analysis with time-varying covariates: extending PyCox to assess COVID-19 antiviral treatments and long-COVID associations.. JAMIA open. 2026.
Predictive discrimination of time-varying versus time-fixed PyCox models.
Added
Time-dependent AUC 0.536-0.609 with modified PyCox versus 0.481-0.519 with time-fixed PyCoxSource [4]
Computational-efficiency comparison among survival-analysis approaches.
Added
Training time: 1 min 14 s for modified PyCox, 3 min 18 s for original PyCox, and 5 min for traditional CoxSource [4]
All 3 findings from this source
Gradient in model-estimated antiviral treatment contrasts when comparing time-varying and time-fixed survival models
Added
3.5-4.0 percentage points for time-varying models, 4.6 percentage points for time-fixed Cox, and 5.7 percentage points for time-fixed PyCoxSource [4]
SOURCE 021 finding
McDuff, Kiera et al.. Battling the Bots and Defending Against Fraudulent Responses in an International Community-Engaged Web-Based Survey With People Living With Long COVID: Methodological Study.. Journal of medical Internet research. 2026.
Responses retained after fraud screening in one international web-based survey of adults living with Long COVID
Added
17% (798/4663) at Time 1; 49% (629/1281) at Time 2Source [2]
SOURCE 012 findings
Balan, Adrian et al.. Efficacy and Safety of Vagus Nerve Stimulation for Hospitalized COVID-19 Patients: A Systematic Review and Methodological Evaluation of Randomized Controlled Trials.. Medicina (Kaunas, Lithuania). 2026.
Browse the 13 original sourcesThe complete bibliography behind this monitor.
[1]
Balan, Adrian et al.. Efficacy and Safety of Vagus Nerve Stimulation for Hospitalized COVID-19 Patients: A Systematic Review and Methodological Evaluation of Randomized Controlled Trials.. Medicina (Kaunas, Lithuania). 2026.
McDuff, Kiera et al.. Battling the Bots and Defending Against Fraudulent Responses in an International Community-Engaged Web-Based Survey With People Living With Long COVID: Methodological Study.. Journal of medical Internet research. 2026.
Kazi, Fatema et al.. Implications for future pandemics from a living systematic review and critical evaluation of observational studies of the effectiveness of COVID-19 vaccination against the Omicron variant.. Vaccine. 2026.
Baik, Seo Hyon et al.. Deep learning survival analysis with time-varying covariates: extending PyCox to assess COVID-19 antiviral treatments and long-COVID associations.. JAMIA open. 2026.
Montoya Monsalve, Ricardo et al.. Methodological quality and clinical applicability of long COVID guidelines and scientific statements: a systematic review using AGREE II and AGREE-REX.. Frontiers in public health. 2026.
Shepshelovich D et al. Concordance between the results of randomized and non-randomized interventional clinical trials assessing the efficacy of drugs for COVID-19: a cross-sectional study. The Journal of antimicrobial chemotherapy. 2021.
Iuliano AD et al. Estimating under-recognized COVID-19 deaths, United States, march 2020-may 2021 using an excess mortality modelling approach. Lancet regional health. Americas. 2021.
Zeraatkar D et al. Consistency of covid-19 trial preprints with published reports and impact for decision making: retrospective review. BMJ medicine. 2022.
Schonhaut L et al. Scientific publication speed and retractions of COVID-19 pandemic original articles. Revista panamericana de salud publica = Pan American journal of public health. 2022.
Read each original paper for its complete methods, population and limitations. Different studies can ask different questions and report different kinds of results. Dates marked “added” describe when a finding entered this monitor, not when the study was published or clinically reviewed.
Download the findings and bibliography displayed in this page snapshot. Values and confidence intervals remain as reported text; this is not a pooled estimate or a systematic review. Product-label observations and editorial takeaways are excluded.
These findings may omit evidence or contain errors and have not been reviewed by a clinician. A citation or download does not establish source verification. Read the original studies for their methods and limitations.
The October 2026 COVID citation audit found errors in this collection. These downloads may still contain affected findings and bibliography entries. Read the correction record before using them.
“Added” dates record entry into the monitor, not publication or clinical review. A null original acquisition or derivation means its historical record is unknown; a current record or verification has its own observation time. The SHA-256 snapshot identifier binds the downloadable public fields. CSV includes finding rows and bibliography-only source rows; JSON includes both lists. Spreadsheet-sensitive CSV cells begin with an apostrophe for safety; JSON preserves the original text.
For attribution, cite ModernDoc and the page URL, data update date and snapshot identifier in the download. Original studies retain their own rights; this download does not grant a licence to their full text.