Multivitamins and COVID: What an 18,000-Person Randomized Trial Found

Multivitamins are among the most-taken supplements on earth and among the least-tested in
randomized trials, mostly because a trial large enough and long enough to answer the question is
expensive. COSMOS is one of the exceptions, and a COVID-era analysis of it has now been published in
the American Journal of Clinical Nutrition.

What the trial was

COSMOS, the Cocoa Supplement and Multivitamin Outcomes Study, ran from June 2015 to
December 2020 as a randomized, double-blinded, placebo-controlled trial. The COVID-19
analysis covers January to December 2020, with 18,205 participants still enrolled as
of 1 January 2020.

That design matters more than the size. Randomized and double-blinded means neither participants
nor investigators knew who was taking the supplement, which removes the two largest sources of bias
in supplement research. Most multivitamin evidence is observational, and people who take
multivitamins differ from people who do not in ways that are difficult to adjust away.

The primary result: no effect on infection

Through 31 December 2020, 382 participants assigned to the multivitamin-mineral
and 404 assigned to placebo reported a COVID-19 infection.

The hazard ratio was 0.93, with a 95% confidence interval of 0.81 to 1.07.

That interval is the whole answer. Because it spans 1.0, the data are compatible with a small
benefit, no effect, or a small harm. The honest reading is that a daily multivitamin did not prevent
COVID-19 infection in this population, and the trial was large enough that a meaningful protective
effect would likely have shown up.

The secondary result, and why it is more fragile

Among 338 cases with symptom data available, the odds ratio for
symptomatic COVID-19 comparing multivitamin against placebo was 0.70 (95% CI 0.44 to
1.11) in the intention-to-treat analysis. Again the interval crosses 1, so this is not a
significant result.

In a per-protocol analysis, the odds ratio was 0.60 (95% CI 0.37 to
0.99), which the authors describe as a promising signal lowering the odds of symptomatic
COVID-19 illness.

Note the upper bound: 0.99. That is significance by the narrowest possible margin, in a secondary
outcome, in a subgroup with symptom data, under a per-protocol analysis. Each of those four
qualifiers weakens a finding, and here they stack.

What intention-to-treat and per-protocol mean

This distinction decides how much weight the symptom result deserves, so it is worth being clear
about.

Intention-to-treat analyses everyone as randomized, whether or not they took the
supplement as instructed. It preserves the randomization, which is the entire scientific value of a
trial, and it tends to underestimate the effect of a treatment because it includes people who did not
really take it.

Per-protocol analyses only those who followed the protocol. That sounds more
accurate and is in fact less reliable, because the people who take a supplement faithfully for years
are systematically different from those who stop. Adherence is a behaviour, and behaviours travel with
other health habits. The moment you filter on adherence, you have partly undone the randomization you
paid for.

So the pattern here, no significant effect on intention-to-treat and a just-significant effect on
per-protocol, is exactly the pattern that should prompt caution rather than a headline.

What this does and does not tell you

Supported: a daily multivitamin-mineral did not reduce COVID-19 infections in a
large randomized trial. There is a weak, unconfirmed hint that it may have been associated with fewer
symptomatic cases among those infected.

Not supported: that multivitamins prevent COVID-19, that they reduce severity, or
that anyone should start or stop taking one on the basis of this analysis. The symptom finding is a
reason to run another study, not a reason to change a shelf.

It is also worth naming what this analysis is: a secondary look at a trial designed for something
else. COSMOS was built to study cocoa extract and multivitamins in relation to cardiovascular disease
and cancer, and COVID-19 arrived in the middle of it. That is not a criticism; using an existing
randomized cohort to answer a new question is efficient and legitimate. It does mean the trial was not
powered for this outcome.

Frequently Asked Questions

Did multivitamins prevent COVID-19?

No. In the COSMOS randomized trial, 382 participants on the multivitamin-mineral and 404 on placebo
reported COVID-19 infection through the end of 2020, with a hazard ratio of 0.93 and a 95% confidence
interval of 0.81 to 1.07. Because that interval includes 1.0, the result is not statistically
significant and is consistent with no effect on infection.

What about severity of illness?

That is where the more interesting number sits, and it needs handling carefully. Among 338 cases
with symptom data, the odds ratio for symptomatic COVID-19 was 0.70 with a confidence interval of 0.44
to 1.11 in the intention-to-treat analysis, which is not significant. A per-protocol analysis produced
0.60 with an interval of 0.37 to 0.99, which the authors call a promising signal. An upper bound of
0.99 is significance by the narrowest possible margin.

Why does per-protocol matter less than it sounds?

Because it analyses only the people who took the supplement as instructed, and adherence is not
random. People who take a daily pill faithfully for years tend to differ from people who drift off it,
often in ways connected to overall health behaviour. Filtering on adherence partly undoes the
randomization that makes a trial trustworthy, which is why intention-to-treat is the primary analysis
and per-protocol is supporting evidence.

How big and how good was the trial?

18,205 participants remained in the trial as of 1 January 2020, and the design was randomized,
double-blinded and placebo-controlled, running from June 2015 to December 2020. That combination is
rare in supplement research and is the main reason this analysis is worth reading. Most multivitamin
evidence is observational, which cannot separate the supplement from the kind of person who takes
one.

Was the trial designed to study COVID?

No. COSMOS was designed to study cocoa extract and multivitamin supplementation in relation to
cardiovascular disease and cancer, and the pandemic occurred during the trial. Analysing an existing
randomized cohort for a new question is a legitimate and efficient thing to do, but it means the study
was not powered specifically for COVID-19 outcomes, and it is one more reason to treat the symptom
finding as preliminary.

Should I stop taking my multivitamin?

Nothing in this analysis is a reason to start or stop. It addresses one narrow question, whether a
daily multivitamin-mineral affected COVID-19 infection and symptoms during 2020, and the answer on
infection was no. People take multivitamins for other reasons, including correcting identified
deficiencies, and those reasons are untouched by this result. Decisions about supplements belong with
a clinician who knows your bloodwork.

What would settle the symptom question?

A trial designed for it in advance, with symptomatic illness as a pre-specified primary outcome and
enough cases to give a confidence interval that does not brush against 1.0. The current result comes
from 338 cases in a secondary analysis of a trial built for something else. That is precisely the
situation where a promising signal either replicates or evaporates, and there is no way to tell which
from the data in hand.

Why do confidence intervals keep coming up here?

Because in this analysis they carry the meaning that the point estimates hide. A hazard ratio of
0.93 sounds like a 7% reduction, but with an interval of 0.81 to 1.07 the data cannot distinguish that
from no effect at all. Reading the interval rather than the headline number is the single most useful
habit when reading trial results, and it is the part that most coverage leaves out.

Sources

• Multivitamin supplementation and COVID-19 incidence and symptom severity in the COcoa Supplement
  and Multivitamin Outcomes Study (COSMOS) randomized trial. American Journal of Clinical
  Nutrition, 2026. PMID 42385272. DOI 10.1016/j.ajcnut.2026.101355. 18,205 participants remaining as
  of 1 January 2020; 382 MVM and 404 placebo COVID-19 cases through 31 December 2020; hazard ratio 0.93
  (95% CI 0.81, 1.07); symptomatic illness odds ratio 0.70 (95% CI 0.44, 1.11) intention-to-treat and
  0.60 (95% CI 0.37, 0.99) per-protocol among 338 cases —
  PubMed record
• COSMOS trial, background on the study design and its primary cardiovascular and cancer outcomes —
  ClinicalTrials.gov

Participant numbers, case counts, hazard ratio, odds ratios and
confidence intervals above are taken from the published abstract of the COSMOS COVID-19 analysis,
checked August 2026.