Read the safety table: what “well tolerated” can leave out

Start with what the event label means

An adverse event is an unfavourable medical occurrence during exposure or observation; the label alone does not establish causation. An adverse reaction involves a causal relationship being considered. ICH E2A also distinguishes severe intensity from seriousness, which relates to outcomes such as hospitalization or threats to life and function.

For an invented trial, a severe short-lived headache and a hospitalization with modest symptoms might occupy different reporting categories. The categories describe different dimensions, so one should not replace the other in a summary.

Write down the actual label used in the table. If the authors report all events regardless of attribution, do not rewrite that as all events caused by treatment.

Count people and episodes separately

Imagine two fictional groups of 1,000 people, all followed for six months. Two hundred people in A and eighty in B experience at least one episode of nausea. The person-based proportions are 20% and 8%, a 12-percentage-point difference.

Now suppose A’s two hundred people report 600 episodes. That count answers a different question. Six hundred episodes do not mean that 60% of the group experienced nausea, because one person can have several episodes.

Look for “participants with at least one event,” “number of events” and “events per person-time.” Keep the unit in every sentence. When observation periods differ, a simple comparison of totals can become especially misleading.

Do not subtract harms from benefits as if they were identical

Add two rows to our invented table: twelve people in A and ten in B have a serious event; fifty and twenty discontinue because of an adverse event. These categories may overlap, and their importance differs from nausea.

If another table reports thirty fewer disease events with A, subtracting the extra nausea cases to produce one “net health score” would assign unspoken values to very different experiences. It could also double-count people who appear in several rows.

A useful article places outcomes alongside one another, with common denominators and time frames where possible. It leaves the value judgement visible rather than hiding it inside an arithmetic total.

Ask how hard the investigators looked

A study that asks a structured symptom questionnaire at every visit may detect different events from one that waits for unsolicited reports. The reporting period, definitions and collection method therefore belong beside the event counts. Cochrane’s adverse-effects chapter discusses these challenges.

For our fictional study, ask whether follow-up continued after discontinuation, whether both groups had similar contact, and whether only events above a frequency threshold appeared in the main paper. A missing row is not evidence of zero events.

Check supplements when available. If the full table cannot be accessed, make that limitation explicit instead of using the abstract’s reassuring phrase as a substitute for the unavailable data.

What zero events can still allow

In a separate invented example, suppose 300 independent participants have complete follow-up and none experiences a particular event. Under a simple binomial model with the same event probability for each person, the probability of zero events is (1−p)³⁰⁰.

Setting that probability to 0.05 gives an exact one-sided 95% upper confidence bound of 1−0.05^(1/300), approximately 0.994%. This original calculation shows that a nonzero event probability remains compatible with observing none under the model.

The calculation does not address delayed effects, unmeasured events or a different population. It is not a personal risk estimate. Its point is narrower: zero in a finite sample is different from a demonstrated impossibility.

Connect trials with longer-term safety evidence

A randomized trial can provide a useful comparison during its observation window. Rare or delayed harms may require other evidence, including large observational studies. Spontaneous reports can raise signals, but without a suitable denominator they do not directly supply an incidence rate.

For the fictional intervention, the next useful question might concern duration: do events accumulate after six months? Another might concern a population excluded from the trial. Label those as gaps, not predicted harms.

When a future paper arrives, state whether it adds a controlled comparison, longer observation or a signal requiring investigation. A useful safety story follows that learning process instead of replacing one blanket adjective with another.

A fictional six-month safety table

Invented counts with complete six-month follow-up. Categories may overlap. No causal attribution, uncertainty estimate or overall benefit–harm verdict is implied.