A safety table can contain correct arithmetic and still invite an incorrect comparison. The denominator, treatment exposure, observation window, and counting rule determine what the percentage represents. These decisions should be visible to the reviewer before a difference between treatment groups is interpreted.
The FDA’s premarketing risk assessment guidance is a useful safety-evaluation reference. The review workflow below is our proposed way of making the meaning of a descriptive safety display easier to inspect.
1. Identify what is being counted
Participants with at least one event and the total number of events answer different questions. Recurrent events may contribute multiple times to an event count but only once to a participant incidence. State the counting rule in both the specification and the display.
For a hypothetical headache summary, ten events among three participants should not be presented as ten participants with headache. Review participant-level contributions for repeated events and confirm that hierarchical summaries do not accidentally count the same participant more than intended.
2. Define the population and observation window
Explain how participants enter the safety population and how treatment assignment is represented. Define the period in which an event is considered relevant to the display, including the treatment-emergent rule. Incomplete dates require explicit handling rather than silent assumptions.
CDISC ADaM provides an upstream analysis-data framework. Our practical recommendation is to retain the inputs behind each population and window flag so that a reviewer can inspect boundary cases without reconstructing the program.
3. Examine unequal exposure
Two treatment arms may have different follow-up or time on treatment. A participant percentage alone does not account for that difference. An exposure-adjusted rate may provide additional information, but its interpretation depends on the time-at-risk definition and the treatment of recurrent events.
Do not present an adjusted rate as automatically resolving confounding or establishing causality. Examine the clinical context, reasons for discontinuation, and the relationship between exposure and event risk. Report the unit of person-time and the rule for stopping its accumulation.
4. Link summaries to patient context
A clinically important imbalance should lead to participant timelines, relevant laboratory findings, concomitant medication, and treatment changes. Those details do not replace aggregate comparison; they help explain what the aggregate includes.
Astraea’s patient and study review capabilities support moving between broad patterns and individual context. The broader operating principle is to keep a summary connected to the records that contribute to it.
Before releasing a safety display, ask a reviewer to explain its numerator, denominator, observation period, and limitations in one paragraph. If they cannot do so from the package, improve the supporting metadata and footnotes. Clear arithmetic is necessary; clear meaning is equally important.
A worked counting distinction
In a hypothetical group of 40 treated participants, three participants experience five events of one type. Participant incidence is 3/40, or 7.5%; the event count is five. If the analysis accumulates 20 participant-years of the specified time at risk, an event-based rate is 5/20, or 0.25 events per participant-year. A first-event rate uses a different numerator and may use a different time-at-risk rule.
These quantities should not be substituted for one another. Record whether exposure stops at the first event, treatment discontinuation, a follow-up boundary, or another specified point. Make unit conversion explicit when reporting rates per 100 participant-years.
This arithmetic example illustrates display semantics only. It does not establish a treatment effect, resolve unequal follow-up, or prescribe the appropriate estimator for a study.
