Understanding Adverse Event Rates: Percentages and Relative Risk in Clinical Trials

When a new drug is tested in a clinical trial, safety isn’t just about whether someone got sick-it’s about how often and under what conditions that sickness happened. Two patients might both experience a headache after taking the same medicine. But if one was on the drug for 30 days and the other for 18 months, the risk isn’t the same. That’s why simple percentages can mislead-and why regulators like the FDA now demand more precise ways to measure adverse events.

Why Simple Percentages Fail in Safety Analysis

The most common way to report adverse events is the Incidence Rate (IR): the number of people who had an event divided by the total number of people exposed. If 15 out of 100 patients got a rash, you say 15% experienced it. Simple. Clean. But here’s the problem: this method ignores how long each person was actually exposed to the drug.

Imagine two groups in a trial. Group A gets the new drug for six months. Group B gets it for three years. If 10 people in Group A get a headache, that’s 10%. In Group B, 20 people get headaches-that’s also 10%. But that doesn’t mean the risk is equal. Group B had 18 times more total exposure time. The IR makes them look the same. They’re not.

A 2010 analysis by Andrade showed IR underestimates true event rates by 18% to 37% when treatment durations differ between groups. That’s not a small error. It’s the difference between thinking a drug is safe and realizing it might be dangerous over time.

Enter Patient-Years: The EIR Method

To fix this, statisticians use Event Incidence Rate adjusted by Patient-Years (EIR). Instead of counting people, you count time. One patient on the drug for one year equals one patient-year. If five patients each take the drug for two years, that’s 10 patient-years.

The formula is straightforward: divide the total number of events by the total patient-years, then multiply by 100 to get events per 100 patient-years. So if 20 headaches happened across 1,000 patient-years of exposure, the EIR is 2.0 per 100 patient-years.

This method works better for recurrent events. If one patient gets a headache five times over 18 months, EIR counts all five. IR only counts them once. That’s more honest. It tells you how often the problem occurs, not just how many people saw it.

But EIR has its own flaw. It can overstate risk if one person has multiple events. A patient who gets 12 nausea episodes in a year makes the rate look worse than if 12 different people each had one. For safety signals, that’s useful. For regulatory approval? It can muddy the picture.

The FDA’s Shift to Exposure-Adjusted Incidence Rate (EAIR)

In 2023, the FDA requested EAIR in a supplemental biologics license application. That was a signal. The agency was done with oversimplified numbers.

EAIR builds on EIR but adds one critical layer: it accounts for event recurrence and variable exposure time and treatment interruptions. If a patient stops the drug for two months due to side effects, that time isn’t counted as exposure. If they restart, exposure resumes. This mirrors real-world use.

EAIR doesn’t have one universal formula-but it’s defined by its purpose: to reflect the actual risk of experiencing an event during the time you’re exposed. It’s not about how many people got sick. It’s about how likely you are to get sick while taking the drug.

MSD’s safety team found that switching to EAIR uncovered previously hidden safety signals in 12% of their chronic therapy programs. These were events that only showed up when exposure time was properly measured-events that IR would have buried.

Relative Risk and Confidence Intervals: Comparing Groups Accurately

You can’t just report numbers. You have to compare them. That’s where relative risk comes in.

The Incidence Rate Ratio (IRR) is the ratio of the IR or EIR between two groups. If Group A has an EIR of 2.5 per 100 patient-years and Group B has 5.0, the IRR is 2.0. Group B has twice the rate.

But is that real? Or just random noise? That’s where confidence intervals matter. The Wilson score method with continuity correction is used for single incidence rates. For comparing two groups, the Wald method is standard. In R, statisticians use prop.test() for rates and riskratio() for ratios.

If the 95% confidence interval for an IRR crosses 1.0, the difference isn’t statistically significant. Many regulators now require this level of precision. A p-value alone isn’t enough. You need to show the range of possible values.

Competing Risks and Why Kaplan-Meier Doesn’t Work

Not all events are created equal. If a patient dies during a trial, they can’t experience a future adverse event. That’s called a competing risk.

Traditional methods like the Kaplan-Meier estimator assume that if someone drops out, it’s random. But death isn’t random-it’s directly related to the disease or treatment. Using Kaplan-Meier here gives false results. It makes the drug look safer than it is because it treats death as a censoring event, not a competing outcome.

A 2025 study in Frontiers in Applied Mathematics and Statistics showed that cumulative hazard ratio estimation-breaking down the hazard function into separate risks for death and the adverse event-improves accuracy by 22% when competing event rates exceed 15%. That’s not a minor tweak. It’s a fundamental correction.

The FDA and EMA both now expect statisticians to acknowledge competing risks in their analyses. Ignoring them isn’t just sloppy-it’s misleading.

Industry Adoption: Progress, Pain, and Pitfalls

The industry is moving. In 2020, only 12% of regulatory submissions included exposure-adjusted metrics. By 2023, that jumped to 47%. CDISC adoption for safety reporting is now at 89% among top pharmaceutical companies.

But adoption isn’t smooth. A 2024 PhUSE survey found that 42% of companies had formatting issues when submitting EAIR to regulators. Medical reviewers didn’t understand it. One Roche report said 35% of reviewers misinterpreted EAIR results at first.

Programming is harder too. Building EAIR in SAS takes 3.2 times longer than IR. Median development time: 14.7 hours vs. 4.5. Common errors? Wrong event dates (28%), ignoring treatment breaks (19%), inconsistent patient-year math (23%).

The PhUSE team released standardized SAS macros for EAIR in March 2023. They’ve been downloaded over 1,800 times. Users report an 83% drop in programming errors. That’s progress.

What You Need to Know for Real-World Safety Decisions

Here’s the bottom line:

• Never rely on simple percentages when treatment durations vary.
• EIR gives you a better picture than IR-but it still counts events, not people.
• EAIR is the gold standard for regulatory submissions today. It accounts for time, recurrence, and interruptions.
• Always report confidence intervals. Don’t just say “it’s higher.” Show how much higher-and how sure you are.
• Competing risks like death must be modeled separately. Kaplan-Meier is outdated for safety analysis.

The FDA’s 2024 draft guidance on exposure-adjusted analysis suggests standardized EAIR methods will be mandatory soon. By 2027, 92% of Phase 3 submissions are expected to include it.

If you’re reviewing trial data, asking for IR alone is like judging a car’s speed by how far it drove-not how long it was moving. You’re missing half the story.

What’s Next for Adverse Event Analysis?

The FDA’s Sentinel Initiative is testing machine learning tools that use exposure-adjusted metrics to detect safety signals earlier. Early results show 38% better detection than traditional methods.

New MedDRA coding terms (version 26.1, 2023) now include 47 new codes for time-to-event reporting. Regulatory templates now require detailed documentation of exposure time calculation methods.

The message is clear: safety isn’t about counting heads. It’s about measuring time, understanding risk over duration, and being honest about what happens when people live with treatment.

The next time you see a headline saying “X% of patients had side effects,” ask: How long were they on the drug? If you can’t answer that, you don’t know the real risk.