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

When you hear that 15% of patients in a drug trial had a headache, it sounds simple. But is that number really telling you how risky the drug is? Not if you don’t know how long those patients were taking it. That’s where adverse event rates get complicated - and why the FDA is pushing for smarter ways to measure them.

Why Simple Percentages Don’t Tell the Whole Story

For years, the go-to way to report side effects in clinical trials was the Incidence Rate (IR): just divide the number of people who had an adverse event by the total number of people in the study. If 30 out of 200 patients got nausea, you say, "15% experienced nausea." Easy. But here’s the problem: what if one group took the drug for 3 months and another took it for 2 years? The group with longer exposure will naturally have more events - not because the drug is more dangerous, but because they were exposed longer. Using raw percentages hides that difference.

A 2010 analysis by Andrade found this method underestimates true event rates by 18% to 37% in trials with uneven treatment durations. Imagine two diabetes drugs. One requires daily pills for life. The other is only used for 6 months. If 10% of patients on the long-term drug get liver enzyme changes, but only 8% on the short-term drug do, is the first drug really worse? Not if you don’t account for exposure time. That’s why regulators are moving away from this outdated approach.

The Shift to Exposure-Adjusted Metrics

In 2023, the FDA requested that a biologics company submit data using Exposure-Adjusted Incidence Rate (EAIR) instead of traditional IR. It was a quiet but powerful signal: the agency no longer accepts simplistic percentages for serious safety assessments.

EAIR counts how many events happen per 100 patient-years. A patient-year means one person taking the drug for one full year. If 10 patients each took the drug for 2 years, that’s 20 patient-years. If 5 of them had a rash, the EAIR is 25 events per 100 patient-years (5 ÷ 20 × 100).

This method doesn’t just count people - it counts time. That’s critical for chronic conditions. A drug for rheumatoid arthritis might cause joint pain in 12% of patients over 6 months. But over 5 years, that same 12% could become 45% - not because the drug got worse, but because people were exposed longer. EAIR captures that trend accurately.

How EIR Fits In - And Where It Falls Short

Another method, Event Incidence Rate (EIR), also uses patient-years but focuses on total events, not unique patients. So if one person gets three headaches in a year, EIR counts three events. That’s useful for recurrent issues like migraines or diarrhea. But it can inflate risk. If 5 people each get 4 episodes of vomiting over a year, EIR says there are 20 events. IR says 5 people were affected. Which tells you more about safety? It depends on the question.

EIR is great for understanding how often side effects occur. But if your goal is to know how many people are at risk, IR still has value. The key is knowing which metric answers which question. The FDA doesn’t demand one method - it demands justification. If you use IR, you must explain why it’s appropriate for your study.

Relative Risk and Why Confidence Intervals Matter

Comparing two drugs? You don’t just look at their individual rates. You calculate the Incidence Rate Ratio (IRR) - the ratio of one group’s rate to the other’s. If Drug A has an EAIR of 30 per 100 patient-years and Drug B has 15, the IRR is 2.0. That means Drug A has twice the rate of the event.

But is that difference real? Or just random noise? That’s where confidence intervals come in. Statisticians use the Wald method to calculate these for IRR. If the 95% confidence interval for that 2.0 IRR runs from 1.1 to 3.8, you can say with confidence that Drug A is genuinely riskier. If the interval crosses 1.0 - meaning it could be 1.0 or lower - then you can’t rule out chance.

The Wilson score method is often used for IR and EIR confidence intervals. These aren’t just math exercises. They’re the difference between making a life-changing decision based on solid evidence - or a misleading number.

Competing Risks and Why Kaplan-Meier Fails Here

What happens when a patient dies before you can see if they’d develop a liver injury? That’s a competing risk. Traditional methods like the Kaplan-Meier estimator assume everyone stays at risk until the end of the study. But in reality, death removes someone from the pool of people who could have an adverse event.

A 2025 study in Frontiers in Applied Mathematics and Statistics showed that using Kaplan-Meier in these cases leads to biased results - sometimes overestimating risk by 20% or more. The better approach? Cumulative hazard ratio estimation. It breaks down risk into separate components: death, and the adverse event you’re tracking. This gives a clearer picture of what’s actually happening.

For drugs treating advanced cancer or heart failure, this isn’t academic. It’s essential. If you misrepresent the safety profile, you could scare doctors away from a life-extending treatment - or worse, let a dangerous drug slip through.

What This Means for Real-World Trials

Companies are catching on. In 2023, 47% of FDA submissions included exposure-adjusted metrics - up from just 12% in 2020. Drug makers are investing in new software, training teams, and rewriting analysis pipelines. MSD found that switching to EAIR uncovered previously hidden safety signals in 12% of their chronic disease programs. Roche, on the other hand, had to train medical reviewers because 35% initially misunderstood EAIR results.

SAS and R programmers now have standardized macros to calculate EAIR. The PhUSE GitHub repository for these tools has been downloaded over 1,800 times. But it’s not easy. One programmer said EAIR took 3.2 times longer to code than IR. Common mistakes? Wrong start/end dates, ignoring treatment breaks, miscalculating patient-years.

CDISC, the global standard for clinical data, now requires both IR and EAIR for serious adverse events in oncology trials. The FDA’s Biostatistics Review Template includes checklists to make sure companies are doing it right.

What You Need to Know - Even If You’re Not a Statistician

You don’t need to write SAS code to understand this. But you do need to ask the right questions when reading clinical trial results:

• Is the adverse event rate reported as a simple percentage - or adjusted for time?
• How long did patients actually take the drug?
• Were the treatment groups exposed for similar durations?
• Is there a confidence interval? Does it include 1.0?
• Are competing risks (like death) accounted for?

If a study only gives you IR without context, treat it like a headline without the article. It might be true - but it’s incomplete.

The Future Is Exposure-Adjusted

By 2027, experts predict 92% of Phase 3 drug submissions will include EAIR alongside traditional IR. The FDA’s 2024 draft guidance is pushing for standardization. Machine learning tools are being trained to detect safety signals using these metrics - and they’re already 38% better at catching early warnings than old-school methods.

This isn’t about making statistics harder. It’s about making safety clearer. A drug might have a higher percentage of side effects - but if those side effects only happen in the first week, and patients stop taking it, the real risk is low. Another drug might have a lower percentage - but if patients stay on it for years, those same side effects could become life-altering.

The goal isn’t to scare people away from new medicines. It’s to make sure we understand exactly what we’re getting into. And that starts with asking: adjusted for what?