Endpoints, Sample Size and Why the Numbers Matter

Part of Clinical Research Coordinator (CRC) Training: The Protocol and Study Design.

About 5 minutes. Free, in your language, no account needed.

Why this matters

Coordinators are often told statistics are not their concern. They are, for one reason: the statistical plan is the thing your daily work either supports or quietly destroys.

A missed primary endpoint measurement, a participant lost to follow-up, a visit outside its window — each of these is a small operational event and a real statistical loss. This lesson explains the connection so you can protect the study without becoming a statistician.

Endpoints

An endpoint is the specific measurement used to answer a study question.

• The primary endpoint answers the main question. There is usually one. The whole study is sized around it.
• Secondary endpoints answer supporting questions.
• Exploratory endpoints look for things worth studying later.
• Safety endpoints describe harms.

Endpoints differ in how directly they matter to a patient.

A clinical endpoint is something a patient feels or experiences: survival, a heart attack, a hospital admission, pain. A surrogate endpoint is a measurement believed to predict a clinical one: blood pressure, cholesterol, tumor size, a laboratory marker. Surrogates make studies faster and smaller, but they can mislead, because changing a marker does not always change what happens to the person.

For your work, the practical point is that the primary endpoint measurement is the highest-value data point in the study. If a visit has to be shortened, the primary endpoint assessment is the last thing to go, not the first.

Sample size, in plain language

The protocol states how many participants the study needs. That number is calculated in advance from four things.

• The effect size the study is designed to detect. A smaller effect needs more participants.
• Variability in the measurement. Noisier measurements need more participants.
• Power, the probability of detecting a real effect if one exists. Commonly set at 80 or 90 percent.
• The significance level, often written as alpha, commonly 0.05, which controls the risk of concluding there is an effect when there is not.

The calculation usually assumes a certain rate of dropout. If more people leave than expected, or more data are missing than expected, the study can end up underpowered: unable to answer its own question even though everyone did their job.

That is why retention is a scientific activity, not a courtesy. Module 6, Lesson 3 covers it.

Bias, confounding, and what randomization cannot fix

Bias is a systematic error that pushes results consistently in one direction. Common kinds you can influence:

• Selection bias — who gets into the study, or into which arm. Randomization addresses this.
• Measurement bias — how a measurement is taken. A blood pressure cuff of the wrong size, or one person rounding readings up, produces bias. Training and standard procedures address this.
• Attrition bias — who leaves. If the people who leave are systematically different from those who stay, the remaining group no longer represents the original one.

Randomization does not protect against sloppy measurement or unequal dropout. Those are site responsibilities.

Missing data and analysis populations

The protocol will describe how the analysis handles participants. Two terms recur.

• Intention to treat analyzes participants in the group they were randomized to, regardless of what they actually received. It preserves the benefit of randomization and is usually the primary approach.
• Per protocol analyzes only those who followed the protocol closely. It answers a different, more optimistic question and is normally supportive rather than primary.

Because intention to treat keeps everyone, data from participants who stopped treatment are still valuable. A participant who stops the tablet has not stopped being part of the study. Ask, in line with the protocol and their consent, whether they will still attend follow-up. Many will.

Statistical significance is not importance

A result can be statistically significant and clinically trivial, especially in a very large study. A p-value below 0.05 says the result is unlikely to be chance alone. It says nothing about whether the effect is large enough to matter to a patient. Confidence intervals and effect sizes carry that information.

You do not have to judge this. You do need to be able to hear a participant say "so it works" and answer carefully.

In practice

VT-201 was sized on the assumption that 15 percent of participants would leave before the final visit. At Northside Clinic the figure is running at 28 percent.

Nadia looks at why. Almost all of the leavers stopped attending after the visit that requires fasting, which is booked in the afternoon at her site. People arrive hungry and irritable and do not come back.

She raises it with Dr. Osei and with Marcus, the monitor. The protocol does not require an afternoon appointment; the site's scheduling did. Fasting visits move to the first slot of the morning. Attrition at the site falls over the following quarter.

Nadia also revisits the three participants who stopped taking the tablet but never formally withdrew. Two agree to keep attending for assessments. Under an intention-to-treat analysis their data still count.

Nobody broke a rule here. The study was being damaged by an appointment time.

Key takeaways

• The primary endpoint is the study's most important measurement, and it is the last assessment to be sacrificed when a visit is under pressure.
• Sample size is calculated in advance from effect size, variability, power and the significance level, and it assumes a dropout rate; excess dropout can leave a study underpowered.
• Randomization fixes selection bias but not measurement bias or attrition bias, and both of those are influenced directly by site practice.
• Intention-to-treat analysis keeps participants in their randomized group, which is why follow-up data from people who stopped treatment remain valuable.
• Statistical significance is not clinical importance, and effect sizes and confidence intervals carry the information about whether a result matters to a patient.