Why is mathematics important to clinical trials? A trial is not simply a group of people taking a treatment and reporting what happened. Health can improve or worsen for many reasons: natural recovery, expectations, other care, different starting risks, measurement error and chance. Mathematics helps researchers turn a focused health question into a design that can separate a treatment effect from those competing explanations.
The key mechanisms are randomisation, comparison groups, outcome definitions, effect sizes, uncertainty, sample-size planning and sensitivity analysis. None of these makes a trial automatically ethical, unbiased or relevant to every patient. They create a disciplined structure within which clinicians, statisticians, participants, regulators and ethics bodies can evaluate evidence.
This is an educational explanation, not medical advice and not a guide for running a real trial. The U.S. Food and Drug Administration’s E9 guidance states that statistics is essential to clinical-trial design and analysis. The later E9(R1) addendum emphasises a structured treatment-effect question and sensitivity analysis. Those are professional documents; students can still learn the underlying reasoning.
Choose the clinical-trial question you want to answer
- Why compare groups?
- What does randomisation achieve?
- How is an effect measured?
- How large should a study be?
- Why can missing data change the story?
- How can students learn this responsibly?
A student interested in probability can begin with random allocation and binomial outcomes. A student interested in algebra can compare risks, differences and ratios. A student interested in science can focus on variables, controls and bias. The important habit is to define the question before calculating the answer.
A treatment effect is a comparison
Suppose 70 of 100 patients improve after receiving a new treatment. The observed improvement rate is 70%. Is that the treatment effect? Not by itself. If 60 of 100 comparable patients would have improved under the existing treatment, the new treatment’s observed absolute difference is 10 percentage points. If 75 of 100 would have improved without it, the new option appears worse. The counterfactual—what would have happened to the same people under another choice—cannot be observed directly.
A randomised comparison uses groups to estimate that missing contrast. One group receives the intervention of interest and another receives a specified comparator, such as standard care or a placebo when ethical and appropriate. The groups are followed under a protocol, and outcomes are compared. The comparator is part of the question; “Does it work?” is incomplete without “compared with what, for whom, on which outcome and over what time?”
This logic is related to ordinary fair comparisons. If two classes use different teaching methods but also have different starting knowledge, lesson time and assessment difficulty, a raw final-score difference cannot be assigned to the method alone. Clinical trials formalise the comparison because health decisions carry serious consequences.
Define the outcome before seeing it
An outcome might be survival time, symptom score, infection, hospitalisation, blood-pressure change or another clinically meaningful measure. Each requires a definition, a measurement time and rules for handling repeated or incomplete observations. Choosing the outcome after looking at results invites selective reporting.
If a trial measures twenty outcomes, one may look unusual by chance even when there is no true effect. This is a multiplicity problem. Researchers can pre-specify primary and secondary outcomes, control error rates and interpret exploratory analyses honestly. Mathematics cannot replace scientific judgement, but it makes the cost of repeated searching visible.
The medical-screening article on base rates asks a different question—how test results update beliefs about a person. A clinical trial asks how outcomes differ under assigned strategies. Both need careful denominators and cannot be reduced to one impressive percentage.
Randomisation balances assignment, not every observed table
Randomisation uses a chance mechanism to assign participants to treatment groups. Its central benefit is not a promise that every characteristic will be numerically equal. It makes assignment independent of participant characteristics under the design, supporting valid probability-based comparisons and limiting conscious or unconscious selection.
Imagine 40 participants assigned by a fair random mechanism, 20 per group. One realised allocation might put 12 older participants in group A and 8 in group B. That imbalance does not prove randomisation failed. Random processes produce variation. The design protects against systematic choice; small studies can still show chance imbalances.
Stratified or blocked randomisation may be used to control balance on important factors or over time. These methods have rules that must be reflected in the analysis. “Random” does not mean casual. A reproducible allocation sequence, protection from foreknowledge and correct implementation are operational parts of the design.
A simple allocation calculation
With simple equal randomisation, each participant has probability 1/2 of assignment to A. If ten independent assignments are made without forcing equal totals, the number assigned to A follows a binomial distribution with n = 10 and p = 1/2. The probability of exactly five in A is C(10,5)/2¹⁰ = 252/1024 ≈ 0.246. Perfect numerical balance is only one possible result.
Many trials use methods that guarantee or encourage closer group sizes. The binomial example is a teaching model, not a claim about every protocol. It shows why a chance mechanism must be described before its output is interpreted.
Blinding addresses a different source of bias. Participants, care providers or outcome assessors may be unaware of assigned treatment when feasible. Randomisation governs allocation; blinding limits ways expectations or decisions can differ after allocation. Neither solves every problem, and some interventions cannot be blinded.
Absolute and relative effects answer different questions
Suppose an adverse outcome occurs in 20% of the comparison group and 15% of the treatment group. The absolute risk difference is 15% − 20% = −5 percentage points. The relative risk is 15% / 20% = 0.75. One might also say the relative risk reduction is 25%.
“A 25% reduction” sounds larger than “5 percentage points”, but both describe the same illustrative data from different angles. The absolute change helps readers understand how many outcomes differ per 100 people under the observed rates. The relative change shows proportional reduction. Responsible communication often provides both and states the follow-up period and population.
The number needed to treat is sometimes calculated as the reciprocal of an absolute risk reduction. With an absolute reduction of 0.05, the simple estimate is 1/0.05 = 20. It should not be treated as an exact personal promise. It inherits uncertainty from the trial, depends on baseline risk and time horizon, and may not transfer unchanged to another population.
Continuous outcomes need a scale
If an outcome is a symptom score, the group difference may be measured in score points. A mean reduction of 3 points is not interpretable without the scale, its variability and clinical meaning. Standardised effect sizes divide by a standard deviation, helping comparisons across scales but becoming less intuitive to patients.
If a treatment group has mean improvement 8 with standard deviation 10, and a comparison group has mean improvement 5 with similar variability, the raw mean difference is 3 points. A rough standardised difference using SD 10 is 0.3. Whether that matters clinically is not settled by 0.3 alone.
Time-to-event outcomes add censoring and survival functions. Repeated measurements add correlations within the same person. Cluster-randomised trials add similarity within hospitals or schools. The mathematical model must match how the data were generated.
Uncertainty belongs beside every effect estimate
An observed difference is affected by sampling variation. A confidence interval gives a range produced by a procedure with a defined repeated-sampling interpretation. It is not simply “the range where the true value has a 95% probability” under ordinary frequentist analysis. That wording may feel natural, but it changes the meaning.
Take an illustrative difference in proportions. If 60 of 100 improve in group A and 50 of 100 improve in group B, the estimated difference is 0.10. A rough standard error is √[0.6(0.4)/100 + 0.5(0.5)/100] = √0.0049 = 0.07. A simple 95% normal approximation is 0.10 ± 1.96(0.07), or about −0.037 to 0.237.
The interval includes zero, so the data are compatible with a range from a small disadvantage to a sizeable advantage under this approximation. That does not prove “no effect”. It shows the estimate is imprecise. Exact methods, model choices and protocol details may change the calculation; the example teaches the relationship among effect, variability and sample size.
Statistical significance is not clinical importance
A tiny effect can be statistically distinguishable from zero in a very large study. A potentially important effect can remain uncertain in a small study. P-values and confidence intervals address aspects of compatibility with a model; they do not measure importance, bias, ethics or truth by themselves.
Predefined error thresholds help control decision processes, but a result just below a threshold is not fundamentally different from one just above it. Look at the effect estimate, interval, outcome quality, design, missing data, multiplicity and consistency with other evidence.
This is one reason the benefits of learning mathematics include judgement, not just calculation. Students learn to ask what a number measures, what assumptions produced it and how sensitive the conclusion is.
Sample size connects effect, noise, power and error
Sample-size planning begins with a question and assumptions. For a simple two-group comparison, planners consider the effect size worth detecting, outcome variability or event rate, allocation ratio, significance level, desired power, expected loss to follow-up and the planned analysis. Changing any assumption can change the required number.
Power is the probability that a specified analysis will meet its success criterion if a particular effect and model are true. It is not the probability that the hypothesis is true. An 80% powered design does not guarantee success, and a non-significant result does not mean the trial had 80% probability of being wrong.
For a simple comparison of two means with equal group sizes and common standard deviation σ, a rough planning formula is n per group ≈ 2(z₁₋α/₂ + z₁₋β)²σ²/Δ². Here Δ is the target mean difference. With two-sided α = 0.05, power 80%, σ = 10 and Δ = 5, use z values about 1.96 and 0.84. Then n ≈ 2(2.8)²(100)/25 = 62.72, rounded upward to 63 per group before allowances.
The square relationship is important
If the target difference is halved from 5 to 2.5 while other assumptions stay fixed, Δ² falls by a factor of four, so the required sample size rises by about four. Detecting smaller effects is expensive because signal is being separated from the same noise.
Reducing the significance level or demanding higher power also increases sample size. Greater variability increases it. Better measurement can reduce noise, but “better” must be real, not assumed. Unequal allocation may be useful for practical or ethical reasons but can reduce statistical efficiency for a fixed total under simple conditions.
Too small a trial may be unable to answer its question and can expose participants without producing useful knowledge. An unnecessarily large trial may also use resources and participants poorly. Sample size is therefore both a mathematical and ethical planning issue, evaluated in the context of the full protocol.
The formula above is illustrative. Binary outcomes, survival outcomes, non-inferiority questions, clusters, repeated measures and adaptive designs use different calculations. Professional trials require qualified statisticians and governance.
Error rates are design choices, not moral labels
In a simplified superiority test, a Type I error occurs when the procedure declares evidence of a treatment effect under a specified null model that is actually true. A Type II error occurs when the procedure fails to reach that criterion under a specified alternative. The significance level α bounds or defines the Type I error rate under assumptions; power equals 1 − β for the chosen alternative.
The labels “false positive” and “false negative” are useful introductions but can hide nuance. Real clinical questions may involve confidence intervals, multiple endpoints, non-inferiority margins, repeated looks at accumulating data and decisions beyond a binary test. The hypotheses must be defined.
If a study examines results repeatedly and stops whenever p < 0.05 without adjustment, the overall chance of a false-positive declaration can exceed 5%. Group-sequential methods and data-monitoring plans account for interim looks. Mathematics permits flexibility only when its consequences are planned.
**Did You Know?** Stopping early for overwhelming benefit, futility or harm may be ethical and efficient, but it is not simply “peek and choose”. Boundaries, roles and decision rules are specified to protect validity and participants.
Missing data is a mathematical and clinical problem
Participants may miss visits, stop treatment, switch treatment or withdraw. Their outcomes may differ systematically from those who remain. Deleting every incomplete record can change the groups being compared and weaken the protection created by randomisation.
Suppose 100 people are randomised to each group. In the treatment group, 20 people with poor early response leave and are omitted, while only 5 leave the comparison group. Calculating success among completers can make the treatment look better even if the true assigned-group contrast is less favourable. The missingness mechanism matters.
Researchers specify strategies for intercurrent events—events after treatment starts that affect interpretation—and for missing outcomes. They may use multiple imputation, likelihood models, inverse weighting or other methods under explicit assumptions. No method recovers unknowable data without assumptions.
Sensitivity analysis asks whether conclusions change under plausible alternative assumptions. This is a powerful lesson for students: a result is more trustworthy when it survives reasonable changes, and uncertainty should widen when assumptions are weak.
The 2021 FDA E9(R1) addendum frames treatment effects through an estimand: the population, treatment condition, variable, handling of intercurrent events and population-level summary. Students need not master the regulatory vocabulary to appreciate the principle—define exactly what difference you intend to estimate.
Bias, confounding and generalisability
Randomisation supports internal validity for the assigned comparison, but other biases can remain. Outcome measurement can differ. Participants can deviate from assigned treatment. Reporting can be selective. Sites can implement procedures unevenly. Data errors can occur.
Generalisability asks whether the result applies beyond the trial sample. A study in one age range, care setting or risk group may not answer the same question elsewhere. A large sample does not automatically make a narrow population representative.
Confounding is especially important in observational comparisons, where treatment choices may be related to prognosis. Statistical adjustment can help when relevant variables are measured and models are appropriate, but it cannot guarantee removal of unmeasured confounding. Randomisation and modelling solve different problems.
The epidemic-modelling article shows how population dynamics depend on assumptions and data. Clinical trials usually focus on assigned interventions within specified participants. Both areas need causal humility.
A complete worked miniature example
Imagine an educational, non-medical trial model comparing strategy A with strategy B for a binary outcome. Two hundred participants are randomised equally. In A, 58 of 100 have the outcome; in B, 46 of 100 do.
The observed risks are 0.58 and 0.46. The absolute difference A − B is 0.12, or 12 percentage points. The risk ratio is 0.58/0.46 ≈ 1.261. The odds in A are 58/42 ≈ 1.381; in B they are 46/54 ≈ 0.852; the odds ratio is about 1.621. These measures are not interchangeable.
A rough standard error for the risk difference is √[0.58(0.42)/100 + 0.46(0.54)/100] = √0.00492 ≈ 0.0701. A normal-approximation 95% interval is 0.12 ± 1.96(0.0701), approximately −0.017 to 0.257.
This miniature study does not establish a clear positive difference under that approximation because the interval crosses zero. The point estimate still matters, and the range includes effects that could be important. The right conclusion is not “A definitely works” or “A has no effect”. It is that the sample leaves substantial uncertainty.
If the outcome were rare, sample sizes small or analysis more complex, this simple approximation might be poor. If participants were clustered or measurements repeated, the independence assumption would fail. The worked example demonstrates arithmetic while modelling caution.
A practical learning path for students
Build a clinical-trial notebook around questions rather than formulas:
- define the population and comparator;
- name one primary outcome and time point;
- draw a random allocation scheme;
- calculate absolute and relative effects;
- attach a standard error or interval;
- list possible missing-data reasons;
- distinguish statistical from clinical importance;
- state what the data cannot establish.
Use invented or openly published aggregate data, not private health information. Start with coins or coloured cards to simulate randomisation. Repeat the simulation to see that group balance and outcome differences vary from run to run. Plot the distribution of differences under no true treatment effect.
Then change one design feature at a time. Double the sample size and watch the standard error shrink roughly with 1/√n. Increase outcome variability and see precision worsen. Introduce selective dropout and observe how a completer-only estimate can shift.
Students who use spreadsheets can generate random numbers, assign groups and compute summaries. Coding is optional; reasoning is essential. Before pressing calculate, predict the direction of change. After calculating, explain the denominator.
What parents and teachers can reinforce
Ask: “What exactly is being compared?” “Were groups assigned fairly?” “How wide is the uncertainty?” “Is the outcome meaningful?” “Who was included?” These questions protect students from being dazzled by one percentage or headline.
Avoid turning an article into personal medical judgement. Decisions about treatment belong with qualified professionals who know the individual case and current evidence. Mathematics supports shared understanding; it does not replace clinical care.
For age-appropriate career exploration, clinical research can involve medicine, nursing, pharmacy, statistics, epidemiology, data management, ethics, regulation and patient engagement. Strong mathematics opens ways to contribute but does not guarantee a pathway.
Misconceptions worth repairing
**“Randomisation makes the groups identical.”** It makes allocation random under the design. Realised groups can differ by chance.
**“A non-significant result proves no effect.”** It may reflect a small effect, wide uncertainty, low power or model limitations. Examine the interval and design.
**“A significant result must be important.”** With enough data, a small effect can meet a threshold. Clinical meaning needs context.
**“A 50% relative reduction means 50 fewer people out of 100.”** Not necessarily. If risk falls from 2% to 1%, the relative reduction is 50% and the absolute reduction is 1 percentage point.
**“More data fixes every bias.”** More biased data can produce a precise biased estimate.
**“One trial settles the whole question.”** Evidence accumulates across studies, settings, outcomes and quality assessments.
Frequently asked questions
Why not give the new treatment to everyone?
Without a suitable comparison, natural recovery and other changes can be mistaken for treatment effects. Ethical design depends on genuine uncertainty, available care and oversight; there is no one comparator for every case.
Is 100 participants enough?
There is no universal number. It depends on outcome type, target effect, variability, error rates, allocation, follow-up and analysis. A calculation should follow a defined question.
Does a 95% confidence interval contain 95% of patients?
No. It describes uncertainty in an estimated parameter under a procedure and model. It is not a range containing 95% of individual outcomes.
Why analyse people in their randomised groups?
An intention-to-treat-oriented analysis helps preserve the benefits of random allocation for the effect of an assigned strategy. Exact approaches depend on the estimand and protocol, so professional guidance matters.
Can artificial intelligence replace trial statistics?
AI may support tasks, but prediction tools do not remove the need for design, randomisation, estimands, bias control and transparent uncertainty. A sophisticated model cannot repair a question that was never defined.
The deeper lesson: fair evidence is designed
Clinical-trial mathematics is not decoration added after data collection. The question determines the design; the design determines what can be estimated; the analysis quantifies effects and uncertainty; sensitivity analysis tests dependence on assumptions. Every stage is connected.
This is why mathematics education matters beyond examinations. It helps a reader resist two opposite errors: believing every technical claim, and rejecting all statistics as manipulation. A mathematically literate reader can inspect definitions, denominators, uncertainty and design.
Read next:
- Why Mathematics? | Medical Screening, Base Rates and Test Results
- Why Mathematics? | Survey Sampling, Margin of Error and Weighting
- Why Mathematics? | Epidemic Modelling, Reproduction Numbers and Exponential Growth
- Why Mathematics? | Drug Half-Life, Pharmacokinetics and Safe Dosing Models
- FDA E9 Statistical Principles for Clinical Trials
Further worked examples: when the denominator changes the claim
Imagine 1,000 participants in each group. An outcome occurs in 10 people under treatment A and 20 under treatment B. Risks are 1% and 2%. The absolute risk reduction is 1 percentage point, while the relative risk is 0.5 and the relative reduction is 50%.
Now imagine risks of 40% and 80%. The relative risk is still 0.5, but the absolute reduction is 40 percentage points. The same relative headline can describe very different numbers of affected people. Baseline risk is therefore essential.
Odds add another layer. In the first example, odds are 10/990 and 20/980, giving an odds ratio of about 0.495. It is close to the risk ratio because the outcome is rare. In the second, odds are 0.4/0.6 and 0.8/0.2, giving an odds ratio of 1/6 ≈ 0.167, far from the risk ratio 0.5. Odds ratios should not be casually described as risk ratios.
Person-time rates
If follow-up differs, researchers may report events per person-time. Suppose one group has 12 events over 600 person-years and another has 18 over 500 person-years. Incidence rates are 2.0 and 3.6 per 100 person-years. The rate ratio is about 0.556.
A rate is not the same as the proportion of people who ever experience an event. One person might contribute different follow-up time, and repeated events may or may not be allowed. Readers should check the event definition and denominator.
Cluster randomisation changes effective information
Sometimes hospitals, schools or communities rather than individuals are randomised. Outcomes within a cluster tend to be correlated. Fifty participants from one clinic do not necessarily provide the same independent information as fifty participants from fifty clinics.
A simple design effect is 1 + (m−1)ρ, where m is average cluster size and ρ is intracluster correlation. If m = 20 and ρ = 0.05, design effect = 1 + 19(0.05) = 1.95. A nominal sample of 390 then has rough independent-equivalent information of 390/1.95 = 200 under this simplified relation.
Even a small correlation matters when clusters are large. Ignoring clustering makes standard errors too small and confidence intervals too narrow. The trial’s unit of randomisation, unit of observation and unit of analysis must be connected.
Simulation shows what random variation looks like
Take a model where both groups have true success probability 0.50. Randomly generate 50 outcomes per group and calculate the difference. Repeat 10,000 times. The differences will centre near zero, but many individual simulations will show non-zero gaps.
The standard error of the difference is √[0.5(0.5)/50 + 0.5(0.5)/50] = 0.10. A gap of 0.20 is about two standard errors from zero and will occur occasionally even when the model has no group effect.
Increase each group to 500. Standard error becomes √0.001 ≈ 0.0316. The distribution narrows because larger samples reduce random sampling variation. It does not eliminate measurement bias or protocol failures.
Now simulate selective dropout: remove outcomes coded as failures from one group with higher probability. The estimated difference shifts even though assignment began fairly. This demonstrates why post-randomisation missingness deserves its own analysis.
Reproducibility needs more than a random seed
A seed reproduces pseudo-random numbers in a program. It does not document which participants were eligible, which outcome definition was used, whether code matched the protocol or which data changes were made. Reproducibility is a chain of provenance.
Students can practise by writing a simulation plan before running it: values of n and p, number of repetitions, statistic to record and expected pattern. Afterward, compare observed mean and spread with theory. This prevents changing the goal after seeing an attractive result.
Non-inferiority asks a different question
A new treatment might offer easier delivery or fewer burdens, so the question may be whether it is not unacceptably worse than an established treatment by more than a pre-specified margin. This is a non-inferiority design, not a failed superiority trial renamed later.
Suppose outcome is improvement and the effect is new minus standard. A margin of −5 percentage points means effects worse than −5 are considered unacceptably inferior for the defined question. If a confidence interval is −2 to +6, its lower end is above −5, supporting non-inferiority under the planned framework. It does not prove equality.
The margin requires clinical and historical justification. Choosing it after seeing data would undermine interpretation. Assay sensitivity—the ability of the trial to distinguish effective from ineffective treatment—also matters. Poor adherence can make groups look similar and falsely favour a non-inferiority conclusion.
This example shows why direction and scale must be written explicitly. If lower values are better, the inequalities reverse. A diagram of the margin, zero and interval often prevents verbal confusion.
From protocol to publication: a mathematical audit trail
Before enrolment, a strong plan connects:
- research objective;
- estimand and treatment contrast;
- eligibility and population;
- randomisation and blinding;
- primary outcome and time point;
- sample-size assumptions;
- analysis set and model;
- missing-data strategy;
- multiplicity and interim analysis;
- sensitivity analyses.
After data collection, a report should show participant flow, baseline description, numbers analysed, effect estimates, uncertainty and deviations. A polished final p-value cannot substitute for this trail.
Students can audit a published abstract with five columns: claim, numerator, denominator, comparison and uncertainty. Add a sixth for limitations. If a column is empty, the claim needs more context before it guides a decision.
The value of mathematics here is transparency. A calculation exposes its dependencies. Readers can disagree about assumptions without pretending the result came from nowhere.
Did You Know? Four design choices that change interpretation
First, allocation ratio need not be 1:1. A 2:1 ratio may provide more experience with a new treatment or help recruitment, but for a fixed total and equal variances it often gives less precision for the simple difference than equal allocation. Practical benefits and statistical cost are balanced explicitly.
Second, superiority, equivalence and non-inferiority are different hypotheses. Failing to prove superiority does not prove equivalence. Equivalence generally asks whether an interval lies within two pre-specified margins; non-inferiority uses a one-sided clinical margin. The question must be chosen before results.
Third, a subgroup finding can be tempting. If many age, sex, site and severity subgroups are searched, some interactions can look striking by chance. Credible subgroup analysis considers pre-specification, interaction tests, biological rationale, consistency and uncertainty—not separate “significant here, not there” labels.
Fourth, harms and benefits may need different time horizons and sample sizes. A trial powered for a common efficacy outcome may be too small to quantify a rare adverse event. Post-market monitoring and multiple evidence sources can therefore remain important.
For a student audit, take one imaginary result and rewrite it four ways: absolute risk difference, risk ratio, number needed to treat and a confidence interval. Then add baseline risk, follow-up time and population. The exercise shows how much context a headline removes.
Mathematics cannot decide what risk a person should accept. It can make the evidence and trade-offs clearer so that ethical, clinical and personal judgement have something honest to work with.
One final check is especially valuable: repeat the conclusion without statistical shorthand. Name the people studied, the strategies compared, the outcome, the follow-up period, the estimated difference and its uncertainty. If that plain-language sentence cannot be written accurately, the analysis is not yet ready to guide a reader.