Clinical research asks questions involving people, but studies can answer very different kinds of question. A carefully described comparison is more informative than a label such as “clinically studied”. Start with what investigators planned to measure, who took part and what comparison supports the conclusion.
Randomised and observational designs
A randomised controlled trial allocates participants to groups using a chance-based process and compares outcomes. An observational study records exposures or experiences without assigning the intervention in that way. Randomisation can reduce systematic differences between groups; it does not make every later analysis immune to bias.
For either design, write down the population, exposure or intervention, comparator and outcome. An observational association can be valuable while still having alternative explanations. In a randomised study, check whether participants were analysed in the groups originally assigned and how departures from the plan were handled.
Blinding and allocation are separate
Clinical trial randomisation concerns assignment to a group. Clinical trial blinding concerns who knows the assignment while a study is conducted or assessed. Participants, care teams and outcome assessors may have different levels of knowledge. The two processes address different potential sources of bias.
Read the operational description instead of relying only on “single blind” or “double blind”. Some interventions cannot be concealed from participants, while an outcome assessor may still be unaware of allocation. Record which arrangements the authors actually used; do not fill an omitted detail with a preferred design.
An endpoint is a stated measurement
A primary endpoint is the main outcome specified to answer a study's central question. Secondary endpoints address additional questions. Both need a measurement definition and time point. A symptom score, a biomarker change and a long-term health outcome are distinct endpoints, even when discussed in the same paper.
Compare the published endpoints with the protocol or registry. A result highlighted in an abstract may be secondary or exploratory. Keep that status in a summary rather than giving every positive result equal evidential weight. Find the effect estimate and uncertainty, not only whether a threshold for statistical significance was crossed.
Registration, protocols and results
Clinical trial registration places key study information in a public record. A clinical trial protocol describes the planned methods; a clinical study report documents a study and its results in more detail. Their purposes overlap but they are not interchangeable documents.
Use the study identifier to connect records, and check version dates. A registry entry confirms that a record exists, not that every finding has been independently established. A protocol supports understanding of the plan, not a claim that the intervention succeeded. Where documents differ, record the difference rather than selecting only the most favourable version.
Missing data changes what can be concluded
Count how many people were recruited, assigned, followed up and included in the analysis. A percentage based only on completers can answer a different question from one based on everyone assigned. Look for explanations of missing measurements, withdrawals and sensitivity analyses.
In a hypothetical trial assigning 100 people to each group but measuring an outcome in 80 and 95, the missing observations deserve attention before comparing success percentages. The numbers alone do not show the direction of any bias; the reasons and analysis matter. This example illustrates reporting, not a result from a real trial.
Build a source trail before drawing a conclusion
Save the identifier, protocol version, registry link, full report and any correction. Distinguish the authors' interpretation from the observations supporting it. If a treatment question is personal, bring it to a suitable healthcare professional instead of using a journal summary as a treatment decision.
Use biomarkers and diagnostics to evaluate a measurement claim, and research reading records to organise the evidence. This introduction does not provide a trial sponsor's regulatory submission, operational protocol or current compliance checklist.
Sources and further reading
Sources checked 25 September 2026: NIHR research glossary; NIHR: what to expect on a study; NCCIH: methods in scientific papers.
Crossover trials: separate sequence, period and carryover
In a crossover clinical trial, the same participant receives interventions in a randomised sequence. In the simplest arrangement, one group follows A then B and another follows B then A. The person contributes a within-person comparison, so two observations from that person are not two independent participants. The CONSORT crossover extension explains why the report needs both the assigned sequence and information about each period.
Period effects and carryover effects answer different questions. A period effect is a change associated with when measurement happens, regardless of the intervention. Carryover means an earlier intervention still influences a later observation. A gap between periods may be intended to reduce carryover; its presence does not prove that every lingering effect has disappeared. UK guidance on crossover studies discusses these distinctions and the difficulty of studying long-lasting effects this way.
Consider an invented two-period study in which scores improve during the second visit under both sequences. That pattern could reflect familiarity with the assessment rather than an advantage of whichever intervention was delivered second. Conversely, a participant who leaves after the first period cannot contribute the same complete paired comparison as someone who finishes both.
When reading the results, keep three records together: sequence allocation, the numbers observed in each period, and the analysis accounting for repeated observations. A pooled A-versus-B chart without this context can conceal how the comparison was constructed. These checks help interpret a reported design; they do not establish that crossover is suitable for a particular treatment.
Adaptive trials: locate the decision rule before the result
An adaptive clinical trial allows specified features to change as information accumulates. The important reading question is not simply whether something changed, but whether the rule for changing it was set out before the relevant interim comparison. The FDA methodological guidance on adaptive designs describes prospective rules, interim analyses and the protection of trial integrity. This is a United States source for statistical concepts, not a statement of UK regulatory requirements.
For example, an invented study might plan an interim assessment after a stated amount of follow-up, with a rule to stop an arm if its results are sufficiently unpromising. That differs from repeatedly checking results and choosing a convenient stopping point after seeing them. Both histories could produce the same final number of participants, so the final sample size alone cannot distinguish them.
A useful report connects four pieces of information: the planned decision, the data available when it was made, the change actually made, and how the final analysis accommodates the design. Repeated opportunities to declare success affect error control. Selecting an apparently promising arm can also affect estimation, so an ordinary-looking confidence interval deserves scrutiny against the specified analysis.
Keep the original plan and dated amendments alongside the results. The word adaptive alone does not establish credibility. These reading questions do not supply the simulations or statistical design needed to run such a trial.
Platform trials: identify the comparison behind each headline
A platform trial can evaluate several research arms within a continuing structure, with arms entering or leaving over time. This does not mean every participant contributes to every comparison. UCL's description of the multi-arm, multi-stage approach distinguishes comparisons against a shared control from the platform's ability to add new arms. The methodological account of testing treatments within one platform provides the development context.
Imagine an invented platform that begins with arms A and B and adds C later. A headline reporting the platform's entire recruitment total is not the sample size for C. To understand C's result, identify its eligible population, recruitment dates, assigned comparator and follow-up window. Someone recruited before C existed could not have been randomised to it at that time.
Concurrent and earlier control observations therefore need to be distinguished. If a report borrows earlier observations, inspect its stated reasoning and analysis rather than assuming they are interchangeable with controls recruited alongside the new arm. Patient mix, measurement or background care may change across a long-running study. Those possibilities are questions for the report, not proof that a particular comparison is biased.
An arm ending recruitment also has a specific meaning. A planned decision about that arm does not automatically announce the outcome of other arms or the whole platform. Record conclusions at the comparison level, with the relevant protocol version and dates. That is a more informative summary than treating a platform as one permanently fixed experiment with one universal result.
What does censoring mean in a time-to-event endpoint?
A time-to-event endpoint needs a starting point, a defined event and a rule for ending observation. When follow-up ends without the event being observed, the record may be censored: it supplies information up to a stated time, rather than an observed event at that time. The FDA guidance on lung-cancer trial endpoints shows why endpoint definitions and censoring rules must be examined together in that setting.
For an invented example, one participant has the defined event after six months. Another remains event-free when observation ends after six months. Their recorded durations match, but their contributions are different. Treating both as events discards the distinction; treating the second as guaranteed event-free indefinitely invents information.
A reading checklist should retain the time origin, event definition, assessment schedule, reasons for censoring and numbers still under observation at reported times. Late estimates based on few remaining participants need their uncertainty and risk-set counts.
Also ask whether ending observation could relate to the event being studied. The statistical assumptions and sensitivity analyses matter when that is plausible. This example explains endpoint records; it does not recommend a universal censoring rule or predict an individual's outcome.
Are adverse-event tables counting people or occurrences?
An adverse-event table may count people with an event, the number of occurrences, or both. These quantities answer different questions. The ClinicalTrials.gov team describes reporting the number affected alongside the number at risk in its original explanation of registry results. A person can contribute more than one occurrence, so event counts cannot automatically be divided by the group size and labelled the percentage of people affected.
In a fictional group of 40 participants, eight people experience a particular event and those people report 14 occurrences altogether. The affected proportion is 8 / 40, or 20%. The 14 occurrences describe a different total; 14 / 40 is not evidence that 35% of the people experienced the event.
For a reporting check, identify the safety population, observation period, method of collecting events and whether repeat occurrences are counted. If groups have different observation durations, preserve that difference rather than assuming equal opportunity to record events.
The presence of an event after an intervention also does not, on its own, establish causation. Keep the authors' relatedness assessment distinct from the count. These are interpretation checks for aggregate reporting, not instructions for classifying or submitting an individual safety case.
Does a missing registry field mean the same thing in every trial?
A blank registry field is a reason to investigate, but its meaning depends partly on the record's age and the field's history. The National Library of Medicine explanation of ClinicalTrials.gov displays notes that today's Researcher View includes fields that did not exist, or were optional, when some older records were first submitted. A current missing-information label alone cannot establish which requirement applied at the time.
Imagine reviewing two fictional records submitted ten years apart. Both lack information in the same field today. Before drawing the same conclusion about both studies, record their initial submission dates and establish whether that field was collected at each date. The missing answer remains missing even when its historical explanation differs.
A short transparency note should distinguish information absent from the registry, information available in another linked document and information that cannot be located. Preserve the record version and access date so another reader can repeat the comparison.
Registration details and posted results also occupy different parts of the record. Finding a protocol summary does not mean summary results have been published. This helps interpret registry documentation without treating registration as endorsement or offering a regulatory filing checklist.
What should a data-correction audit trail let a reader reconstruct?
An electronic research dataset is more informative when its corrections remain traceable. The relevant question is not simply whether software stores a final value, but whether authorised review can reconstruct what changed and why. The FDA electronic-record guidance, question 12, describes time-stamped records containing previous and replacement values, the person making the change and its reason. This is a cited US framework, not a claim about every UK study's legal obligations.
Consider a fictional endpoint value entered as 320 and later corrected to 32 after a transcription query. A final spreadsheet containing only 32 cannot explain that history. A traceable correction connects both values, the source check, the recorded reason and the authorised change without making the earlier entry disappear.
For a data-quality reading checklist, look for the originating record, value and unit, correction rationale, responsible role, timing and dataset version used for analysis. Distinguish a corrected source entry from a later transformation performed by an analysis script.
An audit trail does not by itself prove that the underlying observation was accurate. It makes changes inspectable. Interpreting that evidence also requires the relevant source record and the study's documented review process.
Does withdrawing from a study always erase information already collected?
It should not be assumed. Stopping study activities, stopping new data collection and deciding what happens to existing information are separate matters. The Health Research Authority's guidance on research data rights explains that research provisions and the legal basis for processing can affect what happens after withdrawal. It also stresses the information given to participants and the ethical considerations.
An illustrative clarification message might say: please confirm whether you want to stop visits, further contact, new information collection, or more than one of these; the study team will explain what happens to information already held. This is an example of separating questions, not approved participant wording. Its appropriate form depends on the study documents and the person's wishes.
Look for the relevant participant information, the scope and date of withdrawal, what collection ceased and how existing observations were handled in the reported analysis. Avoid assuming either automatic deletion or unlimited continued use. A specific person's request needs a response from the study team or responsible data-protection contact using that study's circumstances. This explanation is for interpreting research records, not a legal determination about an individual's information.
Is taking part in a study the same as public involvement in its design?
The roles are different. A participant supplies observations through taking part in the study. Public involvement means that people contribute to decisions about the research itself, such as its questions, design or communication. The Health Research Authority's explanation also distinguishes involvement from simply sharing information about research with an audience.
Imagine a questionnaire study with 500 respondents and a separate advisory group that helps make the questions understandable. Advisers might flag an invitation whose phrase follow-up does not explain whether it means a telephone call or a visit. Clarifying that message changes how participation is explained. The respondents supply study data; the advisers influence their collection. The number 500 cannot therefore be used as the number involved in designing the questionnaire. Conversely, an adviser need not also enrol as a participant to make a meaningful contribution.
In a research report, look for who contributed, at which stage, what they were asked to influence and what changed as a result. A sentence saying patients were consulted leaves those details unresolved. Keep the decisions and the contributor roles visible rather than treating involvement as a badge attached to the study. An account of the process can help readers understand the work, but does not by itself prove that every community was represented or that the findings are correct.
What should a plain-language results summary preserve when it simplifies a study?
Simpler language should preserve the study's question, comparison and uncertainty. It should not turn a limited result into a broad promise. The Health Research Authority's lay-summary guidance recommends clear, neutral wording and information about who took part, what happened and what the results showed. Explaining an unfamiliar term is more useful than silently removing a qualification.
In an invented example, a study observes a small change in a questionnaire score but does not measure later hospital use. A summary can explain the score in ordinary language and state the limits of the finding. Saying that the study kept people out of hospital would add an outcome the research never measured. Removing a technical phrase cannot justify adding that conclusion.
Check that the summary identifies the study and links to fuller information, retains the relevant population and time period, and presents unfavourable or uncertain findings as clearly as favourable ones. Readers should also be able to distinguish group results from personal results. Having people from the intended audience review the text can expose unclear wording. Readability and accuracy need checking together; a short text is not automatically an understandable one.
Does a multicentre trial result represent each participating centre equally?
A pooled trial result combines information; it does not mean that every centre recruited equally or observed the same pattern. An original study of analysis methods for multicentre trials examines why centre effects and stratified allocation matter when analysing binary outcomes. The appropriate model depends on the design and data, not simply on adding the word multicentre to a report.
Imagine a trial recruiting 600 people at one centre and 40 at another. Its total of 640 does not describe two equally sized comparisons. If allocation was stratified by centre, the report should connect that design choice with the analysis. A reader should also distinguish an overall effect adjusted for centre from a claim that effects differ between centres: those are separate statistical questions.
Useful methodology notes retain recruitment by centre, how allocation was organised, which centre terms entered the planned model and whether small centres created analysis difficulties. Record exploratory centre comparisons as exploratory. A centre with an imprecise estimate is not necessarily inconsistent with the overall finding. Conversely, a pooled estimate cannot establish identical performance everywhere. These checks support interpretation of a published analysis; they do not choose a statistical model for a new trial.
Why can analysing only participants who followed the plan change a trial comparison?
Random allocation concerns the groups at the start; later adherence is an observed behaviour. Restricting an analysis to people who followed the plan can therefore alter the comparison created by randomisation. CONSORT 2025 reporting guidance asks authors to describe who entered each analysis and in which group, rather than relying on a label such as modified intention-to-treat.
In a fictional trial, 100 people are allocated to each group, but 70 in one and 90 in the other complete the planned activities. Comparing only those completers does not automatically preserve the original comparability of the two groups. Reasons for completion may relate to prognosis or the outcome. Conversely, including everyone by allocation does not magically supply outcomes that were never measured; missing-data assumptions still need examination.
For an analysis-population example, retain the allocated numbers, exclusions and reasons, treatment actually received, available outcomes and rules for handling missing observations. A per-protocol estimate can answer a different question from an allocation-based estimate, but its causal interpretation needs appropriate methods. The safest reading note states exactly which people were compared. An acronym by itself cannot demonstrate that selection and missingness were handled adequately.
Does a hazard ratio of one half mean that half as many people have an event?
That conclusion does not follow from the hazard ratio alone. A hazard concerns the event rate at a moment among those still at risk; cumulative risk concerns the proportion experiencing an event over a specified period. The Cochrane effect-measures chapter distinguishes these measures and describes the proportional-hazards assumption often used in time-to-event analysis.
As a reporting example, a paper gives a hazard ratio of 0.5 but no event proportions at twelve months. Rewriting that as half as many people had the event by twelve months adds information not supplied by the ratio. The absolute comparison depends on the event pattern and follow-up. Nor does the same ratio tell a reader how many additional months a participant can expect without the event.
Keep the time origin, event definition, estimate, interval of uncertainty and model assumptions together. Inspect event or survival probabilities at relevant times when those are provided, with their numbers at risk. If relative hazards vary over time, one summary ratio can conceal that changing pattern. The aim is to report the quantity actually estimated, not to convert every time-to-event result into an unsupported percentage or personal prognosis.
Is failing to find a difference the same as demonstrating non-inferiority?
No. A test that does not establish superiority may simply leave a wide range of differences compatible with the data. Non-inferiority asks whether the evidence excludes a loss larger than a justified, specified margin. The FDA guidance on non-inferiority clinical trials explains the role of that margin and the conditions needed for an interpretable comparison. It is a US methodological source, not a statement of every UK regulatory requirement.
In an invented example, higher scores are better and the prespecified margin permits a difference no worse than minus 5 units. An interval for the new-minus-comparator difference extending from minus 8 to plus 3 includes zero, but also includes losses beyond the margin. The absence of a clear superiority finding therefore does not demonstrate non-inferiority. The example illustrates interval logic and does not justify choosing 5 as an actual margin.
Read the margin's rationale, the direction of the outcome, the analysis populations and how departures from the planned comparison were addressed. Check whether the margin was fixed before relevant results were known. A conclusion should distinguish demonstrated non-inferiority from an inconclusive finding, and both from equivalence. Those statements require different evidence and should not be substituted for each other.
Does matching every database entry to its source establish that a trial was conducted correctly?
Source comparison is valuable, but accurate transcription is only one part of study quality. A value may match its source while a required procedure was late, an eligibility condition was not met or a process was not followed. The FDA risk-based monitoring guidance considers critical data and processes as well as how monitoring findings are addressed. This is a methodological example from a US source.
Imagine an observation recorded as occurring on day 20 in both the source and database. The entries agree. If the study required that assessment within an earlier window, agreement alone has not resolved the departure from the plan. Equally, a missing signature and an incorrect endpoint value are different findings; counting both simply as one query can hide their different implications.
A site-monitoring reading record should identify what was checked, why it mattered, what discrepancy or process concern was found and how it was resolved or followed up. Distinguish evidence of copying accuracy from evidence about conduct, completeness and participant protection. A high percentage of verified fields cannot by itself summarise all those dimensions. These questions do not replace a study-specific monitoring plan or the responsible team's assessment.
Top comments (0)