CONSORT checklist

Items 23 and 24 are the ones with teeth. Trial registration and a publicly available protocol let a reader check whether the outcome you reported is the outcome you planned to report, which is the single most effective guard against selective reporting.

Items 8 to 10 are what make a trial a trial, and they are routinely under-reported. "Patients were randomised" tells a reader nothing. Three separate things need saying.

Item 13 is the one reviewers check with a calculator. Every number must account for every patient, and the most common failure is a gap between how many were randomised and how many were analysed with no explanation of where the difference went.

Enrolment

assessed for eligibility: 520
excluded: 120
  did not meet inclusion criteria: 85
  declined to participate: 28
  other reasons: 7
randomised: 520 − 120 = 400

The three exclusion reasons must sum to the exclusion total. Giving a single figure with no breakdown is the first thing a reviewer will query, because it hides whether the trial struggled to recruit or was selective about whom it accepted.

Allocation and follow-up

The three checks a reviewer makes

What the numbers say once they reconcile

The per-protocol analysis covers 353 of 400, so 11.8% of randomised patients are absent from it. More importantly, 18 were lost from the intervention arm against 11 from the control. Differential loss is a finding in itself, and if the reason relates to the treatment it undermines the per-protocol comparison regardless of how the numbers add up.

Items 8 to 10 are where most trial reports are thinnest. Here is the difference between a sentence that satisfies a reviewer and one that invites questions.

Item 8, sequence generation

Weak: "Patients were randomly allocated to treatment or control."

Strong: "The allocation sequence was generated by an independent statistician using computer-generated random numbers, with permuted blocks of variable size (4 and 6), stratified by centre and baseline severity."

The weak version is compatible with alternate allocation, coin tossing, or allocation by day of the week, none of which is randomisation. Block sizes matter because fixed small blocks let a recruiter predict the last allocation in each block.

Item 9, allocation concealment

Weak: "The study was double blind."

Strong: "Allocation was concealed using a central web-based randomisation service, accessed only after a participant had been enrolled and consented. Site staff could not view or predict upcoming allocations."

These describe different things. Blinding is what happens after assignment; concealment protects the assignment itself. Answering item 9 with a statement about blinding is the single most common error in the whole checklist.

Item 10, implementation

Weak: omitted entirely, which is what usually happens.

Strong: "The sequence was generated by the trial statistician, participants were enrolled by site research nurses, and allocation was assigned automatically by the web system on enrolment."

Three roles, named separately. If the same person generated the sequence, decided who was eligible and assigned the treatment, concealment was not possible however the sequence was produced.

The diagram has four stages: assessed for eligibility, randomised, received the allocated intervention, and analysed. The numbers must reconcile, with every loss accounted for and a reason given.

The usual failure is that the number analysed is smaller than the number randomised with no explanation of the gap. That gap is where the trial's credibility lives: participants who left are rarely a random subset, and silently dropping them turns a randomised comparison into an observational one.

Item 7a asks for the assumptions, not just the answer. Here is the full calculation for a trial comparing two proportions, and what each number in it commits you to.

The four assumptions

The second row is the one that matters and the one most often fudged. It should be the smallest difference worth detecting, not the difference you hope to find. Choosing an optimistic value is how trials end up underpowered for anything realistic.

The calculation

z(α/2) = 1.960, z(power) = 0.842, sum = 2.802

n per arm = 2.802² × [0.24 × 0.76 + 0.15 × 0.85] / (0.24 − 0.15)²
    = 7.849 × 0.3099 / 0.0081
    = 2.432 / 0.0081 = 300.3

So 301 per arm, 602 in total. Always round up.

Allowing for attrition

Divide by the proportion you expect to retain, rather than adding a percentage, which under-inflates.

15% attrition: 301 / 0.85 = 354.1 → 355 per arm, 710 total

Why the target difference dominates everything

Sample size scales with the inverse square of the difference you want to detect. Halving it roughly quadruples the trial.

Item 7a, sample size. The assumed effect, the variability assumed, the power and the significance level, with the source of each assumption. A sample size with no stated assumptions cannot be checked.

Item 12a, statistical methods. The methods for the primary and secondary outcomes. Cluster randomised trials must say here how clustering was accounted for. Randomising clinics and analysing patients as independent is the most common serious analytical error in trials.

Item 12b and 17b, subgroups and multiplicity. Subgroup analyses must be declared as prespecified or exploratory, and the distinction decided in advance. Several outcomes bearing on one conclusion need a stated approach to multiple comparisons, chosen before the results were seen.

Item 17a, estimation. For each outcome, the effect size with a confidence interval. A p-value on its own does not satisfy this, and reviewers increasingly say so.

It comes down to one question: did the researchers assign the exposure?

A non-randomised comparison of two interventions is not a trial, whatever it is called locally, and should be reported with STROBE plus a clear statement of how groups were formed.

The main checklist assumes a two-arm parallel group trial. Anything else has an extension, and using the base checklist for a design it was not written for leaves obvious gaps.

If you are planning rather than reporting, SPIRIT is the protocol counterpart of CONSORT and is worth working through before recruitment opens.

CONSORT describes what to report after a trial ends. SPIRIT is its counterpart for the protocol, and working through it before recruitment opens removes most of the problems CONSORT later exposes.

Register the trial prospectively and make the protocol available. The combination is what allows anyone to confirm that the outcome you reported is the outcome you set out to measure, and it is increasingly a condition of publication rather than a courtesy.