The night-shift randomisation
It’s 2 a.m. The fellow calls you: “We’ve got a new septic shock patient. Do we enrol them in the steroid trial?” You glance at the clipboard: the next sealed envelope says “B”. You open it, find the assignment, and start hydrocortisone.
Later you learn the envelopes were numbered sequentially, kept in a drawer at the nurses’ station, and the night nurse could feel the bump of the next card through the paper. Was that truly random? Was allocation concealed?
These aren’t just methodological nitpicks — they are the difference between a believable effect estimate and one that is inflated by bias.
Why randomisation works (in theory)
Randomisation aims to create two (or more) groups that are exchangeable — meaning that, apart from the intervention, they have the same distribution of known and unknown prognostic factors. If achieved, any difference in outcome can be attributed to the intervention.
Exchangeability relies on two pillars:
- Unpredictable sequence generation – the next assignment cannot be guessed.
- Allocation concealment – the person enrolling the patient cannot know or influence which group the next patient will get.
If either fails, baseline imbalances (known or unknown) can creep in, and the estimated effect is no longer purely causal.
Blinding, while not part of randomisation per se, protects against performance and detection bias after allocation. Together, these three elements form the core internal validity safeguards of an RCT.
Sequence generation – how to make it unpredictable
Acceptable methods (CONSORT 2020):
- Computer-generated random numbers (e.g.,
rand()in statistical software). - Random number tables.
- Coin toss, dice, or drawing labelled chips — only if the procedure is truly random and concealed.
Unacceptable (quasi-random) methods that still appear in the literature:
- Alternation (ABABAB…).
- Allocation by date of birth, day of week, or medical record number.
- Any rule based on patient characteristics that can be anticipated.
If the paper says “patients were randomised” but describes one of the above, treat it as a non-randomised study until proven otherwise.
Quick check in the methods
Look for a sentence such as: > “The randomisation sequence was generated by an independent statistician using block randomisation with variable block sizes of 2, 4, and 6, implemented in REDCap.”
If you see only “patients were randomised” with no detail, that is a red flag.
Allocation concealment – keeping the next assignment secret
Even a perfect random sequence is useless if the person enrolling the patient can peek at the next allocation.
Adequate concealment strategies:
- Central randomisation (phone/web-based system controlled by a trial coordinating centre).
- Sequentially numbered, opaque, sealed envelopes (SNOSE) — provided the envelopes are truly opaque, sealed, and opened only after enrolment.
- Pharmacy-controlled randomisation (the dispensing pharmacist prepares the intervention according to a concealed list).
Inadequate concealment:
- Open random number lists kept at the bedside.
- Envelopes that are translucent or not sealed.
- Allocation revealed to the enrolment team before consent.
Quick check in the methods
Look for phrases like: > “Allocation was concealed by a central interactive voice‑response system.” > “Sequentially numbered, opaque, sealed envelopes were prepared by an independent administrator and opened after baseline assessments.”
If the description mentions “envelopes kept in the ward folder” or “the treating physician opened the next envelope,” that suggests possible lack of concealment.
Blinding – who is unaware of the assignment?
Blinding (masking) aims to prevent knowledge of the intervention from influencing behaviour (performance bias) or outcome assessment (detection bias).
CONSORT encourages reporting who was blinded, not just a yes/no. Typical groups:
Group | What blinding prevents |
Participants / patients | Alters co‑interventions, reporting of symptoms, adherence. |
Clinicians / caregivers | Changes in co‑interventions, timing of assessments, extra vigilance. |
Outcome assessors | Subjective outcomes (e.g., pain scores, delirium assessment) become less prone to bias. |
Data analysts | Prevents subconscious decisions during cleaning or modelling that favour one arm. |
In ICU trials, full blinding of patients and clinicians is often impossible (you can’t blind a vasopressor infusion), but outcome assessors and data analysts can and should be blinded whenever feasible.
Quick check in the methods
Look for a statement such as: > “Outcome assessors were blinded to group assignment; mortality was adjudicated by a blinded committee using source documents.” > “The treating team was not blinded to the fluid strategy, but all laboratory results were processed by a central laboratory unaware of allocation.”
If the paper says “double‑blinded” without specifying who, treat it as insufficient detail.
How these three elements interact in practice
Threat | How it arises | What to look for |
Selection bias (due to poor allocation concealment) | Enroller can predict next assignment and preferentially enroll healthier/sicker patients. | Check concealment method; if inadequate, downgrade confidence in causality. |
Performance bias (due to lack of blinding of patients/clinicians) | Knowledge of intervention leads to different co‑interventions or care intensity. | See if clinicians were blinded; if not, examine whether co‑interventions were protocolised and similar. |
Detection bias (due to lack of blinding of outcome assessors) | Subjective outcomes are assessed differently knowing the arm. | Verify who assessed outcomes and whether they were blinded. |
Attrition bias (unrelated to randomisation but related to conduct) | Differential loss to follow‑up breaks exchangeability. | Regardless of randomisation quality, always check ITT and reasons for dropout. |
Your bedside checklist for randomisation, allocation, and blinding
Before you trust an RCT’s effect estimate, verify:
- Sequence generation: Was it truly unpredictable? (computer-generated, random numbers, etc.) Reject alternation, date‑based, or other quasi‑random methods.
- Allocation concealment: Was the next allocation hidden from the enrolment team? (central system, SNOSE, pharmacy). Reject open lists or unsealed envelopes.
- Blinding: Who was blinded? At minimum, outcome assessors and data analysts should be blinded for subjective outcomes; note if participants/clinicians were not blinded and consider implications.
- Baseline table: Despite good randomisation, check Table 1 for major imbalances (age, APACHE, comorbidities). Large differences suggest either chance imbalance (unlikely in large N) or concealed bias.
- Analysis: Was the primary analysis intention‑to‑treat? Per‑protocol analyses answer a different efficacy question.
Go deeper
- CONSORT 2020 explanation and elaboration (free PDF): sections on sequence generation (item 4a), allocation concealment (item 4b), and blinding (item 5a‑5c). https://www.equator-network.org/reporting-guidelines/consort-2020/
- StatPearls – “Randomized Controlled Trials” (NBK475335): free, concise overview of RCT design, including randomisation and blinding. https://www.ncbi.nlm.nih.gov/books/NBK475335/
- BMJ Statistics Notes – “The odds ratio” (Altman & Bland): while focused on OR, the notes include a short bit on why randomisation matters for causal inference. https://www.bmj.com/content/bmj_stats_notes
- “Randomisation and recruitment in clinical trials” (J Clin Epidemiol 2021, open access): practical guide on implementing and checking concealment. https://doi.org/10.1016/j.jclinepi.2020.10.017 (free via PMCID: PMC8564899)
Next: Chapter 4 — Intention‑to‑treat vs per‑protocol