Randomization List Generator.

Random allocation stops anyone choosing which treatment the next participant receives, which is how bias creeps into the comparison. It also lets you use probability theory to judge whether a difference between groups is larger than chance would produce[1]. ICH E9 adds that randomisation tends to produce groups in which known and unknown prognostic factors are similar[3].

Of these benefits, preventing selection bias at trial entry is usually the most important[2]. Systematic schemes such as alternation, date of birth or hospital numbers are not random and should not be reported as random, because whoever enrols participants can predict the next allocation and choose who to recruit[1][2]. Once participants are randomised, our guide to participant retention covers keeping them in the trial.

The seed is hashed with SHA-256[5] into a master key, and each stratum gets its own key and its own stream of random numbers:

master key = SHA-256(wg-randlist-v1, 0x00, seed)
stratum key = SHA-256(master key, stratum level numbers)
random words = SHA-256(stratum key, 0), SHA-256(stratum key, 1), …

Whole numbers come from these words by rejection sampling, so every outcome has exactly the same chance. A block of size b holds (b ÷ S) × r copies of each arm, where r is the arm’s ratio and S the ratio total, and is shuffled with Durstenfeld’s version of the Fisher and Yates algorithm[6]:

block of 6 at 1:1 = A A A B B B, then for i from 6 down to 2, swap position i with a random position from 1 to i

Simple randomisation draws one whole number per participant and maps it to an arm in proportion to the ratio. The method is published as wg-randlist-v1, so an independent statistician can rebuild a list with any SHA-256 library. We checked this implementation against an independent Python version on fixed test vectors, and its output passed a battery of 37 statistical tests.

ICH E9 asks for blocks short enough to limit imbalance but long enough to avoid predictability towards the end of each block, with staff kept unaware of the length, and notes that two or more lengths chosen at random achieve the same purpose[3]. Schulz and Grimes recommend varying the size at random and using larger blocks, particularly in unblinded trials[4], and found that investigators overuse fixed blocks[23].

Avoid blocks of two. In 179 open trials in four general medical journals, heterogeneity in baseline age, a marker of subversion, was 62% with fixed blocks of four and 40% with variable blocks that included two, against 0% with simple randomisation and with minimisation[10].

Random sizes reduce certainty, not tendency. Under the model of Blackwell and Hodges, random block sizes reduce the bias that relies on knowing the next allocation for certain, but not the bias from guessing the arm that is behind[11][22]. Our calculation shows the size of both effects for a 1:1 trial:

Moving from fixed blocks of 4 to sizes of 4, 6 or 8 drawn at random cuts certain deductions from a third to about 2% in our calculation, yet a recruiter who sees past allocations still beats chance. That is why concealment and blinding matter more than block design. The defaults leave out the smallest size, which the evidence above links with subversion[10], vary at random as ICH E9 suggests[3], and stop at four times the ratio total, because larger blocks control balance less well if a trial stops early or is analysed part way through[7][22]. For a 2:1 ratio the defaults are 6, 9 and 12.

Successful randomisation depends on two steps: an unpredictable sequence, and concealment of that sequence from the people who enrol participants[2]. Without concealment, even a proper random sequence can be subverted[16]. In 250 trials, treatment effects measured as odds ratios were exaggerated by 41% when concealment was inadequate[17], and a later study of 1,346 trials found the bias mainly in trials with subjective outcomes[18].

Central assignment, such as a pharmacy or a central computer or telephone system, is especially desirable[2], and ICH E9 asks for randomisation to be organised centrally in multicentre trials[3]. Before you use a list in a live trial:

• Ask someone who does not recruit, enrol or assess participants to generate the list and keep it.
• Allocate through a central randomisation system, a pharmacy, or another concealment method approved for your trial. Do not share the list, the block sizes or the seed with recruiting staff.
• Store the generation record with the same access controls as the list, because it can recreate the whole sequence.
• Generate the list once. Choosing between lists because one looks more balanced removes the randomness you need.
• Check your protocol and your trial statistician’s plan. This tool supports planning and does not replace either.

Randomisation follows screening and electronic consent, and the staff who run those steps are the people the list must stay hidden from.

ICH E9 asks that a randomisation schedule be reproducible if the need arises[3]. The seed, the settings and the algorithm version rebuild the identical list, and the generation record holds a SHA-256 checksum of each CSV so you can confirm a match byte for byte.

• A shorter list with the same seed is the start of a longer one, so you can extend a list without changing the rows already used.
• Adding a level at the end of a factor leaves the existing strata unchanged.
• Renaming arms or levels changes no allocation, but reordering them does.

• No concealment. The tool does not hide allocations, handle emergency unblinding or number drug kits. Those need a central randomisation system.
• Lists only. Minimisation and response adaptive designs depend on the participants already in the trial, so they cannot be made as a list in advance.
• Plain text. Labels and seeds are limited to plain ASCII characters, so the files and the seed behave the same in every browser and spreadsheet.
• Not validated for regulated use. Ask your trial statistician to review the design, and use a validated system for live allocation in regulated trials.

For other planning questions, see our clinical trial solutions.