Why do EBP people keep insisting on Randomization? I can already see that my device works!

Despite our love for data and extensive education, we have all seen experts in the middle of heated discussions about interventions they really believe in. They can't understand why everybody isn't using them. Lots of people love the latest, shiniest and most expensive device. These experts use it on their patients, and they get better. Of course, their love for the above-mentioned intervention becomes even deeper.

Sometimes, the observations are correct, and the devices or interventions eventually become mainstream. This is true with interventions that clearly change prognosis - someone is bleeding out, you apply pressure, and they stop bleeding... easy to see causation! However, more frequently than not, interventions have small benefits. In that case, it is hard to tell if it was the intervention, time, or any of the other cointerventions that helped. Causation or coincidence? Sometimes, the experts are wrong, the intervention might cause harm. Expert opinion may also prevent the use of an intervention that could have benefits. Multiple examples come to mind, but the clearest one, in my opinion, are beta blockers. When I was in training, giving a beta blocker to a patient in heart failure was malpractice. Who would dare give a medication that will weaken an already weak heart?! All the experts were in agreement.

Fast forward a few years and high-quality data from randomized controlled trials showed that beta blockers are, in fact, one of the most effective drugs to prevent mortality in heart failure - reducing lethal arrhythmias.

• Expertise is good, but systematically collected data with low risk of bias is much better!

Randomization was used for the first time in 1946 to assign people to streptomycin to treat tuberculosis.1 The reason it was used was practical - there were not enough doses to treat everybody! Since then, it has become the best way to test an intervention against the current standard of care.

The limitation to observational studies lies in not knowing what characteristics of the participants we don't know, or if any of these affect our intervention (and therefore, we can't adjust for it).

Randomization's superpower (when done correctly) is the ability to create almost perfectly prognostically equal groups (as long as the groups are large enough). Characteristics we know are important and even the ones we don't think about should be distributed equally. The only differences in prognosis or outcomes will be introduced by the investigator's intervention. At that point, we can really talk about a direct causality and not just coincidence. This is the best way to look at interventions, and sadly, no matter what statistical wizardry we use in observational studies, we will never be able to achieve the same results.

• Randomization aims to ensure all characteristics, both known and unknown, are equal in the different groups

Randomization works well only when we simultaneously do allocation concealment, and we are not allowed to know, predict, or manipulate the group to which each participant will belong. Even good-natured investigators might try to break the code, as a challenge, or because they might want to benefit their patients (unconsciously or less frequently consciously). Other tools that are used during randomization are blocking and stratification.

In blocking, an even number of patients is divided equally to the groups. A block of 2, will have one participant in each group (i.e. Heads or Tails). The first participant determines the second participant's group. That is not a problem, as each of them had a predetermined chance, 50% in this case, of belonging to any of the groups, so randomization is preserved. There are 2 potential sequences in blocks of 2, Tails (T) and Heads (H), and H and T. A group of four will have 2 T and 2 H with 8 possible combinations. A group of 6, will have 3 T and 3 H, with 16 combinations, and so forth. The potential combinations and order increase with the bigger blocks.

Because they divide participants equally, blocks ensure that groups with small samples will have similar numbers of participants in each of them.

To make it much harder for investigators to predict where the next patient will be assigned (and maybe wait to enroll their favorite patient), randomization lists can be created with random-sized blocks. That way nobody can know which group is next!

Stratification is the last tool used to create the random list. If there is a prognostically important characteristic for the study, especially if not very prevalent, and we want to ensure is represented equally in both the intervention and control group, we should stratify for it. Each "strata" or layer will get their own list and blocks. This is usually seen in multicenter studies with each center having its own list. That prevents center A, which enrolled only a few patients, from having everybody end up in one of the groups and the other group empty or very unbalanced.

But beware, like everything in life, moderation is the key: too many layers will mess up randomization, so use sparingly!

Advanced questions in randomization:

Don't be tempted by pseudo-randomization! This is when they use a known characteristic to assign a group, such as a phone number or SSN. There is no randomization there, you can only belong to one of the groups, since the characteristic is usually set beforehand. That might introduce bias. It is better than observations alone, but worse than randomization. Since we are already spending the money and doing the work, why not use the best system available - real randomization.

And what about 2:1 or 3:1 randomization? As long as you have a predetermined chance of belonging to the group and it is not preset, it is still random. Then... why do it this way? Maybe you want to have more patients in the intervention group, so physicians are more comfortable using your device. Maybe even though we don't have data showing it works; you are convinced the intervention is better and you want to help more patients. It is not a problem, however, be aware of the ratio and look at both the absolute numbers and the relative ones.

1. Bhatt A. Evolution of clinical research: a history before and beyond james lind. Perspect Clin Res. 2010 Jan;1(1):6-10. PMID: 21829774; PMCID: PMC3149409.