YouTube Instagram TikTok How it works FAQ

The fastest way to turn talking videos into clean, searchable text. Free, private and no account required.

About

Built by Rishav Raj with one goal: free transcripts for everyone.

rishav@yttotranscript.com

Transcribe

YouTube to text
YouTube with timestamps
Instagram Reels
TikTok videos

Learn

How it works
FAQ
YouTube transcripts guide

More

Chrome extension
Contact
Privacy policy
Terms of service

© 2026 YTtoTranscript. Made for creators & researchers.

Not affiliated with YouTube, Instagram or TikTok. Respect creators’ copyright.

IPPCR 2019 Issues in Randomization Part 3 of 4 — Transcript · YTtoTranscript

Home/Transcripts/YouTube

YouTube Transcript

IPPCR 2019 Issues in Randomization Part 3 of 4

NIH VideoCast · 3,048 words · 15 min read · EN

Watch on YouTube

Share X LinkedIn WhatsApp

Below is the complete, readable transcript of IPPCR 2019 Issues in Randomization Part 3 of 4 by NIH VideoCast on YouTube. Read the full text, copy any part you need, or generate a transcript for any video with our free tool.

>> Paul Wakim: Hello, I'm Paul Wakim. I am chief of the Biostatistics and Clinical Epidemiology Service at the NIH Clinical Center, and this is part three on randomization in clinical trials. And in this part, we're going to continue our discussion about how to randomize. And if you remember, we talked about the first two methods

that are listed here in part two. And in part three we're going to talk about stratified, cluster and adaptive randomization. So, in stratified randomization a stratification variable or factor is an important baseline characteristic that may influence or predict the participant outcomes, and this is going to get a little bit clearer as we go.

So, strata -- so I'm giving you a little bit of terminology under the stratification. So, stratification variable would be like sex. The strata are the levels of the stratification variables. So, it would be males and females, two strata, two levels. So, example of stratification variables, you could have site and multi-site in clinical trials --

that would be a site could be a stratification variable -- sex, age group, severity of disease. And in stratified randomization is actually very simple. The randomization is performed separately within each stratum to achieve treatment assignment balance within each stratum. So, think about running the trial for separately for the different strata. So, again, how do you implement stratified randomization?

So, let's say you have a block randomization, a block size of four, and we talked in part two with a stratification by site, so the clinical center is the stratification variable, the different sites are the different levels of the stratification variable. So, it's actually pretty simple. You think of running a trial at every site separately.

Transcribe another video

Paste any YouTube, Instagram or TikTok link to get a free transcript.

Free · No sign-up · Unlimited

Add to Chrome — transcribe in one click

And you do your bock randomization as if each site was a separate trial. So, you run the blocks -- the random blocks of size four for site one, with completely different series of blocks for site two and a completely different series of block of size four for site three. So, it's nothing really new.

The only thing new is that you do it separately for the different stratification levels. So, now let's say we have -- we do stratification by sex, how do we run this? Well, that's what we're going to do. So, this is permuted-block randomization, block size of four, stratified by sex. So, we have the same 32 participants that we had before,

but now what we are going to do is we just going to separate the women from the men. And, you know, separate is all virtual, we're not separating anything. This is only virtually in terms of programming. We separate the women from the men and we basically run as if we were running two separate trials.

So, the women are going to get their block randomly selected and assignment and the men are going to get their blocks randomly selected and assigned. So, basically, we do what we did before except on two separate parallel tracks. And so, the assignment happens to be like this. But now, look, look at what has -- what has happened.

We have the same number of people in treatment A as in treatment B, which is what we had in block randomization. But now, what we also have is the same number of women in treatment A as in treatment B and the same number of men in A as in treatment B. So, now we have taken care of both problems.

We have the same numbers in A and B and we have a balance of women and men in both treatments. And we solve now both problems that we had with simple randomization. So, the pros is that within each stratum, male or female, there is balance between the treatment groups. There is as many men --

there are as many men in treatment A as in treatment B, and as many women in treatment A as in treatment B. But please don't confuse with saying there are as many men as women in treatment A and as many men as women in treatment B. We don't control how many men and women consented

and are ready to be randomized. So, we can't put as many men as women in treatment A and as many men as women in treatment B. We can't do that. We don't control that. What we can control is to balance between treatment -- the women between treatment A and B and the men between treatment A and B.

So, that statement is not correct. One of the disadvantages of stratified randomization is the number of cells. You know, people sometimes go wild with the stratified randomization, but the number of cells can increase very rapidly with just a few stratification factors. Let me give you an example, so let's say you're running a trial

and you have four stratification factors, which is really -- I mean look at what they are, they're not really far-fetched. You have a multi-site clinical trials. Let's say you have three sites. So, you definitely want to stratify by site. You have sexes, two levels, two sexes, so you want to stratify by sex

because that may be different for men and women the intervention may have -- the efficacy may be different for men and from women. And let's say you think age has something to do with the outcome. So, you want to stratify with age. And you just pick three age groups, not, you know, not a lot.

And you want to stratify by severity of disease and just the minimum is, you know, low severity and high severity. So, it sounds like, you know, it's pretty, pretty decent. But look what happens. How many "cells," and I put quote cells, do you have? I put 36. Now, remember I had men and women

and you're running it separately. Well, now you have 36 tracks running parallel and separately. And, of course it's okay because computer programs are doing all this from a technical point of view that's not an issue. The issue is, do you have enough people in each one of those cells and if you don't that may cause a problem.

So, remember, randomization is conducted separately within each cell. So, if we put all these pieces together, I'm just showing you an example, just to tell you that this is real. People are using these designs. It's not just concept and theory. This is a trial, a clinical efficacy and safety of a novel tetravalent dengue vaccine

in healthy children in Asia, the phase three randomized observer-masked, placebo-controlled trial. And in this trial, healthy children aged two to 14 years randomly assigned 2:1 to dengue vaccine or placebo. So, two -- one third, I mean one third went to the dengue vaccine, and one third went to the placebo via permuted blocks

of six stratified by age and site. So, again, this is just to illustrate that, you know, this is real stuff. People are using these methods. Okay, so, now we did the stratified randomization, let's talk about cluster randomization. There's nothing very deep about cluster randomization, the only difference, cluster randomization instead of randomizing a person you randomize a group.

For example, they call it social units as an example, a cluster of individuals. So, a cluster randomization trial is one in which intact social units or clusters of individuals rather than individuals themselves are randomized to different intervention groups. What are examples of clusters? You can have a household where you randomize the whole household,

not the individuals inside the household, the whole household is randomized to different groups. The neighborhood, a classroom, a school, a worksite, a physician, a clinic or a hospital. These are clusters of people that the whole cluster gets randomized. Now, here's a little bit the trick with cluster randomizations, it's a little bit tricky because the outcome measure

is usually at the participant level, but the unit is being randomized. So, there is a little bit of, not a disconnection, but a little bit -- they're off, and we're going to mention very briefly about that. So, here's an example, a real example of cluster randomization, it's the ring vaccination trial, a novel cluster randomized controlled trial

designed to evaluate vaccine efficacy and effectiveness during outbreaks with special reference to Ebola. So, in this case, and I thought it's genius what they -- once they recognize a person who's newly diagnosed with the Ebola virus disease, they form a ring around that person and that becomes the cluster. And then they randomize the whole cluster,

those who are eligible within the cluster, 1:1 to either immediate intervention or a 21-delay vaccination. So, the cluster is this ring around the person who was newly diagnosed with the Ebola virus. And then if you look at the title, it talks about efficacy and effectiveness. The efficacy, they compared the hazard ratio in those enrolled in the study.

So, those in the cluster that were eligible and enrolled in the study, and the effectiveness is all of the people in the cluster, those who were enrolled and not, they compared those in terms of the hazard ratio.

So, why do we cluster randomize instead of just doing a regular individual randomization? Well, it could be -- individual randomization may be impractical or even impossible. Think for example about a teaching method. You know, you can't -- maybe within a school you could have different classrooms having different teaching methods, but you can't have within a class,

different teaching methods. So, that's just impossible. It could be also impractical, it complicates things, if you do it within the cluster -- you randomize within the cluster. Cluster randomization prevents contamination. So, think of a physician for example. A physician, the whole -- all the patients that see that physician can be a cluster.

But let's say you don't use them as cluster you randomize them individually, so the physician is basically saying, "Oh, this patient is part of treatment A, so I'm going to treat that person as A, and this patient is treatment B, so I'm --" and therefore there could be some -- we call it bleeding of the treatments

from A to B and from B to A. So, it's called contamination. And cluster randomization prevents that because it's randomizing the whole cluster. And it may be just convenience or cheaper to cluster randomize, so. So, how to implement cluster randomization. It's basically the same thing, it's the same as individual randomization except that the unit of randomization is the cluster.

So, all the methods we talked about with individual randomization can apply to cluster randomization, except that the only difference is that there's a matching of clusters before randomizations. So, let's say for example, we're going to randomize hospitals. So, a hospital is the cluster. And we have a list of hospitals that agree to participate in this trial.

Okay, so we have a list of hospitals and they're going to be randomized. Hospitals are going to be randomized. Well, you may want to match the hospitals in terms of size for example. So, you get the biggest two hospitals up there, and then you randomize the two biggest hospitals in A or B.

And then you get the next two in terms of size and you randomize. So, you're basically matching them in terms of size in this example of hospitals. So, again, sort of a little bit of a repeat here, but cluster randomization, the pros, it's more feasible in some cases, minimizes contamination, more convenient, and cheaper.

The cons are, it's the unit of analysis. It's sort of -- what I was talking a minute ago is you're randomizing clusters, but you -- in some cases, you randomize individuals. The unit of analysis should be the unit of randomization. But that's not really what you're doing, you're analyzing individuals. So, there is a little bit of --

it's a little bit off and some things need to be done to take care of that sort of in-alignment or un-alignment, or something like that. The number of participants in each cluster can be a problem. So, if you have a hospital that's -- or two hospitals that are huge compared to the others,

their results may dominate the results of all the others. So, the number of participants in each cluster -- I mean it's not a question of randomization, it's not a question that, yeah, you have as many, you have the two biggest hospital, one in A and one in B, that's fine, but these two hospitals may dominate the results

of all the others that come after it. And then you have the intra-class or intra-cluster correlation and that's very important to take into account the effect of an intra-class correlation. What is the intra-class correlation? Basically, it says that within a cluster there are similarities. They are all within that cluster. They are not from another cluster.

So, therefore, there is a correlation among those in that cluster that is not -- they are independent, or they are not correlated with another cluster but within that cluster there is some correlation. They are under the same physician. So, obviously, a good physician, there is some correlation. A bad physician, there is some correlation.

So, that need really needs to be taken into account, the ICC, what's called the ICC. So, so far, we have talked about fixed allocation algorithm. What does that mean? It's not fixed versus random. Fixed allocation algorithm means that the randomization scheme is decided up front and this does not change. So, you can have the whole string of assignments

before even you start the trial. That's fixed, and it doesn't change. That's the way it's going to be and that you do the whole trial with that scheme, that sequence of assignments. That's fixed allocation algorithm, versus adaptive methods, or dynamic allocation, and we're going to talk about that, and that's where the chances of assignment change

as you are enrolling. And that's what's called adaptive or dynamic. So -- and there are at least two, but I'm going to talk about at least two adaptive randomization. Minimization was a covariate-adaptive allocation, basically you change the allocation probabilities. So, you're tweaking those probabilities. You know, we sent one half to A, one half to B.

But here, after you see where, like the next man, the men went A versus B, or the women went to A versus B, after you see that, you change the probability. So, it's not really one half, one half, you tweak them so that they kind of fix the balance. Not in a deterministic way but in a probabilistic way.

So, you change the allocation probabilities to minimize any imbalances that have occurred by chance with respect to important baseline factors. So, you're observing what's going on as the trial is going, and you keep tweaking those probabilities to keep the balance. And then there are the response-adaptive allocation. Now, remember the first one is on baseline factors.

So, sex, age, severity of disease, et cetera, et cetera, socio-economic status, et cetera, baseline factors. In the response-adaptive allocation you are tweaking the probabilities based on their response, the outcome of the trial, so responses of participants enrolled to date change the allocation probability of the next assignment. So, how do you implement adaptive randomization?

It's actually pretty complex. You can imagine there's this dynamic thing that goes on and you really need expertise and software. For minimization, that's the first one, baseline characteristics need to be quickly measured and entered into the database. So, what this is talking about is saying basically, okay, if you're going to be tweaking those probabilities,

you have to keep track and as fast as possible. So, okay, what was the next assignment? Who was it? Where did they go? Let's put that in the database, so we can change the probabilities of assignment as fast as possible. With response-adaptive, assessments need to be quickly measured and entered to database. The first one is the baseline characteristics. You randomize.

You look at their baseline characteristics. You enter that in the database. But in response to that you have to wait until their outcome, their primary outcome, is measured -- is assessed and measured, and then enter it into the database and then change the probabilities. So, that needs to be done quickly for response-adaptive.

So, what are the pros in minimization? It may be more efficient than stratified randomization when sample size is small relative to the number of stratification factors or cells. And these things are changing, I mean, people are understanding minimization better, and so these statements -- you have to take these statements a little bit with a not too definitive

because things are changing. The research on adaptive randomization is really still going on and it does change. Response-adaptive allocation, it maximizes the number of participants on the more promising treatment because you can say, okay when I see that this treatment is doing well, people are responding well to this treatment, I want to increase the probability

of sending people to that treatment which is better. So, that's the advantage. What are the disadvantages, the disadvantages of randomization? The implementation and the statistical analysis are more complex. And there are these agencies -- they have some concerns. I don't want to sound too definitive about them because, again, like I said, research is changing,

but there is some concern about adaptive randomization, particularly, particularly the response-adaptive randomization. And that's it for this segment. In summary, in stratified randomization, randomization is performed separately within each group to achieve treatment assignment balance within each group. Cluster randomization is when groups of units or individuals are entirely, as opposed to individually, randomized to the different study treatments.

And in adaptive randomization, allocation probabilities for the next treatment assignment are changed after each executed treatment assignment. And my questions to you, is what is the difference between fixed and dynamic allocation? And what is the main benefit of stratified randomization? Thanks for watching.