Cluster Randomized Trials

Join Dr. David Reuben, from University of California, Los Angeles, and Dr. Subashan Perera, from the University of Pittsburg School of Medicine, as they discuss cluster randomized trials. They touch upon the main characteristics, statistical advantages, and disadvantages, and potential ethical or operational issues of conducting these types of trials.   

To view a transcript click here then select the transcript tab. 

Transcript

Copy Transcript Copied

David Reuben, MD: Hello, I'm Dr. David Rubin, Archstone Professor of Geriatrics at the David Geffen School of Medicine at UCLA. And I'm here with Dr. Subashan Perera, who is a Professor of Medicine and Biostatistics, in the Division of Geriatric Medicine at the University of Pittsburgh. 

Dr. Perera, can you tell me a bit about yourself and what you do?

Subashan Perera, PhD, FGSA: Sure. My areas of expertise are biostatistics, clinical trials, and research methodology. And in the past, my focus has been in research, enrolling mainly older adults. Many of them are living with multiple chronic conditions. A biostatistician's role in this environment is very broad, not confined with data analysis at the completion of a research study.

We are all involved from the very beginning in formulating hypotheses, key design decisions, finding ways to conduct the study within feasibility and resource [00:01:00] constraints, randomization, communications with regulatory entities such as DSMBs, and of course data analysis, and interpretation and preparing findings for publication.

David Reuben, MD: Thank you very much. In my personal research, I live by the biostatisticians, so they're most valuable players indeed. 

To begin, can you tell us what clustering is? Why it is important to consider? And what are the main characteristics of a randomized clinical trial that makes it a cluster randomized trial?

Subashan Perera, PhD, FGSA: Sure. Clustering occurs when individuals from a certain group tend to be more similar to each other compared to individuals from a different group, usually due to reasons that are not fully and readily measurable. 

For example, we can think of a cluster as older patients of a geriatrician. Due to a range of factors- perhaps the [00:02:00] physician's training, whether the physician follows a conservative or aggressive strategy- due to range of factors, the outcomes among the geriatrician's patients tend to be more similar compared to patients of another geriatrician. This similarity is called clustering. Sometimes we also say patients are nested within the geriatrician. 

Clustering is important to consider because it has many important implications. For example, most statistical methods assume independence among observations. This is violated by clustering or similarity of individuals within the same cluster. They are in some sense, dependent on the care received by the geriatrician. If you ignore this clustering during trial planning, then that can lead to overestimating of the statistical sensitivity and underestimating of sample size requirement. And that would [00:03:00] increase the likelihood of concluding treatments are similar, when they are in fact different. 

So on the other hand, if you ignore clustering during trial data analysis, that can lead to underestimating the amount of noise in the data. In practical terms that would increase the likelihood of concluding treatments are different when there is insufficient evidence to say they're different.

The main characteristic of a cluster randomized trial is that whole clusters rather than individual patients, are randomized to treatments, but the outcomes are collected at the individual level, despite the randomization at the cluster level. And the outcomes are analyzed at the individual level while accounting for clustering.

In the previous example, you can imagine geriatricians being randomized to treatment practices, rather than individual patients, but outcomes being collected from individual patients and [00:04:00] analyzed. 

David Reuben, MD: This makes so much sense. Because there's a certain culture among a practice. There's certain resources that are available. [In] the practice, there are customs in which workflow and, and other things that, that make physicians or or other providers who are in the same practice or in the same group perform very similarly. So this is a very helpful clarification. 

When do we commonly see cluster randomized trials conducted? 

Subashan Perera, PhD, FGSA: As a general rule, we typically see cluster randomized trials in some organized setting. It may be a place where older adults reside or receive care and services.

We already mentioned the cluster of patients of a geriatrician. Other examples ,which are more residential in nature, are nursing homes, assisted and independent living facilities, and senior apartment [00:05:00] complexes. It doesn't have to be a residential setting. So for example, patrons attending a senior community center or another daycare center can also be considered clustered.

Also, let me add that cluster randomized trials generally tend to be large and more pragmatic. They also tend to examine health systems, or quality control interventions, or implementation strategies. Especially in the case of implementation strategies, we commonly see stepped wedge designs where you would roll out a certain implementation strategy across facilities of practices. They also tend to be clustered randomized trials. 

David Reuben, MD: So you spoke a little bit about reasons why there's a need for cluster randomized trials, but what are some of the advantages and statistical disadvantages of conducting these?

Subashan Perera, PhD, FGSA: I think there are three main reasons for conducting a [00:06:00] cluster randomized trial.

One is to preserve internal validity, mainly by preventing cross-contamination between the treatment groups. For example, think about a trial examining the impact of a more extensive physical examination of patients in chronic low back pain. If individual patients are randomized, a physician cannot perform the extensive evaluation for one patient and not the next patient, simply because the second patient is not randomized to the intervention group. It is unlikely that the physician would be able to do that without carrying over some of the concerns that he or she expressed towards the first patient. So it makes more sense to randomize all patients of a physician to one intervention in that scenario.

A second reason is that some interventions are naturally designed to be delivered to groups rather than individuals. [00:07:00] For example, if you can imagine the group exercise program trial, conducted in an independent living facility. That's more conveniently conducted as a cluster randomized trial because of the nature of the intervention.

Perhaps the third advantage is a cluster randomized trial may mitigate participant disappointment. We want to avoid that if possible, because that can result in participants dropping out, which can bias data. For example, you can imagine someone might be disappointed about being randomized to the control group and they may be more likely to drop out of the trial, especially if they see someone near to them being assigned to the active intervention perceived as more beneficiary.

So those are some of the, the advantages. There are statistical disadvantages to conducting a cluster randomized trial despite its advantages. So one of the first considerations is whether [00:08:00] the cluster randomized trial is actually needed, and whether any of the advantages can be realized. If not, the, the disadvantages outweigh the advantages.

For example, a trial of a medication, a pill where everyone involved can be effectively blinded with the placebo. That may not be a suitable place for a cluster randomized design. 

Given that a cluster randomized trial is needed. The main statistical disadvantage is the need for a greater number of participants compared to an individually randomized study. Another disadvantage is that statistical analysis of the data is more complex. 

David Reuben, MD: Yeah. I think this, those are all really good points. One of the things just to reiterate is the whole idea of contamination where people in our trial are getting both interventions because their [00:09:00] doctors can't differentiate or whoever's administering the intervention can't differentiate between one intervention versus another. They, they get a little bit of both. So that contamination issue is, is very important as well. 

So how is the extent of clustering and its influence on the methodological aspects of a trial assessed? 

Subashan Perera, PhD, FGSA: The extent of clustering is quantified with what we call the intraclass correlation coefficient or ICC for short.

It is basically like any other correlation. It's a number between zero and one, and it indicates how large the between cluster variation is in an outcome relative to total variation. So the larger the number, the greater the extent of the cluster. In planning a study sample size requirement is first assessed ignoring clustering, assuming that it's an individually randomized [00:10:00] study.

It's common to have a certain level of statistical power to detect a certain magnitude effect under an anticipated participant retention rate. So that's a standard computation and usually compute the sample such requirement ignoring cluster. 

Next, what we do is the computer sample size requirement is inflated by what we call a design factor. And this design factor depends on intraclass correlation, average cluster size, and variability cluster size. Cluster size data is usually readily available because you know in which clusters that you will conduct the study.

But the intraclass correlation for this computation has to come from a previous multi-cluster study. That also means that it has to come from some, from a somewhat larger prior study rather than a small pilot study. [00:11:00] And what we find is that even seemingly small values of intraclass correlation, say 0.05, can make a substantial difference in the sample size requirement, and statistical power. 

After you're done with the study, when you're analyzing the data, statistical methods that can accommodate clustering and non-independence of individuals have to be used. Commonly used techniques include mixed models, shared frailty survival models, and generalized estimating equation models. And there are others. It is very important to publish interclass correlations from a cluster randomized trial so that others can use them for rigorously planning any subsequent studies. 

David Reuben, MD: So just to summarize the effect of the interclass correlation, the higher it is, [00:12:00] the effect on the sample size needed would be?

Subashan Perera, PhD, FGSA: Yes. Everything else being equal, the design factor, which you should inflate, the sample size, is greater with the greater intraclass correlation.

David Reuben, MD: Yeah. So the, the higher the interclass correlation, the larger sample size you would need given the design factors were equal. Is that correct? 

Subashan Perera, PhD, FGSA: That- that is correct. And also a little bit different: once in our prior work we have found objectively measured outcomes tend to have smaller intraclass correlations compared to self-reported measures of mood and, and, and so on. And that, that makes sense because participant interaction affects subjectively collected self-report types of outcomes more than objectively measured outcomes. 

David Reuben, MD: So what [00:13:00] are the randomizations, ethical, or operational issues [that] need to be considered when conducting a cluster randomized trial?

Subashan Perera, PhD, FGSA: Yeah. There are several other considerations. For example, regarding randomization in a large individually randomized trial: simple random assignment almost always works to balance out the two treatment groups characteristics at baseline. But in the cluster randomized trial, even with a large number of participants- even with the- what's being randomized are clusters. So the number of clusters being randomized can be small, and that increases the chance of an imbalance at baseline between the treatment groups. So a randomization scheme in the cluster randomized trials should really be stratified by important cluster characteristics for the study. 

Regarding ethical considerations, it's important to note that a scientifically flawed trial cannot be ethical. [00:14:00] So, even more important to ensure sufficient attention is paid to the methodological complexities that, that we just talked about. 

Also, conducting a cluster randomized trial by itself is not the justifiable reason for not obtaining informed consent from individuals. In fact, the opposite might be true in older adults with chronic conditions.

So for example, in a nursing home cluster randomized trial, informed consent is needed not only from individual residents or their representatives, but also cluster gatekeepers such as facility administrators or state regulatory authorities. 

Regarding operations, there are some difficulties as well to maintain the same level of operational rigor compared to an individually randomized trial. It is difficult to keep individual participants and cluster gatekeepers [00:15:00] blinded when you deliver an intervention to the whole cluster. Also, it may not be possible to complete all baseline assessments before randomization, as is usually done in the individual randomized trial. This is because some planning ahead needs to happen and scheduling to deliver a cluster level intervention.

David Reuben, MD: Well, we're, we're getting close to our time, but I would like to ask you if you could summarize, what would your take home message or messages about cluster randomized trials- what would they be? 

Subashan Perera, PhD, FGSA: In summary, I would like to reiterate that several important and unique issues must be considered when planning and conducting a randomized trial. And also [the] same is true when consuming the findings of a cluster randomized trial, and critically evaluating the results of one that you might read in the literature. 

I think the most important issue is the justification and rationale [00:16:00] for cluster randomized design. Given there is a reasonable justification, it is important that clusters are chosen appropriately, and clustering is incorporated into study planning and data analysis. And also careful attention should be paid to ethical and operational challenges involved with conducting a cluster randomized trial. 

These considerations are more formally compiled in several sets of published guidelines and checklists. One is the Consolidated Standards of Reporting Trials or CONSORT Statement. And another is the Ottawa Statement on the Ethical Design and Conduct of Cluster Randomized Trials. They can serve as convenient references for, for best practices. 

David Reuben, MD: So one of the things sometimes people say, and I just wanted to hear your response to them, what it would be is. They, they say that the cluster randomized trials [00:17:00] aren't as strong evidence as a individually randomized trial.

What would you, what would you say back to them? 

Subashan Perera, PhD, FGSA: I think it depends on the purpose of conducting the trial. So for an initial investigation, that may be true because a, a small trial realistically cannot be a cluster randomized trial, and an initial trial may be a feasibility or efficacy trial, and it's probably more important to not go too far towards the pragmatic end of the, the spectrum. And also cause of the size of the study, it can be con- conducted under more controlled conditions. But larger, more definitive studies, we would like them to be as realistic as possible compared to real world scenarios. And we would like to [00:18:00] have those studies conducted so they have internal validity and external validity. And when you get to that point, I think cluster randomized trials are almost necessary for the reasons that we mentioned. 

David Reuben, MD: That was a very clear explanation. Thank you for sharing your thoughts and for developing this module. And thanks everybody for being patient listeners and please look forward to actually completing the module.

Thank you very much. 

Subashan Perera, PhD, FGSA: Thank you. Thank you for having me.