I’m thrilled to dive into a topic that has been untouched in the past 300 episodes – crossover studies. It’s a classical design that offers unique advantages, yet it’s often overlooked.
Joining me today is Markus Savli, a biostatistician with a diverse background in biomedical engineering, mathematics, and epidemiology. Together, we’ll explore the intricacies of crossover studies and their applications in scientific research.
A crossover study is a longitudinal study where subjects receive multiple treatments in a sequence. This design allows subjects to serve as their own controls, enhancing the ability to address confounding factors. Crossover studies are particularly valuable in scenarios where treatments need to be compared, and their effects measured over multiple periods.
But they also have some pitfalls, that we talk about in this episode.
Listen to this episode now while we walk through the following key concepts:
Key Concepts:
- Crossover Studies: Longitudinal studies where subjects receive multiple treatments in a sequence
- Advantages: Subjects serve as their own controls, addressing confounding factors
- Randomization: Essential for balanced distribution of subjects across different treatment sequences
- Order Effects: Influence of treatment sequence on study outcomes, mitigated by randomization
- Carryover Effects: Treatment effects persisting into the next phase, controlled by washout periods
Challenges with Multiple Treatments:
Case Study: Non-Inferiority Study in Allergy Field:
- Utilized a controlled environmental chamber for allergen exposure
- Rigorous design included active controls, placebos, and non-treatment groups
- Meticulously controlled conditions: temperature, humidity, and allergen exposure ensured reliable results
This episode highlights the critical significance of comprehending crossover studies in the realm of scientific research. These studies enable a profound analysis of treatment effects, fostering innovative discoveries and ultimately enhancing patient outcomes.
Understanding the complexities of crossover designs is pivotal for researchers, providing a robust framework to delve into intricate treatment nuances and pave the way for impactful advancements in the field.
Listen to this episode now and share this with your friends and colleagues!
Never miss an episode!
Join thousends of your peers and subscribe to get our latest updates by email!
Get the
Learn on demand
Click on the button to see our Teachble Inc. cources.
Featured courses
Click on the button to see our Teachble Inc. cources.
Markus Savli and his company Biostatistik Consulting
Statistical Services for Clinical Studies and Data Analysis
Our seasoned professionals bring over 20 years of experience in pharmaceutical and biotech services. We’ve partnered in numerous projects in clinical trials, quality control, and scientific support, with 100% satisfaction shaping the future of healthcare.
Schedule a free, no-binding call to explore how our expertise can boost your projects: +41 79 833 22 22
E-mail: markus.savli@biostatistik-consulting.ch
Learn more about his company here: https://www.biostatistik-consulting.ch/en/biostatistics/
Transcript
Cross-Over Studies – When Do They Offer a Great Advantage?
[00:00:00] Alexander: Welcome to another episode of the effective statistician and today I’m super excited to talk about a topic that I haven’t touched on in over 300 episodes and I can’t believe it. It is such a classical design. Although I need to confess personally. Never worked on this because, well, we’ll get into the reasons later on.
So, and for this, I have Markus here. Hi Markus, how are you doing?
[00:00:34] Markus: Hi Alexander. Great. I’m doing fine. Thank you for the invitation.
[00:00:38] Alexander: Nice to have you. So let’s start introduce yourself. What are you doing? What’s your company doing?
[00:00:45] Markus: Markus Savli is my name. I’m a biostatistician by training. I have a background in biomedical engineering and mathematics as well as epidemiology.
I work and, I am at the founder of biostatistics and consulting. We are a consulting office in Zurich in Switzerland. And as the name already indicates we offer biometric services for pharmaceutical and biotech, medtech companies, but furthermore for universities, for example, we can do scientific writing like meta-analysis or similar services. Programming dashboards and so on. Right?
[00:01:29] Alexander: With your background that is so broad much broader than, you know, most statisticians come with, who have spent a whole career only in clinical trials to biomedical engineering programming, really kind of a broad data science backgrounds are pretty cool.
[00:01:48] Markus: Yeah. Thank you.
[00:01:49] Alexander: So today we will talk about crossover studies and you recently published one and we’ll go into this a little bit later, but let’s, in simple terms, what’s a crossover study?
[00:02:05] Markus: A crossover study, in simple terms is a longitudinal study where subjects receive multiple treatments in a sequence. This can be an observational study, where it’s less controlled, or, and this is what we are more interested in the medical or in the scientific area, to have a controlled experiment. So, then it’s randomized, usually, and it’s balanced. That means the same number of subjects receive the same number of treatments and the same number of periods.
[00:02:41] Alexander: In the most simple case, if we have two treatments, treatment A and treatment B. Half of the patients first get treatment A and then treatment B and the other half gets first treatment B and then treatment A. So we have two treatments and two periods.
[00:03:02] Markus: Right. This is you explained already the randomized part of it, right? So you don’t start with the entire number of patients with treatment A and then go over to cross over to treatment B. You randomize them. That’s correct.
[00:03:17] Alexander: So, that first kind of appears like yeah, looks like an unusual design. Why not just kind of compare, you know, A and B, have the first period, that’s it. Why do the second period?
[00:03:32] Markus: This comes with certain, properties. First you can consider when you take a subject and have him getting one treatment and then the other or the placebo treatment, the same subject serve as its own control. And the advantage of this is that you can influence The, the confounding right in a better way than you can do in a randomized or just different groups, right.
Then in a parallel design, for example, it’s a bit less uncontrolled. And given this circumstances, it gets along with the higher statistical power and higher power means that you can get along with fewer subjects. And this is a matter of cost in the end. Right.
[00:04:15] Alexander: Yeah. So especially if you have high variability. Between the different subjects. Yeah, that is clear that, you know, the within subject control gives you much more additional comparisons, because you don’t just compare treatment A and treatment B between the different subjects in period A. And period B actually, but also you compare period A to the first period to the second period and also have these comparisons.
So yeah, cool. Very good. So you have instead of, well, just one comparison, A versus B at the beginning, you have basically four comparisons. Now, A versus B in the first period, A versus B in the second period, and A versus B in both groups, so to say. Do you call them sequences or groups? How would you call that? Patient.
[00:05:17] Markus: Yeah. Sequence. Yeah. Sequence is good. Right? Period. In the ideal world, there should be no difference actually. When you start with an and go cross over to B or you start with B and cross over to a, this is actually what you want. We wanna find out in this study whether this is really the case.
[00:05:36] Alexander: Yeah. I think that is one of the assumptions that you need to make. Isn’t it? So if. It does matter whether you first go to B and then to A versus A, and then to B. So the treatment difference between A and B depends on the period. What? Why is that a problem?
[00:06:02] Markus: Yeah, you don’t want to actually, you don’t want to experience this and just study then you’re really are a bit in a mess for the point is for some study designs, there’s actually no other option than do a crossover design.
For example, bio equivalent studies, right where the guidelines actually tells you that you have to do this and then you need to control for these effects. We talk about it. like an order effect. What you mentioned before, this is when the sequence may affect the outcome. An example could be that, let’s say the first treatment has a, produces a more side effects or adverse effects, right?
In the patient, and then in the second layer later treatment where there is less side effects, but the patient is more susceptible or more kind of framed for the side effects and is perhaps reporting more adverse effects than they actually are. Right? So this is what we don’t want to have. And that’s why we do randomization.
Where we can check, is it from A to B or from B to A, right? Is there a difference, right? And if this, it should be balanced. Actually, there should be no difference between these sequences.
[00:07:15] Alexander: This is a carryover effect?
[00:07:19] Markus: Not is the order effect. And the carry over effect is another effect. It is where the treatment itself has last long into the next phase. So it confounds, the effect of the next phase. In a good crossover design, there’s sufficient time between the treatment effects of treatment periods, and we call it a washout phase so that there is no interference, and they say, well, in the guide, I say more or less five half lives, right?
Should be sufficient, right? And you need to check this in your data, right? It’s the case, right?
[00:07:57] Alexander: So that already has a couple of different implications. So first it only works for treatments that have. reasonably short half lives. Yeah, so if you have a half life of several months or something like this, it becomes a very, very long study.
Second is, of course, that you have some kind of chronic Stable disease. Yeah. If you cure the patients with your treatment, that’s probably not the right, the right way because then you will always have a carryover effect. So, it needs to be some kind of symptomatic treatment. And yeah, so you need to have this kind of, you need to more or less.
Get back to the first baseline, you know, so basically you have two baselines, isn’t it? And the settings and you have a baseline for each period.
[00:08:54] Markus: That’s right there. You, you mentioned a couple of points, right? That this crossover isn’t, it’s not suitable for all situations. And the one, when you cure the patient in the first.
The treatment phase the baseline for the second phase is different of course, then, and there is no more improvement for the patient in the second phase. So the second phase in will be worse in the end in the calculation, but then the study design is not appropriate for that. But the biostatisticians should know about it and about the defect.
So in the curative situation is perhaps not the ideal study design. But for the chronic ones where you apply certain treatments and then the patient benefits during this treatment phase in a certain way. But then when you stop this treatment and the baseline returns to its original level, then you have a good situation where you can independently check both treatments.
It could be the active verum and the placebo situation, right? Or in the case of bioequivalents, where you apply kind of a single tablet or a single application of a drug, right, where you check the pharmacokinetics. So after a certain half-lifes, the pharmacokinetics returns to its baseline.
This is the case when, I mean, Bio equivalence has, its. Has its value because, for example, a company decides to change its formulation from a suspension to a tablet or vice versa, right? The properties are similar according to their production, but still the government and the authorities require that you need to check this.
And then. With a single formulation, right? You do the pharmacokinetics, blood AOC, and so on, and then you compare this. So, in that situation, the baseline is established, and you can go ahead with it.
[00:10:47] Alexander: Okay, awesome. Let’s go into the problems are very simple. Two treatments. What happens if we have multiple treatments?
So let’s say three treatments. Let’s make one step further. Do we then always have also three sequences? So, you have basically treatment A, B, and C. So you have A, B, C, A, C, B, and all these kind of different combinations.
[00:11:21] Markus: Yeah, the simplest. The sign is the two period, two sequence crossover, right? It’s the most simple when you have a, let’s say, the active and the placebo are two different ones. But when you go for three ones, it could be different dosages like this. Then it’s not three you get six treatment sequences because of this, you want to randomize all the sequences, and want to check and avoid actually any, or you don’t want to miss any effect, right?
For a certain sequence. So you need to randomize of this six treatment sequences, right? And for the, to assign the patients, you do a block design. And what we do is that we have a complete design. That means every patient finishes all the sequences, right? And we have a couple of patients in every of each of these sequences, right?
[00:12:13] Alexander: So it gets, the study gets longer and of course you have. Yeah, you need a certain number of patients for each of these six sequences. Okay. I guess that’s probably the maximum that you would use, going to four or five, it gets more and more complex. Or have, have you seen crossover studies with even more treatments?
[00:12:40] Markus: I myself was not involved in such studies, but the three periods is not so uncommon, I would say. Four or five, I mean, theoretically. Should work, but I cannot remember when I saw one. Actually, I haven’t paid attention to this higher sequences. But as you said is for sure more complicated and you need to pay attention to the details then right. Yeah. I think, theoretically, it should work.
[00:13:12] Alexander: The studies get longer and longer and bigger and bigger. And of course, you increase the chance of something going wrong like dropouts. So, that is what was one of my first, you know, questions when I saw this design. Okay, what happens to… How do you analyze patients that drop out between the two sequences two periods, for example?
[00:13:35] Markus: Yeah, this is always the case when you have a longitudinal study design that you which often happens and is unfortunate for the study design and for the analysis in a the guidelines when, when we go for the, let’s say, for the bioequivalence guidelines of the FDA, of the EMA they require a minimum of around 12 patients, right?
And so, when you have around 12, 18 patients right they prohibit, actually, any extra patients, right? When you miss one, and then you just load another one into this study. No, this is not allowed. So, but in other hands, they say… Or do you offer the option that you can account for dropouts. When at the beginning of the study before you start right it is likely or you assume Oh, there will be a kind of a dropouts because of this peculiar treatment what you apply to them You can account for these dropouts and you know add this 10 15 20 percent whatever, Patients the beginning, right?
So that when they drop out during the study that you have still the sufficient high number of patients, because when you, when one period is missing, then you don’t have this comparison, right? This is missing actually. Yeah. And this is what we do. We just add account for the dropouts.
[00:15:00] Alexander: You would, if the patient drops out. You wouldn’t consider them in the analysis? Because he doesn’t have all the information?
[00:15:10] Markus: When that’s a problem and that’s why we have this ITT and protocol options right for and sensitivity analysis options right. But when you do a kind of a mean comparison, right, between the two treatments, then it is in difficult, right?
Because you don’t have one and it’s not a, it’s not like in the parallel design. Okay. You can make the mean in the one group is one or two or X patients more than any other one. But in the parallel design where you want to make use of different statistical methods. I mean, this one is missing, right?
[00:15:47] Alexander: So it’s a, it’s a much bigger problem compared to the simple parallel group design. Okay. Very good. Let’s go back to the studies that I mentioned at the beginning that you recently published and tell us a little bit about this crossover study because that actually had three different treatments.
[00:16:09] Markus: Correct. That was not a bioequivalence study. It was a non inferiority study in the field of allergy. It was a medical device product, kind of an over the counter drug. So less stringent than other pharmaceuticals and the sponsor wanted to compare his product to already marketed competitor and in the allergy area of Phil is.
They are very, the patients are very sensitive rights to the environment and also to the treatment. So the treatment effect is small and highly variable, right? And that makes it really difficult to assess these treatment effects and the, well, we don’t say drug, the over the counter is a kind of a nose spray, right?
It’s applied into the nose to relieve certain allergic effects, right? Or the idea was to check what was the potential for the for the allergic relief to get this relief. And they compare, wanted to compare to this, this other drug and also to kind of a non treatment. So the idea is, okay, if this is so highly variable, is this pure chance or is this what, how does the patient fare off when he’s doing nothing?
And how did we do this allergic study or check this treatment effects? This, they usually do it in winter time. Why? Because then there is this pollen, there’s no pollen season. So the, it’s during the day. The temperature is cold and there’s no other grasses or pollen around or less and then they have a The patient in a controlled environmental chamber that is you can say kind of a small room in the laboratory ,and the patient sits there and this room is air conditioned, it is temperature controlled, it is humidity controlled, and whatever other parameter, environmental parameter controlled, and via the air condition, they expose the patient with a certain set of pollen and other allergens.
Okay. And the patient sits there and then it’s a The exposure lasts from three to six hours, depending on the study. And the patient sits there and then he, he’s there and creates the allergy. He gets the allergy and then every 15 minutes he has to report his symptoms like sneezing, itchy nose, watery eyes respiratory, whatever.
There’s a list of symptoms that he has to report. And and then it’s done. And then after a washout period, he returns for the next one and so on, right?
[00:19:04] Alexander: Okay. Wow. So highly controlled environment in which you can also control the exposure. Okay. That’s really, really interesting. So the exposure in terms of the allergens, not the exposure in terms of the treatment, which you can,
[00:19:23] Markus: it’s even, it goes in that far, the, the, the way the tissues, right. When you have a running nose and the blow to your nose, they collect the tissues and weigh the amount of what came out of the nose. Right.
[00:19:35] Alexander: Well, yeah. Okay. That’s really an interesting lab setting. Yeah. And a really great setting for these types of studies because you have these repeatable conditions. people in winter get out of the chamber again, they don’t have any exposure to allergens anymore, get back to normal. You can have a washout period.
And these are really, really good conditions for these crossover studies. And yeah, it’s, it’s really good that you have these golden standards, so to say approach here with active. Control and active control and also placebo control. So you can see, have all these things that you very often need for non inferiority study, because very often your non inferiority margin.
Depends on the difference between active, active and non active and placebo. Very good. Awesome. We’ll put the link to the study also into the show notes, so you can have a look there. Where can people find you, Markus?
[00:20:49] Markus: That’s very easy. I’m of course in LinkedIn. That’s one option can connect to me on LinkedIn and drop me a message or to take a look at my website. And see what else we are doing in the team, and by email, that’s the traditional old, but still working way.
[00:21:10] Alexander: Yeah. We’ll put links to these kind of different things into the show notes. So you can very, very easily connect to Markus. A couple of interesting things Markus sits in Switzerland, and sometimes there’s a lot of concern in terms of Swiss people about Swiss data not leaving the country.
So if you have If you need to work with, you know, Swiss biotech company or Swiss hospital or these kind of areas, this is very often a concern. And then having a Swiss consultant there that can do all these kinds of different things without the data leaving the country. Then, for sure, reach out to Markus.
And then, of course, well, although he is from Switzerland, he actually speaks German. So also for all kind of German training, German HTA submissions, all these kind of different things. If you belong to the part of the world where your mother tongue is not German. Which. Yeah, biggest part of the world, and you need to have someone said is a good German speaking statistician.
I also highly recommend you reach out to Markus. Thanks so much, Markus. Is there any final things that you would like to give to the listener?
[00:22:39] Markus: I can add Spanish. To the options language options.
[00:22:43] Alexander: Okay. Okay. Very good. So we talked about crossover studies the benefits, the problems compared to traditional designs have a look into this and increase your knowledge around these things and for sure Connect to Markus, always great to improve your network, to strengthen your network.
I always recommend this to people because your network is your net worth very often. You know, the more people you are connected to, the more and easier you will find maybe new colleagues, new partners, new employers in the future to connect with Markus. If you haven’t connected with me already, of course, connect with me as well.
Thanks so much, Markus, for being on the show.
[00:23:36] Markus: It was a pleasure. Thank you, Alexander, for inviting.
Join The Effective Statistician LinkedIn group
This group was set up to help each other to become more effective statisticians. We’ll run challenges in this group, e.g. around writing abstracts for conferences or other projects. I’ll also post into this group further content.
I want to help the community of statisticians, data scientists, programmers and other quantitative scientists to be more influential, innovative, and effective. I believe that as a community we can help our research, our regulatory and payer systems, and ultimately physicians and patients take better decisions based on better evidence.
I work to achieve a future in which everyone can access the right evidence in the right format at the right time to make sound decisions.
When my kids are sick, I want to have good evidence to discuss with the physician about the different therapy choices.
When my mother is sick, I want her to understand the evidence and being able to understand it.
When I get sick, I want to find evidence that I can trust and that helps me to have meaningful discussions with my healthcare professionals.
I want to live in a world, where the media reports correctly about medical evidence and in which society distinguishes between fake evidence and real evidence.
Let’s work together to achieve this.