We are the Association for Child and Adolescent Mental Health, or ACAMH for short. And this is ACAMH Learn. Hello, everyone. My name is Jorien Treur, and I am an associate professor at the Amsterdam UMC, the academic hospital in Amsterdam, the Netherlands. And today, I'm going to present for you about triangulation, evidence triangulation, smoking, vaping, and mental health. So the title of my talk is investigating causality using evidence triangulation. And I'm going to apply this approach of triangulation to the topic of smoking, vaping, and mental health. And my aims for today and for this lecture are twofold. So first of all, I would like to highlight and emphasise the approach of evidence triangulation as a way to obtain more reliable answers to difficult causal questions. And then the second aim of this talk is to also show what evidence triangulation looks like and what it can tell us about smoking, vaping, and mental health. I'm going to go into more detail about that later on. But let's start with talking about causality. So causal inference is a very tricky business. It's quite difficult to prove a causal effect. So in this case, if we want to know whether a certain variable x actually influences, causally influences in outcome variable y, this is difficult. So first of all, this is the effect that we're after, a causal effect of x on y. And if we know that this is indeed true, that this is the case, it helps us because it would let us know that when we intervene on this variable x, it would help prevent or somehow regulate outcome y. So it can be of interest for policy prevention and treatment. But unfortunately, there always is a bias that comes from confounding. So what we call confounders are other variables that influence and increase both the x variable in this question and the y variable in this question. So that is what we need to get rid of. But also, in many cases, it can be true that the outcome variable actually has an effect on the exposure variable. So then you would be looking at reverse causality, which sometimes there's directionality in both directions, or sometimes there's bidirectional causality. But it complicates matters further when we want to know whether this variable x has an effect. So for the topic of today, we can focus on this question of how can we reliably determine if smoking, cigarette smoking has a causal effect on mental health problems. Well, first, what we would like to establish if there is an association between the two. So is it the case that more smoking is associated with more mental health problems. Well, this is true. So this figure is a hypothetical figure. But we know from a lot of literature that people who smoke and people who smoke more are on average also at a higher risk of developing anxiety, depression, psychosis, and a range of poor mental health outcomes. But as we mentioned in the previous slides as well, this, of course, can also be the result of confounding. So in this figure, you can see what that might look like. So if we introduce something as socioeconomic position, which we know correlates with both smoking and mental health problems, you will see that across these different strata of socioeconomic position, the association is actually very different. So these are things that if you don't take them into account, merely showing an association is not enough for causality. And it's important to mention that this can also occur in longitudinal studies. So longitudinal research does tell us more because you can look at the temporal nature of the association, whether the x increase comes before the y increase. But still confounding can affect those types of studies too. So when we talk about causal inference, one of the things that people of course, think about and which is the gold standard of proving causality is randomization, randomising people into different groups, which is what is shown here in this figure. But this is not always possible for every question that we might have. So I'm sure most of you will know what this type of study design does. But just very briefly, you start out with a population of interest. In this case, the colours, again, relates to the confounder or potential confounder of socioeconomic position. And then by randomising them, so randomly assigning them to one of two different groups, we make sure that the distribution of those confounders is equal or at least randomly determined. Then we expose one of the groups to the intervention or variable that we're interested in. And the other group is the control group. And then we assess the effect it has. But there are many potential essential weaknesses here or limitations. So for example, selection bias, when we cannot get a representative group or it's difficult to get certain groups into these types of studies. Attrition bias, where people will drop out of the study in specific ways that also introduce bias. But importantly, for something like smoking especially, it's not always possible to randomise someone. It might not be ethical, for example, to make somebody start smoking or smoke more, or it might not be possible for one way or another. The control group might also be more difficult to determine. And it would often require a long follow up duration. So, for example, if we are interested in learning about a risk factor for the onset of mental illness, you would have to have a year long-- years long follow up time, and that might not be feasible. So in conclusion, using randomised trials to learn about causality for this question of smoking and also later vaping on mental health is largely impossible. Although there might be some possibilities with looking at smoking cessation. So stopping smoking, we'll get back to that later on as well. So going from these more conventional methods. So we have observational methods, where we associate these variables with one another longitudinally, potentially to learn more. And we have the randomised trials. Those are not enough for this question. So what researchers do for the question of smoking and mental health, but also other complex causal questions, is to look at more advanced methods to do so. And there are many different ones. So I picked a few here to show as example because these are being conducted often in this field fields and because these methods have told us things about smoking and mental health. But of course, there are many more. So very briefly, one example is a random intercept cross-lagged panel model. This is a type of model that uses longitudinal data. As you can see, on the different time points, time point 1, 2 and 3. And the strengths of this method is that it introduces a random intercept. So this is an intercept or a latent variable that captures the differences between people. So there are just very many individual differences between people that can be caused by all-- a range of different things. And when we capture those in this model, that means that what we're left with is only within person effects. And that is what we're interested in for causality, because we want to know whether in one individual, for example, smoking increases feelings of depression. Another example that's interesting for this topic is the regression discontinuity design, where researchers use a certain external variable or event that takes place and which is not associated with the exposure variable or is not determined by the exposure variable. And in essence creates two different groups. So in this example, it could be, for example, a policy change, such as an age-- an increase in the age of sales for cigarettes. And then what you would see is that there is, based on year of birth, two different groups in the population, one that is affected by this policy change and one that is not. And that's completely by chance, basically just because they were born a bit earlier. And that leads to a quasi randomised design, where you can compare these groups and see whether it has an effect for example, mental health. And then finally, the third example is one that I also personally conduct a lot of studies on but also it's becoming more common in the fields. And this method is Mendelian Randomization or MR. Basically, this method, which can also be seen as a quasi experimental method, uses genetic variation, SA randomizer or as an instrumental variable, one can also say. And in this picture, you can see it alongside the randomised trial design because that's a convenient way to explain it mostly. And instead of in a randomised trial, the randomization occurring through the researchers that are assigning people to different groups, in an MR design, it should say MR on the bottom. So the bottom one is where we see allele segregation. Basically, what we're assuming is that people will randomly inherit a certain genetic variant because we know during meiosis that this allele segregation process is random. So by chance, certain people have a higher risk variance for an exposure. For example, for smoking, certain people are genetically more liable to smoke or to become dependent on smoking. And other people have a low risk variance. So in the population, this will also mean that there are semi-randomised groups of people. And we can compare those and also determine if it affects the outcome. So this was very briefly on more complex methods. But after showing you all of these fancy and complex methods, it's important to realise that basically every method has its own limitations, and every method has sources of bias that can lead to spurious findings. So one can never really get a full answer with using one method alone. And that is where triangulation comes in. So basically, in its simplest form, triangulation means that we use different methods. So a more than one method to look at a complex causal question, such as is something a risk factor for developing mental health problems? And all of these methods will have different strengths and different biases. So they give a different viewpoint and a different perspective on the question. And basically, when we see that these methods align, so if it is the case that they all conclude that there is evidence for causal effects, it would mean that there's much more robust evidence for causality. Because it's very unlikely that all of these methods that have different biases and different strengths would come to the same answer if in fact, there is no true causal effect. So that's the theoretical idea around it. And triangulation, as a term, has been-- or as a word, has been used for many decades across different fields in slightly different ways. But more recently, this definition of triangulation was posed by Lawler Tilling and Davey Smith, which states that it is the practise of strengthening causal inferences by integrating results from several different approaches, where each approach has different and assumed to be largely unrelated key sources of potential bias. So it's important to note here that the sources of bias should be different and preferably uncorrelated. And besides this definition that has been posed, there is just simply also a lot more attention for this idea of triangulating and of replicating evidence and effects across different contexts, and that this needs to be a bigger part in developing more robust evidence for difficult questions. So a number of things to note here that are distinct about the triangulation that I'm talking about here, is that it is a really deliberate choice of methods that complement one another. And as you can see, from the definition here as well, that complements one another in their biases. So you would use one method to alleviate the bias that the first method has, for example. It's also a combination of different types of methods. So in that sense, it is different than, for example, a meta analysis in a review. In a systematic review, one might also combine evidence from different studies, but those generally are the same types of studies. So if there is a certain bias in it, it would not be combining them, would just give you a more precise estimate but not a more unbiased estimates. And most cases of triangulation, the way it is interpreted is that it is a prospective process. So first, you start your study plan. You choose and combine the methods, and then you conduct them alongside one another. But retrospectively, it's also an option and is also being done. So that would mean that people are adhering to these triangulation principles and this philosophy of looking at different research lines by looking at existing evidence. It just means it's a little bit easier or it's more difficult to really integrate the evidence because the studies have already been conducted. But that is something that is being done as well. So for prospective triangulation, which means really combining studies in one project and conducting them and publishing them together, we, on this paper, published a guide recently to give a more pragmatic guideline and a step by step plan of how to do that, because there wasn't much clarity at the moment yet. So I think just very briefly, showing this step by step plan shows you a bit more also about what that is triangulation. So to get the full picture, I would recommend reading this guideline. But very briefly, you would start out with, of course, developing your causal question. And important here is that you are as specific about this as possible. So also state the population, and the context, and all of these things surrounding the causal question that you are after. And then the second step is to draw something called a Directed Acyclic Graph or DAG. It's quite simply drawing out the relevant variables and including arrows towards where the effects are hypothesised to go. And it helps you to clarify what your assumptions are, but also identify the sources of bias that you would need to account for in one way or another. Then the third step is to identify what resources and samples are available to you. So it's of course, also partly a pragmatic process. You might want to have a certain data sets or data collection, but maybe this is not feasible for one reason or another. Or maybe there's no funds for it. So the context within which you are able to conduct your triangulation study or the boundaries that you have would need to be identified. And then within those boundaries, we suggest that people should really aim to select methods that, in combination, are able to mitigate the sources of bias that are the most problematic in a certain research question. And also, choose methods that contrast one another in their strengths and weaknesses. In step five, we included this because one usually starts out with a certain causal question. But depending on the methods and the data that you end up having, the causal questions per method might differ slightly or more than slightly. And this is not necessarily a problem. But if you want to interpret the outcomes, carefully, you will have to also take this into account. So it's a good idea to be specific for every method which causal question you are, in fact answering. Then step six is to make it explicit what you expect the potential biases to be, or in which direction you expect them to work. And finally, also to make it clear beforehand what you expect to see under causality or under no causality. So which differences across the methods which you expect. And then always, it's a good idea to preregister. But definitely for triangulation studies, it's a good idea to pre-register all of this in an extensive plan. So now that we've talked a bit more about triangulation in those principles of combining evidence to tackle difficult questions, I'm going to come back to the topic of this lecture and of the editorial perspective that me, Jentien Vermeulen, and Margot van de Weijer wrote, which is all about smoking, and vaping, and mental health problems. And this is definitely one of those very difficult causal questions, where we have in that same figure that I showed for a hypothetical x and y, in this case, we have the cigarettes smoking as the potential risk factor or exposure, and then mental health outcomes as the outcome variable in question. And well, there's numerous confounding factors that you might be able to think of here. And I included a few. So things like stress, or trauma, or lack of social support, these are all variables that will increase the risk of smoking or smoking more, but also increase the risk of mental health problems. And if we don't account for those, we won't know if there's true causality. And that's why it is so difficult to show this. So what then you could do is to think about those principles of triangulation, and look at the literature, and take different research lines with different types of strengths and weaknesses, and combine those to get to a more reliable answer. So when we look at the observational evidence, this is for smoking. Just to confirm, again, for cigarette smoking, all of these results. We can see from a very large body of literature that there is indeed an association between smoking and mental health problems. And this is a specific nice systematic review that was done a while back, but had included a lot of studies, where more than one third of these studies consistently show that smoking, at baseline, led to more anxiety and depression at a later time point. And these were then, of course, corrected for confounders as much as possible. So that's a part of the observational evidence. But then we know that confounding cannot really be fully adjusted for. There might be measurement error. There might be confounders that were not available in the data set. So this is one part of the story. Then we have another part, which is randomised evidence. So I talked about that a bit in the beginning that it's not always possible to randomise. For example, we would not randomise people to start smoking. But in this very nice and big Cochrane systematic review, about 102 studies were included, and there were 170,000 participants in here in which they looked at studies where people were assigned to quit smoking. And as a secondary aim, they also looked at whether their anxiety and depression changed. And actually here, it was very consistently linked to improvement. So when people quit smoking, anxiety and depression dramatically decreased. So this was, of course, a strong finding. But here, there were some other side notes where there's a lot of heterogeneity. For example, in these effects, and there might be attrition bias. So these are things that the randomised trials cannot really fully answer. So then we bring on in our last piece of evidence, which is the genetic studies, the Mendelian randomization. And this is a systematic review I did myself with my colleagues a couple of years ago in which we reviewed all of the available studies that looked at smoking and any type of mental health outcome. And there, we also showed very consistently, across sensitivity methods, across many different methodologies, that more smoking increases the risk of mental health problems and disorders. So are there causal effects? Well, currently, the consensus on this is yes, for smoking, we know that it increases mental health problems. This is my triangulation of evidence across studies. But there are also studies that triangulate methods within the same study. But what we don't know as well yet is what the precise causal mechanisms are. So this is another part of our editorial, where we go into these pathways. Our assumption here is that what most likely is the case is that if smoking affects mental health problems, it probably affects the brain in one way or another. So in that green part, we can see different types of brain changes that might happen. So brain morphology, connectivity, or function is probably part of the causal pathway. Then there are several hypotheses around how these changes might occur. And the most prominent at the moment are that there is some type of inflammation happening. So smoking increases inflammation and oxidative stress. We know that it does. And we also know that mental health problems can be preceded by these processes. But no-- but we don't know for sure yet if that is the causal chain here. But then other than these biological pathways, there might be other pathways that smoking affects mental health problems, other pathways that are biological or that are not biological. So this is what we know for smoking. But then when we continue to vaping, there's a lot less evidence. So as a short intro on vaping is that overall, when we look at the health effects, it is thought to be less harmful than regular smoking. And that is why for people that do smoke already and wish to quit, it can be seen as a valid smoking cessation tool. And this is mostly because a lot of the carcinogenic chemicals in cigarette smoke are found in very low levels, or not at all in vapour. So it dramatically decreases the risk of developing cancer, which is one of the more lethal effects of smoking. There are, however, well established other health effects of vaping, for example, on the lungs, and respiratory health. And of course, important to mention and clarify that non-smokers should not vape. And then when we go to the mental health effects, it is very unclear at the moment what's the causal nature is. So we know that vaping, like very many risky lifestyle behaviours, is associated with poor mental health. But there is no convincing evidence yet on causality here, and this is because there's very little and limited data available also for quality data. And the risk of confounding and reverse causality is very high, just as it is for smoking. And then of course here, smoking can also be a confounder of the vaping to mental health relationship. So very briefly, just to show that when we look at the evidence now, this figure shows the pathways for smoking to mental health problems. When we look at vaping, there are some indications that it does not increase inflammation and oxidative stress as much as smoking. So probably, the most likely pathways, if there are causal effects of vaping, would be nicotine. Because of course, there is nicotine in vape and quite a lot. So the nicotine binding to nicotinic receptors and potentially thereby changing brain function or connectivity is a hypothesised pathway, where vaping might be able to affect mental health. And there might be other biological or non biological pathways. So I'm coming to towards the ends of this talk. What does this mean? So the evidence now around smoking and vaping, what does that mean for youth and children? Well, first of all, of course, any nicotine use or dependence is unwanted in youth. So we should definitely still push a lot on prevention of both smoking and vaping in children and youth. But it's also important to better understand the causal effects, as well as the mechanisms of smoking and vaping on mental health disorders. And why is this important? Because one might say, well, we already know it's bad for a number of things. Why do we still need to know this? I think, for example, when we want to have more evidence-based expectations of what policies and interventions might lead to. So if there is a policy developed to prevent mental health problems, and part of that is to prevent smoking or vaping in this case, maybe. But in fact, there is no causal effect, then it's important to know what to expect beforehand for those policies. And also prioritisation for future policies. Understanding the mechanisms better. So assuming that there are causal effects. Well, we know that there are effects for smoking if there's also effect for vaping. Knowing better what the mechanisms could also help us to mitigate these effects more. And in general, understanding these very complex questions better will lead to more evidence-based health messages towards the public and individuals. And yeah, just to end up with saying that in this very difficult topic, where we do know a lot already, but we don't yet know everything that we might want to understand about this relationship. There are a couple of things which as the authors of the editorial, think would help us in new studies. So, first of all, having sufficiently powered longitudinal studies that explicitly connects these variables. So we have hypotheses around how smoking and vaping might affect the brain, and the body, and how that might affect mental health. But there hasn't really been any research that explicitly connects these. And given with all of the large scale studies that are out there, this is now becoming a possibility. Besides having appropriate data for all of these variables, applying even more sophisticated analytical methods is always a good idea to more suitably correct for confounding, which is a very big problem of bias here, and to reach reliable answers, which is the general theme of this talk. I think it's very important to strive for converging lines of evidence. So using triangulation within studies is important. Having a few more good triangulation studies out there, but also being more explicit and careful around integrating evidence from different lines that exist already, and using the triangulation principles to do so. I've come at the end of my talk, and then I would just like to finish with thanking my lovely colleagues and collaborators and my funders. [AUDIO LOGO]