Christoph: Hello and welcome to episode five of the Define and Conquer podcast the Special Interest Group on Formalization and Measurement by the European Health Psychology Society recorded ⁓ on the 8th, 2026. am Christoph from the University of Salzburg.
Olga: I'm Olga Persky from Stockholm University in Sweden.
Christoph: And today our guest is Dr. Louise Popper. And yeah, hi Louise. Thank you very much for joining us, for being on the podcast on this nice and sunny day. And yeah, why don't you just start introducing yourself and telling us your origin story.
Louise Poppe: Okay, well first of all thank you for having me. So yes, I'm Louise Poppe. I'm a postdoc ⁓ at University. ⁓ I currently work at the of Public Health and Primary Care. And I have background, ⁓ well I did bachelor and master's in experimental and theoretical psychology. I am here to talk about ⁓ directed acyclic grafts. which I'm really interested in. So I started learning about them already seven years ago. together with two colleagues of mine, ⁓ de Pape and Jelle van Koenberg, ⁓ already gave many workshops on them trying to spread the word.
Olga: Thank you. always like following up on interesting things. I was just wondering, mean, theoretical psychology as a kind of topic for your ⁓ ⁓ sounds very cool. ⁓ you mind telling us a little bit more about what theoretical psychology is?
Louise Poppe: Yes, most people selected clinical psychology, ⁓ but you also had ⁓ experimental and theoretical psychology. We were a very small group. think ⁓ 30 ⁓ were ⁓ called nurse because it's more ⁓ methodological, statistical It's about, for example, language psychology. ⁓ how do when we see some Yeah, just black and white and we can make letters out of it, how it works with the neurons and so on. And also how we like if we feel an emotion, is it first cognitively or is it more emotionally first? And this was all covered in experimental and theoretical psychology. I don't remember that much anymore about what exactly the theories behind it. ⁓ it was a really nice, yeah, were really nice courses. But I remember that like the majority of that group, then did a PhD afterwards, so it was really ⁓ research-focused, the other one was more ⁓ clinically
Christoph: Very interesting. so today our topic are directed acyclic graphs. You already said you got introduced to them seven years ago, but do you remember why ⁓ you got interested and what the first like topical context was?
Louise Poppe: it was on a research meeting and Anik was presenting it. And ⁓ in beginning I did not really see the merit. I was like, okay, this is just something special. It's very niche. ⁓ But after a while... Yeah, I really saw the mirror and realized it. And then I also met Jelle and together we formed a group that was really interested about it. And then we started giving these workshops to convince our colleagues about it.
Olga: So then that takes us to thinking a bit more about core problem. So we try to make inferences about cause and effects and maybe we can't do an experiment and we have observational data. So, yeah, can maybe tell us a bit more about, so what is the core problem here and maybe the DAGs try to address?
Louise Poppe: Yeah, well, if I'm going back to my background, so I did experimental psychology. So we did many, many experiments, like very standardized conditions, like ⁓ temperature was at the same level whole time and so ⁓ on. then in reality, ⁓ often have observational data sets. And then the between ⁓ a causal association a non-causal association gets a bit more vague. Then in observational data when we see a relation between two variables, it can be a causal relation, but there are also many other things that might explain why we're seeing a relation between the two. And in the workshops we always explain it like, for example, imagine that you collect the data among older adults and you find that there is a relation between their level of physical activity and their cognitive performance. might be causal. So there are studies indicating that physical activity can have an impact on your cognitive performance via many pathways, for example via sleep or more social interaction and so But it could also be the other way around. So if you're having cognitive problems, it might be more difficult for you to set up physical, ⁓ or to in ⁓ physical activity, course. And then ⁓ another reason might be that there is confounding so that there is common cause of both variables that has an impact for example ⁓ the people ⁓ more physically active ⁓ having better cognitive performance ⁓ be younger and if you do not control for that then you have confounding ⁓ and a last one and I was not really ⁓ aware of that one and it's called collider bias and ⁓ it's when you select on a consequence of both ⁓ So for example, imagine that you ⁓ decided collect these data in an elderly care facility because you thought like, this will be really easy data collection. ⁓ are older adults there, so I'll just go to an elderly care facility. ⁓ Well, you probably will find a negative relation between physical activity and cognitive performance. ⁓ And because if you have ⁓ someone that elderly care facility ⁓ who ⁓ functioning ⁓ quite ⁓ then will have some physical problems because otherwise these people would not be there. And the other way around. if you have someone who is very physically who can still do a lot of things, then this person is very likely to have ⁓ a cognitive problem ⁓ again, this person would not be there. just illustrating with observational data, it gets tricky if you want to ⁓ make ⁓ causal conclusions that actually helps us to be more aware about it. People sometimes think they can solve it, but it's just to at least become aware about it.
Christoph: Very interesting. mean, observational studies are quite complex in that sense. And before we talk about what directed acyclic graphs or DAGs actually are, I was wondering, are they then not relevant if I'm doing an experimental study? Can I just randomize and then be sure that I can talk about causality or should I also think about this in that context?
Louise Poppe: ⁓ It might be relevant for example selection issues like if you're just in intention to treat, that's randomized. Whether people ⁓ the intervention or not, it's randomized. ⁓ if you're interested in the protocol effect, is ⁓ people actually engaged with the intervention and did it, ⁓ that's ⁓ then things have an impact. For example, my age might have an effect on how I engage with the intervention. And then a DAG ⁓ can relevant, so they can also be ⁓ relevant. for experimental research, but until now I mostly use them for observational research. It's just a way to order your way of thinking as well.
Olga: Yeah, that's very interesting. I also recently came across actually at my university, Stockholm University, how they teach of the ⁓ ⁓ statistics at master's level. And I liked it that they also included kind of, yeah, really thinking about directed acyclic ⁓ graphs for of when experiments go wrong or sort of when people don't necessarily. exactly according to the randomization procedure, for example.
Louise Poppe: it's nice that they included it least, that they are including DAX because it's not happening ⁓ often that I that they include it. But I think it's really good because otherwise you just hear that thing like ⁓ correlation is ⁓ causation ⁓ and then stops. But if the conversation stuff is there, people are still to interpret it. They're just going to use other words like associational or risk factor or whatever.
Olga: Yes, I agree. so, yeah, you kind of gave us some really nice examples here of sort of, well, if you find an association, but then trying to understand kind of, which variables might you need to adjust for, but maybe also some things you shouldn't adjust for, because that could maybe also interrupt kind of some sort of causal interpretation. So was wondering, do you know of any examples of kind of a well-intentioned analysis, but where things were actually made worse because people adjusted for the wrong variables?
Louise Poppe: there is this, there's the well-known birth weight paradox. So let's that... low weight children in groups who show a high prevalence of low birth weight. They typically have lower mortality rates than ⁓ low weight infants in groups a low prevalence of low birth weight. And just to give you like ⁓ an example. when there was a study looking at the impact of maternal smoking, so during pregnancy, and they that smoking had... Yeah, of course it costs higher mortality, infant mortality, except in mothers who... ⁓ babies with a low birth weight. So then you could conclude that smoking be if you have a baby with a low birth weight. And that's of course not correct. ⁓ And the mistake is in that ⁓ they often have this variable birth weight, it's often collected and then they stratify for it but that's not yeah it's not good to do that because it's a mediator so your smoking has an impact on birth weight and birth weight can have an impact on mortality and by correcting for this mediator you can create selection bias and sometimes when I give this example to students because we also included it in the statistics scores here, the dex, they find it weird. But then I tell them, imagine that I give you a baby with low birth weight and how would you perceive this baby's mortality risk? If I tell you like, this is a baby with a low birth weight and it's because his mother smoked versus it might be because of genetic factors, for example. And that's something completely different. So if you have a low birth weight, just because your mother smoked, you're probably better off than when you have a low birth weight because of maybe genetic reasons, for example.
Christoph: Right. So observational research is complicated. Sometimes we should adjust for things, but we shouldn't blindly adjust for things. yeah, quite complicated indeed. And ⁓ guess the solution are directed acyclic graphs, right? That's we're here. But ⁓ before we about how they actually solve this in ⁓ concrete terms, you could very generally explain DAGs are for those who... have never seen one. I mean, it's a little bit difficult to explain a graphical method on a purely audio podcast, but I hope you can paint a picture.
Louise Poppe: Yeah Yeah, I'll do my best. So, Directed Graph is like ⁓ the itself says it's a graph. So, ⁓ and this graph, ⁓ you your... your exposure and your outcome, so your independent and dependent variable, you put arrows between them, so then you draw an arrow from the exposure to the outcome. But you're also going to include other variables that might have an influence on this exposure and this outcome, and ideally you put them ⁓ chronologically, so what happens before can have an impact on what happens later. And by doing so, it becomes clear ⁓ variables you should include. include in the analysis and which ones you should not include. you will see which ones can be confounders, ⁓ be mediators, ⁓ ones can be colliders. And it's really, like as I before, it's not going to suddenly make your analysis. Like ⁓ suddenly you will able to have, you only have the effect. That's not what it's going to do, but you're going ⁓ to be able to a more informed, to make more informed decisions about your analysis and the variables that you include. to conclude.
Christoph: Okay, and you use the term exposure and outcome and also independent and dependent. Is that basically interchangeable or does exposure mean something more specific than independent varying?
Louise Poppe: Yeah, well, in that terminology, we often use the word exposure, but I know like in other domains, it's not often used much. Then we often say like more independent ⁓ variable. So would say like the variable that you're interested in that might have an effect on the other, ⁓ that's exposure independent variable. And then your outcome, yeah, the one that you think will be influenced by the other variable. ⁓ Yeah.
Christoph: And think one ⁓ of DAGs that people always struggle with a little bit at first is this acyclic nature of them. Can explain a bit more what that means?
Louise Poppe: Well, the acyclic nature just means that you cannot change the past. That's a bit what the underlying assumption is. So it means that cannot start in a certain variable and if you then follow all the arrows in the correct direction, you cannot end up in the same variable. And it's because they show this temporality. They show how time passes and which variable has an influence on which one. ⁓ it just shows that something in the future cannot affect something ⁓ the past. And that's an assumption ⁓ we make that I think is okay make. It's not the biggest assumption. But sometimes people say, yes, but I'm interested in bidirectional effects. Then ⁓ how should I it? But if you think about it, this bidirectional is in fact unidirectional if you put it over time. So for example, if you're interested in the effect of self-efficacy on physical activity, which could be bidirectional, then you could say you have self-efficacy at time point 1, which is affecting your physical activity at time point 2, and this physical activity at time point 2 is affecting your self-efficacy at time point 3. So your bidirectional effect, if you're looking at it from a temporal perspective in fact unidirectional.
Christoph: Very interesting. Yeah, that actually is something I was thinking about these days because I'm just working on a deck for the relation between physical activity and depression. there I was also interested in the bi-directional association, but yeah, it makes sense that previous physical activity has an effect on current depression maybe previous depression has an effect on current physical activity. But if I'm depressed in the moment, ⁓ I also move less, right? So it still feels like there's a temporal, like at the same time, effect to me. No?
Louise Poppe: Well, but then you would, I think you have to define it a bit closer what the temporal effect is, then it could be like really, like if it's more like, I have a feeling that's a bit EMA related, maybe. And then it might have an influence on what you're going to do on the next moment. So you can change this temporal order according to what you're interested in.
Christoph: Okay, so if my feeling is that it is at the same time, then actually I just need to change my of time and think a smaller scale of time, basically.
Louise Poppe: That's something that you could do. We wrote a paper about that. I can send it to you if you're interested.
Christoph: Yeah, I mean my context is actually also epidemiological at the moment, so it's just a two-wave study. ⁓ So I can't really change the time scale there, right?
Louise Poppe: Yes? No, no. That's not really a problem, it's just about making this assumption and how you're drawing the deck is showing what assumption you are making. If you think that, for example, that is happening at the same time and they don't have, for example, an impact on each other, or it's only one way, you can ⁓ it like that in ⁓ deck, but this arrow is kind of showing that there is an ⁓ over how short this period of time can be. ⁓
Christoph: I think ⁓ further important element of DAX that we didn't cover yet ⁓ unobserved variables, right? Because in this context, if I have two waves, then I'm kind of missing what's ⁓ happening right? To then explain this seemingly relationship the time, right? So how can I ⁓ incorporate things that ⁓ didn't measure, but that are kind of necessary to fully draw the causal relationship in a directed acyclic graph.
Louise Poppe: Yeah, so in your deck you would... put all the relevant variables in there, whether we have them or not. ⁓ you can see which variables that you ⁓ ⁓ have information on. But itself, it's already important to provide that information. So when people read your study, that they can learn from it and say, OK, these are indeed important variables. We should collect information that. then there are to, if you do not information on these variables, ⁓ there are ways to and to see how much bias these variables would ⁓ give it would change your estimates. So that's that you can do and if they are more ⁓ like distal it might have ⁓ less of impact but if they're ⁓ proximal it be more But it's about ⁓ transparent. ⁓ I think that's really ⁓ merit of DAX, is that we are transparent about the things that we are missing.
Christoph: Yeah, absolutely. I think that's one of the main advantages, but also maybe one of the main challenges of it. just you to be honest about what you have and what you don't have.
Louise Poppe: Yes, I already had a study that we did not conduct in the end because we noticed that it would just not solve the question by the deck. And I still see it as one of my major accomplishments in science is that I did not spend on a study that would not be able ⁓ to the question in the end. So ⁓ I that's a good outcome. And we it down and what you learned about it.
Christoph: Can you give this example in detail? It sounds quite interesting.
Louise Poppe: It was an EMA study and that we were looking at the impact of ⁓ activity ⁓ cognitive performance. if you're more physically active in the next hours, you going to... ⁓ is your cognitive performance going to be better ⁓ before? ⁓ then I sat together with ⁓ experts then we were looking at all the things that could have an impact on both your physical activity at the moment and your cognitive performance. And there are just many variables that can ⁓ have impact on both. Like, for example, if you're feeling unwell. It was in all the adults it was also medication use that could have an impact their emotions and and we realized that there were just too many variables that we would have to measure and and ⁓ even we might not capture might not be able to to capture and everything ⁓ and we realized like this is this is not going to answer the question and while it's even an EMA study you really have this temporal ordning like you're assessing it right before ⁓ the ⁓ but that's one piece of the puzzle it helps but it does not you to make causal conclusions and ⁓ we a lot ⁓ about and then you have a study or there is a paper ⁓ about you can solve it by using a quasi experimental design and this like an encouragement design and then people are encouraged to be physically active on random ⁓ moments and they these these prompts and and that would be a nice because then you won't have the impact of these time-varying confounders. So we wrote it out in the paper and gave some explanations. Yeah, learned a lot about it ⁓ by it.
Olga: sounds like a very nice way of really thinking deeply about of time and ⁓ all of these complex But then also nice that you wrote what you learned from it. Maybe we can link to it in the show notes as well, would be very nice. And I that also nicely takes us to ⁓ another one our questions. ⁓
Louise Poppe: Yeah, it'd be nice.
Olga: Which is, because so it sounds like when, you're sitting down to develop a DAG, it quite similar to maybe thinking about a conceptual framework or, well, maybe a structural equation model or path model. So can you maybe help us ⁓ navigate a bit So is it, ⁓ is a big between a DAG and the conceptual framework, for example?
Louise Poppe: Yes. Yeah, so they're both models and they help you to think about the order behind the reality of what we see. a theoretical model, I'm thinking about the behavior change model now, is just giving you, is helping you explain reality. So you make some assumptions like if we reach these determinants and then the behavior will show for example. While a DAQ ⁓ that you create, like you're not going to create a DAQ because that be the output. Like it never ends with a DAQ. The DAQ is always an input ⁓ for specific ⁓ study for its specific data analysis. And it ⁓ focuses on one variable ⁓ another variable. ⁓ in ⁓ conceptual model, you will have many variables that you could target and then to see the effect. ⁓ But example, if you would have the health action process approach, you might be interested in the effect of ⁓ outcome expectancies. then you would draw a diagram between outcome expectancies and your behavior of interest ⁓ and add variables that could also have an effect on its confounders and so So you could start from a theoretical model, but then if you're really interested in one ⁓ aspect, like is the causal effect between ⁓ variables of it, then you would have to draw a DAQ to design your study and to do the data analysis. And related to ⁓ structural ⁓ modeling, DAQs sometimes described as non-parametric sums because we do not make any ⁓ parametric assumptions ⁓ in DAQs, so not going to say, for example, if the... relationship is nonlinear for example. But structural equation modeling really belongs to the ⁓ phase, ⁓ a DAG belongs to the identification phase in which you think like can I really identify the causal effects ⁓ ⁓ DAG can be used for many different estimation methods and they're really inputs ⁓ and structural equation model ⁓ sits this estimation phase ⁓ and they be used for causal and non-causal questions. there are mainly outputs, right? So you're not going to create one to guide your analysis. They kind of are the analysis.
Olga: Yeah, so thank you so much for explaining that. That really sort of makes a nice distinction between kind of the DAG and some of the statistical estimation methods. But I was also wondering, ⁓ mean, linked to this. So ⁓ there a big difference then also between a DAG and a theory or maybe also an ontology? So we had a previous episode with Polina Schenk where we... talked a lot about ontologies.
Louise Poppe: ⁓ Well, it's good to have a clear theory in mind before you draw your deck, that always helps. But an adag still differs from an ontology because in an ontology you... I'm not an ontology expert, want to say that first. But I believe that in an ontology you have relationship between your variables that are not necessarily causal. You could say it is a part of and so on. So it helps you to gather a better understanding of a domain and what it entails. I think it can also be an output that you create an ontology and then people ⁓ use it. ⁓ Again, ⁓ a is more when you would focus on one variable on the next. It's input, it is something that you create for design and analysis. ⁓ they all look, yeah, of course, graphed and they all look very similar. You ⁓ get confused in the end, but really have, yeah, ⁓ there's a reason for using them.
Olga: said that it's ⁓ to have a theory in mind when developing a DAG, but do you then mean that a theory can help you think about ⁓ or sort of exposures that are linked to outcomes? ⁓ did you mean? To theory to inform your DAG?
Louise Poppe: Well, I- For example, of the topics that I'm working on now is the impact of green exposure on health outcomes. There is theory about how green space ⁓ impacts your health. It says it goes via air pollution, encouraging physical activity, encouraging social interactions and ⁓ stress reduction and on. This theory helps me to know I should not include, for example, air pollution. confounder because it's a way of how green space impacts my health so if I would include it as a confounder that would be wrong then and of course it all depends on your research question but if you make like if you think like we're starting from this theoretical model and we believe this is correct then it kind of helps you to build your your deck.
Christoph: Okay, yeah, I think that already part of the next question, which would be how do I actually develop a DAG, right? So, theories can inform our DAGs, but yeah, what other sources of knowledge can I use? And is there a specific process, like ⁓ steps ⁓ arrive at your DAG that you can recommend?
Louise Poppe: Well, I had this question about, like, I think... I wanted to create my own DAG and I wrote it in the proposal and it would take me three months I assumed. took way longer and had this huge imposter syndrome because I had to create a DAG and I wanted to be good ⁓ and were no really guidelines in literature so you want to do it good but there were no guidelines and that's when we did our scoping review and ⁓ we kind of all these guidelines and talked about it a lot with all of these experts and then they wrote their own ⁓ Something really helps is just first like into the literature. it's good to have a good idea of how one variable impacts the what other variables can have an impact on that. And like also other studies that are examining the effect, they're going to show for which variables they controlled ⁓ and then you kind of have idea of what their causal model was, how it the variables they thought were confounders and by drawing a you can see whether it actually makes sense ⁓ not because sometimes ⁓ are just added because people sometimes have the idea like the more you control ⁓ things the better. So literature really helps. ⁓ there is a paper by Ferguson and colleagues which helps you ⁓ a DAQ based on literature ⁓ it's quite labor-intensive I must say. Then to To experts and your colleagues, I have never been able to show a DAG without creating a big discussion it with people. you say this is my analysis plan, often have that people are like, yeah, that's OK, that's OK. But if you show a DAG, ⁓ then discussion really starts. And that helps. And also, can with in the field. And we also ⁓ now have a study the impact of depression. ⁓ symptoms on post-surgical neuropathic pain among women with breast cancer and then interviews were also conducted with a ⁓ who had ⁓ cancer to what variables she might be important and also with ⁓ an oncologist. it's like drawing a deck be quite a lot of work but it also helps you to get really a better idea about your exact research question. I so it's really not lost time and ⁓ in end it will save you a lot of time for designing your study and doing the data ⁓
Christoph: Yeah. And to me, it feels like this should be things you should be doing anyways, right? Like you should look at the literature, you should talk to your colleagues before you run a study, but maybe a is just a more formal way to do this. And it can help you like see for yourself whether you already did enough of that. Right? So if you don't know whether you should add a path there or not, then maybe you should read a little bit more. And if there are still a lot of discussions in the team, which I often have, if I bring a DAG to our team meetings.
Louise Poppe: Thank you.
Christoph: And I also learned, okay, maybe this isn't as clear as I thought at first, but a is a great way to visualize that, right? And to have something to talk about. Okay. So those are kind of the sources of knowledge that you can use. But would you say like a systematic literature review, is that necessary? Or does that depend on like how great of a DAG I want it to be? A minimal DAG, is that also possible with just the usual literature search?
Louise Poppe: Yes. Yeah, I think it also depends on how well you know the literature. If you have been in it for several years, kind of... know the literature. According to the paper of Ferguson and Collex, you would have to do a systematic literature review. I have done a scoping review for another project and then for even another project we just looked at ⁓ variables that we thought were most relevant based on the expertise that we have and based on the expertise ⁓ our colleagues and other experts in the field. ⁓ don't think that you always need to do a systematic but I would just encourage you to discuss it with as many people as you can and then you will see that other things really come up.
Christoph: Yeah, and maybe AI tools can actually speed up this process or linking back to ontologies. there is a well formulated ontology, maybe you can more easily screen the literature to ⁓ what's there already. Right. So now we have the knowledge base ⁓ for DAX, what are like the five steps if there are five to construct them?
Louise Poppe: I believe in the paper that we wrote that the first one was to start on time. Like, don't say like right before the analysis like I'm going to draw a quick deck. because that's not really going to work. Just try to do it as early as possible. then you should clearly define exposure and your outcome and the more ⁓ precise the And then it would be good add common causes of the exposure and the outcome. and then to see which arrows could be left out because leaving one out is always a bigger assumption than leaving one in. So if you're not sure, you just leave the arrow in and otherwise it's quite a big assumption to leave one out. So I kind of prefer to like if I have a common cause I will just connect it with everything because that's like that I'm not making any assumptions about it, it cannot have any effect. And then you could add mediators, it's not really necessary because you will not have to ⁓ control for them if you're interested in the total effect. But it can help you like this. you see a variable that you can say, ⁓ no, no, it's really a mediator, then you can add it, of course, to the deck just for clarity. You add selection notes based on the design of your study. you even ⁓ yeah, you represent measurement bias in it, but that's like expert level. If you also ⁓ want to do that so it's something that you consider. It is really relevant of course but I think if you have not worked with it before it might be interesting to start with the bias and selection bias and try to incorporate that in
Christoph: there are layers to it. And I think to understand why you only need to include common causes and maybe not that just causes the exposure or that is downstream consequence of ⁓ the of interest, think it's important to know how you work with a DAG, right? Because if I'm interested in one causal connection, ⁓ in causal path, and I have ⁓ a cost that only like another variable that only affects the exposure, then don't need to condition on it because ⁓ no connection to the outcome. So ⁓ think at least for me with limited knowledge of DAX, it's always easiest to just look at the backdoor criterion, Which is just, or maybe you want to explain it better. Yeah. I think that's, that's just the easiest way to know what to ⁓ and what to condition on. Yeah.
Louise Poppe: Yeah, yeah. Yes, yes. Yeah, so if you can go from your exposure to your outcome via a common cause, you have confounding bias, it's a backdoor criterion, ⁓ you should control for that variable. And in theory, yes, you can have a variable that only affects your exposure, but it usually takes place the past. And then from practical... Yeah, just like... I would always like draw a arrow anyway because I think it might have an impact. I cannot think of variables that really would not have an impact on my outcome but maybe there are ⁓ could of course be.
Christoph: So this backdoor criterion is maybe the main or tool for it. But after I drew my DAG, there other tools I can use to work with it and to get answers from it?
Louise Poppe: Well, do you mean like software that you could use? ⁓
Christoph: For example, I mean Pearl, right? He did a lot ⁓ decks ⁓ there's this term, ⁓ ⁓ something like this, which I never fully understood. There's some ⁓ things around decks as well, which we don't need to cover in detail, but what are the ⁓ useful tools with a deck?
Louise Poppe: Yes, yes. Yeah, so You need to understand the rules a bit. You can control for certain variables. Controlling means that you adjust for them in the analysis or you use stratification. When you control for a confounder, it of stops the... You could think about it as electric current going between variables. So if you control for a confounder, you make sure that this electric current is no longer going through your confounding variable. And then, this non-spurious association between your exposure ⁓ your outcome gets gets controlled for, statistically controlled for. It's really difficult to explain it in a podcast, I must say. But, and ⁓ you have collider is when you have a variable in which ⁓ two arrows collide, come together. ⁓ I'm showing it my hands, but that's really not relevant now. And then, will stop the current unless control for it. So it's not good to control for a collider. And that's what happened. ⁓ That's happened. Like, for example, if you're interested in the effect of physical activity and cognitive performance ⁓ and recruit in an elderly care facility, because ⁓ those variables, physical activity, cognitive performance, ⁓ both an impact on ⁓ whether not you're in the elderly care facility and you for it by only recruiting people from the elderly care facility. ⁓ and that is then collider bias. So you have these different rules that you can follow and in the beginning it might be a bit difficult. I also found it a bit strange to do but after a while it becomes a second nature to interpret them.
Christoph: Okay, so.
Louise Poppe: like software like Daggerty ⁓ just show you when there is a problem like you can say like adjust for these and these variables ⁓ then they will indicate if there is a problem or or if you should adjust for more variables ⁓ to the non-causal ⁓ links non-causal paths so that help ⁓
Christoph: Yeah, but it is easier if you can just follow the connections visually, right?
Louise Poppe: Yes, yes, I prefer to just write them on paper but yeah you can also do it via taggity or what so ever.
Olga: So it sounds a bit like useful to be aware of some of these common ⁓ and sort of how to, yeah, ⁓ it looks visually maybe before you get started. that sort of that you need some like pattern recognition?
Louise Poppe: Yeah. No, it's not really pattern recognition, it's just following a set of rules. There's just a bit of learning about it. In the beginning it might be a bit complex, but after a while... ⁓ Yeah, you see when there is a problem and when there's no problem. like confounding is quite easy to see when there is a problem. For collider bias, have to make sure that you're not conditioning on things happening later. also by drawing them chronologically, ⁓ gets clear which variables might be problematic include in the analysis or which ones you should include. Yeah, it's just as I was saying before it just helps you to visualize your assumptions and people sometimes say like, yeah, it's a bit too much work I'm not going to draw a deck but then I think yeah, okay, but you're still going to make these assumptions You're just not going to be transparent about them But it's not like if you're not drawing a deck that you're not making any assumptions. It's just You just keep them for yourself
Olga: Yeah, so it definitely a useful tool for kind of transparency and also, yeah, making sure that your analysis when ours, well, good they can be for estimating causal effect. So it also sounds like, I mean, for a novice in this space, there's quite a few kind of things to get your head around. Can you think of the most home and beginner mistakes or maybe when developing a DAG for the first time.
Louise Poppe: I a really one is that before you create deck you need to have a really specific idea about what research question is. that I sometimes notice that it's not really clear, like people are drawing a DAG but don't really know which categories they would like to compare. And one of the assumptions in causal inference is also this consistency assumption. for example, if you would have... I want to the effect of on depression, for example. ⁓ the of BMI on depression... That's not really a good research question. Because you should focus on what could be an intervention, for example. Because if you change your BMI via diet or physical activity or like chopping off your arm, it will have completely different effects on your depression outcome. And that's the whole consistency assumption. So there should not be different... varieties of your treatment. there should only be like one way to go from one to the other. And ⁓ also think that this is sometimes a bit difficult in psychology. Like for example, if you think about determinants, ⁓ are many ways that you could change them. ⁓ yeah, also, ⁓ it's something that you need to be aware of. So ⁓ first, have a really clear research question. I think that's the first one. ⁓ A really paper about this is What is your Estimand by ⁓ Lundberg and colleagues I can also ⁓ send it to you ⁓ second one is that people like there to be more people having ⁓ imposter syndrome their deck and they ⁓ a question often is like like when should I stop like I added so many variables like when does it stop and are just ⁓ I mean they are simplification of and and After a while you should just be able to say like, I really think I included the most relevant ones. I discussed it with my colleagues. I discussed it with experts. And this is where it stops. But then at least you're again transparent to reviewers and so on. But you cannot include everything. mean, then it gets, if you include the most proximal ones and the ones that you believe are important. But not the ones that you happen to have in your data set because that's also something. that we see, you should still include the ones that you also did not measure. ⁓
Olga: So then it sounds like, there is a different topic ⁓ well, but maybe also how to set up your analysis plans for kind of secondary data analysis before you've looked at the data or kind of, so it sounds like even if you know ⁓ you're going to use a secondary data set and not collect primary data to maybe develop the DAG first ⁓ you have a detailed look.
Louise Poppe: Yes, that would be really a good idea and then you can also have an idea like... of whether you should maybe merge different datasets to get all the information you need. So always better because it might be tempting to have look at the dataset and see which variables that you have and then do not think further of any other variables that might have an influence. it's, yeah, it's, ⁓ I mean, belongs to the design phase and you should really see it as like specifying, conceiving your question.
Olga: So yeah, that's really interesting and also makes us think about who can use DAGs or can use DAGs to inform their work or are they mainly used in research?
Louise Poppe: Yeah, I think it's more research related. Of course you can draw a causal model of how you think ⁓ will have an impact on people, yeah, it's mainly to inform your design and your data analysis. it's really research ⁓ As I said, they are inputs, they're not meant to be outputs. So yeah, you create them for designing studies and analyzing your data.
Christoph: But in any case, it's good to be more transparent, right? Also for practitioners, so for us as researchers to be more transparent also helps practitioners and the public, think. I think that brings us kind of to the end. Are there any ⁓ tools or resources you still want to recommend to the listeners?
Louise Poppe: Yes, yeah, yeah, yeah, yeah, indeed. I talked about Degety which can help you draw a deck and see which you should control for. There is the free online book of What If by Hernan. And there's also an edX course ⁓ by Hernan. ⁓ which is also really good. It's just, yeah, you see these video clips and it shows you how many minutes and then you get questions and so on afterwards. So we usually ask PhD students that are just starting that they start just ⁓ going the course. And there's also... one by Peter Tennant about the use of tags and I also think this is really accessible because sometimes you can have quite difficult papers ⁓ the causal inference literature ⁓ yeah you have to look for a bit for the ones that are a bit more applied. And I also think, it's a bit self-promotion of course, but I also think that with our scoping review we kind of try to put the basis and explain the basis and how you can start with it. you also have the papers by Julienne Rode, ⁓ also provides a really nice introduction. I can also ⁓ send link to that ⁓ paper and I think it might really help.
Christoph: Yeah, your scoping review is a great resource. I read that in preparation for this episode and I definitely learned a lot that I can just apply right away to my own practice. So I would highly recommend that. But we will put all the resources in the show notes. Seems like there's a lot out there, which is great. Right. Do you have anything else you want to say?
Louise Poppe: And no, thank you for inviting me.
Christoph: Thank you for coming on the show and enjoy the sunny afternoon.
Louise Poppe: Thanks.
Olga: I'm Olga Persky from Stockholm University in Sweden.
Christoph: And today our guest is Dr. Louise Popper. And yeah, hi Louise. Thank you very much for joining us, for being on the podcast on this nice and sunny day. And yeah, why don't you just start introducing yourself and telling us your origin story.
Louise Poppe: Okay, well first of all thank you for having me. So yes, I'm Louise Poppe. I'm a postdoc ⁓ at University. ⁓ I currently work at the of Public Health and Primary Care. And I have background, ⁓ well I did bachelor and master's in experimental and theoretical psychology. I am here to talk about ⁓ directed acyclic grafts. which I'm really interested in. So I started learning about them already seven years ago. together with two colleagues of mine, ⁓ de Pape and Jelle van Koenberg, ⁓ already gave many workshops on them trying to spread the word.
Olga: Thank you. always like following up on interesting things. I was just wondering, mean, theoretical psychology as a kind of topic for your ⁓ ⁓ sounds very cool. ⁓ you mind telling us a little bit more about what theoretical psychology is?
Louise Poppe: Yes, most people selected clinical psychology, ⁓ but you also had ⁓ experimental and theoretical psychology. We were a very small group. think ⁓ 30 ⁓ were ⁓ called nurse because it's more ⁓ methodological, statistical It's about, for example, language psychology. ⁓ how do when we see some Yeah, just black and white and we can make letters out of it, how it works with the neurons and so on. And also how we like if we feel an emotion, is it first cognitively or is it more emotionally first? And this was all covered in experimental and theoretical psychology. I don't remember that much anymore about what exactly the theories behind it. ⁓ it was a really nice, yeah, were really nice courses. But I remember that like the majority of that group, then did a PhD afterwards, so it was really ⁓ research-focused, the other one was more ⁓ clinically
Christoph: Very interesting. so today our topic are directed acyclic graphs. You already said you got introduced to them seven years ago, but do you remember why ⁓ you got interested and what the first like topical context was?
Louise Poppe: it was on a research meeting and Anik was presenting it. And ⁓ in beginning I did not really see the merit. I was like, okay, this is just something special. It's very niche. ⁓ But after a while... Yeah, I really saw the mirror and realized it. And then I also met Jelle and together we formed a group that was really interested about it. And then we started giving these workshops to convince our colleagues about it.
Olga: So then that takes us to thinking a bit more about core problem. So we try to make inferences about cause and effects and maybe we can't do an experiment and we have observational data. So, yeah, can maybe tell us a bit more about, so what is the core problem here and maybe the DAGs try to address?
Louise Poppe: Yeah, well, if I'm going back to my background, so I did experimental psychology. So we did many, many experiments, like very standardized conditions, like ⁓ temperature was at the same level whole time and so ⁓ on. then in reality, ⁓ often have observational data sets. And then the between ⁓ a causal association a non-causal association gets a bit more vague. Then in observational data when we see a relation between two variables, it can be a causal relation, but there are also many other things that might explain why we're seeing a relation between the two. And in the workshops we always explain it like, for example, imagine that you collect the data among older adults and you find that there is a relation between their level of physical activity and their cognitive performance. might be causal. So there are studies indicating that physical activity can have an impact on your cognitive performance via many pathways, for example via sleep or more social interaction and so But it could also be the other way around. So if you're having cognitive problems, it might be more difficult for you to set up physical, ⁓ or to in ⁓ physical activity, course. And then ⁓ another reason might be that there is confounding so that there is common cause of both variables that has an impact for example ⁓ the people ⁓ more physically active ⁓ having better cognitive performance ⁓ be younger and if you do not control for that then you have confounding ⁓ and a last one and I was not really ⁓ aware of that one and it's called collider bias and ⁓ it's when you select on a consequence of both ⁓ So for example, imagine that you ⁓ decided collect these data in an elderly care facility because you thought like, this will be really easy data collection. ⁓ are older adults there, so I'll just go to an elderly care facility. ⁓ Well, you probably will find a negative relation between physical activity and cognitive performance. ⁓ And because if you have ⁓ someone that elderly care facility ⁓ who ⁓ functioning ⁓ quite ⁓ then will have some physical problems because otherwise these people would not be there. And the other way around. if you have someone who is very physically who can still do a lot of things, then this person is very likely to have ⁓ a cognitive problem ⁓ again, this person would not be there. just illustrating with observational data, it gets tricky if you want to ⁓ make ⁓ causal conclusions that actually helps us to be more aware about it. People sometimes think they can solve it, but it's just to at least become aware about it.
Christoph: Very interesting. mean, observational studies are quite complex in that sense. And before we talk about what directed acyclic graphs or DAGs actually are, I was wondering, are they then not relevant if I'm doing an experimental study? Can I just randomize and then be sure that I can talk about causality or should I also think about this in that context?
Louise Poppe: ⁓ It might be relevant for example selection issues like if you're just in intention to treat, that's randomized. Whether people ⁓ the intervention or not, it's randomized. ⁓ if you're interested in the protocol effect, is ⁓ people actually engaged with the intervention and did it, ⁓ that's ⁓ then things have an impact. For example, my age might have an effect on how I engage with the intervention. And then a DAG ⁓ can relevant, so they can also be ⁓ relevant. for experimental research, but until now I mostly use them for observational research. It's just a way to order your way of thinking as well.
Olga: Yeah, that's very interesting. I also recently came across actually at my university, Stockholm University, how they teach of the ⁓ ⁓ statistics at master's level. And I liked it that they also included kind of, yeah, really thinking about directed acyclic ⁓ graphs for of when experiments go wrong or sort of when people don't necessarily. exactly according to the randomization procedure, for example.
Louise Poppe: it's nice that they included it least, that they are including DAX because it's not happening ⁓ often that I that they include it. But I think it's really good because otherwise you just hear that thing like ⁓ correlation is ⁓ causation ⁓ and then stops. But if the conversation stuff is there, people are still to interpret it. They're just going to use other words like associational or risk factor or whatever.
Olga: Yes, I agree. so, yeah, you kind of gave us some really nice examples here of sort of, well, if you find an association, but then trying to understand kind of, which variables might you need to adjust for, but maybe also some things you shouldn't adjust for, because that could maybe also interrupt kind of some sort of causal interpretation. So was wondering, do you know of any examples of kind of a well-intentioned analysis, but where things were actually made worse because people adjusted for the wrong variables?
Louise Poppe: there is this, there's the well-known birth weight paradox. So let's that... low weight children in groups who show a high prevalence of low birth weight. They typically have lower mortality rates than ⁓ low weight infants in groups a low prevalence of low birth weight. And just to give you like ⁓ an example. when there was a study looking at the impact of maternal smoking, so during pregnancy, and they that smoking had... Yeah, of course it costs higher mortality, infant mortality, except in mothers who... ⁓ babies with a low birth weight. So then you could conclude that smoking be if you have a baby with a low birth weight. And that's of course not correct. ⁓ And the mistake is in that ⁓ they often have this variable birth weight, it's often collected and then they stratify for it but that's not yeah it's not good to do that because it's a mediator so your smoking has an impact on birth weight and birth weight can have an impact on mortality and by correcting for this mediator you can create selection bias and sometimes when I give this example to students because we also included it in the statistics scores here, the dex, they find it weird. But then I tell them, imagine that I give you a baby with low birth weight and how would you perceive this baby's mortality risk? If I tell you like, this is a baby with a low birth weight and it's because his mother smoked versus it might be because of genetic factors, for example. And that's something completely different. So if you have a low birth weight, just because your mother smoked, you're probably better off than when you have a low birth weight because of maybe genetic reasons, for example.
Christoph: Right. So observational research is complicated. Sometimes we should adjust for things, but we shouldn't blindly adjust for things. yeah, quite complicated indeed. And ⁓ guess the solution are directed acyclic graphs, right? That's we're here. But ⁓ before we about how they actually solve this in ⁓ concrete terms, you could very generally explain DAGs are for those who... have never seen one. I mean, it's a little bit difficult to explain a graphical method on a purely audio podcast, but I hope you can paint a picture.
Louise Poppe: Yeah Yeah, I'll do my best. So, Directed Graph is like ⁓ the itself says it's a graph. So, ⁓ and this graph, ⁓ you your... your exposure and your outcome, so your independent and dependent variable, you put arrows between them, so then you draw an arrow from the exposure to the outcome. But you're also going to include other variables that might have an influence on this exposure and this outcome, and ideally you put them ⁓ chronologically, so what happens before can have an impact on what happens later. And by doing so, it becomes clear ⁓ variables you should include. include in the analysis and which ones you should not include. you will see which ones can be confounders, ⁓ be mediators, ⁓ ones can be colliders. And it's really, like as I before, it's not going to suddenly make your analysis. Like ⁓ suddenly you will able to have, you only have the effect. That's not what it's going to do, but you're going ⁓ to be able to a more informed, to make more informed decisions about your analysis and the variables that you include. to conclude.
Christoph: Okay, and you use the term exposure and outcome and also independent and dependent. Is that basically interchangeable or does exposure mean something more specific than independent varying?
Louise Poppe: Yeah, well, in that terminology, we often use the word exposure, but I know like in other domains, it's not often used much. Then we often say like more independent ⁓ variable. So would say like the variable that you're interested in that might have an effect on the other, ⁓ that's exposure independent variable. And then your outcome, yeah, the one that you think will be influenced by the other variable. ⁓ Yeah.
Christoph: And think one ⁓ of DAGs that people always struggle with a little bit at first is this acyclic nature of them. Can explain a bit more what that means?
Louise Poppe: Well, the acyclic nature just means that you cannot change the past. That's a bit what the underlying assumption is. So it means that cannot start in a certain variable and if you then follow all the arrows in the correct direction, you cannot end up in the same variable. And it's because they show this temporality. They show how time passes and which variable has an influence on which one. ⁓ it just shows that something in the future cannot affect something ⁓ the past. And that's an assumption ⁓ we make that I think is okay make. It's not the biggest assumption. But sometimes people say, yes, but I'm interested in bidirectional effects. Then ⁓ how should I it? But if you think about it, this bidirectional is in fact unidirectional if you put it over time. So for example, if you're interested in the effect of self-efficacy on physical activity, which could be bidirectional, then you could say you have self-efficacy at time point 1, which is affecting your physical activity at time point 2, and this physical activity at time point 2 is affecting your self-efficacy at time point 3. So your bidirectional effect, if you're looking at it from a temporal perspective in fact unidirectional.
Christoph: Very interesting. Yeah, that actually is something I was thinking about these days because I'm just working on a deck for the relation between physical activity and depression. there I was also interested in the bi-directional association, but yeah, it makes sense that previous physical activity has an effect on current depression maybe previous depression has an effect on current physical activity. But if I'm depressed in the moment, ⁓ I also move less, right? So it still feels like there's a temporal, like at the same time, effect to me. No?
Louise Poppe: Well, but then you would, I think you have to define it a bit closer what the temporal effect is, then it could be like really, like if it's more like, I have a feeling that's a bit EMA related, maybe. And then it might have an influence on what you're going to do on the next moment. So you can change this temporal order according to what you're interested in.
Christoph: Okay, so if my feeling is that it is at the same time, then actually I just need to change my of time and think a smaller scale of time, basically.
Louise Poppe: That's something that you could do. We wrote a paper about that. I can send it to you if you're interested.
Christoph: Yeah, I mean my context is actually also epidemiological at the moment, so it's just a two-wave study. ⁓ So I can't really change the time scale there, right?
Louise Poppe: Yes? No, no. That's not really a problem, it's just about making this assumption and how you're drawing the deck is showing what assumption you are making. If you think that, for example, that is happening at the same time and they don't have, for example, an impact on each other, or it's only one way, you can ⁓ it like that in ⁓ deck, but this arrow is kind of showing that there is an ⁓ over how short this period of time can be. ⁓
Christoph: I think ⁓ further important element of DAX that we didn't cover yet ⁓ unobserved variables, right? Because in this context, if I have two waves, then I'm kind of missing what's ⁓ happening right? To then explain this seemingly relationship the time, right? So how can I ⁓ incorporate things that ⁓ didn't measure, but that are kind of necessary to fully draw the causal relationship in a directed acyclic graph.
Louise Poppe: Yeah, so in your deck you would... put all the relevant variables in there, whether we have them or not. ⁓ you can see which variables that you ⁓ ⁓ have information on. But itself, it's already important to provide that information. So when people read your study, that they can learn from it and say, OK, these are indeed important variables. We should collect information that. then there are to, if you do not information on these variables, ⁓ there are ways to and to see how much bias these variables would ⁓ give it would change your estimates. So that's that you can do and if they are more ⁓ like distal it might have ⁓ less of impact but if they're ⁓ proximal it be more But it's about ⁓ transparent. ⁓ I think that's really ⁓ merit of DAX, is that we are transparent about the things that we are missing.
Christoph: Yeah, absolutely. I think that's one of the main advantages, but also maybe one of the main challenges of it. just you to be honest about what you have and what you don't have.
Louise Poppe: Yes, I already had a study that we did not conduct in the end because we noticed that it would just not solve the question by the deck. And I still see it as one of my major accomplishments in science is that I did not spend on a study that would not be able ⁓ to the question in the end. So ⁓ I that's a good outcome. And we it down and what you learned about it.
Christoph: Can you give this example in detail? It sounds quite interesting.
Louise Poppe: It was an EMA study and that we were looking at the impact of ⁓ activity ⁓ cognitive performance. if you're more physically active in the next hours, you going to... ⁓ is your cognitive performance going to be better ⁓ before? ⁓ then I sat together with ⁓ experts then we were looking at all the things that could have an impact on both your physical activity at the moment and your cognitive performance. And there are just many variables that can ⁓ have impact on both. Like, for example, if you're feeling unwell. It was in all the adults it was also medication use that could have an impact their emotions and and we realized that there were just too many variables that we would have to measure and and ⁓ even we might not capture might not be able to to capture and everything ⁓ and we realized like this is this is not going to answer the question and while it's even an EMA study you really have this temporal ordning like you're assessing it right before ⁓ the ⁓ but that's one piece of the puzzle it helps but it does not you to make causal conclusions and ⁓ we a lot ⁓ about and then you have a study or there is a paper ⁓ about you can solve it by using a quasi experimental design and this like an encouragement design and then people are encouraged to be physically active on random ⁓ moments and they these these prompts and and that would be a nice because then you won't have the impact of these time-varying confounders. So we wrote it out in the paper and gave some explanations. Yeah, learned a lot about it ⁓ by it.
Olga: sounds like a very nice way of really thinking deeply about of time and ⁓ all of these complex But then also nice that you wrote what you learned from it. Maybe we can link to it in the show notes as well, would be very nice. And I that also nicely takes us to ⁓ another one our questions. ⁓
Louise Poppe: Yeah, it'd be nice.
Olga: Which is, because so it sounds like when, you're sitting down to develop a DAG, it quite similar to maybe thinking about a conceptual framework or, well, maybe a structural equation model or path model. So can you maybe help us ⁓ navigate a bit So is it, ⁓ is a big between a DAG and the conceptual framework, for example?
Louise Poppe: Yes. Yeah, so they're both models and they help you to think about the order behind the reality of what we see. a theoretical model, I'm thinking about the behavior change model now, is just giving you, is helping you explain reality. So you make some assumptions like if we reach these determinants and then the behavior will show for example. While a DAQ ⁓ that you create, like you're not going to create a DAQ because that be the output. Like it never ends with a DAQ. The DAQ is always an input ⁓ for specific ⁓ study for its specific data analysis. And it ⁓ focuses on one variable ⁓ another variable. ⁓ in ⁓ conceptual model, you will have many variables that you could target and then to see the effect. ⁓ But example, if you would have the health action process approach, you might be interested in the effect of ⁓ outcome expectancies. then you would draw a diagram between outcome expectancies and your behavior of interest ⁓ and add variables that could also have an effect on its confounders and so So you could start from a theoretical model, but then if you're really interested in one ⁓ aspect, like is the causal effect between ⁓ variables of it, then you would have to draw a DAQ to design your study and to do the data analysis. And related to ⁓ structural ⁓ modeling, DAQs sometimes described as non-parametric sums because we do not make any ⁓ parametric assumptions ⁓ in DAQs, so not going to say, for example, if the... relationship is nonlinear for example. But structural equation modeling really belongs to the ⁓ phase, ⁓ a DAG belongs to the identification phase in which you think like can I really identify the causal effects ⁓ ⁓ DAG can be used for many different estimation methods and they're really inputs ⁓ and structural equation model ⁓ sits this estimation phase ⁓ and they be used for causal and non-causal questions. there are mainly outputs, right? So you're not going to create one to guide your analysis. They kind of are the analysis.
Olga: Yeah, so thank you so much for explaining that. That really sort of makes a nice distinction between kind of the DAG and some of the statistical estimation methods. But I was also wondering, ⁓ mean, linked to this. So ⁓ there a big difference then also between a DAG and a theory or maybe also an ontology? So we had a previous episode with Polina Schenk where we... talked a lot about ontologies.
Louise Poppe: ⁓ Well, it's good to have a clear theory in mind before you draw your deck, that always helps. But an adag still differs from an ontology because in an ontology you... I'm not an ontology expert, want to say that first. But I believe that in an ontology you have relationship between your variables that are not necessarily causal. You could say it is a part of and so on. So it helps you to gather a better understanding of a domain and what it entails. I think it can also be an output that you create an ontology and then people ⁓ use it. ⁓ Again, ⁓ a is more when you would focus on one variable on the next. It's input, it is something that you create for design and analysis. ⁓ they all look, yeah, of course, graphed and they all look very similar. You ⁓ get confused in the end, but really have, yeah, ⁓ there's a reason for using them.
Olga: said that it's ⁓ to have a theory in mind when developing a DAG, but do you then mean that a theory can help you think about ⁓ or sort of exposures that are linked to outcomes? ⁓ did you mean? To theory to inform your DAG?
Louise Poppe: Well, I- For example, of the topics that I'm working on now is the impact of green exposure on health outcomes. There is theory about how green space ⁓ impacts your health. It says it goes via air pollution, encouraging physical activity, encouraging social interactions and ⁓ stress reduction and on. This theory helps me to know I should not include, for example, air pollution. confounder because it's a way of how green space impacts my health so if I would include it as a confounder that would be wrong then and of course it all depends on your research question but if you make like if you think like we're starting from this theoretical model and we believe this is correct then it kind of helps you to build your your deck.
Christoph: Okay, yeah, I think that already part of the next question, which would be how do I actually develop a DAG, right? So, theories can inform our DAGs, but yeah, what other sources of knowledge can I use? And is there a specific process, like ⁓ steps ⁓ arrive at your DAG that you can recommend?
Louise Poppe: Well, I had this question about, like, I think... I wanted to create my own DAG and I wrote it in the proposal and it would take me three months I assumed. took way longer and had this huge imposter syndrome because I had to create a DAG and I wanted to be good ⁓ and were no really guidelines in literature so you want to do it good but there were no guidelines and that's when we did our scoping review and ⁓ we kind of all these guidelines and talked about it a lot with all of these experts and then they wrote their own ⁓ Something really helps is just first like into the literature. it's good to have a good idea of how one variable impacts the what other variables can have an impact on that. And like also other studies that are examining the effect, they're going to show for which variables they controlled ⁓ and then you kind of have idea of what their causal model was, how it the variables they thought were confounders and by drawing a you can see whether it actually makes sense ⁓ not because sometimes ⁓ are just added because people sometimes have the idea like the more you control ⁓ things the better. So literature really helps. ⁓ there is a paper by Ferguson and colleagues which helps you ⁓ a DAQ based on literature ⁓ it's quite labor-intensive I must say. Then to To experts and your colleagues, I have never been able to show a DAG without creating a big discussion it with people. you say this is my analysis plan, often have that people are like, yeah, that's OK, that's OK. But if you show a DAG, ⁓ then discussion really starts. And that helps. And also, can with in the field. And we also ⁓ now have a study the impact of depression. ⁓ symptoms on post-surgical neuropathic pain among women with breast cancer and then interviews were also conducted with a ⁓ who had ⁓ cancer to what variables she might be important and also with ⁓ an oncologist. it's like drawing a deck be quite a lot of work but it also helps you to get really a better idea about your exact research question. I so it's really not lost time and ⁓ in end it will save you a lot of time for designing your study and doing the data ⁓
Christoph: Yeah. And to me, it feels like this should be things you should be doing anyways, right? Like you should look at the literature, you should talk to your colleagues before you run a study, but maybe a is just a more formal way to do this. And it can help you like see for yourself whether you already did enough of that. Right? So if you don't know whether you should add a path there or not, then maybe you should read a little bit more. And if there are still a lot of discussions in the team, which I often have, if I bring a DAG to our team meetings.
Louise Poppe: Thank you.
Christoph: And I also learned, okay, maybe this isn't as clear as I thought at first, but a is a great way to visualize that, right? And to have something to talk about. Okay. So those are kind of the sources of knowledge that you can use. But would you say like a systematic literature review, is that necessary? Or does that depend on like how great of a DAG I want it to be? A minimal DAG, is that also possible with just the usual literature search?
Louise Poppe: Yes. Yeah, I think it also depends on how well you know the literature. If you have been in it for several years, kind of... know the literature. According to the paper of Ferguson and Collex, you would have to do a systematic literature review. I have done a scoping review for another project and then for even another project we just looked at ⁓ variables that we thought were most relevant based on the expertise that we have and based on the expertise ⁓ our colleagues and other experts in the field. ⁓ don't think that you always need to do a systematic but I would just encourage you to discuss it with as many people as you can and then you will see that other things really come up.
Christoph: Yeah, and maybe AI tools can actually speed up this process or linking back to ontologies. there is a well formulated ontology, maybe you can more easily screen the literature to ⁓ what's there already. Right. So now we have the knowledge base ⁓ for DAX, what are like the five steps if there are five to construct them?
Louise Poppe: I believe in the paper that we wrote that the first one was to start on time. Like, don't say like right before the analysis like I'm going to draw a quick deck. because that's not really going to work. Just try to do it as early as possible. then you should clearly define exposure and your outcome and the more ⁓ precise the And then it would be good add common causes of the exposure and the outcome. and then to see which arrows could be left out because leaving one out is always a bigger assumption than leaving one in. So if you're not sure, you just leave the arrow in and otherwise it's quite a big assumption to leave one out. So I kind of prefer to like if I have a common cause I will just connect it with everything because that's like that I'm not making any assumptions about it, it cannot have any effect. And then you could add mediators, it's not really necessary because you will not have to ⁓ control for them if you're interested in the total effect. But it can help you like this. you see a variable that you can say, ⁓ no, no, it's really a mediator, then you can add it, of course, to the deck just for clarity. You add selection notes based on the design of your study. you even ⁓ yeah, you represent measurement bias in it, but that's like expert level. If you also ⁓ want to do that so it's something that you consider. It is really relevant of course but I think if you have not worked with it before it might be interesting to start with the bias and selection bias and try to incorporate that in
Christoph: there are layers to it. And I think to understand why you only need to include common causes and maybe not that just causes the exposure or that is downstream consequence of ⁓ the of interest, think it's important to know how you work with a DAG, right? Because if I'm interested in one causal connection, ⁓ in causal path, and I have ⁓ a cost that only like another variable that only affects the exposure, then don't need to condition on it because ⁓ no connection to the outcome. So ⁓ think at least for me with limited knowledge of DAX, it's always easiest to just look at the backdoor criterion, Which is just, or maybe you want to explain it better. Yeah. I think that's, that's just the easiest way to know what to ⁓ and what to condition on. Yeah.
Louise Poppe: Yeah, yeah. Yes, yes. Yeah, so if you can go from your exposure to your outcome via a common cause, you have confounding bias, it's a backdoor criterion, ⁓ you should control for that variable. And in theory, yes, you can have a variable that only affects your exposure, but it usually takes place the past. And then from practical... Yeah, just like... I would always like draw a arrow anyway because I think it might have an impact. I cannot think of variables that really would not have an impact on my outcome but maybe there are ⁓ could of course be.
Christoph: So this backdoor criterion is maybe the main or tool for it. But after I drew my DAG, there other tools I can use to work with it and to get answers from it?
Louise Poppe: Well, do you mean like software that you could use? ⁓
Christoph: For example, I mean Pearl, right? He did a lot ⁓ decks ⁓ there's this term, ⁓ ⁓ something like this, which I never fully understood. There's some ⁓ things around decks as well, which we don't need to cover in detail, but what are the ⁓ useful tools with a deck?
Louise Poppe: Yes, yes. Yeah, so You need to understand the rules a bit. You can control for certain variables. Controlling means that you adjust for them in the analysis or you use stratification. When you control for a confounder, it of stops the... You could think about it as electric current going between variables. So if you control for a confounder, you make sure that this electric current is no longer going through your confounding variable. And then, this non-spurious association between your exposure ⁓ your outcome gets gets controlled for, statistically controlled for. It's really difficult to explain it in a podcast, I must say. But, and ⁓ you have collider is when you have a variable in which ⁓ two arrows collide, come together. ⁓ I'm showing it my hands, but that's really not relevant now. And then, will stop the current unless control for it. So it's not good to control for a collider. And that's what happened. ⁓ That's happened. Like, for example, if you're interested in the effect of physical activity and cognitive performance ⁓ and recruit in an elderly care facility, because ⁓ those variables, physical activity, cognitive performance, ⁓ both an impact on ⁓ whether not you're in the elderly care facility and you for it by only recruiting people from the elderly care facility. ⁓ and that is then collider bias. So you have these different rules that you can follow and in the beginning it might be a bit difficult. I also found it a bit strange to do but after a while it becomes a second nature to interpret them.
Christoph: Okay, so.
Louise Poppe: like software like Daggerty ⁓ just show you when there is a problem like you can say like adjust for these and these variables ⁓ then they will indicate if there is a problem or or if you should adjust for more variables ⁓ to the non-causal ⁓ links non-causal paths so that help ⁓
Christoph: Yeah, but it is easier if you can just follow the connections visually, right?
Louise Poppe: Yes, yes, I prefer to just write them on paper but yeah you can also do it via taggity or what so ever.
Olga: So it sounds a bit like useful to be aware of some of these common ⁓ and sort of how to, yeah, ⁓ it looks visually maybe before you get started. that sort of that you need some like pattern recognition?
Louise Poppe: Yeah. No, it's not really pattern recognition, it's just following a set of rules. There's just a bit of learning about it. In the beginning it might be a bit complex, but after a while... ⁓ Yeah, you see when there is a problem and when there's no problem. like confounding is quite easy to see when there is a problem. For collider bias, have to make sure that you're not conditioning on things happening later. also by drawing them chronologically, ⁓ gets clear which variables might be problematic include in the analysis or which ones you should include. Yeah, it's just as I was saying before it just helps you to visualize your assumptions and people sometimes say like, yeah, it's a bit too much work I'm not going to draw a deck but then I think yeah, okay, but you're still going to make these assumptions You're just not going to be transparent about them But it's not like if you're not drawing a deck that you're not making any assumptions. It's just You just keep them for yourself
Olga: Yeah, so it definitely a useful tool for kind of transparency and also, yeah, making sure that your analysis when ours, well, good they can be for estimating causal effect. So it also sounds like, I mean, for a novice in this space, there's quite a few kind of things to get your head around. Can you think of the most home and beginner mistakes or maybe when developing a DAG for the first time.
Louise Poppe: I a really one is that before you create deck you need to have a really specific idea about what research question is. that I sometimes notice that it's not really clear, like people are drawing a DAG but don't really know which categories they would like to compare. And one of the assumptions in causal inference is also this consistency assumption. for example, if you would have... I want to the effect of on depression, for example. ⁓ the of BMI on depression... That's not really a good research question. Because you should focus on what could be an intervention, for example. Because if you change your BMI via diet or physical activity or like chopping off your arm, it will have completely different effects on your depression outcome. And that's the whole consistency assumption. So there should not be different... varieties of your treatment. there should only be like one way to go from one to the other. And ⁓ also think that this is sometimes a bit difficult in psychology. Like for example, if you think about determinants, ⁓ are many ways that you could change them. ⁓ yeah, also, ⁓ it's something that you need to be aware of. So ⁓ first, have a really clear research question. I think that's the first one. ⁓ A really paper about this is What is your Estimand by ⁓ Lundberg and colleagues I can also ⁓ send it to you ⁓ second one is that people like there to be more people having ⁓ imposter syndrome their deck and they ⁓ a question often is like like when should I stop like I added so many variables like when does it stop and are just ⁓ I mean they are simplification of and and After a while you should just be able to say like, I really think I included the most relevant ones. I discussed it with my colleagues. I discussed it with experts. And this is where it stops. But then at least you're again transparent to reviewers and so on. But you cannot include everything. mean, then it gets, if you include the most proximal ones and the ones that you believe are important. But not the ones that you happen to have in your data set because that's also something. that we see, you should still include the ones that you also did not measure. ⁓
Olga: So then it sounds like, there is a different topic ⁓ well, but maybe also how to set up your analysis plans for kind of secondary data analysis before you've looked at the data or kind of, so it sounds like even if you know ⁓ you're going to use a secondary data set and not collect primary data to maybe develop the DAG first ⁓ you have a detailed look.
Louise Poppe: Yes, that would be really a good idea and then you can also have an idea like... of whether you should maybe merge different datasets to get all the information you need. So always better because it might be tempting to have look at the dataset and see which variables that you have and then do not think further of any other variables that might have an influence. it's, yeah, it's, ⁓ I mean, belongs to the design phase and you should really see it as like specifying, conceiving your question.
Olga: So yeah, that's really interesting and also makes us think about who can use DAGs or can use DAGs to inform their work or are they mainly used in research?
Louise Poppe: Yeah, I think it's more research related. Of course you can draw a causal model of how you think ⁓ will have an impact on people, yeah, it's mainly to inform your design and your data analysis. it's really research ⁓ As I said, they are inputs, they're not meant to be outputs. So yeah, you create them for designing studies and analyzing your data.
Christoph: But in any case, it's good to be more transparent, right? Also for practitioners, so for us as researchers to be more transparent also helps practitioners and the public, think. I think that brings us kind of to the end. Are there any ⁓ tools or resources you still want to recommend to the listeners?
Louise Poppe: Yes, yeah, yeah, yeah, yeah, indeed. I talked about Degety which can help you draw a deck and see which you should control for. There is the free online book of What If by Hernan. And there's also an edX course ⁓ by Hernan. ⁓ which is also really good. It's just, yeah, you see these video clips and it shows you how many minutes and then you get questions and so on afterwards. So we usually ask PhD students that are just starting that they start just ⁓ going the course. And there's also... one by Peter Tennant about the use of tags and I also think this is really accessible because sometimes you can have quite difficult papers ⁓ the causal inference literature ⁓ yeah you have to look for a bit for the ones that are a bit more applied. And I also think, it's a bit self-promotion of course, but I also think that with our scoping review we kind of try to put the basis and explain the basis and how you can start with it. you also have the papers by Julienne Rode, ⁓ also provides a really nice introduction. I can also ⁓ send link to that ⁓ paper and I think it might really help.
Christoph: Yeah, your scoping review is a great resource. I read that in preparation for this episode and I definitely learned a lot that I can just apply right away to my own practice. So I would highly recommend that. But we will put all the resources in the show notes. Seems like there's a lot out there, which is great. Right. Do you have anything else you want to say?
Louise Poppe: And no, thank you for inviting me.
Christoph: Thank you for coming on the show and enjoy the sunny afternoon.
Louise Poppe: Thanks.
