Christoph: Hello and welcome to episode seven of the Define and Conquer podcast by the European Health Psychology Society and its special interest group, Conceptualization, Formalization and Measurement. My name is Christoph Bamberg and I am the host for this episode today because Olga Persky is not our host this time, but our guest of the episode. So I will ask her some questions. She will talk about her work. So welcome Olga. You're currently based at the Karolinska Institute in Sweden, and we're going to talk about computational and formal models. But before we dive in, feel free to give us an overview of your background story and how you arrived at this topic.
Olga: Yeah, thank you so much, Christoph, for having me as the guest today. so that's very exciting. Also, since we recorded the last episode, I have also moved institutions. So I was at Stockholm University and now my group and I we have relocated to the Karolinska Institute and the Department of Clinical Neuroscience. But yes, it's a it's a good question how I sort of got to formal and computational modelling. and really when I well, so I so I did my PhD and a few years of postdoc in health psychology at University College London. And then I became really interested in just in time adaptive interventions, which we haven't really talked much about on the pod yet. but these are kind of health behaviour change interventions that try to provide the right type of support to each person when and where they most need it. And then I was thinking, well, so can we use our current theories to inform the development of these interventions? And very quickly I realized that actually many, if not most, of our health psychology theories, they're not really up to the job. because they're very imprecise. and then that sort of took me down a path of trying to to think about well, can we develop theories that are perhaps more a bit precise specifically in terms of time scales and aspects that are needed to to develop just-in-time adaptive interventions. And that took me to a Marie Curie Global Fellowship, which I spent at the University of California, San Diego working with Dr. Eric Heckler and his group, and really thinking about formal and computational models and how they can be used to inform just-in-time adaptive interventions. And as part of that work, I mean I will tell you a bit more about that in this episode. So we were working on a formal and computational model of relapse, particularly in smoking cessation. And then now, fast forwarding to today, I have just launched my research group here now at the Karolinska Institutes, and we call ourselves the Precision Health Psychology Research Group. and we use these methods, to study addictive behaviors and to develop just in time adaptive interventions. so yes, and no, I will tell you a bit more about that work too.
Christoph: Yeah, very interesting. Thank you for this detailed overview. basically you started with a more applied research program with JIT-ties, right? And then you noticed that the basis just wasn't there and then kind of took a step back. I feel like we had some other guests who had a similar kind of realization in their research career and then went more into formal stuff like theories and so on. That's quite interesting.
Olga: Yeah, for sure. And I think that that's often how we become methods geeks as well. That's kind of the there is a nice kind of dialogue between, yeah, really wanting to solve some applied questions and help people, but then, like you said, realising that well, we actually have to go back to the methods because we can't quite solve the problems that we want to solve.
Christoph: Yeah, maybe to pick up the listeners a little bit more. So we already talked a lot about the reproducibility crisis and also the theory crisis, which kind of seems to be at the basis of the reproducibility crisis, maybe. Could you describe a little bit what the problem is there and the solution that you're offering in your work?
Olga: Yeah, so I mean the reproducibility crisis, maybe just to say something quickly to iterate for our listeners, and that's that well, many results in psychological science but also beyond psychological science, many other disciplines, that first of all the statistical analyses and workflows are not reproducible, which means that if we were to run the same Code again, we don't get the same results. and then we also have the replic replicability problem that if we have kind of what we think is a robust effect or phenomenon that we have found in an experimental study, for example, that if another team is to try to replicate that, they will not find the same result. and then many scholars have argued that There are some deeper issues. Of course, I mean, we cannot necessarily reproduce or replicate our results, but that stems from maybe some deeper issues in that our measurements are poor, so the measurement crisis, and then also our theories are poor in that well, they have been described, for example, by Paul Meal as being weak. I tend to prefer using the term imprecise, which means that most theories, not just in health psychology, but in psychological science, they're typically described using fairly vague natural language descriptions. Or maybe also what we tend to see are kind of box and arrow diagrams of how different constructs influence. one another. But the problem with theories represented at that level is that it leaves many things open to interpretation. So for example, I already mentioned timescales, but n it's also really about construct definitions, what are the relationships between different variables in the model, what's their functional form. and without specifying this more precisely, it's virtually becomes impossible to actually test a theory. and really becomes difficult to, yeah, derive pr precise predictions. I just wanted to give sort of more of a concrete example, to make it yeah less abstract. So I mean
Christoph: Yeah, please.
Olga: one very popular theory in health psychology, but also mm other psychological domains is self-determination theory. It's a very complex theory which has sort of sub-theories within it. but it broadly proposes that human motivation is dependent on or well when we satisfy three different conditions, so the need for competence, the need for relatedness, and the need for autonomy. and there's actually really nice example, so I will link to it in the show notes as well. but if we just zoom in on, for example, the need for competence, the way that that's described verbally in the theory, it means that if we really try to pinpoint, well, w what do we mean by competence? in this nice paper led by Sebastian de Terding and colleagues. So They kind of make the point that, well, I mean, f we could define competence as just successful task performance. So it means that just by me observing that I can perform some intended activity or a task, maybe to some specified standard, that means that I achieve competence. but then a very different interpretation of that. would be that actually competence is capacity growth. So it's not just seeing that you can perform the task, it's growth in some skill that evolves over time. So technically speaking, the better we become at something and achieve competence, then maybe we stop meeting that need for competence and we might need to move to a different skill in order to achieve that need. Which would make sense for kind of human motivation and flourishing more broadly, that maybe over the lifespan that's why we'd maybe keep developing lots of different skills. But this is just the highlights that within that theory, and this is just a tiny part of this big theory, and even within that yeah, there are competing interpretations of what it means. So how Do we even go about testing this then empirically and comparing results across studies? yeah.
Christoph: Yeah, sounds already quite complex, but I human behavior is complex, right? So our theories also have to have some degree of complexity. But so if natural language descriptions, so kind of the terms we use in our everyday life, if that doesn't work or is not sufficient for a theory, what is the solution then that you propose?
Olga: Yeah, so I mean this is certainly not my proposition. I think I'd I'm somebody who has jumped on the bandwagon a few years ago, but I think that there are some really compelling arguments for why formal and computational modelling can perhaps not solve the theory crisis, but at least take us forwards. so when I say formal Model, computational models. I mean these terms are used in many different disciplines. people use the terms differently. So I think yeah, it could be useful to just try to distinguish them from yeah, maybe some other terms. Also in our podcast, we had a a nice episode with Paulina Schenk already where we talked about former ontologies. which are also formal models. so that would kind of be my starting point maybe for describing what they are. so I mean, if we look back to that episode, so a formal ontology is really kind of a tool for structuring knowledge in a given domain, and comes from information science. When I use the term formal model, I really mean formal theory. So the key thing here is that it's linked to explanatory theories. and then I think we first need to unpack, well, what is a theory before we can say what is then a a formal theory and how does it link to computation. and so I mean there are many definitions of what a theory is, but one is that well it's a logically coherent set of propositions and together they aim to explain one or ideally several empirical phenomena. So for example, I just talked about effects maybe that we see in experimental studies, like the Stroop effects or something like that. That might be a an empirical phenomenon that we then want a theory to explain. and then we have well these phenomena, they are not the same as data or observations. The observations and data are evidence for the existence of phenomena. so the phenomena are then relatively stable features of the world. so it's also a little bit up to debate perhaps what counts as a an empirical phenomenon. But then if we sort of are happy with that as a starting point for what a theory is, and then that's also what's the point of theories? Well, we need them to understand phenomena, predict them and also influence them. So this comes back to the intervention part. so in order to for example, build just in time adaptive interventions and really influence the phenomena, maybe we need to understand them better. with the help of a theory. and and then, well,
Christoph: Okay.
Olga: now the formal part,
Christoph: Mm-hmm.
Olga: perhaps, formal and computational. So what is the proposition then? Well it's to instead of representing a theory verbally, as we typically do, is to use some some formalization. So it could be a mathematical framework, it could be logical propositions, formal logic, to represent the the propositions in the theory. so that's kind of part one. And then the n I mean, many scientists would maybe choose to use a mathematical framework with equations to to represent these propositions. And if we do that then has the benefits that we can also instantiate this as a computational model, which means that in the equations we then have free parameters, so they are the same as for example, in a statistical model, regression model, we also have these free parameters. So then we can choose a range of values and simulate the behavior of of the system or the whole theory given different parameter values. So it kind of gives us a nice way of of also checking different theoretical assumptions and see what are the actual consequences. Of course there's more to it than this, but yeah, that's yeah.
Christoph: there was already a lot in there. A lot of new terms and concepts. So I like the explanation of theory as what was it? A set of logically coherent propositions? Is that correct? And then the
Olga: Yeah. Yes.
Christoph: propositions we can translate into mathematical formulas. Okay. Yeah. So far so good. Maybe we can get a bit more practical how to do this and watch these mathematical formulas could look like. yeah, I mean, you did some empirical work on this or some applied work. Maybe that's the better term in the context of smoking. Maybe you could introduce this example and then kind of what aspect of formalization and computational modeling fits to smoking in what way.
Olga: Yeah, so to give one example from my own work, I mean there are also many nice examples in the literature, for those who are interested, mainly from the domain of clinical psychology. so for example, a formal and computational model of panic disorder. Led by Don Robineau and colleagues. But in our work, because I study addictive behaviors, our first attempt at developing a formal model was in the domain of smoking cessation. But really, the phenomena that we want to explain are within person patterns that we then know from lots of research, clinical experience, lived experience, that they can be sort of abstracted as a person wanting to quit smoking but then either remaining abstinent over time or experiencing a few lapses that then sort of set off additional lapses, which means that the person goes back to smoking regularly. or what we would call relapse. So these are sort of the stable features that we've observed across lots of different contexts. so we wanted to then develop a a formal theory of relapse in smoking.
Christoph: And to embed this more in the broader context and the theories that are out there, they are not strong enough or not precise enough, as you would say, to explain the phenomena and to predict relapse and so on. so like developing a just-in-time adaptive intervention would fail then. Is that it?
Olga: Right. Yes. So of course, I mean this is also an area that has been studied a lot. so there are many different theories. and there's some really great work done in this area. for example, the dynamic model of relapse. v but yes, I suppose the starting point for our work was that none of these existing theories have been formalised. And once we were at it, we also started thinking a bit broader about it. So of course, a starting point, and this is something that I've discussed extensively with colleagues in different contexts, when you do this work, would you why not start from an existing theory and formalise it, make it more precise? but it could also be that why not broaden it up a little bit further and really think about well, what are potential propositions or explanatory principles here for how these phenomena come about. So that's n the approach that we chose for this project. But there's also many other projects where maybe the starting point really is to formalize an existing theory. But
Christoph: Mm-hmm.
Olga: yes, the need I I think is still there for most theories that they would not be Described precisely enough.
Christoph: Okay. And so you already did this for smoking and there's a paper out already. I think it's in peer community in the journal.
Olga: Yes.
Christoph: And so maybe you could describe the process of arriving at this computational model with which you ended up in the end.
Olga: Yeah, and and this was also, I mean, something that I've been thinking a lot about over the last few years is also really how to do this work. So it's kind of the sort of methods part in addition to the actual work and applying it. so we've done a couple of scoping reviews, trying also to look at what methods have other scholars been using to develop formal and computational models. also yeah linking to them in the show notes. so if you're interested there's a lot to read about how to develop these models, how to evaluate them also across scientific fields, because of course formal modelling is used a lot in disciplines like physics, ecology I mean in our own discipline there's also a lot of formal modelling in cognitive science, now also in psychiatry. but that was a long-winded way of saying that the sort of framework that I've focused on a lot and that I like a lot, which also fits quite well with the goals of the modelling here, is something called the theory construction methodology. by Danny Boris Bauman colleagues. And what's quite nice about the theory construction methodology is that it provides sort of a higher level roadmap really for how to go about doing this work. So within the theory construction methodology, that sort of has five bigger steps. It's not overly prescriptive About exactly how to go about sort of informing each step, which is also why, sort of in our smoking project and working with the complapse model, as we call it, we also chose to sort of complement that framework with also some guidance from, for example, computational modelling in cognitive science, but also participatory systems mapping. which is perhaps used more in public health. but this theory construction methodology starts off with well, we need to identify the empirical phenomena that we want our formal theory to explain. it's quite a big step, or maybe there already are quite well defined phenomena that are sort of Worthy of explanation. I think in this case with the smoking model, we already knew that, I mean, what we want to explain are these different trajectories. abstinence, prolapse, relapse. and then the second step in the theory construction methodology is to use abduction to develop what they call a proto-theory. So that's a couple of terms that I think I should unpack. so I think on the podcast we've talked about induction, deduction maybe a bit, but abduction, I mean ab or abductive reasoning basically means inference to the most likely explanation. And the reason why they think this is a good place to start is that well when we're developing a so called proto theory, we're not quite ready to sort of close the deductive loop because we need to be a little bit more specific before we get there. and then
Christoph: So it's a pragmatic approach in a way.
Olga: In a way, yes. Well
Christoph: Okay.
Olga: pragmatic, but I also think philosophically tenable
Christoph: Mm-hmm.
Olga: approach. and then yes, we use this to develop a proto-theory, and this is effectively well, box and arrow diagrams that we're familiar with, or sort of we try to explicate these propositions or explanatory principles of the theory.
Christoph: Mm-hmm.
Olga: what we did in the Complaps project with the smoking model was that we did want to use prior knowledge, of course. I mean that's what also we can do in this ad abductive reasoning process as well. thinking about well, we know a lot about how these phenomena come about already. so we did a couple of informal literature reviews of different theories. we also did stakeholder interviews. so we had really nice input from other researchers, policy makers, clinicians and people with lived experience of trying to quit smoking. Then also, I mean, feel free to interrupt me if you want to ask some questions before I talk through all the five steps.
Christoph: Yeah, sure. I found it interesting that you used the stakeholder interviews as well. Would you say that that's always necessary to have a well-founded theory? Or maybe there are also some areas where there's already enough evidence, enough theories there that you don't need this anymore and you can be completely deductive? Or is that unrealistic?
Olga: Yeah, so it's a good good question. And I think also in the literature right now, I think it's more common not to v include sort of an interest holder or participatory mapping stage. I think most examples that I've seen in the literature would be maybe an interdisciplinary team though, so that some folks would have more experience maybe with the clinical population and others maybe have more mathematical engineering knowledge. But I think always when we are going through this process, I mean we need to draw on background knowledge. So whether that comes from kind of expertise in the author team or we do formal or informal literature reviews or speak to different interest holders. I think we n well, we need the knowledge and it's more sort of sometimes I think that can be maybe more informative than going into all of the literature. It could also be resource constraints that we don't have time to read the entire literature. So it's in a sense also more efficient maybe to speak to people. but then I also think I mean the more this maybe comes more into my new project, the Time Matters projects, where we are also I mean we have a whole work package as well about concept definitions and measurements. And here I think it also becomes more and more important with interest holder. inputs because then we also need to think about well how can we measure these things and can people actually respond to the items? so a PhD student in my team, Charlotte Run now is working hard on on some of these aspects. and I think yeah, the the sort of interest holder perspective is really key here.
Christoph: So it depends on the topic and how thorough you want to be and how serious you would take the endeavor. Interesting. Okay. So the first step was specifying the phenomenon, which the theory should explain, right? Second step, formulating like a candidate theory, a proto theory. And then the third step is the translation into a formal model. And I guess that's the key step in a way, right? Could you explain how this works?
Olga: Yeah, so then once we think that's we're quite happy with sort of the explanatory principles, maybe we've also well, in our case we also wanted to think about timescales, feedback loops because we know that these phenomena are dynamic in the sense that they're caused by sort of the interaction between multiple components over time. so then yeah, what the theory construction methodology proposes is to use abstraction then to to formalize the theory but also the phenomena themselves because in order in order to know that or to do some checks on our theory, we also need to quantify the phenomena. or sort of make them make it testable somehow. So this f abstraction or formalization process, I think also the theory construction methodology is not necessarily very prescriptive with regards to how to do it, but there is one thing to keep in mind. So they talk about something that they refer to as analogical abduction. So again abduction comes in here as well. And that's so if for example there are other theories that have already been formalized and that there are already equations that have been found to capture a phenomenon in another field, maybe in ecology or some other domain, maybe infectious diseases or something like that. Then what they propose is to try to borrow some of that those formalisms and see well with 'cause it would be quite neat as well, right? So thinking in terms also of generalizable theories that if we can have equations or formalisms that explain lots of phenomena in psychology, that would be beautiful, right? maybe it is sort of a reinforcement learning type of equation that we can use here. so what we did in the Complaps project, yeah. Of course.
Christoph: Sorry, if I may jump in there just to wrap my head around all these new terms. So could you give an example of a phenomenon that maybe in ecology works in explaining things and that could also work in psychology? Like spontaneously I thought maybe about saturation or something that if there's an island, then at some point... I don't know, like it's full of birds and then there's no more space for new nests and then the island is saturated. Maybe with craving
Olga: Exactly.
Christoph: there's something like this. Is it on this level? Okay. Okay.
Olga: Exactly. Yeah, so d see intuitively, yeah, makes sense, right? So e exactly it could be something like there is logistic growth to
Christoph: Mm-hmm.
Olga: a maximum, which is a very typical one, yeah, your island example, or maybe with population growth or something like that. so that could be a useful formalism. Could also be something like A predator-prey type of dynamic. So there's the Lotka Volterra model that's already good at sort of giving rise to those types of dynamics. So it might be that psychologically maybe there is competition between I don't know, hunger and eating or something like that.
Christoph: Yeah, I did already try to apply this model to craving and hunger
Olga: Okay.
Christoph: data, but I don't know. For me, it didn't really translate so well. But okay, I think I understand the principle.
Olga: No, I mean maybe one would have to think more about that. But yeah, I I have also seen I mean taking ex sort of formalisms from biology to psychology, so maybe also equations that can give rise to sort of spontaneous outbreaks, kind of outbreaks in well, either sort of a virus or something like that. Yes. So these are the types of examples.
Christoph: So we kind of borrow building blocks from other disciplines to construct the formal models then.
Olga: Exactly, yeah. And then also how we did it in the Complaps project. So Jim Allen, who's a mathematician and the collaborator on the project, we also spent a lot of time, him and I, sort of talking about almost standard sort of motives, in maths or maybe that map on to kind of biological processes, or maybe how we would model sort of exposure to a stimulus and decay something like that. so those are also the types of formalisms that that we've embedded in our model. but really it's sort of a I think back and forth as well between
Christoph: Mm-hmm.
Olga: the sort of verbal propositions and how to describe that mathematically.
Christoph: Yeah, this approach.
Olga: And it really requires a kind of knowledge of I mean, you need to have a bit of a toolbox to draw on as well. So obviously if I mean, I'm a psychologist, I don't have a maths background, but working closely with collaborators, I think it's i it's quite nice. We you can get quite far.
Christoph: Yeah. Okay. So I think the principle is relatively clear to me now. And then can you give an example of such a, I'm just going to use the term building block then how it applied to smoking in your example.
Olga: Yeah, so so actually Jim and I have also written a short piece about this for the health psychology community that was published in the European Health Psychologist or magazine. the example that we use there I think is Q reactivity. so sort of exposure to cigarette cues and then how that might impact cravings. Or the desire to smoke. And there we use exactly this formalism of sort of that's there is one part of the equation, which means that if you're exposed to a Q, then there'll be sort of a a spike in your craving, and the exact magnitude of that can vary from person to person. And then there'll be a decay part of the equation, which means that that sort of decays over time. but then also if you quickly encounter another cue, they will add up, or it will sort of so that's one example.
Christoph: Okay. Maybe just so I can visualize it a little bit. So in the end, it's still a function, right? So you map a y to an x in a way, right? Like you have the left-hand side of the equation and the right-hand side of the equation. And the thing I want to predict is smoking, whether someone smokes or not. Yeah. And in the right-hand side, I have these aspects of the theory then that is formalized. So Q reactivity is one part of kind of the right hand side explaining the probability of smoking overall or something like this. Or is that too simplistic?
Olga: Yes. So that's no no, that's yeah, that's pretty much it, apart from that we I mean, maybe this should have been made clearer earlier on as well, that when we choose a mathematical framework to work with, I suppose most psychological phenomena are dynamic and having a mathematical framework that can account for characteristics of dynamical systems like feedback loops, autocorrelation, and aspects like that is useful. So we have represented our system as a dynamical system, which means that yes, we are, I mean it sort of functions the way that you just described. But it's a I mean the system updates every five minutes in our case, which means that the equations for each of the kind of constructs in our model they update and feed into one another. So yes, we are we we can use the whole system of equations to understand the state of the system at every time point, which is maybe then five minutes. and then have the yeah probability of smoking every five minutes. but yeah, w I mean we can also trace the sort of evolution of the system over time.
Christoph: Hey, and I guess then they're also kind of nested equations, right? That one equation explains something else in another one and then they are interdependent. So it gets complex quite fast. I assume.
Olga: Yes.
Christoph: Which I think in the article, you explained that you more or less first identify all possible building blocks and how to formalize them. And then I think you wrote, you, turn them off and then one by one, you switch them on. or something like this to then see how the phenomenon or the prediction of the formalized theory, how it changes depending on what aspects you add to the function. Is that correct? Could you explain that a little bit more? This kind of iterative or step-by-step process of building the function.
Olga: Yeah, so d so and that kind of takes us to I suppose I mean we already talked about step three of the process and kind of using abstraction to formalize because the next step is then also to use mathematical analysis and simulation to really check kind of the well Danny Borisbowman the colleagues referred to it as the explanatory adequacy of of your formal. model. and really that means that well we want our model to or our formal theory to explain the phenomena. And the way to check that, which is quite neat, is to see, well, can the formalized version of the theory can it grow the phenomena that it was designed to explain? So that kind of provides an initial sense check
Christoph: Mm-hmm.
Olga: That the model is good. And when thinking about sort of criteria for how do we know that this is a good formal model or not, that's sort of a really crucial step. So then when when we build up the model, yeah, the approach to take, or one approach to take is to formalize every sort of building block, simulate it with. A few different parameter values, or maybe draw from a distribution of plausible values. and then check to see, well, visually does this seem to make sense. maybe for folks with a stronger mathematics background, you can of course also do I mean more mathematical analysis when building up the model. but then really it's about kind of adding all of the components together and then see can can the model generate the phenomena but then also to run sensitivity analysis and see okay what if i s now mm switch off or or I set everything in the model to kind of the middle of the the plausible parameter range. And then I systematically vary one component. Can I flip the whole system from trajectory of abstinence to relapse, for example? So then
Christoph: Mm-hmm.
Olga: we're more looking at kind of qualitative differences in swapping between the phenomena, I suppose, rather than precise quantitative differences.
Christoph: Yeah.
Olga: Of course, you can also run checks in terms of Yeah, the because now it's also switching in, I mean it becomes more and more technical as well. So also one way to check I mean, is to do an exercise called parameter recovery, which also means that we simulate the model with known parameter values, and then we specify some sort of cost function, maybe a likelihood function, maybe we can't use likelihood-based methods. That's for another time.
Christoph: You
Olga: and and then we basically see by fitting our model to kind of the output data generated from the model, can we recover the known parameters? And that's also a check on, I mean, is this model structurally identifiable? So there are many, many different ways of kind of the yeah, checking is the model any good? And this is before we even fit it to any empirical data.
Christoph: Yeah, to me, this sounds somewhat analogous to developing a statistical model with a Bayesian perspective. Or maybe that's also just because I'm a big fan of Bayesian statistics. But you also specify the equation, right? And then kind of what likelihood functions and so on. And you kind of build the model. And then you can also do prior predictive checks to see whether the assumptions in the model make sense a priori, right? Before you even look at the data. So yeah, there's a
Olga: Yes.
Christoph: interesting correspondence there.
Olga: Yeah, and I think that there are many parallels to Bayesian inference. I mean, this also maybe we need to do another episode on this more specifically. but of course also we can take a Bayesian approach to estimating the parameters in this type of model
Christoph: Mm-hmm.
Olga: too. and there's a method called approximate Bayesian computation, which is also when we can't use the likelihood based approaches then Bayesian approaches to parameter estimation seem very useful.
Christoph: But one conceptual difference would be that in your model, kind of the logic is not based around uncertainty, right? There's no fuzzy logic or anything there, but the equation is assumed to be true in a way. And in Bayesian statistics, it is a logical language, but just not having two states, zero or one, but there's a spectrum in between, which is the probability, right? So there are some fundamental differences.
Olga: Yeah, but I think I think it could be yeah, but conceptually I think it's also I mean because saying that there's I mean, from a Bayesian perspective that n there's no kind of true parameter value. I mean, if that represents the mean or variance or I mean here in these models it represents the parameters represent slightly different things.
Christoph: Mm-hmm.
Olga: But I think it it also works to say that there's a distribution so that there's no kind of true fixed.
Christoph: Hmm. Okay. I see.
Olga: So I think there's n quite a few different parallels. I was also thinking that there are parallels between sort of the stage of incorporating prior knowledge into the proto theory with Bayesian and
Christoph: Mm-hmm.
Olga: priors and choosing your priors, and I've seen some quite nice papers where people use qualitative methods to inform priors for a later statistical model. So multiple
Christoph: Yeah, interesting.
Olga: parallels, I think.
Christoph: Yeah, mean, I guess the one idealized way to do patient statistics is also to run simulations of the data a priori and then inform the model by that. So yeah, this kind of simulation aspect is also baked into that. But I think that's enough of a tangent into patient statistics. Sorry for that. So now assuming that we have specified the computational modeling and it's running and we get some kind of output that we can then visualize and then check. guess you did this for the smoking behavior that you wanted to predict. I was wondering whether there was anything surprising or was it just we expected this model to behave like this and then you observed this behavior or was there some behavior, if you change a certain parameter that was unexpected, but that made sense thinking about it again.
Olga: Yeah, so I think probably d not in that sense because I think I mean from sensitivity analysis we can see that yes, if n well, for example, mm reactivity to cues, the magnitude goes up a lot, so then people are more likely to lapse and then relapse. Or we also have I mean self-efficacy in there so if the person remains abstinent their self-efficacy will grow sort of through mastery experience but then also if that growth parameter is quite large then the person effectively never lapses so I think that the I think those are more sort of sense checks that it's operating as we would want it to operate. I think maybe b mm other things that are more s they're not really surprising either, but I think it's more when we formalize the phenomena, like abstinence, for example, we then have no events or zero probability. Maybe the probability isn't zero, but it's below a certain threshold, maybe. and again, not surprising, but I think we it becomes very hard to fit these models if we don't have any event. So I think this is maybe something that's more yeah, a bigger question. Also I think this would generalize to lots of other health behaviours where the focus is on quitting something unhealthy. v and I mean what we want to support people to do it would be maybe to have zero events or very, very few events. so this
Christoph: Yeah.
Olga: becomes tricky.
Christoph: Yeah, I think I was of course wondering about how I could apply this to eating behavior, which is more my research field. And I think there the interesting complexity is that we have to eat, right? So, zero abstinence. Yeah. I mean, you could abstain from eating chocolate, unhealthy snacks or so on, but I think it's less clear whether that's always good, right? Also like hedonic goals and so on are also important, right? so I guess it translates better to drinking maybe, but yeah.
Olga: Yeah, and also I mean yeah, exactly. And I would say also in addiction, I mean the goal is not always abstinence and there's a really nice literature on recovery and sort of thinking about yeah, maybe it's more multifaceted so that it's not just the behaviour but also other indicators of social emotional connection, well being that sort of would be part of of the phenomenon. so I think there's a lot more work to be done there as well.
Christoph: Mm-hmm.
Olga: again, we have another kind of well work package on this in the ERC funded Timeatus project, which
Christoph: Yep.
Olga: is focused more on spending some time trying to to characterize the phenomena further.
Christoph: Okay, so assuming that we have this model running then, or this formalized theory, then we can use it as an input to just in time adaptive intervention, kind of to tie back to the original motivation for this three, four, five year research project, right? So then if I ask participants, did you encounter a Q? How is your self-efficacy in the day? kind of plug this in as values and then I get out the probability. I guess it's a continuous prediction then, right? So I get the probability of smoking in five minutes, in 10 minutes and so on. Yeah, that's the application then. you say that describes it? And is this something, is this realistic now already or is there still some work that needs to be done? for you to apply this to real data.
Olga: Yes. So d so I suppose I mean before maybe we Have something that can be reliably used to inform interventions. yes, I think there are a few additional steps. so I mean, maybe we do want to be able to successfully recover parameters, for example, in these kind of more simulation exercises, and then probably also fit the model to empirical data. Which in a way I think should be w once you've done the parameter recovery, then I mean that's sometimes the harder step than fitting it to real world data. but then I suppose the good thing is that to now that we have I mean a f a formalized version of the theory, it also means that we can be quite precise about different types of interventions and how would they impact kind of the dynamics in the system. So what we're planning to do, for example, is first to simulate that and then we can run within person experiments to actually test that empirically. And here also to really flag some of the work that Eric Heckler and Daniel Rivera have been leading for the past decade, or even more longer than that, is to use methods from the control systems engineering toolbox to do some of this work as well. and effectively what we're working towards here as well is that we want to use well simulations then within person experiments to build what control engineers refer to as a controller, which is an algorithm that uses a formal dynamical systems model as basis, and then becomes sort of a decision-making algorithm to decide i is an intervention needed for this person at this moment. so there is a whole the
Christoph: Yeah.
Olga: well, additional pipeline of course to to build this work or to build adaptive interventions in this way. so really the formal modelling part is kind of one step in that
Christoph: Mm-hmm.
Olga: to more applied work and to really have sort of truly adaptive interventions that can sort of fit to each person.
Christoph: Yeah, interesting how transdisciplinary this whole work is. That seems quite promising. All right. I think we could talk a lot more about this topic still. I still had some questions about, because this is on the single person, single individual level, how to compare different individuals. But you also write about this in the article that we will link in the show notes. So I think it's already well addressed there. right. Where would you say someone who found this a little bit daunting or challenging, to apply to their own work? Where, where should they start? maybe it already helps to just think about simulating, data, what, what it might look like before you, collect data or something like this. Like what would be a baby step, in this direction that someone could do?
Olga: Yeah, so I think and also the I think the intention here is of course to invite the community to start thinking more about well the work that we do. And it's not to say that all of the theories we have are not used for or that everyone should be doing this type of work. so I think the real take home home is just to To consider some of these aspects. And yes, I think your point about I mean, even with a theory that a lot of people use all the time, that's maybe a bit easier to simulate data from than, for example, the self-determination theory. Because I think as hopefully I highlighted, I don't think it's that straightforward how how to simulate data from it. Or Maybe somebody who's an expert on that theory can come back and email us about it and say, here is how I would do it. That would be excellent.
Christoph: That would be great.
Olga: but maybe with some, yeah, like the theory of planned behavior or some less complex theory, that yes, a good first step might be to start simulating some data and really think about timescales. Does the theory say something about time scales? and yeah, and and really also think about if we can, I mean to be more precise, not just about the theoretical structure, but also construct definitions, measurements, so linking back to I mean, other episodes of this podcast, I think that really takes us a long way.
Christoph: Yeah, in the end, it's all tied together, right? We need better theories, we need better concepts and better measurements and everything is just one building block. I think that's a good end to this episode, to episode seven. Thank you very much, Olga, for explaining your work. I'm looking forward to seeing more of it in the future. Thank you for coming on the show.
Olga: Thanks so much.
Olga: Yeah, thank you so much, Christoph, for having me as the guest today. so that's very exciting. Also, since we recorded the last episode, I have also moved institutions. So I was at Stockholm University and now my group and I we have relocated to the Karolinska Institute and the Department of Clinical Neuroscience. But yes, it's a it's a good question how I sort of got to formal and computational modelling. and really when I well, so I so I did my PhD and a few years of postdoc in health psychology at University College London. And then I became really interested in just in time adaptive interventions, which we haven't really talked much about on the pod yet. but these are kind of health behaviour change interventions that try to provide the right type of support to each person when and where they most need it. And then I was thinking, well, so can we use our current theories to inform the development of these interventions? And very quickly I realized that actually many, if not most, of our health psychology theories, they're not really up to the job. because they're very imprecise. and then that sort of took me down a path of trying to to think about well, can we develop theories that are perhaps more a bit precise specifically in terms of time scales and aspects that are needed to to develop just-in-time adaptive interventions. And that took me to a Marie Curie Global Fellowship, which I spent at the University of California, San Diego working with Dr. Eric Heckler and his group, and really thinking about formal and computational models and how they can be used to inform just-in-time adaptive interventions. And as part of that work, I mean I will tell you a bit more about that in this episode. So we were working on a formal and computational model of relapse, particularly in smoking cessation. And then now, fast forwarding to today, I have just launched my research group here now at the Karolinska Institutes, and we call ourselves the Precision Health Psychology Research Group. and we use these methods, to study addictive behaviors and to develop just in time adaptive interventions. so yes, and no, I will tell you a bit more about that work too.
Christoph: Yeah, very interesting. Thank you for this detailed overview. basically you started with a more applied research program with JIT-ties, right? And then you noticed that the basis just wasn't there and then kind of took a step back. I feel like we had some other guests who had a similar kind of realization in their research career and then went more into formal stuff like theories and so on. That's quite interesting.
Olga: Yeah, for sure. And I think that that's often how we become methods geeks as well. That's kind of the there is a nice kind of dialogue between, yeah, really wanting to solve some applied questions and help people, but then, like you said, realising that well, we actually have to go back to the methods because we can't quite solve the problems that we want to solve.
Christoph: Yeah, maybe to pick up the listeners a little bit more. So we already talked a lot about the reproducibility crisis and also the theory crisis, which kind of seems to be at the basis of the reproducibility crisis, maybe. Could you describe a little bit what the problem is there and the solution that you're offering in your work?
Olga: Yeah, so I mean the reproducibility crisis, maybe just to say something quickly to iterate for our listeners, and that's that well, many results in psychological science but also beyond psychological science, many other disciplines, that first of all the statistical analyses and workflows are not reproducible, which means that if we were to run the same Code again, we don't get the same results. and then we also have the replic replicability problem that if we have kind of what we think is a robust effect or phenomenon that we have found in an experimental study, for example, that if another team is to try to replicate that, they will not find the same result. and then many scholars have argued that There are some deeper issues. Of course, I mean, we cannot necessarily reproduce or replicate our results, but that stems from maybe some deeper issues in that our measurements are poor, so the measurement crisis, and then also our theories are poor in that well, they have been described, for example, by Paul Meal as being weak. I tend to prefer using the term imprecise, which means that most theories, not just in health psychology, but in psychological science, they're typically described using fairly vague natural language descriptions. Or maybe also what we tend to see are kind of box and arrow diagrams of how different constructs influence. one another. But the problem with theories represented at that level is that it leaves many things open to interpretation. So for example, I already mentioned timescales, but n it's also really about construct definitions, what are the relationships between different variables in the model, what's their functional form. and without specifying this more precisely, it's virtually becomes impossible to actually test a theory. and really becomes difficult to, yeah, derive pr precise predictions. I just wanted to give sort of more of a concrete example, to make it yeah less abstract. So I mean
Christoph: Yeah, please.
Olga: one very popular theory in health psychology, but also mm other psychological domains is self-determination theory. It's a very complex theory which has sort of sub-theories within it. but it broadly proposes that human motivation is dependent on or well when we satisfy three different conditions, so the need for competence, the need for relatedness, and the need for autonomy. and there's actually really nice example, so I will link to it in the show notes as well. but if we just zoom in on, for example, the need for competence, the way that that's described verbally in the theory, it means that if we really try to pinpoint, well, w what do we mean by competence? in this nice paper led by Sebastian de Terding and colleagues. So They kind of make the point that, well, I mean, f we could define competence as just successful task performance. So it means that just by me observing that I can perform some intended activity or a task, maybe to some specified standard, that means that I achieve competence. but then a very different interpretation of that. would be that actually competence is capacity growth. So it's not just seeing that you can perform the task, it's growth in some skill that evolves over time. So technically speaking, the better we become at something and achieve competence, then maybe we stop meeting that need for competence and we might need to move to a different skill in order to achieve that need. Which would make sense for kind of human motivation and flourishing more broadly, that maybe over the lifespan that's why we'd maybe keep developing lots of different skills. But this is just the highlights that within that theory, and this is just a tiny part of this big theory, and even within that yeah, there are competing interpretations of what it means. So how Do we even go about testing this then empirically and comparing results across studies? yeah.
Christoph: Yeah, sounds already quite complex, but I human behavior is complex, right? So our theories also have to have some degree of complexity. But so if natural language descriptions, so kind of the terms we use in our everyday life, if that doesn't work or is not sufficient for a theory, what is the solution then that you propose?
Olga: Yeah, so I mean this is certainly not my proposition. I think I'd I'm somebody who has jumped on the bandwagon a few years ago, but I think that there are some really compelling arguments for why formal and computational modelling can perhaps not solve the theory crisis, but at least take us forwards. so when I say formal Model, computational models. I mean these terms are used in many different disciplines. people use the terms differently. So I think yeah, it could be useful to just try to distinguish them from yeah, maybe some other terms. Also in our podcast, we had a a nice episode with Paulina Schenk already where we talked about former ontologies. which are also formal models. so that would kind of be my starting point maybe for describing what they are. so I mean, if we look back to that episode, so a formal ontology is really kind of a tool for structuring knowledge in a given domain, and comes from information science. When I use the term formal model, I really mean formal theory. So the key thing here is that it's linked to explanatory theories. and then I think we first need to unpack, well, what is a theory before we can say what is then a a formal theory and how does it link to computation. and so I mean there are many definitions of what a theory is, but one is that well it's a logically coherent set of propositions and together they aim to explain one or ideally several empirical phenomena. So for example, I just talked about effects maybe that we see in experimental studies, like the Stroop effects or something like that. That might be a an empirical phenomenon that we then want a theory to explain. and then we have well these phenomena, they are not the same as data or observations. The observations and data are evidence for the existence of phenomena. so the phenomena are then relatively stable features of the world. so it's also a little bit up to debate perhaps what counts as a an empirical phenomenon. But then if we sort of are happy with that as a starting point for what a theory is, and then that's also what's the point of theories? Well, we need them to understand phenomena, predict them and also influence them. So this comes back to the intervention part. so in order to for example, build just in time adaptive interventions and really influence the phenomena, maybe we need to understand them better. with the help of a theory. and and then, well,
Christoph: Okay.
Olga: now the formal part,
Christoph: Mm-hmm.
Olga: perhaps, formal and computational. So what is the proposition then? Well it's to instead of representing a theory verbally, as we typically do, is to use some some formalization. So it could be a mathematical framework, it could be logical propositions, formal logic, to represent the the propositions in the theory. so that's kind of part one. And then the n I mean, many scientists would maybe choose to use a mathematical framework with equations to to represent these propositions. And if we do that then has the benefits that we can also instantiate this as a computational model, which means that in the equations we then have free parameters, so they are the same as for example, in a statistical model, regression model, we also have these free parameters. So then we can choose a range of values and simulate the behavior of of the system or the whole theory given different parameter values. So it kind of gives us a nice way of of also checking different theoretical assumptions and see what are the actual consequences. Of course there's more to it than this, but yeah, that's yeah.
Christoph: there was already a lot in there. A lot of new terms and concepts. So I like the explanation of theory as what was it? A set of logically coherent propositions? Is that correct? And then the
Olga: Yeah. Yes.
Christoph: propositions we can translate into mathematical formulas. Okay. Yeah. So far so good. Maybe we can get a bit more practical how to do this and watch these mathematical formulas could look like. yeah, I mean, you did some empirical work on this or some applied work. Maybe that's the better term in the context of smoking. Maybe you could introduce this example and then kind of what aspect of formalization and computational modeling fits to smoking in what way.
Olga: Yeah, so to give one example from my own work, I mean there are also many nice examples in the literature, for those who are interested, mainly from the domain of clinical psychology. so for example, a formal and computational model of panic disorder. Led by Don Robineau and colleagues. But in our work, because I study addictive behaviors, our first attempt at developing a formal model was in the domain of smoking cessation. But really, the phenomena that we want to explain are within person patterns that we then know from lots of research, clinical experience, lived experience, that they can be sort of abstracted as a person wanting to quit smoking but then either remaining abstinent over time or experiencing a few lapses that then sort of set off additional lapses, which means that the person goes back to smoking regularly. or what we would call relapse. So these are sort of the stable features that we've observed across lots of different contexts. so we wanted to then develop a a formal theory of relapse in smoking.
Christoph: And to embed this more in the broader context and the theories that are out there, they are not strong enough or not precise enough, as you would say, to explain the phenomena and to predict relapse and so on. so like developing a just-in-time adaptive intervention would fail then. Is that it?
Olga: Right. Yes. So of course, I mean this is also an area that has been studied a lot. so there are many different theories. and there's some really great work done in this area. for example, the dynamic model of relapse. v but yes, I suppose the starting point for our work was that none of these existing theories have been formalised. And once we were at it, we also started thinking a bit broader about it. So of course, a starting point, and this is something that I've discussed extensively with colleagues in different contexts, when you do this work, would you why not start from an existing theory and formalise it, make it more precise? but it could also be that why not broaden it up a little bit further and really think about well, what are potential propositions or explanatory principles here for how these phenomena come about. So that's n the approach that we chose for this project. But there's also many other projects where maybe the starting point really is to formalize an existing theory. But
Christoph: Mm-hmm.
Olga: yes, the need I I think is still there for most theories that they would not be Described precisely enough.
Christoph: Okay. And so you already did this for smoking and there's a paper out already. I think it's in peer community in the journal.
Olga: Yes.
Christoph: And so maybe you could describe the process of arriving at this computational model with which you ended up in the end.
Olga: Yeah, and and this was also, I mean, something that I've been thinking a lot about over the last few years is also really how to do this work. So it's kind of the sort of methods part in addition to the actual work and applying it. so we've done a couple of scoping reviews, trying also to look at what methods have other scholars been using to develop formal and computational models. also yeah linking to them in the show notes. so if you're interested there's a lot to read about how to develop these models, how to evaluate them also across scientific fields, because of course formal modelling is used a lot in disciplines like physics, ecology I mean in our own discipline there's also a lot of formal modelling in cognitive science, now also in psychiatry. but that was a long-winded way of saying that the sort of framework that I've focused on a lot and that I like a lot, which also fits quite well with the goals of the modelling here, is something called the theory construction methodology. by Danny Boris Bauman colleagues. And what's quite nice about the theory construction methodology is that it provides sort of a higher level roadmap really for how to go about doing this work. So within the theory construction methodology, that sort of has five bigger steps. It's not overly prescriptive About exactly how to go about sort of informing each step, which is also why, sort of in our smoking project and working with the complapse model, as we call it, we also chose to sort of complement that framework with also some guidance from, for example, computational modelling in cognitive science, but also participatory systems mapping. which is perhaps used more in public health. but this theory construction methodology starts off with well, we need to identify the empirical phenomena that we want our formal theory to explain. it's quite a big step, or maybe there already are quite well defined phenomena that are sort of Worthy of explanation. I think in this case with the smoking model, we already knew that, I mean, what we want to explain are these different trajectories. abstinence, prolapse, relapse. and then the second step in the theory construction methodology is to use abduction to develop what they call a proto-theory. So that's a couple of terms that I think I should unpack. so I think on the podcast we've talked about induction, deduction maybe a bit, but abduction, I mean ab or abductive reasoning basically means inference to the most likely explanation. And the reason why they think this is a good place to start is that well when we're developing a so called proto theory, we're not quite ready to sort of close the deductive loop because we need to be a little bit more specific before we get there. and then
Christoph: So it's a pragmatic approach in a way.
Olga: In a way, yes. Well
Christoph: Okay.
Olga: pragmatic, but I also think philosophically tenable
Christoph: Mm-hmm.
Olga: approach. and then yes, we use this to develop a proto-theory, and this is effectively well, box and arrow diagrams that we're familiar with, or sort of we try to explicate these propositions or explanatory principles of the theory.
Christoph: Mm-hmm.
Olga: what we did in the Complaps project with the smoking model was that we did want to use prior knowledge, of course. I mean that's what also we can do in this ad abductive reasoning process as well. thinking about well, we know a lot about how these phenomena come about already. so we did a couple of informal literature reviews of different theories. we also did stakeholder interviews. so we had really nice input from other researchers, policy makers, clinicians and people with lived experience of trying to quit smoking. Then also, I mean, feel free to interrupt me if you want to ask some questions before I talk through all the five steps.
Christoph: Yeah, sure. I found it interesting that you used the stakeholder interviews as well. Would you say that that's always necessary to have a well-founded theory? Or maybe there are also some areas where there's already enough evidence, enough theories there that you don't need this anymore and you can be completely deductive? Or is that unrealistic?
Olga: Yeah, so it's a good good question. And I think also in the literature right now, I think it's more common not to v include sort of an interest holder or participatory mapping stage. I think most examples that I've seen in the literature would be maybe an interdisciplinary team though, so that some folks would have more experience maybe with the clinical population and others maybe have more mathematical engineering knowledge. But I think always when we are going through this process, I mean we need to draw on background knowledge. So whether that comes from kind of expertise in the author team or we do formal or informal literature reviews or speak to different interest holders. I think we n well, we need the knowledge and it's more sort of sometimes I think that can be maybe more informative than going into all of the literature. It could also be resource constraints that we don't have time to read the entire literature. So it's in a sense also more efficient maybe to speak to people. but then I also think I mean the more this maybe comes more into my new project, the Time Matters projects, where we are also I mean we have a whole work package as well about concept definitions and measurements. And here I think it also becomes more and more important with interest holder. inputs because then we also need to think about well how can we measure these things and can people actually respond to the items? so a PhD student in my team, Charlotte Run now is working hard on on some of these aspects. and I think yeah, the the sort of interest holder perspective is really key here.
Christoph: So it depends on the topic and how thorough you want to be and how serious you would take the endeavor. Interesting. Okay. So the first step was specifying the phenomenon, which the theory should explain, right? Second step, formulating like a candidate theory, a proto theory. And then the third step is the translation into a formal model. And I guess that's the key step in a way, right? Could you explain how this works?
Olga: Yeah, so then once we think that's we're quite happy with sort of the explanatory principles, maybe we've also well, in our case we also wanted to think about timescales, feedback loops because we know that these phenomena are dynamic in the sense that they're caused by sort of the interaction between multiple components over time. so then yeah, what the theory construction methodology proposes is to use abstraction then to to formalize the theory but also the phenomena themselves because in order in order to know that or to do some checks on our theory, we also need to quantify the phenomena. or sort of make them make it testable somehow. So this f abstraction or formalization process, I think also the theory construction methodology is not necessarily very prescriptive with regards to how to do it, but there is one thing to keep in mind. So they talk about something that they refer to as analogical abduction. So again abduction comes in here as well. And that's so if for example there are other theories that have already been formalized and that there are already equations that have been found to capture a phenomenon in another field, maybe in ecology or some other domain, maybe infectious diseases or something like that. Then what they propose is to try to borrow some of that those formalisms and see well with 'cause it would be quite neat as well, right? So thinking in terms also of generalizable theories that if we can have equations or formalisms that explain lots of phenomena in psychology, that would be beautiful, right? maybe it is sort of a reinforcement learning type of equation that we can use here. so what we did in the Complaps project, yeah. Of course.
Christoph: Sorry, if I may jump in there just to wrap my head around all these new terms. So could you give an example of a phenomenon that maybe in ecology works in explaining things and that could also work in psychology? Like spontaneously I thought maybe about saturation or something that if there's an island, then at some point... I don't know, like it's full of birds and then there's no more space for new nests and then the island is saturated. Maybe with craving
Olga: Exactly.
Christoph: there's something like this. Is it on this level? Okay. Okay.
Olga: Exactly. Yeah, so d see intuitively, yeah, makes sense, right? So e exactly it could be something like there is logistic growth to
Christoph: Mm-hmm.
Olga: a maximum, which is a very typical one, yeah, your island example, or maybe with population growth or something like that. so that could be a useful formalism. Could also be something like A predator-prey type of dynamic. So there's the Lotka Volterra model that's already good at sort of giving rise to those types of dynamics. So it might be that psychologically maybe there is competition between I don't know, hunger and eating or something like that.
Christoph: Yeah, I did already try to apply this model to craving and hunger
Olga: Okay.
Christoph: data, but I don't know. For me, it didn't really translate so well. But okay, I think I understand the principle.
Olga: No, I mean maybe one would have to think more about that. But yeah, I I have also seen I mean taking ex sort of formalisms from biology to psychology, so maybe also equations that can give rise to sort of spontaneous outbreaks, kind of outbreaks in well, either sort of a virus or something like that. Yes. So these are the types of examples.
Christoph: So we kind of borrow building blocks from other disciplines to construct the formal models then.
Olga: Exactly, yeah. And then also how we did it in the Complaps project. So Jim Allen, who's a mathematician and the collaborator on the project, we also spent a lot of time, him and I, sort of talking about almost standard sort of motives, in maths or maybe that map on to kind of biological processes, or maybe how we would model sort of exposure to a stimulus and decay something like that. so those are also the types of formalisms that that we've embedded in our model. but really it's sort of a I think back and forth as well between
Christoph: Mm-hmm.
Olga: the sort of verbal propositions and how to describe that mathematically.
Christoph: Yeah, this approach.
Olga: And it really requires a kind of knowledge of I mean, you need to have a bit of a toolbox to draw on as well. So obviously if I mean, I'm a psychologist, I don't have a maths background, but working closely with collaborators, I think it's i it's quite nice. We you can get quite far.
Christoph: Yeah. Okay. So I think the principle is relatively clear to me now. And then can you give an example of such a, I'm just going to use the term building block then how it applied to smoking in your example.
Olga: Yeah, so so actually Jim and I have also written a short piece about this for the health psychology community that was published in the European Health Psychologist or magazine. the example that we use there I think is Q reactivity. so sort of exposure to cigarette cues and then how that might impact cravings. Or the desire to smoke. And there we use exactly this formalism of sort of that's there is one part of the equation, which means that if you're exposed to a Q, then there'll be sort of a a spike in your craving, and the exact magnitude of that can vary from person to person. And then there'll be a decay part of the equation, which means that that sort of decays over time. but then also if you quickly encounter another cue, they will add up, or it will sort of so that's one example.
Christoph: Okay. Maybe just so I can visualize it a little bit. So in the end, it's still a function, right? So you map a y to an x in a way, right? Like you have the left-hand side of the equation and the right-hand side of the equation. And the thing I want to predict is smoking, whether someone smokes or not. Yeah. And in the right-hand side, I have these aspects of the theory then that is formalized. So Q reactivity is one part of kind of the right hand side explaining the probability of smoking overall or something like this. Or is that too simplistic?
Olga: Yes. So that's no no, that's yeah, that's pretty much it, apart from that we I mean, maybe this should have been made clearer earlier on as well, that when we choose a mathematical framework to work with, I suppose most psychological phenomena are dynamic and having a mathematical framework that can account for characteristics of dynamical systems like feedback loops, autocorrelation, and aspects like that is useful. So we have represented our system as a dynamical system, which means that yes, we are, I mean it sort of functions the way that you just described. But it's a I mean the system updates every five minutes in our case, which means that the equations for each of the kind of constructs in our model they update and feed into one another. So yes, we are we we can use the whole system of equations to understand the state of the system at every time point, which is maybe then five minutes. and then have the yeah probability of smoking every five minutes. but yeah, w I mean we can also trace the sort of evolution of the system over time.
Christoph: Hey, and I guess then they're also kind of nested equations, right? That one equation explains something else in another one and then they are interdependent. So it gets complex quite fast. I assume.
Olga: Yes.
Christoph: Which I think in the article, you explained that you more or less first identify all possible building blocks and how to formalize them. And then I think you wrote, you, turn them off and then one by one, you switch them on. or something like this to then see how the phenomenon or the prediction of the formalized theory, how it changes depending on what aspects you add to the function. Is that correct? Could you explain that a little bit more? This kind of iterative or step-by-step process of building the function.
Olga: Yeah, so d so and that kind of takes us to I suppose I mean we already talked about step three of the process and kind of using abstraction to formalize because the next step is then also to use mathematical analysis and simulation to really check kind of the well Danny Borisbowman the colleagues referred to it as the explanatory adequacy of of your formal. model. and really that means that well we want our model to or our formal theory to explain the phenomena. And the way to check that, which is quite neat, is to see, well, can the formalized version of the theory can it grow the phenomena that it was designed to explain? So that kind of provides an initial sense check
Christoph: Mm-hmm.
Olga: That the model is good. And when thinking about sort of criteria for how do we know that this is a good formal model or not, that's sort of a really crucial step. So then when when we build up the model, yeah, the approach to take, or one approach to take is to formalize every sort of building block, simulate it with. A few different parameter values, or maybe draw from a distribution of plausible values. and then check to see, well, visually does this seem to make sense. maybe for folks with a stronger mathematics background, you can of course also do I mean more mathematical analysis when building up the model. but then really it's about kind of adding all of the components together and then see can can the model generate the phenomena but then also to run sensitivity analysis and see okay what if i s now mm switch off or or I set everything in the model to kind of the middle of the the plausible parameter range. And then I systematically vary one component. Can I flip the whole system from trajectory of abstinence to relapse, for example? So then
Christoph: Mm-hmm.
Olga: we're more looking at kind of qualitative differences in swapping between the phenomena, I suppose, rather than precise quantitative differences.
Christoph: Yeah.
Olga: Of course, you can also run checks in terms of Yeah, the because now it's also switching in, I mean it becomes more and more technical as well. So also one way to check I mean, is to do an exercise called parameter recovery, which also means that we simulate the model with known parameter values, and then we specify some sort of cost function, maybe a likelihood function, maybe we can't use likelihood-based methods. That's for another time.
Christoph: You
Olga: and and then we basically see by fitting our model to kind of the output data generated from the model, can we recover the known parameters? And that's also a check on, I mean, is this model structurally identifiable? So there are many, many different ways of kind of the yeah, checking is the model any good? And this is before we even fit it to any empirical data.
Christoph: Yeah, to me, this sounds somewhat analogous to developing a statistical model with a Bayesian perspective. Or maybe that's also just because I'm a big fan of Bayesian statistics. But you also specify the equation, right? And then kind of what likelihood functions and so on. And you kind of build the model. And then you can also do prior predictive checks to see whether the assumptions in the model make sense a priori, right? Before you even look at the data. So yeah, there's a
Olga: Yes.
Christoph: interesting correspondence there.
Olga: Yeah, and I think that there are many parallels to Bayesian inference. I mean, this also maybe we need to do another episode on this more specifically. but of course also we can take a Bayesian approach to estimating the parameters in this type of model
Christoph: Mm-hmm.
Olga: too. and there's a method called approximate Bayesian computation, which is also when we can't use the likelihood based approaches then Bayesian approaches to parameter estimation seem very useful.
Christoph: But one conceptual difference would be that in your model, kind of the logic is not based around uncertainty, right? There's no fuzzy logic or anything there, but the equation is assumed to be true in a way. And in Bayesian statistics, it is a logical language, but just not having two states, zero or one, but there's a spectrum in between, which is the probability, right? So there are some fundamental differences.
Olga: Yeah, but I think I think it could be yeah, but conceptually I think it's also I mean because saying that there's I mean, from a Bayesian perspective that n there's no kind of true parameter value. I mean, if that represents the mean or variance or I mean here in these models it represents the parameters represent slightly different things.
Christoph: Mm-hmm.
Olga: But I think it it also works to say that there's a distribution so that there's no kind of true fixed.
Christoph: Hmm. Okay. I see.
Olga: So I think there's n quite a few different parallels. I was also thinking that there are parallels between sort of the stage of incorporating prior knowledge into the proto theory with Bayesian and
Christoph: Mm-hmm.
Olga: priors and choosing your priors, and I've seen some quite nice papers where people use qualitative methods to inform priors for a later statistical model. So multiple
Christoph: Yeah, interesting.
Olga: parallels, I think.
Christoph: Yeah, mean, I guess the one idealized way to do patient statistics is also to run simulations of the data a priori and then inform the model by that. So yeah, this kind of simulation aspect is also baked into that. But I think that's enough of a tangent into patient statistics. Sorry for that. So now assuming that we have specified the computational modeling and it's running and we get some kind of output that we can then visualize and then check. guess you did this for the smoking behavior that you wanted to predict. I was wondering whether there was anything surprising or was it just we expected this model to behave like this and then you observed this behavior or was there some behavior, if you change a certain parameter that was unexpected, but that made sense thinking about it again.
Olga: Yeah, so I think probably d not in that sense because I think I mean from sensitivity analysis we can see that yes, if n well, for example, mm reactivity to cues, the magnitude goes up a lot, so then people are more likely to lapse and then relapse. Or we also have I mean self-efficacy in there so if the person remains abstinent their self-efficacy will grow sort of through mastery experience but then also if that growth parameter is quite large then the person effectively never lapses so I think that the I think those are more sort of sense checks that it's operating as we would want it to operate. I think maybe b mm other things that are more s they're not really surprising either, but I think it's more when we formalize the phenomena, like abstinence, for example, we then have no events or zero probability. Maybe the probability isn't zero, but it's below a certain threshold, maybe. and again, not surprising, but I think we it becomes very hard to fit these models if we don't have any event. So I think this is maybe something that's more yeah, a bigger question. Also I think this would generalize to lots of other health behaviours where the focus is on quitting something unhealthy. v and I mean what we want to support people to do it would be maybe to have zero events or very, very few events. so this
Christoph: Yeah.
Olga: becomes tricky.
Christoph: Yeah, I think I was of course wondering about how I could apply this to eating behavior, which is more my research field. And I think there the interesting complexity is that we have to eat, right? So, zero abstinence. Yeah. I mean, you could abstain from eating chocolate, unhealthy snacks or so on, but I think it's less clear whether that's always good, right? Also like hedonic goals and so on are also important, right? so I guess it translates better to drinking maybe, but yeah.
Olga: Yeah, and also I mean yeah, exactly. And I would say also in addiction, I mean the goal is not always abstinence and there's a really nice literature on recovery and sort of thinking about yeah, maybe it's more multifaceted so that it's not just the behaviour but also other indicators of social emotional connection, well being that sort of would be part of of the phenomenon. so I think there's a lot more work to be done there as well.
Christoph: Mm-hmm.
Olga: again, we have another kind of well work package on this in the ERC funded Timeatus project, which
Christoph: Yep.
Olga: is focused more on spending some time trying to to characterize the phenomena further.
Christoph: Okay, so assuming that we have this model running then, or this formalized theory, then we can use it as an input to just in time adaptive intervention, kind of to tie back to the original motivation for this three, four, five year research project, right? So then if I ask participants, did you encounter a Q? How is your self-efficacy in the day? kind of plug this in as values and then I get out the probability. I guess it's a continuous prediction then, right? So I get the probability of smoking in five minutes, in 10 minutes and so on. Yeah, that's the application then. you say that describes it? And is this something, is this realistic now already or is there still some work that needs to be done? for you to apply this to real data.
Olga: Yes. So d so I suppose I mean before maybe we Have something that can be reliably used to inform interventions. yes, I think there are a few additional steps. so I mean, maybe we do want to be able to successfully recover parameters, for example, in these kind of more simulation exercises, and then probably also fit the model to empirical data. Which in a way I think should be w once you've done the parameter recovery, then I mean that's sometimes the harder step than fitting it to real world data. but then I suppose the good thing is that to now that we have I mean a f a formalized version of the theory, it also means that we can be quite precise about different types of interventions and how would they impact kind of the dynamics in the system. So what we're planning to do, for example, is first to simulate that and then we can run within person experiments to actually test that empirically. And here also to really flag some of the work that Eric Heckler and Daniel Rivera have been leading for the past decade, or even more longer than that, is to use methods from the control systems engineering toolbox to do some of this work as well. and effectively what we're working towards here as well is that we want to use well simulations then within person experiments to build what control engineers refer to as a controller, which is an algorithm that uses a formal dynamical systems model as basis, and then becomes sort of a decision-making algorithm to decide i is an intervention needed for this person at this moment. so there is a whole the
Christoph: Yeah.
Olga: well, additional pipeline of course to to build this work or to build adaptive interventions in this way. so really the formal modelling part is kind of one step in that
Christoph: Mm-hmm.
Olga: to more applied work and to really have sort of truly adaptive interventions that can sort of fit to each person.
Christoph: Yeah, interesting how transdisciplinary this whole work is. That seems quite promising. All right. I think we could talk a lot more about this topic still. I still had some questions about, because this is on the single person, single individual level, how to compare different individuals. But you also write about this in the article that we will link in the show notes. So I think it's already well addressed there. right. Where would you say someone who found this a little bit daunting or challenging, to apply to their own work? Where, where should they start? maybe it already helps to just think about simulating, data, what, what it might look like before you, collect data or something like this. Like what would be a baby step, in this direction that someone could do?
Olga: Yeah, so I think and also the I think the intention here is of course to invite the community to start thinking more about well the work that we do. And it's not to say that all of the theories we have are not used for or that everyone should be doing this type of work. so I think the real take home home is just to To consider some of these aspects. And yes, I think your point about I mean, even with a theory that a lot of people use all the time, that's maybe a bit easier to simulate data from than, for example, the self-determination theory. Because I think as hopefully I highlighted, I don't think it's that straightforward how how to simulate data from it. Or Maybe somebody who's an expert on that theory can come back and email us about it and say, here is how I would do it. That would be excellent.
Christoph: That would be great.
Olga: but maybe with some, yeah, like the theory of planned behavior or some less complex theory, that yes, a good first step might be to start simulating some data and really think about timescales. Does the theory say something about time scales? and yeah, and and really also think about if we can, I mean to be more precise, not just about the theoretical structure, but also construct definitions, measurements, so linking back to I mean, other episodes of this podcast, I think that really takes us a long way.
Christoph: Yeah, in the end, it's all tied together, right? We need better theories, we need better concepts and better measurements and everything is just one building block. I think that's a good end to this episode, to episode seven. Thank you very much, Olga, for explaining your work. I'm looking forward to seeing more of it in the future. Thank you for coming on the show.
Olga: Thanks so much.
