 Well, it's really a pleasure to be here. I was asked to give sort of an introductory overview of a field that I've been involved in since I started, which is something called ambulatory assessment. And that's sort of the broader term that's come into fashion over the past few years. And I'll talk a little bit more about what exactly ambulatory assessment contains. But if you've heard of things like experience sampling, ecological momentary assessments, things like that, those both fall under the umbrella term of ambulatory assessment along with wearable biosensors that are now becoming a big deal in many fields, in many health-related fields. So what I'm going to do today is talk you through what ambulatory assessment is. A little bit about the variety of techniques that it subsumes. Talk a little bit about the motivation, some of the methodological advantages. And at the same time, I'm going to try to be really honest so this isn't just an advertisement, right? I'm going to try to also look at the flip side of each of those so called advantages and think about what are the disadvantages, right? Because we have to make sacrifices in order to do this kind of work and to get at ecological processes as they happen. And then I think one of the best ways to sort of see the potential of this kind of work is to go through some actual research examples. So after I go through the overview, then I'm going to present some examples of work that we're doing in my lab to really sort of get at some of the processes related to adolescent and young adult substance use and problem behavior. What are the social, interpersonal, and affective circumstances that are associated with use, behavior, and risk for use, okay? So as I mentioned, ambulatory assessment has become kind of an umbrella term. And I'd put this on the reading list, I'd really recommend it if you're interested in this kind of thing. Troll and Abner-Pramer in the annual review of clinical psychology, talk about what ambulatory assessment is. And here I'm stealing their language, they say it's a suite of methods that allow us to study individuals in their naturalistic environments, okay? Now there are other ways to study people in their naturalistic environments, obviously ethnography. But what you tend to find in an ambulatory assessment type approach is we're taking quantitative methods, quantitative assessments. And we're trying to embed them in people's day to day lives, okay? And we are getting basically a range of reporting strategies. So self-report has sort of reigned supreme and we're able to do this more and more effectively now with the proliferation of mobile technologies and the ease of administering mobile surveys. There's been some work in observational coding and then biological and physiological processes are all included. One of the oldest ways of doing this was something called the Daily Diary Study. And the feature here was once daily reporting. And it was really about giving me, as the experimenter or the observationalist, a sort of retrospective report of your day, okay? And so you might tell me what were the most stressful things that happened to you that day, what was your mood in general? Did you have social interactions? Or did you drink? Things like that. And essentially what these studies are trying to get at is what you might think of as a good day, bad day effect, right? How are people different from themselves on good days versus bad days? And you could define a good day or a bad day in a variety of ways. But so one thing I wanna point out is that even though these methods, there's a lot of excitement around them right now, particularly in how they can intersect and inform mobile health interventions. They're not new, right? They've been around for quite a long time. And one of the, as I mentioned, oldest ways of doing this was to actually get paper and pencil diaries that you would mail to people and they'd mail them back to you, okay? You'd sort of trust the participant to fill them out when they said they did, which, if you look in the literature, there were some problems with that, as you might imagine. Telephone calls, people would sort of call people at the end of the day. There's a very famous study called the National Study of Daily Experiences where people from the Midas sample were actually called every single day for about seven days to sort of really get at stress and affect in day-to-day life and this concept of affective reactivity. And then, oops, and then web-based computer surveys. I'll actually present a study where this method was used. And then, so I, I mean, this is a function of my own training. I tend to use the EMA terminology, Ecological Momentary Assessment, but you may have been trained in an area that experienced sampling methodology was used, but by and large, they're about the same thing, okay? And the feature here is that what you're getting is multiple reports. And you're getting these as people are, I mean, the goal is to get people in the moment and truly random prompting schedules, you want to catch them unaware. You don't want them to be prepared for the fact that you're about to ask them a question about their mood. Because if they're aware that you will be asking that kind of question at a certain time, they might stop what they're doing, sit down, change their environment, and respond. So the goal here is to really catch them in the moment and try to understand what they're feeling, who they're with, how they're doing, okay? Right, and so you're trying to get at not only what the shifts and context are, are they moving from inside to outside? Social situation, solitary situation. But how shifts and things like context are related to their shifts in emotion, their shifts in how they behave. And as I've mentioned, you're doing this in the flow of day to day life, okay? And so with the proliferation of mobile technology, we've been able to do this well. Now actually, even when I started doing this, which was about 12 years ago, we were using things like palm pilots, okay? But now with smartphones and tablets, we're able to do this a lot more quickly, a lot easier. And now we're starting to see a lot of studies that use wearable sensors, okay? And the clear advantage here is that you can get at some of these psychosocial processes or some biological factors that you're interested in without having to rely on self reports. Because one of the downsides of this method, especially when you're using self report is that you're asking people about the same things repeatedly. And that may or may not change what they report, okay? So if to the extent you can, people are starting to build in these wearable sensor technologies that hopefully people will just forget about, okay? And so there are many different varieties. This one came out about four or five years ago and it was piloted in a study where they had adolescents riding around in their neighborhoods and wearing this sort of sensor on the strap of their backpack that detected pollution levels in their neighborhood. And so they were sort of trying to quantify from an objective and experiential point of view how much pollution were youngsters exposed to in their daily commutes. You also have sensors that will do this for noise level. Gary Evans has done a lot of work in this area looking at the relationship between the noise of your neighborhood and how well you do in school, how well you pay attention. Even some immune function parameters he's shown are sensitive to this. And then bio-behavioral sensors are also really in fashion right now. And so there's some studies that are using these consumer grade wearables to study things like sleep, so things like Fitbits and all that. There's also been a long history of actigraphy where we're trying to understand how much people are moving in their day to day lives and maybe we're trying to get them to move more. And one sensor that I've started using a lot is an alcohol monitoring sensor, we can actually detect alcohol intoxication through the skin. Which is interesting if what you're interested in is studying alcohol use in day to day life, but you don't necessarily trust people's self reports in the moment, right? And I'll explain that a little bit more. Now this sensor here is unfortunate in that it's the gold standard right now. It's the best one that we have from a scientific standpoint, but it's rather unattractive and somewhat stigmatizing. So we're looking for one that's a little bit smaller that you can actually wear like a Fitbit, and that device is coming. But anyway, the point here is that there's a lot of options here for wearable sensors. Okay, so designing one of these studies, running one of these studies requires a lot of equipment, it makes you go through a lot of trouble, right? So why would you do it? Well, there are three sort of arguments that are trotted out. And these are sort of the things like when you want to justify this in a grant application, you sort of put these three things down. But so essentially what we're trying to do is we're trying to get real world data. That is, we're trying to get assessments that are collected in the environments that people actually inhabit. And maybe we're trying to do that as they're in those environments as well. Okay, and the idea here is that what you're trying to do is enhance ecological validity, right? So you want data that represent what is actually happening in people's lives, okay? Now, of course you're making a trade-off here because a lot of what you find from these studies, you won't be able to say too much with regard to internal validity, but what you gain in ecological validity might be so great that it doesn't matter. So for example, if you're interested in stressors and stress processes, you can stress somebody out in a lab, but the stress that you control and you provide for them doesn't have the same meaning as when something happens in their day-to-day life, okay? So again, trade-offs here. We're also using real-time assessment, which allows us to collect reports as close in time as we can manage, or as close in time as is reasonable to when our events of interest occur, okay? And the argument here is that what we're going to do is we're going to reduce the potential for recall bias. So instead of me asking you how angry you are, you're going to how angry have you been on average this past week? That's going to be some strange mix of how angry you were recently, how angry you were when you'll remember how angry you were at your moment of highest anger, and you'll do some sort of strange conglomeration of those and spit out an average, right? And the idea here is if we get you with enough frequency and we believe our assessments are representative of your lived experience that we can actually figure out, from a more empirical basis, how angry you were relative to somebody else. And even in the moments, what kind of moments you were angriest, which is another important issue, right? And so I just sort of touched on this, but another way to say this is that this allows the estimation of empirical as opposed to remembered averages about what you're typically like, okay? And then the final bit is that we are getting data over time. So we are getting repeated assessments across a variety of situations. And this allows us to calculate and discover a number of dynamic processes, right? We could look at temporal dynamics. Maybe we're interested in your anger experiences, your craving experiences if you're a cigarette smoker related to time of day because we want to know when to send you that message that's going to prevent you from smoking, okay? When is it going to have its maximum impact? When are you most at risk? We can discover that for each individual with these kinds of studies. Day of week processes, which might be very important for things like alcohol use. We can even look at effective processes leading up to a binge drinking episode and start to look for clues of when we might intervene or when people are at their highest risk. Another powerful way, or another powerful use of this data is to look at within person processes where you're essentially comparing each person to him or herself across time. And you've got multiple reports to do this. So you can start to ask questions using each person as their own comparison, okay? Which, if you think about it, removes all of those common factors, all of those factors that are unique to that person across time. All stable confounders are sort of removed because you're using each person as their own comparison. So it automatically differences out things like gender, right? Race, ethnicity, anything that will remain stable within your measurement scheme. Another way I like to think about this is that these procedures allow us to zoom in, okay? And if you think about a longer term longitudinal study, if you're interested in the development of some process, like how, what is the risk trajectory for major depression across life? You can do a longitudinal study, you can watch people grow and change. And this will give you a view from, I like to say 10,000 feet. You can see the big picture, you get some detail, but you lose some precision, right? And that's largely because if you want a long term trajectory, you're limited in the number of times you can assess somebody. And so there are at least months to years between each assessment. So we lose out on some of the precision. And through ambulatory assessments, what we're able to do is zoom in, potentially key points in that trajectory and start to understand why some people have an elevation in their rate of depression versus others. Maybe they're exposed to different kinds of experiences, right? Maybe they're exposed to more stressors than others. Maybe they're also more reactive to those stressors, okay? And so these are things that we can look at in an ambulatory assessment study, right? So changes in context we can look at. We can examine links between exposures, context and behavior in the flow of daily life. And there's been a lot of excitement around this type of work because this might help us identify entry points for intervention led change. Are there specific scenarios that recur in a certain person's life that lead to negative outcomes in the moment and increase their risk of substance use relapse, okay? And if we can find those and find the individualized predictors, can we use these in some sort of intervention? As you might imagine, so I've talked about, we're getting a sample of individuals, we're assessing them many, many times. And so what these designs often do is, they present us with a lot of opportunity. But they also present us with a significant challenge. And that is, we'll end up with a massive amount of data, okay? And so the data that we get from ambulatory assessment designs is something that's being referred to in the literature is intensive longitudinal data, okay? And a loose guideline that's been put out there in the literature is, it's data with many measurements over time. Some have suggested, well, maybe more than 30, okay? The point is a lot, and more than you would get from a quote unquote traditional longitudinal study. And often these are measurements that are condensed with respect to time. So you get a lot of them quickly, okay? And so the processes that you're looking at are more sort of micro-longitudinal processes, right? They're things like variability, they're short term diurnal cycles, things like that. Another feature of this data that makes it difficult is that it's often sampled at irregular time points, right? So I mentioned that idea of sampling somebody randomly, right? So that they don't expect our assessments. Well, that creates issues where the spaces between data points are not necessarily uniform, which can be somewhat of an issue if you are used to something like time series modeling and you are forced to make the assumption that each observation is equally spaced from each other, okay? So we have to find a way to break that assumption. Sometimes the data are actually the timing at which the data are collected is subject dependent. Like if I'm interested in studying people when they're drinking, then I usually have to find a way either to prompt them when they're drinking. I have to find some way to predict when that happens, which is very difficult. Or I have to ask them to initiate a certain survey sequence, which is something we're trying to do right now. And so somebody who drinks more than another will give us more data, right? And so essentially we have irregular spacing and a regular number of observations. And this presents a challenge for us. Now, you have many approaches to analyze these kinds of data, generalize estimating equations, do well. Time series analysis, you can modify to do certain things, etc. But what I want to just touch on briefly and really briefly is multi-level modeling, which tends to be the dominant approach here. And the reasons why are that the multi-level model does a good job of mapping on conceptually to often how we think about these data. Right, cuz this type of method is ideal for exploring two kinds of variability simultaneously. Okay, so we can look at how individuals differ from themselves across time, okay? And we can look at how people differ from each other on average. And in a multi-level setup, we can predict both of those sources of variability at the same time, okay? Which is a powerful approach. I'm gonna give you some examples of why that's important. Because I think this is a somewhat simple but difficult issue. It also allows us not only to describe what's happening in our sample on average, right? So what is the within person dynamic for the average person? But it also allows us to individualize the estimates we get from the model, okay? So maybe the association on average, if we look within a person between stress and drinking, what's the average increase in drinking on a high versus low stress day, we can find that. And maybe it's slightly positive. But if we were to individualize it, for some, it might be flat, right? For some, it might be highly positive. For some, it might even be negative. They drink less when they're experiencing stress. And so this modeling approach with this kind of data would allow us to appreciate those differences. And of course, it handles the clustering of observations within individuals. Okay, so, I like to visualize the design of these studies like this. Where you basically see that what we end up doing, so it's not just an idiographic type approach where we focus on one person, right? We focus in depth on each individual and we do this multiple times. And that's where we get, so the fact that we do this repeatedly for many different individuals, gives us this between person variation, okay? And in multi-level parlance, you'd often talk about this as level two. This is the second level, the higher level, where we are sort of looking at the average of these reports for person one, and we're comparing them to the average of the reports for person two, all the way to person 200, right? And so this is often where our standard set of statistical methods is focused, we're sort of comparing people to each other, typically, okay? Then at level one, what we're able to do is take a look at person one and we say, okay, if moment two was a stressful moment, moment one was not, what's the difference in craving for cigarettes for this person, right? And we can sort of calculate, well, what's the average difference in craving between high stress and low stress moments for person one? And we can compare that to person two, person three, right? So basically we get within person effects from level one. We can then evaluate how they're different at level two. And I like to show this because I think it makes this a bit more concrete. So this is some simulated data that I use a lot when I teach statistical workshops. And here we're just sort of assuming that we've measured cigarette smokers and their level of craving. And we've measured this multiple times, up to about 30 plus observations across a number of days, okay? And so each dot is a different craving measure. And so we can see that it varies all over the place for many of these individuals. And then what we can do is we can calculate the average for each person. And so if you look at that red line for each person, that's the average amount of craving that we got from their self reports, okay? And so when we're talking about between person differences, what we're talking about is the difference in the height of those red lines, okay? And when we're talking about within person effects, we're talking about the spread of the blue dots around the red line for each individual. We're trying to understand why this person here, person 10 is elevated in their craving compared to when they're not, right? So why are they so much higher than their average? We could also ask questions about why person 11 tends to have higher craving than person 10. And we could even get really fancy and look at the amount of variability. And ask why are some people fluctuating so much more than others? And this is especially important if you're interested in things like bipolar disorder where affective variability tends to be a defining characteristic. This is something that you could capture empirically in this way. And this has been done in the work of Tim Troll. He's done an excellent job with this kind of investigation. Okay, so I've talked a little bit about this. But so essentially another way I like to talk about this is between person questions are who questions, right? So you want to know how people differ from each other on average. Are people who experience a lot of stress more likely to engage in heavy drinking? And then when questions are within person questions. So you're trying to ask when a person experiences a stressor, are they more likely to engage in drinking compared to themselves when they're not? Okay, so that's the comparison here. The comparison is people to each other when we're talking about the who questions. For the when questions, we're talking about comparing people to themselves. The important point about this that I think sometimes gets lost is that if you know the answer to one of these, you do not know the answer to the other, okay? And you can make arguments for this on both conceptual and statistical grounds. But knowing something about one of these actually tells you nothing about the other. So let me give you a couple of examples of that, okay? So typing speed and spelling errors, I like this one, okay? So if you look and you compare people to each other, what you'll often find is that people who type faster than others tend to commit fewer spelling errors, okay? And the conclusion there is maybe they have more training as typists, perhaps they're just better typists. But when a person types faster than they usually do, they tend to commit more errors, right? And so you can see that we're talking about two entirely different causal processes with something as simple as this, right? So I've chosen ones here where the effect is opposite. They don't need to be opposite, but I just find that this points it out pretty nicely. So then let's look at another one, exercise and blood pressure. Now, if you compare people to each other, people who exercise more regularly, they tend to have lower blood pressure than people who do not exercise as regularly. I think you can see where I'm gonna go with this, right? When a person is exercising, their blood pressure tends to be higher than when they are not exercising, okay? So again, this sort of shows you the independence of these two levels of analysis and why we need to consider both of them. And this is not in the readings, but I think it's interesting. John Nezlek does a really good idea, he does a really good job of talking about this idea and he essentially shows that you can have relationships in one direction. So if we're looking at the dashed line, we're looking at the between person relationship. And you can see that the average for person three is lower than the average for person two and person one, right? So the mean of X is positively associated with the mean of Y. But if we look at each of the individual lines, we can see that as X gets higher for that person, Y decreases. So this is an opposite relationship, right? You can have an opposite relationship in the opposite direction, okay? And you could also have a situation where you see a between person relationship. So as the mean of X goes up, the mean of Y goes up. But if you look within people, there's no relationship between these two at all. And this has been an unfortunate situation in some studies of behavioral science where we'll make comparisons. So we'll sort of look at people who are exposed to more stress than others. And we find that they have higher levels of depression. And we might conclude, okay, well, being exposed to stress leads people to develop depression, but we haven't actually shown it. And we will need to sort of look within that individual in order to make that kind of claim. And if we don't, what we're doing is we're committing what's called an ecological fallacy. Where we're looking at a relationship from a higher level and trying to extrapolate it downward to an individual, okay? So we're looking at a group of people and trying to extrapolate that to a person. And it doesn't always work, okay? So what I'm gonna do is take you through a few examples of the work that we've been doing. And a lot of these research examples are in substance use and substance use processes. But I think that they will illustrate some of these concepts pretty well. So the first that I wanted to mention was a study where we looked at witnessing substance use. So whether or not adolescents saw others in their immediate vicinity using alcohol or drugs and whether that was related to them engaging in some sort of antisocial behavior. So antisocial behavior being things like vandalism, petty theft, other sorts of rule breaking behavior in their day to day lives. There's also a longitudinal study that has embedded daily diary measurements where we sort of looked at in college students' day to day lives the relationship between stressors across a variety of domains and drinking. And then we looked to see if those students who increased their drinking more strongly on high versus low stress days were more likely to develop alcohol problems later. So we sort of took within person information, pulled it up to the between person level and tried to understand if there was something about those dynamics that were informative for their risk for alcohol use. And then finally, I'm going to talk to you about the study where that's in the field now where we're using alcohol monitoring devices and ecological momentary assessments to try to understand something about the social and effective milieu that surrounds heavy drinking episodes. And to try to understand how well people do at actually reporting alcohol use behavior in the moment. Which is something people are interested in. But it remains to be seen how well people can do that as they're becoming intoxicated. I'm going to skip over this stuff because I think we've gone over that. So this is the first paper I want to talk to you about. So essentially what we did here, one part of it that I didn't quite mention is that so we're looking at the relationship between witnessing others use substances and antisocial behavior on the same day. We also looked at a gene environment interaction in day to day life. So looking at one of these sort of within person processes and seeing how it differed at the between person level based on adolescence genetic background, okay? And so the main research question here that we wanted to know was, we wanted to ask if adolescents were more likely to engage in antisocial behavior, okay, on days that they witnessed other people using substances compared to themselves on non-witnessing days. Which again, so we're looking at a within person association. We're comparing adolescents to themselves across different types of days. And we wanted to know if this association was stronger for adolescents with the seven repeat allele of the DRD4 gene. We chose this gene largely because of the work in developmental psychopathology showing that children and adolescents with this allele are more sensitive to their social environments. And from some experimental work that shows similar things. And so we wanted to see, first of all, if we could capture a between person difference in that within person association and see if DRD4 was a predictor of that. So to give you some background, you can see across a number of nationally representative surveys that adolescents in the United States are frequently exposed to substance use in their homes, schools, and neighborhoods. So this is from the Center for Adolescent Substance Abuse. And it says that 60% of high school students in the US are in schools where students are using or selling drugs. Data from the National Survey of Drug Use and Health tells us that about 11% of US youth under 18 are living with a parent who's suffering from alcoholism. And we know from developmental psychopathology that adolescents who are exposed to a lot of parental and peer substance use are more likely to engage in antisocial behavior. And as part of that pathway, eventually use substances at an earlier time point and escalate more quickly into problematic patterns of use. And most of the work that I just mentioned is really examining the relationships across years. Whereas I've mentioned before, the long measurement lags in between each of the assessments can miss some of the action here. Because we might be really talking about a day to day process that happens, where you're exposed to some environment, you react to it, and it instantiates a behavior pattern, right? And so we applied EMA to capture some of these exposure contexts in the flow of daily life, giving us reduced recall bias, enhancing ecological validity, as I've mentioned. And also letting us examine this process at a daily level, okay? So I wanna talk a little bit about why we went with DRD4 as a marker of sensitivity. There's a significant amount of work that has shown that the seven repeat allele of the DRD4 gene is something that's associated with a variety of outcomes on sort of the novelty seeking in attention spectrum. So it's associated with ADHD, impulse control, low impulse control. As well as sensitivity to parenting environments, peer environments. Both in developmental and in lab studies. And this lab study by Larson and colleagues was particularly informative where they showed that young adults with the seven repeat allele, if put in a heavy drinking situation, were more likely to escalate their drinking. Compared to those without the seven repeat allele. Okay, the mechanisms for how this works are a little bit sketchy right now. But we're sort of intrigued by this and wanted to see if we could observe the sensitivity in day to day life. So not only did we wanna know if adolescents were, on average, more likely to engage in deviant behavior when they're around others using substances, but were there some adolescents who were especially at risk? So the design, the study that we ran was the MyLife study. So my advisor, Candice Oggers, was the lead on this. This was a data collection effort that I was involved in heavily during graduate school. And what we did was we collected data from about 150 adolescents. And they were between the ages of 12 and 14. And in order to be entered into the study or to be eligible for the study, they had to indicate or their parents had to indicate, rather. A number, three or more markers of risk for early substance use. So they had been diagnosed with attention deficit hyperactivity disorder. They had parents who had a substance use problem. A family member who, or another family member who had a substance use problem. Or had some instances of substance use themselves. The study started with a baseline questionnaire. This is often how these studies go. You sort of, we got the child and the parent to come into the lab. We asked them a range of questionnaires. So we wanted to understand a little bit about the child's behavior history, the neighborhood they were living in, what was the level of collective efficacy. A little bit about the family background. So that we could contextualize the information that we would get from the field study. In the field study, we did 30 days of assessment, three times a day, with a mobile phone that we provided. Now that part's tricky now, because everybody has their own phone. But when we did this, not everybody did. So this was done back in 2009, 2010. The challenge now is really trying to integrate these protocols seamlessly into the technology that people already carry. So there's a big movement toward creating apps that participants download. And we're moving toward that as well. It just requires a decent amount of infrastructure. Okay, but the goal here is we're getting real life assessment versus in the lab. And then we had a follow up study where we got more questionnaires from the child about 18 months later. And then we all, this is when we did our DNA collection. Sort of thought of it after the fact and then this would be a really interesting idea to take a look at how some of the dynamics we capture here might differ based on genetic background. So I'm just gonna walk you through what I did here. So the first, you can think about the results I'm gonna present to you as sort of having been analyzed in two steps. So the first was to estimate the within person association. Now, engagement in antisocial behavior was highly skewed. They either did or did not that day. So we dichotomized it. And essentially what we're looking at here is, here's our mixed model where we have an intercept that is the odds of antisocial behavior on a day they didn't witness substance use. Plus the change on a day that they did versus did not. So that beta one parameters of interest and then these are our random effects down here, particularly interested in the random slope. And this is what allows this beta one relationship to be stronger for some adolescents than for others. So we estimate that relationship and essentially you can think about it as running a regression for each adolescent. But that's all done in one modeling step. And then the next thing we do is we add this interaction term. This beta three where we have DRD4 genotype predicting that within person relationship, okay? And what we're essentially trying to do is explain the variability in the association based on genetic background. So the first thing we saw was that there was a positive association, okay? The odds increased about threefold on days that adolescents witnessed others using substances versus days they did not. We wanted to see if they use substances themselves but there just wasn't enough on day to day variability on that. Particularly because the adolescents were probably too young. Even though they were high risk. So we couldn't really look at that. And we also found that this slope did vary randomly. So you can see that random slope estimate at the bottom. That's the variability. That's a variance term, which is significantly greater than zero. So we can say that this association varies randomly between adolescents, okay? And more so than we would expect by chance alone. Now here's the interaction that we found, okay? So on the y-axis, you've got the probability of antisocial behavior. On the x-axis, you've got whether or not they witnessed versus did not witness substance use. So the first thing I wanna say is that for those without the seven repeat allele that's here, there was still a significant increase, okay? It's just modest in comparison to what we saw for those with the seven repeat allele, okay? So we did find confirmation of this. Now this is a small sample size. And I think this is another thing to point out is that often times with these studies because you're sampling so intensively from each individual, it's difficult to get a large number of people. So in a genetic study, we would typically want a larger sample size than this. But we're sort of forced to make these kinds of trade-offs. But the benefit here is that this relationship that we're seeing is a relationship that comes from a within-person assessment. So we really are removing a lot of confounding by using each person as their own comparison, okay? So again, trade-offs, yeah? Can you or do other studies you're aware of that being able to get a large enough sample size that you could also do between area difference, like neighborhoods, or testing, about that, or? Yeah, so there's actually a study that our group is doing now where we'll have EMA from about 500 adolescents. And their neighborhoods were measured a lot more in depth. And they came from a larger population of about 2,000. So that's coming. I recently just read something from Christopher Browning's group where they're doing something called geospatial EMA, GEMA. And they tend to do things like measure activity space using GPS sensors. And so because that's a passive way of measuring, you can get more adolescents. So they had a pretty respectable sample size as well. Anytime we ask for self-reports, it becomes difficult because you're really sort of asking a lot of participants. But yeah, good question. So essentially, though, the point I'm trying to make here is that these associations were different for all adolescents. And we were able to predict some of that variability. In the next paper, we did something similar where we, again, estimated a within-person association. But this time used it as a predictor of something else. And this really sort of gets at the ideas that have been around for a long time that intra-individual within-person variability is important and meaningful and trying to extract information above a person's average, which we tend to be pretty fond of using. Maybe there's something in the variability that we can get that would add to our predictions over and above a person's typical level of a certain behavior. And so that's what we tried to do here. So we looked at what we called stressor-related drinking and its ability to predict future alcohol problems in a sample of university students. And we were really kind of informed by models of alcoholism that talk about alcohol as part of a coping process for some individuals. There's theory going back to the 1950s. It's called the Tension Reduction Hypothesis. It talks about one of the reasons that some people become alcohol dependent is that they start to rely on alcohol as a means of coping with day-to-day life. And its ability to remove tension creates a reinforcing pattern that causes them to abandon other more positive or healthier coping methods. And so we were sort of interested in whether or not we could capture that in the day-to-day lives of young adults. And the thing that's important to remember is that people differ greatly in this tendency. You might expect some people to really increase their drinking when they're stressed, and others not so much because everybody has a different way of coping with the challenges of day-to-day life. But this idea that alcohol use as a related to stress has predictive ability above how much a person drinks was what really motivated us to do this. So I've talked through this part. So yeah, we were really interested in trying to figure this out in day-to-day life and see if we could do this with a sample of young adults. And so to tell you a little bit about the study, it's called the University Life Study. The goal was to, overall, was to just sort of look at the links between motivations, day-to-day activities, and experiences among college students. This was run by Jennifer Maggs, who's a faculty member in human development and family studies at Penn State. It's a large sample. It's about 725 students, ethnically diverse, African-American students, and Hispanic students were oversampled, given that Penn State is a largely white student body. And so this study is actually what is referred to as a measurement burst study. And that's typically where you embed an ambulatory assessment study of some type in a more traditional longitudinal study. So you've got not only differences between people that you can look at, but you can look at sort of burst level processes that you would look at in a traditional longitudinal study. So at each assessment, you could look at how a trajectory is unfolding across years. And then at each of those time points, you could look within the burst and try to understand some of the contextual dynamics that are sort of propping up those trajectories. That's how I like to think about it. They're sort of pushing them up or down. So essentially what was done was that there were 14 day daily diary bursts. And so each semester, students were asked to, for 14 consecutive days, fill out a daily diary that asked them a number of things. The things that we were mostly interested in were stressors and their alcohol use, and look at the day-to-day associations between these. And so the study went from freshmen to senior year, and the retention was rather high up to 85% of students were retained throughout. So just to give you a sense of the stressors that were asked, so this was taken from Dave Almeida's work in the MITA study that I referenced earlier, where he has developed a stressor inventory that has been used extensively in daily life with young adults and with older adults. And it taps stressors in a variety of domains. So essentially they were asked each day, did they have an argument or a disagreement? We called this one interpersonal tensions where they could have argued about something but decided to let it pass to avoid disagreement. Anything that sort of happened in the professional or academic realm, anything happened where you live that most people would consider stressful, friend or family stressors, anything happening to a close friend. And then just the sort of general stress question, did anything else happen to you? And if you look across days, you find that 20% of days had at least one of these occur. And of those days, 60% were single stressor days. So we did use a count of stressor severity. But we did find that about one out of five days there was some sort of stressful event that the young adults reported. So the first thing that we did, so I'm just going to show you, I'm going to pick out the results from the model that we ran. First thing you see here is that students were more likely to drink on high stress days compared to themselves on low stress days. This effect, so this is an odds ratio of 1.08. So you're really talking about an 8% increase in the odds, which is pretty small, but it is significant. But remember, that's the average. And what we also saw was that there was variability in that estimate. So for some adolescents or for some young adults, it was a lot stronger than for others. And for some, you actually saw it was positive, others negative. So we were sort of interested when we saw this in trying to turn this into a variable in and of itself. And to try to see, well, are the students who are more reactive to stressors in terms of their drinking, are they at greater risk for future alcohol problems than those who are less reactive? The theory that informs that kind of thing that we were interested in is this idea of behavioral signatures that was offered a while back by Walter Michele. And it really came out of the person situation debate and personality, where the person side of the argument was saying that traits really differ between people and what matters is that one person is more neurotic than another. And that the extreme version of this is that they'll be more neurotic across all situations than the person who is less neurotic. Whereas the situationists are saying, no, no, no. It's all about the situation that you're in. That determines when people are neurotic and when people are not. And at the time, Walter Michele's idea was to say, well, it's really about how you shift, your characteristic shift across situations. And he talked about it as what he called a stable if then contingency. So if I become stressed, I tend to increase my drinking. Whereas you could talk to somebody else and they'd say, if I become stressed, I tend to decrease my drinking. And he was sort of saying that this is the heart of where individual differences lie, is really in how people react to their environments or interact with their environments is probably a better word on average. And so we thought of that variability in the association as perhaps representing a behavioral signature. And so we felt that it plugged right into this theory pretty nicely. So I won't belabor this too much, but I just want to show you what we did so that it's transparent. So here we have a model where we are predicting the odds of whether or not you drank that day. And we're primarily interested in the number of stressors you experience that day. So the average relationship, again, is this. Where we're essentially saying, for the typical person, what's the increase in the odds of drinking on a high stress day versus a low stress day? And then we also have a random effect where we allow that association to be stronger for some students than it was for others. So we can think about the slope for each person as being the sum of those two things. So if the average is here and that random term, essentially, if this term is negative for person K, it'll drop that slope down. And that will be the association for that person. Maybe it's flat, maybe it's negative. If it's a positive term, it'll kick that slope up and make it higher than average. So if we visualize this, this is a random sample. This is about 10% of the sample, actually. So about 70-something students. And this black line here is the average slope. That's the slope for the typical person. And then each of these lines is the individualized prediction. And so you can see that if we pick out this person, we're seeing a steeper slope. This person tends to increase their drinking when they become stressed. So that's the number of stressors go up. Law gods of drinking increases. If we pick out this one here, we see an opposite. This person tends to decrease drinking. And then if you look at the bottom here, this person, their drinking behavior is not associated with stress at all. So essentially, we're not asking how much they drink. We're asking what their drinking is related to. Yes, do you have a question? Is this all the data? Because there's just one subject. This is a random sample of about 10%. Yeah, so it's only about 70-something participants. I did want to hit all of them. But 700 lines, you can't pick anything out. This is pretty representative, actually. Most increase. And that's what we saw when we looked at the average slope. But what you do see is that there are some people who are pretty flat. And so putting those in with these kind of reduced the average to the point where it was a modest increase at the average effect. But if you actually unpacked, you see that for some people. That's why you're just talking to me. Yeah, so we thought about other drug use. This sample, though, and I mean, this might not be true anymore. But at Penn State, and these are Penn State students, drinking is really what people do. I mean, it's like there's a whole football drinking culture. And so they do tend to specialize, I think, a little bit more than other areas of the country. So we weren't all that surprised, but it's interesting. I mean, actually, I don't know for sure. I can't prove it yet. But I have a suspicion that these people are maybe people who, I don't know, hunker down and fix the problem and whatnot. But another thing that we're interested in that's highly related to how you would react to it is what domain of life is it in. So if this is a controllable school work stressor day, then it's fixable. And maybe you would drink less on those days. Maybe if it's a health related stressor, you drink less on those days, because you don't want to mess up treatment or whatnot. But that's in the future. We're going to do that in the future. But I think this raises a lot of interesting questions. Yeah. I may have listed. Did you explain the why access? Why are they on native? Yeah, thank you. So this is log odds. And the reason is that because it's a logistic model, in order to get the lines, you have to take the log of the odds. So it's a little bit clunky. But what we ended up doing when we used this as a predictor was just took a z-score of this. So basically, your line is defined by a slope coefficient for each person. So whatever rate of change is this line, that's your score. And then when we make the z-score, it's really are you above or below the mean and by how much? That's also the problem with substance use outcomes, is they never give you, there's no normal distribution. So you have to do stuff like this. But anyway, so we looked at those stressor related drinking slopes in the first three years of college and then predicted the alcohol use disorder identification test severity score. So we're essentially trying to understand. And we actually took out items that were related to heavy drinking because we didn't want drinking predicting drinking. We left in items that were more related to things like dependence, like did you drink in the morning? Did you have trouble stopping drinking? Things like that. And we actually found that even after putting in some pretty powerful controls, so we just put in what's your drinking probability when you're not stressed to try to capture the difference between heavy drinkers and non-heavy drinkers. What's your alcohol problem history? Do you have a history of alcohol problems? And you can see these things have pretty big effects. So you're looking at a 71% increase in the severity with each standard deviation increase in your history. But that drinking slope is still predicted. And this was also interesting to me. It wasn't how much stress you were exposed to. It was what you did in relation to it. So I touched on this briefly before. But now I'm going to transition into this data collection effort that we're doing where we're combining self-reports of alcohol use with an enclet that measures intoxication. And we started doing this. I remember we were in the methodology center, which is a center I'm part of at Penn State. We have these brown bag seminars. And it somehow came up in discussion that there's all this interest in using EMA for alcohol use. And we're going to ask people to report their drinking behavior while they're drinking. Doesn't that sound like it's a problem? As they become more and more intoxicated, should we trust those reports? Do we want those reports? But when is too much? When can we trust them? When can we not? And if you're interested in young adult alcohol use, you have to be especially concerned about this, I think, because a heavy drinking episode is not a 4-5-plus drink episode. It tends to be about double that. This may even exacerbate the problem. And so I haven't seen much writing about this. I mean, you might expect that people under-report, stop reporting. They might over-report. They might exaggerate. But we don't really know. And so it just seemed like an area that was sort of ripe for trying to understand just basic sort of methodological questions, like how good are people at this age group in reporting how intoxicated they are, how much they've had to drink? Does that change throughout the course of an episode? And it turns out there's been a decent amount of work in continuous alcohol monitoring. And some people, so there are a number of sensor-based devices. You can wear them on the wrist. You can wear them on the ankle. And what they do is they take intoxication readings through the skin. They can do this in the field. And some people are starting to call them a gold standard if you want to measure alcohol intoxication in the field. And there was some suggestion in 2015 that, hey, maybe we should consider using these as some sort of way to benchmark the self-reports that we get from people's day-to-day lives. And so this is the device that we have been using. This is so that secure continuous remote alcohol monitor or SCRAM for short. And it's worn on the ankle or just above the ankle like this. And as you can see, it's a pretty big and clunky device. We went with this one because there is the most extensive literature establishing its validity. It tends to do better than wrist-worn devices because of the fact that wrist-worn devices move around a lot more. And this device has been used extensively in forensic settings, where if individuals sort of, they get in trouble with the law for drunk driving or for other alcohol-related problems, they'll often be given the option of instead of serving a longer sentence or being on a longer form of probation, they can wear one of these devices. They can demonstrate their abstinence for a certain period of time and then get a lighter sentence. So they've been used extensively in that way. And they work on the fact that about 1% of the alcohol you consume actually is eliminated through the skin. So there's a fuel cell sensor that sits on the inside of here. It collects that data every half an hour so we can get nice alcohol curves for each episode. And then once the device comes back to it, we upload all that data. And we have often a number of alcohol-related episodes for each participant. So we embedded this sensor in an EMA design. So the Xs are the frequency of measurement. And this is just to show you that the anklet is always measuring when it's being worn. But the actual frequency is about every 30 minutes on average. So it's the densest measurement. We also got three times daily EMA reports where we're asking them in the morning, what's their mood? What are their expectations for the day? A little bit about the drinking episode the night before if they indicated that there was one. And then we did sort of a check up in the afternoon and then again in the evening. And the I think most interesting and perhaps riskiest part of the study is this. We are asking them to initiate a survey sequence when they start drinking. And so we train them in the lab that we sort of say as soon as you take the first sip, we want you to press this button. So on our app, there's a button that they press and it launches a survey sequence. And that's sort of what we're doing here with the beer. That indicates you've started. And then the assessments that follow, we want them timed at a similar frequency to the anklet. And the whole idea here is to get both anklet reports and self-reports from the episode. So the challenge here is we have to keep these really short. And so we do. We only ask a few questions. We ask them how many drinks have they had in the last 30 minutes? How intoxicated they feel? We sort of ask them how buzzed are you? How drunk do you feel? And they're just sort of likeert scales. We ask them a little bit about where they are, who they're with, so we can understand a little bit about the context. Are they in a bar? Are they at a friend's house? Are they by themselves? And then a few positive and negative affect items to start to understand the micro-level interplay between mood and drinking behavior. And so this is hypothetical data, but it sort of illustrates the concept. So here is alcohol concentration measured in TAC units. So TAC stands for Transdermal Alcohol Concentration. That's what you get from the device. And it's analogous to a BAC rating. So if you're above 0.08, you're over the legal limit. And we plot it against the time since the start of the drinking episode, which ideally we're going to get from their self-report. And so a lot of the drinking episodes that we've seen so far look kind of like this. There's a rise. They hang out at the maximum for a while. And again, how long they stay there varies. And then there's a decay. And typically the increase is a bit more complex than the decay, partly because decay is controlled tightly by liver function. So this is what we might get from the sensor. With the self-reports, we can actually calculate an estimated BAC. And there's formulas in the literature that do this. And the idea that has sort of led us to do this was, well, are there points at which the two stop tracking? So as you can see here, at higher levels of intoxication, and again, this is made up data, but the estimated BAC is just it's sort of hanging out at the same spot. They're not telling us how much. They're not telling us the same information that the anklet is telling us. And this is one of many ways it could go. They could overestimate, so we could start to see an overshooting. There's a lot of different things that we could see. But the idea is we want to know not only where their self-reports are not trustworthy, but where they are. And additionally, so we're going to have, if we go back, we're going to have these curves here. So we'll know a lot more than how intoxicated they got. We'll know how quickly they got there. We'll know how long they stayed within a certain margin of that. We'll know how long it took them to decay, how many peaks in the distributions that they have multiple episodes linked together. And so we could use quantitative methods to sort of smooth out these curves and pull features from them to sort of think about not just how intoxicated they got, but the dynamics of the curve and how these sort of add to our prediction of alcohol-related consequences that we might get from the morning after. So here, I've started with a simple example, just like intoxication level, we'll obviously put that in. Maybe it'll be the maximum intoxication level. Maybe it could be some area under the curve, scaled by the number of hours they were drinking to sort of think about intoxication output, so to speak. We could also think about the rate of change at any time point. What's the average speed, average velocity at which they're becoming intoxicated? We can start to, and we're really scratching the surface of this, but we can start to unpack this curve and pull many dynamic features from it and try to come up with models that do a good job of predicting whether or not this is going to end up being a bad episode. And I have a particular interest in trying to do this as early as possible. Maybe we only take the first hour or two of the curve. We'll have less data, but maybe there are parameters we can pull out that can tell us early on this is going to end up being a bad drinking episode versus not. Now, we can't do this in real time. And if we did do this in real time, this would raise a whole bunch of ethical issues, I think. But knowing this information, I think, could be useful for prevention efforts. But again, I still think we're a pretty long way off from it. Some other things that we're doing in the study, we are measuring cigarette and other substance use, even things like Red Bull caffeine, because we're interested in that disjunction between your subjective intoxication and your actual intoxication, and thinking about that as a danger zone. If you are drinking a lot of Red Bull vodka, for example, you might feel a lot more alert and a lot less intoxicated than you actually are. Is that disjunction? When does it appear? Is it associated with poor outcomes? How big does it get for whom? All sorts of interesting things like that. I'm also interested in the social contextual and affective factors prior to drinking episodes. When I first started doing this, I thought that maybe we could use these sensors to inform some sort of real-time intervention. I'm not sure that that's possible anymore, partly because by the time the sensor picks up a reading, it's already gone through your system. So it's about 30, 45 minutes to an hour past. You actually being that intoxicated. So it's actually a little bit delayed or lagged because the alcohol has to get eliminated from your system for the device to register it. So we might already be too late. But perhaps there are other factors that we get from the self-reports before they even drink that are associated with a dangerous drinking episode and maybe change in those factors, particularly something like affect where we're starting to see that there's anticipatory effects of affect before people even start drinking. You tend to see spikes in positive affect and dips in negative affect. So maybe there's certain intensities of those dips that can help us predict when drinking episodes are going to happen and when they might be dangerous drinking episodes. And then always looking at sort of individual differences at baseline as moderators of the dynamics that we see in the day-to-day life studies. So things like drinking history or alcohol use risk. Personality, coping strategies, you name it. So I just want to finish by saying that we've got about 65 participants so far. Now we are screening for participants who tell us that they engage in a heavy drinking episode at least once a week. And that's probably why we have about 230 something drinking episodes from them, even though we only get about four or five days worth of data from each person. So we do have a lot of information to work with. We're going to continue to go up to about 150 participants. We've seen some very long drinking episodes. And this is actually going to end up being an analytic issue that we're going to have to sort out, especially if we want to talk about the dynamics of the drinking episode and what one means compared to the other. We have to deal with the fact that there's a tremendous range in the amount of time that people spend drinking or that each episode lasts. So I'm just going to finish by saying that, ideally, so there's a couple of companies that are starting to make alcohol trackers that would be worn on the wrist like Fitbits and would fit a little bit more snugly than traditional ones. We've been waiting for these because these just sort of seem like the next step. I've always been a little bit skeptical about that anklet, not in terms of its ability to measure, but in terms of how it might change a situation, how it might change a person's drinking patterns. I don't know if I put this in the talk, but I didn't. But we asked participants at the end of the study to tell us a little bit about how the anklet made them feel when they wore it, how they think it affected their drinking behavior. And one comment that really stuck out is that one person said that the anklet kept drinking at the forefront of their mind. And I think even though it only came from one person, it's something that I want to take seriously. Because I think that one of the challenges with this kind of work is that by the very act of measuring it in such an intensive way, you may be changing the behavior that you're trying to study. So we need to be very careful and thoughtful about that. So I think I'm going to stop there. Quick questions or comments? And then ideas to make this page escape? I don't want to shoot. I don't want to shoot the video. Thank you. Yeah. I'm just curious. What do you consider to be a drinking session? Because over there, I saw like 91 hours. Like how? I don't know how that's possible. Are you drinking for two days? Yeah. So what that means is, let me see if I can go back. OK. So it's not necessarily that they're staying up and drinking the whole time. I hope not anyway. What that means is that it never hits zero for two or more readings. So basically, that person drank and then there was an elimination curve. And then it hit zero, but then immediately spiked back up. So at some point, maybe around here, I'm not sure they woke back up, started drinking again. And so on some of the weekends, you see data that looked like that. And so that also presents the challenges of how we treat that. Do we treat that as one episode? Do we treat it as multiple? Where do we cut it? But that's what that means. Yeah. Just briefly, I really enjoyed your talk. Oh, thank you. Yeah, I benefited a lot from it as well. I'm curious that what you see is the future. I just want to zero on a one piece, the candidate gene approach. Because these approaches have been critiqued in various ways. And I was just curious, from your perspective, what's the future of this kind of approach? I was also curious how you made, I understood your rationale for using the VRB-47, but I was just some other possibilities that he popped to mind. And you could potentially have chosen a number of them to test and so forth. So just a little bit of comment on that candidate gene approach. Absolutely. Yeah, we, so at the time, I was interested in Jay Belsky's work on differential susceptibility. And he had a list of VNTRs, which are polymorphisms that have a certain segment of base pairs that's repeated a certain number of times. And so he had the DRD4, he had the serotonin transporter, he had a few others, and sort of usual suspects. And so we looked at his list, and he sort of made the most compelling case, I thought, for DRD4. So it was really chosen on conceptual theoretical grounds. And it was heavily influenced by his work, and then some of the work by Tom Boyce and Bruce Ellis on the biological sensitivity to context model where they mentioned some of those things. And they eventually, their theoretical perspectives kind of merged. And so that's where I got the seven repeat allele from. And the lab studies were, I also was sort of noticing this convergence across study designs, which to me seemed like good evidence that this might actually be a good lead, a good place to go. The future of this work, so I have colleagues that do a lot more genetic work than I do. And there's a lot of discussion about what is the next, what is the way to approach this? Because often with the candidate genes, the effect sizes are so small on average that anything you find is probably an outlier. And it might not be representative, which our study suffers from too. That could absolutely. And I would be concerned about that. And I think that the thing that differentiated our study from most is that we took this special care to measure environment in a certain way. We looked at a within-person process. And so I think it helped us out. But I think if we tried to do it another way, we would have had some issues. I think so the way that it's being discussed right now is to take hits from a GWAS. That's exactly, yeah. Michael, you need to know some of that. Yeah. You look at those become new candidates as they confirm older candidates. And then you can start theorizing them in different ways. That's your thinking as well. Yeah, I think that's a good way to go. But I think the criticism that that approach takes often is that it's a theoretical. And I'm not saying that that's my criticism. But this is the criticism I hear a lot from colleagues who do more of this work than I do. And there are some people that are starting to say, well, maybe we should look at regions of brain activation, figure out which genes are expressed within those regions, and make scores from that. If there's a region that's associated with a behavior or an experience or whatnot. Thank you. I appreciate it.