this is a recording of a live online workshop on time during effect modeling or TV the workshop took place on February 21st 2017 and was presented by Stephanie Lanza scientific director of the methodology Center and professor of bio behavioral health at Penn State the recording consists of a one-hour presentation followed by a one-hour question-and-answer session during which dr. Lanza answers questions posed by workshop participants questions that you submit on your chat window and so we’ll just go through those at a nice pace after after I’m finished going through the slides so really happy to be here today and we’re also interested in your feedback as Bethany said on the format of this type of session so let’s see so just to begin with I know you can’t see me but but I will pop in later when we’re doing a Q&A so that you can see me talking with you but my name is Stephanie I am a professor of vitamin in rural health at Penn State and I am the scientific director of the methodology Center and we are largely funded by NIH in the night at the National Institute on Drug Abuse in particular to advanced statistical methods for Public Health Research so again welcome we’re going to talk about time varying effect modeling to study developmental and dynamic processes so first we’re going to just do a little bit of background sort of how we came to think about Tevan time varying effect modelling and and put it to use in in real in the real world and I’d like to demonstrate the time-varying effect modeling through two studies the first study is going to be focused on dynamic change and what I want to really focus on nicotine addiction in the end the process of withdrawal when you quit smoking so that’s going to be a very intensive dynamic process under study in study two I’m going to show a totally different use of the time varying effect modeling that will have a developmental perspective and we’re going to look at the use of e-cigarettes and traditional cigarettes in adolescents across across age and then I’d like to just kind of conclude with a few thoughts about next steps for 40 them and related methods okay so let’s talk about the background Tootsie them just a little bit time varying effect modeling oh it’s Auto advancing okay so um so first of all human behavior as it relates to health is dynamic we’re gonna set it turn off the auto advance hang on second okay great so so human behavior as it relates to health is in part dynamic its dynamic in in many ways and so this orientation to thinking about health human health as dynamic is relevant when we think about behavior change so how behaviors change across age as we develop and mature and also how we think about how behavior changes across real time but in addition to the actual behavior changing across age or time we may be very interested in changes in processes across age or time for example we might be interested in thinking about risk factors for adolescent substance use but those risk factors may be more salient when you’re 12 years old than they are say at age 17 and so the time during effect modeling approach allows us to unpack this kind of question and a third reason why I think this dynamic perspective on the human health and behavior is a really important is that many of us have analyzed data from behavioral interventions or actually run some of them and we know that an intervention effect is the intervention is on or off but the effect of the intervention is not black and white it’s not on and it’s effective and then it

stops working or it stays working forever we know that intervention effects are actually dynamic and they work better for certain individuals at certain age and they work as a function of time real time whether it’s time in terms of minutes or time in terms of years so let’s start with one motivating example and this is coming back to the smoking cessation study I alluded to so we know that negative effect are bad mood and craving to smoke are tightly linked among addicted smokers and we know that when addicted smokers stop smoking they go through withdrawal and hopefully move toward recovery this process of recovery is not it doesn’t happen suddenly you have to go through this process and we can use time during effect modeling or team them to to take a closer look at this recovery process so questions we might use T them to answer related to smoking cessation and withdrawal are is and is one’s a negative effect differentially associated with their craving at various points in this looking cessation process so you have a bad mood you get into a bad mood you want to smoke a cigarette we know this happens after you quit smoking but in non-addicted smokers that effect is absent and so presumably when people quit smoking that link that association between a negative effect and craving to smoke dissipates with time and eventually as they return to normal it’s no longer associated that’s the kind of question t them can be used to address also if you are a randomizing people to different smoking cessation interventions and then they’re asked to quit smoking on a target quit date then we might be interested in how the process I just described above the time varying effect of negative effect and craving and how that unfolds during a smoking quit attempt that might look better for people getting the bupropion medication compared to those who are only going to counseling for example and the effect of the intervention may also unfold with time so we’ll take a closer look at that exact example in a little while so first let’s think about a thought exercise so oh sorry why is it he said okay so great so let’s just do a thought exercise for a moment so suppose an intervention was conducted a behavioral intervention that was designed to reduce alcohol abuse and you’re gonna perform a two-armed randomized controlled trial right so we know how to do this we conduct a behavioral intervention people are randomized to get treatment or to not get treatment after the treatment is administered let’s say that after the treatment we have one post baseline follow up and see well did the intervention work yes or no or maybe we follow them twice after the intervention and we see did the intervention work at the first time post baseline did it still did that effect still exist to that the second time post baseline and maybe we actually have many time points measured post baseline and we could think about the intervention effect unfolding over time it’s sort of in a sense of a moving window of time so let’s look at that graphically so in this in this graph oh my gosh so where’s that chocolate here okay so um so in this intervention lets us again this is a heavy drinking reduction behavioral intervention and the placebo group at time one after baseline so after after the post-test assessment the placebo group has a rate of heavy episodic drinking in adolescents of safe point forty five forty five percent are engaging in the behavior and in the intervention group we observed that only about 25% are engaging in the behavior and let’s say that that is a statistically significant difference so great we conclude that our intervention worked and maybe we take it to scale let’s suppose we followed them an additional follow-up time point now we have time one and time two and so we have the same data from time one where we have about forty five percent in the

placebo group still engaging in binge drinking at time one post baseline and we see at time two that eats up a little bit to about forty seven percent engaging in heavy upside drinking and in the intervention group we see that there’s the intervention effect the time line as we had observed and we see that the intervention effect weakened slightly by time – so now we have a full thirty five percent in the intervention group engaging in the behavior and so the difference is smaller and it may or may not be still significant at time two okay so this is a time varying intervention effect now let’s suppose that we actually observed individuals adolescents at ten time points after the intervention maybe we we checked in with them every month in the year after the intervention was was conducted and we see the same data at times one and two and then if we track it over time we follow it up at time three four five six seven eight and we see that the intervention effect and weakens until there’s no effect at all between the intervention and the placebo groups but perhaps a real-time a is when they are graduating from high school say and then they’re heading into the summer and lo behold we see that there’s an intervention effect at time nine and at time 10 so it rien so maybe there’s this delayed effect of the intervention when in the adolescence context changed and if we had not looked for a time-varying intervention effect we would never understand the process of the intervention effect unfolding over time so as I said we see the same data at times one and two the intervention effects seem to go away and then it re-emerged after in this case a transition out of high school okay so let’s talk a little bit more about the time varying effect model so in a nutshell teve M is a direct extension of regression all of you I certain have used multiple regression analysis many times in your careers and if you understand linear regression you understand T them and you will within the next 15 minutes you’ll know just what it is it’s a it’s a really natural and intuitive extension of linear regression okay so the the work for developing team M was funded by our p50 Center grant from the National Institute on Drug Abuse and so first of all why do we collect longitudinal data at all well if you ask any behavioral scientists likely they will say that we want to capture and quantify temporal changes in some sort of outcome and potentially changes in time varying covariance okay great I’m a big fan of much two-nil data it’s how we want to study the emergence of behavior problems but also I would argue it may be very natural for us to expect that not only do the outcomes and the covariance change over time but the associations between covariance and outcomes may change over time the association’s may change over time and in regression we’re talking about regression coefficients rate slopes so the time varying effect model is designed specifically to evaluate whether and how associations between variables change over time so let’s look at it mathematically really quickly here so regression coefficients as we know from linear regression Express the associations between variables between X’s and the Y so a traditional bivariate regression predicting our outcome Y from our Cabiria X looks like this where we have Y expressed as a function of an intercept and a slope that’s associated with X and an error term so we fit this with the goal of interpreting and hypothesis test for the significance of beta 1 so how do we interpret beta 1 in linear regression okay this is a great quiz for my regression class students write the slope is interpreted as the expected change on the Y corresponding to and one unit change on X and so it’s a single number summary of the association between x and y in our sample it’s a single number summary T of M takes those coefficients and regression and allows them to be dynamic so I’m showing you the regression equation now for Team M so we’re starting again with bivariate regression the association between x and y is what is of interest but if we believe that

that association that beta 1 is dynamic and changes over time then we can express antivenom a pair of parameter estimates to be estimated as a non parametric continuous function of time so I’ll show this to you in a little bit more concrete way next so this blue curve on the figure shows a hypothetical regression coefficient that is estimated in TEM and again it’s a non parametric continuous function of time does this mean that we have time continuously measured no we don’t in real life we never can assess individuals in streaming real-time unless they’re wearing some sort of wearable device for Pasadena capture but this is I’m talking about serving data maybe it’s intensive longitudinal data or maybe some individuals in this study were assessed at times 1 6 and 14 and others were assessed at times 6 13 and 19 so but we have coverage across that time and the most important thing is that our research question embodies an interest in the Association changing with time so Tifa musta makes regression coefficients as flexible functions of continuous time and I’m talking about the intercept changing as a continuous function of time and also the slopes or the associations between X’s and why okay and so remember how in linear regression I said that the slope for the effect of X on Y was a single number summary in T them I’m showing you here a sample regression coefficient function so we know we no longer estimate regression coefficients we estimate regression coefficient functions so instead of putting a an estimate for beta 1 in our tables that we love to do in our manuscripts instead we are publishing lots and lots of figures and this figure represents one parameter coefficient in the in the model it just happens to be varying with time so you see time is on the x-axis whatever however time is coded in your data and the regression coefficient is in the is in the y-axis okay let’s step back for a minute I just want to catch all of you up to speed of your hearing today about time varying effect modeling it’s a really new procedure it is it is starting to be used more and more it’s taking off very rapidly but I want to give you a sense of sort of the history here since we’re able to UM capture it were so early in the process so in the 1990s there was a technique of functional regression analysis that was introduced in the statistical literature and it never made its way into the literature that behavioral scientists read but a lot of the fundamental the fundamental work in this area of tevonn was done in the 1990s by statisticians um it wasn’t until 2010 when the methodology Center released its first SAS macro for estimating time during effect models under the direction of our esteemed colleague Rennes Ely and you’ll see a picture to the right on your slide of the methodology centers software download page for T them so it’s you can get there from methodology psu.edu if you want to check it out then you can click on free software and you’ll see the page 40 men and our software is free to download if you’re a sax user it’s very very simple to use so this was the first time the TBM was that become available to behavioral scientists now in the early years in our early 2010s we were motivated to use T them to analyze intensive longitudinal data so in 2012 we started to put the word out there for how team M could be used to to estimate regression coefficients as they unfold across intensive time using data from ecological military and so we began to to really work in this area primarily in the area of smoking cessation nicotine and tobacco research and then in 2013 we began to push this direction a little bit further and we started to introduced even for other applications as well and it turns

out our belief is that even is has a very very broad broad applicability to all of our behavioral and health research that goes far beyond the analysis of ecological momentary assessment and we published a special s supplemental issue of mixing and tobacco research that was focused on EMA methods and this was the first year we started to see uptake of Tevan by other behavioral and health researchers so this is an exciting time for Team M and then we offered our first Summer Institute on innovative methods on this topic in 2015 and so the software continues to evolve we added random effects we’re working hard and getting ready to release some version of team m4 that includes complex serving weights so the software development continues as through the application areas that we see 40 them so let me walk through two applications with you the first is talking about smoking cessation recovery as a dynamic process so this is a study of nicotine addiction and it’s in recovery from it we know that tobacco use is a leading cause of is the leading cause of preventable deaths globally and this is important you know people try to quit all the time 95% of cessation attempts end in relapse with withdrawal symptoms being listed as the primary reason so we argued that an improved understanding of withdrawal symptoms and how treatments can alleviate them and of course as you can imagine I’m thinking about the dynamic process of withdrawal might lead to new treatments or inform tailored treatments and treatments that could be tailored to individuals with certain characteristics as well as tailored to certain times in their quit attempt so the overall goal of this particular application was to apply innovative methods to existing data from a randomized control trial on smoking cessation to gain knowledge that could inform the next generation of more nimble and adaptive tailored smoking interventions this isn’t the broad goal of the project so let me tell you about the data that we were analyzing and and just for the record to NIH is very interested in funding secondary data analysis grants and that’s how we funded this part of the project so it’s a great way to think about I’m answering new questions in your research using TM to try to get that funded by NIH so this is data from the Wisconsin smokers Health Study in a comment wishes and and this is a very wonderful data set that was collected by our colleagues at the University of Wisconsin on Meghan Piper Tim Baker Mike Furai and so they had a very large sample 1500 for daily smokers enrolled in a cessation trial there was a randomized trial funded by the National Cancer Institute and so for the purposes of today I’m going to talk about two groups there were six treatment groups but that we’re going to just talk about two groups placebo only group and the treatment group which had some combination of bupropion medication the nicotine lozenge and the nicotine patch and importantly in this study they did real-time assessments of the dynamic phenomena of interest the withdrawal symptoms mood behavior context smoking bans anything that might be relevant to their momentary risk of smoking lapse were assessed in real-time assessments on a mobile device so here’s a depiction of the of the study design so um after the orientation session they started a three three week countdown to the target quick day and at week two weeks prior to their target quit date are planned quit day they received their their wearable device their mobile device and started to provide EMA you look ecological momentary assessment on their device four times per day and they continued that through the target quit date and through two weeks after the target quit date and so you could imagine the very interesting dynamic processes that might be leading up to a very stressful quit day and coming out of the quit day through the recovery of withdrawal and then they they also collected and followed up people for twenty up to twenty six weeks and I think even beyond that and the EMA that we’re going to be analyzing here I said they were assessed

for occasions per day as soon as they woke up they initiated a survey they were surveyed at two random times during the day they were beeped so that they were sampled in randomly and time in space in their natural lives and then at bedtime and this was done every day for several weeks so the specific goals of this study were first to study the underlying dynamics of craving during a smoking cessation attempt and to to estimate the effect of an intervention on not just reducing craving but the effect of the intervention on decoupling craving from its key drivers for example negative effect and details are appear in the published articles and from 2014 so the measures in our study I’m something you’ll see in both of these studies is that the variable list for both of these studies is quite short that the datasets are very small but the results are quite nuanced so the variables in this study there were four of them the outcome which is your craving to smoke and this was captured during the first two weeks after quitting smoking and it was intensively assessed so for 14 days more times forgetting we had two predictors one of them was baseline nicotine dependence so this is a this is a static characteristic of the individuals at time of randomization to treatment they were their level of nicotine dependence before they quit was assessed so again the predictor is not time varying but its effect certainly can be time varying and we’ll see an example of that in a moment the other predictor was in fact time varying and it was their momentary moved their negative effect specifically okay and then we had a moderator that was the placebo versus the intervention group so we have four variables next we specified our team M our time varying effect model and so we had to think about within each intervention group what do we really think varies with time we have to specify this as our scientist right so first of all of course we think that our mean craving over the two-week period after quitting changes with time steps our intercept so we are not going to estimate an intercept we’re going to estimate an intercept function as a function of time second negative effect itself that is changing with time and we would like the effect of negative effect to be changing with time so we’re going to estimate a slope function or a coefficient function for the effect of negative effects on craving and finally we have baseline dependence and we want the effect of baseline dependence to be allowed to unfold with time during the quit attempt so a second slope function okay so that’s how we wouldn’t think about specifying our model then we have to actually understand what team M is estimating and so the equation looks something like this we have craving for individual time T is expressed as a function of it an intercept that varies flexibly with time plus slope as an effect for negative an effect for individual I at time T and again that slope that’s allowed to vary flexibly with time and finally we have an effect of baseline dependence DEP for dependence for individual I it’s static across time but the slope is allowed to unfold flexibly with time and we have our error term so if you’re used to looking at multi-level models this is going to look quite similar for repeated measures but it’s not the same thing and we are we are actually here estimating linear regression where the coefficients are a non parametric functions of time so here’s one of the results this is the effect of negative effect our momentary negative effect on craving and you will see that I’m so the placebo group coefficient function the placebo group is in red and the coefficient function for the treatment group is in blue so let’s just orient ourselves to this figure so these are these are regression coefficients on the x-axis we have the number of days since quick dates so it goes from 0 to 14 and on the y-axis we

have the effect of negative effect so negative effect and craving were continuous variables so these are regression coefficients and betas and so we can interpret them so let’s let’s try to interpret them so for the placebo group on day 0 at the very beginning or right after they quit the association between negative effect and craving is significantly different from 0 and in fact the coefficient is almost 2.0 so forever you want you to hire a negative effect they are expected to be nearly two units higher on craving in contrast for those in one of the treatment groups they’re grouped together here the association between negative effect and craving right after they quit smoking so at day zero that association is much smaller it’s about 1.3 and you’ll see by the non-overlapping confidence intervals that the effect of negative effect on craving was significantly stronger to the placebo group for a day or two and then they were the same so we didn’t think about the difference between placebo and treatment effect varying with time then we would have never found excuse me but would have never seen this time hearing effect excuse me here’s one more finding so this is for baseline dependence remember your baseline dependence is a single number that everyone has at the beginning of the study and you know we expect that the effect of baseline dependence on craving may be pretty static and in fact for the treatment room those receiving some sort of treatment and large pharmaceutical or nicotine replacement treatment the Association is significantly different from zero so the more you depend the higher your dependence level at baseline the higher your craving and in that relationship is very very studied throughout the day two-week period so what we see is in the placebo group those not receiving treatments after the first week we see a man elevation in this association so the placebo group actually their baseline dependence is very intermingled with their craving their momentary craving after that first week of burnin time and they’re still abstinent they’re struggling more they’re likely struggling more if they have higher baseline dependence and they are in the placebo group so implications we drew from this first step one thing that I like to focus on is I think that TV this shows how we can use T them to think differently about intervention effects so with time intervention can change and intervention can change relationships between variables it doesn’t just reduce the outcome but it might change key relationships over time and in this case we see that the intervention diffused to the role of negative effect a key driver of craving very early in the quit attempt and it was only active and it was only an active effect for about two days but those two days could be critical crucial for individuals to get over the hump and they’re quitting attempt and broader implications I think that this shows how we can think about the effects of static variables baseline variables it could be race ethnicity variables anything that doesn’t change with time but the effects can can evolve over time and I think this is very much the case for GE genetic variables and gene expression can change over time and age also the effect of a treatment in a standardized a standard RCT can be time varying as well so we don’t just say does the treatment work yes or no we say when does it work and it allows us to model a potential intervention process that we might posit and again I think that this information could help get us one further step down the road toward tailoring treatments to individuals with carrots certain characteristics and also to time okay great so let’s step back out for a minute that was an example of how team M can shed light on dynamic processes in this case modeled with data that were intensively assessed within individuals

now we’re going to step back and look at me cigarette use in adolescents today and take a very developmental perspective and I say that with a smile on my face because we’re going to be using cross-sectional data but it’s a very much a developmental process that we’re that’s under study here so this work was funded by an ro1 awarded by NIDA to analyze existing national data to advance our knowledge of substance use and it’s corliss from again the etiology or a developmental perspective okay so why eat cigarettes well this is a these slides are about a year old and so there’s they’re already badly out of date as the e-cigarette literature emerges very rapidly ISA grits were developed originally as a reduced term product so they were considered a safe alternative to traditional cigarettes that was that was the assumption that was their goal these are inhalation activated devices and so when you inhale there’s a heat produced which turns a solution into vapor what is in the solution is is questionable the more I learned about it the more I learned it’s it’s typically not nicotine but the idea was that these devices weren’t meant to be used for nicotine and possible with other flavors and there are of course other toxins in there as well the good news is that these cigarettes eliminate the combustion or the smoke however their long term health effects are still to be determined and and the more than they’re studied the more I think they are highly problematic for particularly for youth but the rate is rising very rapidly easy cigarettes are used significantly more more commonly than traditional cigarettes by today’s adolescents until last August they lacked FDA regulations and a question that you know that we have been wondering is whether cigarettes might actually be a gateway to traditional cigarettes maybe kids are nice cigarettes I have nicotine in them and then they’re getting hooked on nicotine and then they’re going to be driven to smoke the carcinogen Laden traditional cigarettes so we went into to examine this a little bit so this study uses data from the National Youth tobacco study so very straightforward and useful cross-sectional data set these were data from 2014 this was commissioned by the Center for Disease Control and Prevention to assess tobacco related beliefs attitudes behaviors and exposures to campaigns and other influences by by adolescents today so it’s also a very large sample we had 20 2007 u.s. middle and high school students so very much sample and the students in this sample spanned ages 11 to 19 okay so that it gives you a flavor and oh into our thinking right now we were interested in looking at correlates of e-cigarette use across ages 11 to 19 with cross-sectional data and it was nationally representative data so it matched on population characteristics pretty well so the goals of this specific study were to understand a little bit more about the etiology of traditional Andy cigarette use and Co use so first we wanted to estimate disparities in the rates of use across adolescents and when I say disparities I mean sex differences and race racial ethnic differences in the rates of use of the two products and excuse me and we would use t them for that and then importantly in goal two we wanted to estimate the rate of use of both products as a function of age and so maybe some students only use e-cigarettes some only use combustible some use both so we wanted to estimate the rate of their individuals using both and so these are these results are unpacked in in a recent manuscript that’s impressed that addictive behaviors which we would be very happy to share with anyone online today okay again we see our measures section again we have a very short list of variables and this is cross-sectional data so it really couldn’t be simpler in the measures Department here we have a measure a binary measure for current traditional

cigarette smoking yes/no 6.4 percent were currently using traditional cigarettes we had a single binary variable for current each cigarette smoking yes/no nine point two percent were reporting a cigarette smoking currently okay importantly we had age in our data set well in this data set we have age to the nearest year this age variable is our time metric in this analysis it’s that we’re interested in looking at rates and associations across age across time across age so so for the age variable you always want to to retain a measure of age in this case or whatever your time variable is with the greatest precision possible so when we analyze the I’d health data and look at developmental change that add health assessed age to the nearest month and so we will keep the age to the nearest month in there and we will code that to the decimal place but here it was just to the nearest year which was fine we had a binary variable for sex and a three-level variable for coding racial ethnic group okay so again we’ve got our data set we’ve got our research questions now we have to specify our model on here we’re going to be looking at rates of a behavior so we’re going to use logistic t-bone right so just think about your logistic regression analysis and add the time varying elements to your coefficients and you’ve got logistic Teva so what varies across age in our study when we believe the probability cigarette useful very then the probability of e-cigarette use will vary so in two separate models predicting each of those outcomes we would specify age varying intercept values obviously sex and race ethnicity are unlikely to change across a age and in fact this is cross-sectional data so they absolutely could not but their effects could be age varying and then finally we are very interested in the intersection of use of the two types of cigarettes within individuals so our aiya lessons who are using one type of cigarette at an advanced risk of using the other one if so how by how much and at what age so that association between cigarette use and ether great news they’re both binary variables so we’re talking about an odds ratio that odds ratio is specified to be age variant so here is one of the models that we fit so you’ll see at the logit the log odds of cigarette use is expressed as a function of an intercept which is age varying because the we posit that the probability of cigarette use varies with age plus the effect of sex on the log odds of cigarette use and we believe that the effective sex the sex differences with a rate of set of the rate of use across sex groups might vary with age so we specify both coefficients to vary with age here’s another example of predicting a cigarette use from traditional cigarette use so let’s take a peek at the results so this is this depicts the sex differences in e-cigarettes on the left and traditional cigarette use on the right across ages 11 to 19 and so males are using it higher rates they’re depicted in the blue lines and females are using a lower rates in middle and they’re depicted in red and so with Tina and you’ll see these are these are prevalences and there’s a point estimate that is a function over time that’s the solid lines and there are for any age you can look vertically above that age period say e cigarette use at age 15 look vertically and see okay the males are about 15% or expected to use e-cigarettes and vertically you see the 95% confidence intervals at that age so we have age specific confidence intervals here two females at age 15 about 10% were expected to use e-cigarettes at that age and so what do we see here what what you know what do we want to really think about interpreting here well the the there were no significant differences in the rate of e-cigarette or traditional cigarette use across sex until age 14 or

15 but between ages 15 and 18 males were using at high rates and then around 18 to 19 where we had a little less data remember these were high school kids the association was no longer significant interestingly we also interpreted racial ethnic differences in the rates of use across ages 11 to 19 the first look at the right panel traditional cigarette use and we see in red the rate across age for black middle and high school students and in blue we see the rate from white so as with many many prior studies we see confirmation that weights are using at higher rates than blacks and in particular this is significant between ages around 14 to 18 and a half Hispanic what you see around ages 15 to 18 Hispanics are smoking traditional cigarettes at a rate that’s right in the middle between blacks and whites that’s also what we’ve seen in many many prior studies but let’s now focus on the age range of 11 to 13 for traditional cigarettes and you see that Hispanics are actually smoking traditional cigarettes today Hispanics age 11 to 13 are using at even higher rates than whites this is a historically a very recent trend and you see it in a cigarette use as well between let’s see the green and the blue curves don’t overlap between around 11 and a half to 14 so middle school Hispanic youth are using e-cigarettes and traditional cigarettes at higher rates than whites and and certainly blacks okay and finally what is the association between using one product and using the other product so we are showing here again on the x-axis we have age 11 and 19 and on the y-axis these are odds ratios and that is an odds ratio of 45 how do we interpret this so these are age specific odds ratios so let’s take age 12 as an example for interpretation purposes so among those aged 12 adolescents using e-cigarettes are at greater than 40 times odds of using traditional cigarettes compared to 12-year olds not using e-cigarettes so you can get a sense from this plot about the co-occurrence of using both products as a function of age so some implications we drew from this second study I think are relevant maybe for policy and prevention the identification of key ages of risk we were able to identify them using t them and those could inform targeted age appropriate interventions we find that traditional and ecig abuse really does go hand-in-hand particularly in very early adolescence and then I think importantly this early use of e-cigarettes and also to combustible cigarettes so significantly more likely among Hispanic youth only during Middle School but suggesting greater risk I believe for future nicotine dependence and future health disparities between a Hispanic and white and black individuals okay so let’s just have just a few concluding remarks some next steps future directions you know one of the things that we and the methodology centers spend a lot of time thinking about and working on is is how does it have it draw new information from complex contemporary data sources and you know intensive longitudinal data is one of those areas of a big focus for us and ILD intensive longitudinal data this you know can come from ecological momentary assessment swear you are actually serving individuals intensively or it could come through some sort of wearable device that is gathering data more passively in either scenario you know these data are complex and sometimes we have data coming in at different rates of intensity over time and it can be very complicated to figure out how to get scientific information from the data there are other types of complex data sources today electronic medical records genetic data we can’t forget social media data but big data and complex data are places of high opportunity not just

opportunity for scientific productivity and grants but opportunity for college gaining new knowledge and I think that I think that there’s many health areas that are understanding well poised to to move quickly in this direction I believe T them specifically can be used to unlock new knowledge from existing data sources you know we can we can use T them to understand complex processes that unfold with time in many ways that we can’t even yet theorize about because our behavioral theories are not on that advanced and specific we can use T them to estimate the dynamic effects of interventions so that we know a little bit better about when they work for whom t them can tell us about developmental processes and associations over time maybe you know revisiting the the rich literature on adolescent risk factors for substance use and placing them into context of developmental age more specifically I believe Teva MS has a lot of promise for understanding associations and how they haven’t thought across the historical time and we have used data for monitoring the future to do just that we’ve used data from 1976 until today to look at how a perceived risk of smoking and actual rates of smoking have evolved over historical time maybe having policy implications and I think the complex link between the age of onset of a behavior or a phenomenon can be studied and with respect to health outcomes and probably behaviors and so on so we’ve done a little bit of work linking behaviors and looking at them as a continuous function of not time and not age but actually across age of onset so there’s a lot of new of questions that could be unpacked with t-bond I would like to give a shout out to my close collaborators in the methodology Center at John dear friends Ely Michael Russell and Service Alenko are four of the collaborators I’ve been working most closely with over the years and that is Ren Seeley original statistical developer of team M so I think our field is indebted to him and I would really like to thank the National Institute of Institutes of Health or all their generous funding of innovative methodological research that we really do try to use to advance public health so with that I if you are still with us I’m grateful for your attention and I want to thank you very much I would love to stay for an hour and take questions so I will be here I will probably sit down in front of the computers so you can have a face-to-face conversation and so we’ll open it up for questions yes hello everybody can see 70 now right there wait we’re gonna make a second in Aaron our moderator here over and this window is gonna ask your questions please feel free to keep those questions coming via the chat and we’ll try to keep up as much as we can and please don’t forget to leave your feedback with you if you have to go we understand if you really like this session please let us know and feel free to suggest topics we can’t promise anything but we will try we’ve lost almost no one okay people are still on so I’m guessing you all have questions I won’t know all the answers so I will tell you when I don’t know this is a really this is a really active area of methodological work and in also application to two types of data and so it’s a little bit like the Wild West out there but where we’re having a blast and a lot of people are getting involved with using team em so Aaron do you have any questions to serve us off yeah there are we had a good set of questions that have come in there there probably is time for more than we have right now so if people submit them we can still try I’ll start with the one question that was asked twice which was is there a way to Duty them in our hmm well that’s a great question um there that is a question that my colleague John dziak who was on the slide at the end could answer definitively for the group and maybe we can add that to an FAQ there is a there is a package and our for fitting

due for doing functional data analysis and so I believe some the functionality does exist in our today was not developed by our team or documented by our team but if you really understand the model you want to specify I think it would be possible to do much of it in in our so we we developed syntax wonder if I have that slide here we developed syntax is assess macros and if I’m able to return to my slides I bet I can but your psychoactive yeah yeah I might have sample syntax to show you I can certainly pull it up and the the sample syntax is extremely simple and SAS but yeah so the question was about our and there’s an arm package that does some varying coefficient models yes that’s right so but I can’t speak to the features of it specifically so we’re gonna just we’re just gonna pause for one second so we we do not have you currently large on the video do you want to keep your slides up in case people ask questions yeah can you all comment in the chat box is this working okay Stephanie’s a small figure is that Alright for everybody yeah so I’ll try and go through them basically in the order that they came in just in the sense of something of a fairness exactly okay so a question about covariance in the first example you use they were very few I am used to working with men does this first example use if you protect predictors to just to simplify the example or is there a larger reason thank you yes yes it’s using it to simplify the presentation in the example and you will see that most of the early papers we’ve published with T them to date and again you know they they only started coming out in 2012 so I’m most of early papers also do not have a lot of covariance in the model and it’s really because it’s challenging to convey what you’re doing in one paper when people are not yet familiar with t BEM so if you’re going to start using it in your research you need to be you know really thoughtful about how much you have to explain what the model is doing in how much you have to focus on interpreting the figures for people who are not familiar with T them so small is good simple as good in that respect having said that remember that at any particular age so think about a vertical slice in the plots at one particular age or time you are fitting nothing more fancy than multivariate regression and so you could think about having blocks of control variables interaction terms we do interactions a lot any kind of variable you would have in there as a predictor of your outcome in multivariate regression you may have in T them and then the the interpretations would follow so you would estimate a slope coefficient for the effect of say height on weight and for people age 18 you would say for every one unit higher in height you are seven pounds larger in weight and and if you had control variables you would say controlling for XY and Z or holding XY and Z constant if you had an interaction term say height multiplied by sex biological sex then you would have at even estimating height as height as a function of I’m sorry weight as a function of height sex and height by weight and you can interpret at what ages height and weight are significantly moderating sorry height and unsex are significantly moderating each other’s effect on the outcome and so some of our work right now is to really focus on conveying how t them can do non-parametric or time or age varying moderation so sex moderates an association when for whom great thank you very much so we also had a pair of questions about multi-level modeling I’ll ask them together okay the first one just didn’t quite grasp the similarities and differences as you were explaining and the second one was in terms of similarity in similarity to

multi-level modeling how a similar would it be if we were in the context of MLM to interact time with negative effect when predicting craving so right okay so so I’m gonna use for the second one first I think so I tried to do this with my poor memory so so if we put in so in our first example time is the x-axis we could have alternatively looked at the association between negative effect and craving and we could have put in as our covariance negative effect time and negative effect times time okay so that would be as pretty as close as you could get to fitting t them in a multi-level model and in that case you’d also have to accommodate the multi-level structure of the data with random effects or a GM GE approach so you still have to take into account the nesting of course within repeated measures within individuals so so what would this give you this will give you an overall estimate marginalize the overall estimate of the association between negative effects and craving and it would give you so no intercept function but you get one marginalized association between negative effect and craving and you would get a main effect of time which would tell you whether craving how might change linearly across the two-week window and then a test for whether the association between negative effect and craving changed linearly with time as well so you are imposing your linearity assumptions within time and your regression prediction and importantly I think over time as well so you are pre pre assuming that the developmental change the developmental course of craving during a quit attempt is a straight line and that the Association is changing in a linear way as well and so I think that you know if if you’re looking in a small time window or you have everyone in your sample was only assessed at months 1 3 5 7 & 9 and those are your those are your times I think you can really do quite a lot with multi-level modeling I use MLM a lot but if you have theoretical reasons or very nuanced data that would allow you to dig deeper into time unfolding processes unfolding with time tetum is a wonderful exploratory approach for that one could even use T them to first see if the associations are varying flexibly with time and then if you really wanted to you could especially if if they were fairly linear then you could switch back over to multi-level modeling and write it up that way but you wouldn’t know until you look it’s kind of like an interaction term we don’t know if an interaction is significant until we estimated and look okay so the other question I think was very related is multi-level modeling yeah how it’s different I think I think I covered that but feel free to write in follow-up questions yeah yeah I think you covered it too again if I’m incorrect please please hit us with another chat question about the time metric used in the addiction the Nathan edition study since they were for assessment points per day personally they know if it was measured in hours and then presented in terms of days oh great so it was measured in seconds it was truly real time and so the mobile device they this was it’s a little bit dated they were using Palm Pilots and the device is actually time stamped the every assessment and every point every day and we kept time to the nearest seconds or even maybe maybe greater detail and but for presentation purposes it was estimated as a continuous function of time our data were almost continuous in time but I was simply labeling the x-axis with anchor points so that we could follow along and really make sense of what was what was unfolding when and that in that two-week window great question and then there was a follow-up to that which I thought fit in if there is a different and at different time points because some people wake up at 9 a.m. and other

people wake up at 6 a.m. is there a lower limit in terms of how many or how few participants are needed for getting appropriate estimates yes oh great questions so um I think power analysis for Team M is not um not well established yet in general and I think but more fundamentally sort of when we think about estimating a team M we need data support it all general points in that time axis we need we need the support of data to estimate a coefficient at that particular time okay that said there’s no easy way to answer how many data points you need at a particular time and so on and again if it’s continuous time then Nolan will have assessments at the exact same time necessarily you know if it’s randomly but continuously in time um but there are there are a couple of things that we can fall back on and rely on one is if we have age period or a time period where for some reason we didn’t have a lot of data to support it in that particular period what you’ll see is your associations can still be estimated and the confidence intervals will appropriately get larger where there’s less data to support the estimate so our confidence intervals are point wise confidence intervals that reflect the confidence that we have at that time or age and part of the that determination is based on the sample size that you have at that age roughly in that age range on which to inform the estimate so your confidence intervals will alert you to that I think the perhaps them oh and something else we can rely on to us is an understanding and it could be a general understanding at first but um team ms relies on a spline spline estimation and what that basically means is that there’s there are smoothers being applied in the background to these coefficient functions and so the estimate at one particular moment is actually borrowing information from nearby moments to triangulate on the best estimate at that particular time or age and so if you actually are missing data at one age completely but you have data and surrounding ages it’s not a problem you can still estimate it and the spline regression will handle that for you and will say you know if you have a longitudinal panel study like add house and maybe there is selective dropout if the people in present in the study at age 12 are fundamentally different in important ways from the people who are still in the study at age 32 and that’s something that you need to think about maybe you want to do case wise deletion and then weight the data up to the national estimates that would be a good way to go or maybe you just want to do you know a description of who’s dropping out of the study and so that the results can be taken into context thank you now we’re going to be advanced question people love it you’re doing great can’t even publications that examine moderation the author’s usually estimate separate models for example for treatment and control groups which has the effect of interacting the groove with all other variables in the model if I wanted to examine moderation with selected variables rather than all variables in the model and include a multiplicative interaction term in the full sample model could you provide some guidance on how to interpret the test of that interaction as shown in the graph for that multiplicative interaction term Thanks okay so this is sorry I’m a big fan of moderation analysis in Team M so this is this is a great question I think team M is an outstanding way to look at moderation you are correct that most of the previous studies published studies that have looked at moderation like full moderation for example treatment versus control particularly our earlier papers we would separate the data into two separate data sets and look at these you know very nuanced dynamic processes and model them for cheating group and model them for control group and sometimes we would plot them together so there are pros and cons to this approach versus a full full sample moderation analysis one Pro is that it’s much easier to digest the information and to produce plots

that allow a parallel comparison of the two groups and so that is so those are the kinds of results that I showed you today for the the treatment and placebo smoking cessation study so that was it’s easier and it’s straightforward and interpretable and in fact in many times that’s where I still start today I will separate the data understand the processes in the two groups there’s there’s a no reason why you have to do that again we you put the full dataset together and I actually more and more I have been leaning toward doing this process where you put it all in the model together you estimate the effect of X and the effect of the moderator and the effect of x times the moderator and you can have that for many coefficients in the model a moderator can moderate three coefficients effects and that’s fine so again anything you could express at a particular time point as a moderation question you can express antietam as a time varying dynamic moderation question it’s so awesome the kinds of questions we can now address and so you would have in the results if you Finity done like that the challenge as with any regression analysis with interactions the challenges on interpreting those interactions so we’re actually writing a book on the basics of tea time right now and that’s one of the big things that I want to cover so that it’s more straightforward for people to get their head around but so one of the but one of the first things you see if you fit up a model a linear a tea gun model with moderation of say sex differences in the association between mood and craving over age you’ll get a coefficient function for the effect of mood you’ll get the main effect of mood you’ll get a coefficient function for the main effect of sex and you’ll get a coefficient function for the interaction term and so that interaction term coefficient function may be significantly different from zero at certain ages and not at others and so now you’re getting a very nuanced test hypothesis test of age varying moderation and so the coefficient function for the interaction term is very important to interpret in terms of periods age ranges or time ranges of significance so when does moderation occur when doesn’t so that’s super helpful then at that point the burden becomes how do I plot this and convey it in figures so that my reviewers are going to accept it and that that’s that’s the challenge and so interpreting interactions is always is always challenging but plotting the results is the answer usually ok thank you can’t even be used in cases where the treatment or explanatory variable is time varying for example when looking at workforce status over time an individual may go from working to unemployed and back to working thanks yes great question and also yes my colleague Rebecca Evans Pulse recently had a paper accepted at Acer alcoholism journal that was the goal was to examine the the effect of being in college versus not being in college on heavy episodic drinking as a function of age throughout young adulthood late adolescence and early adulthood and so what we wanted to be true to in that model this is an example of how you can do this well we want it to be true too and that model is the fact that individuals between ages say eighteen and thirty anyone in that rage range may be in college at the time enrolled in college so we understand college being a risk for a risk setting for frequent binge drinking well that’s true but it’s probably not true in all ages it’s probably true at the normative ages of eighteen to twenty-two or twenty-three but what if someone’s in college at age twenty five twenty seven and so the predictor in that model was college enrollments and it was time varying people could go in and out of college in this case it was cross-sectional data so at whatever age people were they either were in college or they weren’t but if it was longitudinal data they could go in and out of that status and we can still estimate the age varying or time varying effect of that time varying covariant so absolutely time-varying covariates

have a big a big place Antigua just everyone’s information just a second if you see Stephanie or I staring off the side of the screen it is because we are literally right beside one another so the next question is just is it possible to assess interactions of predictors interactions of predictors or moderators for the e-cigarette study for example is it possible to get estimates for the interactions between sex and race oh sure so yes so you would have to have that research question and then think about when you’re getting ready for model specification how you want to specify it so you might have a main effect of sex the main effect of race ethnicity might which might be a series of dummy coded variables and their interactions and then once you figure out that part of it you have to decide which of all of those you want to look at as age varying coefficients and so you could absolutely predict each cigarette use from sex as a dummy variable race ethnicity with you know two or three dummy indicators and then all of their interaction terms and all of those can be age varying effects and again the burden then becomes how do you present those results but I think you know I as an as an analyst myself I think that die model is is the same thing as separating the data and running it by by group to get the results it’s essentially the same thing it’s not exactly the same thing but it gives the analyst an opportunity to look at the effect of race ethnicity for males and then look at the race ethnicity effect for females you can eyeball it see where the action is you can go back to the full model where you have it all interacted you can test very nuance you can test age varying interaction terms and then and then figure out how you want to present the story from your full model you can derive any graphic depiction of your results that you wish just like in regression when you have to unpack an interaction term and produce figures to interpret the the the associations the complex associations you can do exactly the same thing in T them I like to pull my final parameter estimates into an excel file and from there I can generate any hypothetical people and mean curves and so on and so team M will dump out a SAS data file with parameter estimates at across the time axis broken at 100 time equal time intervals and so you can create your publication ready figures right in Excel and for me that’s the easiest way to do it some of my colleagues like to generate the plots right in SAS but I find that to be a little bit difficult to control Thanks so a longitudinal example question the gaps between the lines and the confidence intervals between two groups eg sexes or race ethnicities indicating significant differences or a varying size across age starting small and then getting larger can we make statistical inferences about the extent of the differences okay great so if I understand this question properly is this a great question um so if we want to look at sex differences in the rate of e-cigarette use across age so we’re going to go back to that the first figure I showed in a cigarette example you can you can you can estimate the age varying rate for males and you can estimate the age varying rate for females and you can plot them together on the same plot and that’s meaningful so then I was in the talk and this is a little bit sloppy on my part but I was roughly interpreting the age periods when the intervals do not overlap as statistically significant sex differences that’s not incorrect but it’s an overly conservative test of sex differences the proper way to test it and to be less conservative ie you’ll find more differences if you do this is to go back to a full sample model and put sex in as a predictor of the rate of

e-cigarette use and so that doesn’t give you the nice output of the rates over age that you can just stick into a figure but what it does give you is proper hypothesis tests at and then we show 95% confidence intervals so you get age specific 95% confidence intervals at every point in time or age and so you could if you wish you could produce the 99% confidence intervals or the 90% confidence intervals you can also in the data set that’s created with your final estimates you’ll get the parameter estimate and the standard error so so you could you should be able to get a p-value for the test of the hypothesis of whether it’s significantly different from zero at that particular age yeah so I think the hypothesis testing of group differences is is properly done in a full model where you have it in as a as a correlate and you evaluate the specific ranges at when say sex significantly predicts a cigarette use and that is gonna correspond to where those Kearse did not overlap but it’s a little bit more of a statistically higher power to test when you do it in a full sample analysis question about sample size is there an ideal sample size uh that’s a great question we don’t know yet so you know I think you know you again what you’re gonna you’re going to see in team M that at time points or age points at which you have less data your intervals are going to be wider right just as in they would be in a linear regression with a smaller sample size and so your confidence intervals intervals will be your first clue with small samples it’s not that you won’t be able to run T them it’s just that nothing will be significant so in much the way that it is in linear regression so by all means try it with smaller sample sizes your if you have a whole bunch of adolescents and only a handful of young adults and you want to look at age varying effects try it but your intervals will be very tight for adolescents and then they’ll get very wide because you just won’t have the precision of estimation where the sample size is smaller so you know in I don’t have any answers it depends on how complex your research questions are in terms of your your the model you’re specifying at a particular age how many predictors do you have how many interactions you have are they binary or they continuous are they very variable or not variable so that that those all factor in the same factors that play into power analysis for linear regression playing to t them as well but in in addition then you have to be thoughtful about do I have enough data to make inferences at this point and at this point I’m working on a paper right now that is looking at some early marijuana use and the association between being black black youth and being an attaining achieving a college education by age 25 and we’re looking at it as a function of age of first marijuana use and what we see is that if for users below ages 12 to 14 hour intervals are so wide we can’t say anything about it but I’m leaving it in the paper because why are the confidence intervals so wildly outrageously wide at the marijuana users less than HH 14 it’s not that we don’t have marijuana users who started between ages 12 and 14 it’s that we don’t have black college educated individuals who started using marijuana at age 12 13 14 and so again it’s it’s statistical power the more the longer I’m in research the more I feel as statistical power is so complicated and and so I think your your confidence is will expand as as you’re a man of information you have to answer the questions at that particular age well that you have and so remember your confidence bands can wide in and shrink as you have more or less information on which to base your estimates over time and that’s a good thing right pardon me a question about data do we

have TV in Stata the answer today is no we also develop software for late class and late transition analysis as SAS procedures and we have for these doses Stata plugins but for now Teva ms is assess macro and we don’t have an immediate plan to release its data but i think that we may package our teeth on procedure and as an r package for users I just we understand exactly how some of the nuances in the data are being handled when we program it ourselves things like Bayes random effects handling clustering of repeated assessments within individuals and so we may release in our package but I don’t think Stata is is is on the plate right now and for anyone who uses sass at all um TM really is legitimately not very difficult to run that doesn’t mean it’s simple or that it’s easy to interpret what you get but in terms of if you have data getting results out of T them there’s not a steep learning curve for that right I was gonna puke and see if I have a there was a question about if you can show kind of yeah yeah I would like to show you some sample syntax and sass um but go to methodology psu.edu to the TM page there’s at even research area page and there’s also a TM software free software page and somewhere in there we convey the simplicity of the SAS programming one thing that I get a question about a lot is is oh you’re using add health for a lot of TM analyses to study development but I hopefully has four ways so how is it possible and so this slide was meant to convey how individuals could have only one or even in this case four time points of measurement and it’s nationally representative data was a big cohort sequential sample and it had about 12,000 individuals in the in a sample followed over time and but if you if you if you look at it in terms of how many data points you have it’s 12,000 by you know a three and a half or so on average time points per person we’re talking about thirty five thousand time points person times in our data to support this inference of developmental change and so if you oops if you think about the data point this is actually plotting the points of assessment in ad house and if you know anything about ad heltah everyone who sees this is surprised that we have such incredibly dense good average across age from a four wave panel study and so this is a great way to think about code your data support a team M analysis if it’s cross sectional or panel data well take a look at all your person times and figure out look at a histogram or box plot or something like this that shows how well you have what’s what kind of data support you have across the different age periods maybe it’s only a subset of ages that you have good data for and so you would restrict your team m – just that age range yeah but I don’t have sample syntax right here yeah it’s historical examples and Oh H of onset so this is a this is just a slide showing really briefly how we use T them to link in the age of onset to an outcome and then instead of saying if you start smoking regularly prior to age 14 you’re at risk and if you didn’t you aren’t at risk you know but thinking about the age of onset in a more nuanced way in the population what we found is that the you see that this this is the functional form of the association between the age of onset of first regular smoking in adult nicotine dependence and so this aged age of onset trend gives you much more information I view about how the age of onset is associated with with dependence and

adulthood and we further looked at it separately for females and males in the the pattern is the same but much stronger for females who start smoking regularly at younger ages so this is an example of modeling coefficients as a function of age of onset so some of the time metrics that we’ve explored we’ve explored of course real time and that was example one in today’s talk we explored age age overtime developmental age and and that was the second example today again cross-sectional data but we were looking at age trends we panel data to do that as well we’ve looked at age of onset like I’m showing you right here on the x-axis and we talked just briefly about historical time and this is just some rates of co use of multiple substances from 1976 to 2013 and so you know embedded and many of the data sources we have there is an opportunity to look a little bit deeper release ourselves from the functional forms of linear trends over time and importantly again it’s take-home message from today is that we don’t have to measure we don’t have to model outcomes as linear functions of time they can be nonparametric and most importantly is that the effects of variables can vary flexibly with time we shouldn’t always assume that there is a particular level of risk associated with X in predicting Y it may be for you know at certain times or certain ages that it is a significant risk factor other questions there are there just a few more I was just answering a few people so I wasn’t just attentive to you for that last one as I normally am okay so a question about including time invariant control variables and it’s even model so again this is another more advanced one hanging tight there will be less advanced this is not for you so if I want to control for race race ethnicity it’s not clear to me if I include dummy variables for all categories or eliminate one dummy variable as I would do in an OLS regression which becomes the reference by eliminating one dummy variable how would I interpret the coefficient functions for my variables of interest Thanks okay was that two questions or the race/ethnicity example was is yes controls that’s it okay okay really read that are you alone I think there’s no there was a lot I think embedded in there so one question is categorical predictors so let’s talk about race ethnicity in and I I didn’t have time to get into the details but but we looked at ISA great use for black white and Hispanic youth middle school in high school and you may have looked at that and presumed that there was another racial ethnic category that was left out in fact they were not included in the analysis and so those models were run with two dummy variables indicating three racial ethnic groups white black and Hispanic and and so I could fit those regression models with two dummy variables included I can interpret than differences I could determine from those coefficient functions significant ages when say whites and blacks and Hispanics and blacks may have different rates of behaviors or different associations at what ages and so it’s just as with linear regression you would want to code your variables so that you have one fewer dummy variables then you have categories in your predictor variable your categorical predictor variable so one more question about that though that was embedded in there I think was you know if you just want to control for race ethnicity or SES or whatever you can put in additional covariance in tevonn no problem interestingly though you have as the analyst the option of including a control variable in a way that it has just a simple non varying effect just a classic control variable which is an easy thing to do and an overview or anywhere would criticize you for doing that but you

also have the option of allowing your control variables to have time varying effects along with your other covariance of interest and so that neither one is right or wrong but you know how you interpret your coefficients of interest given that your control variables maybe have time varying or non time varying effects I think is the burden is on us to all of us to make sure we’re interpreting things as clearly as we possibly can I any other there are yeah yeah a couple more so this was an important questions I thought just to clarify this point yeah the question are there instructions on creating these graphs either in SAS or in Excel oh let me think yes so let me think well so one thing I will say is that most of our papers using T them check the appendix we almost always include our syntax in the appendix so that you can it’s meant so that you can take it and adopt it for your purposes in future studies and I’m trying to think you know for well for Excel the only trick if you want to produce figures in Excel is that you have to be comfortable enough in SAS to save the parameter estimate file which is produced automatically for you but to somehow save it in an opening in Excel so that’s the only trick once the data are in Excel it’s as with any other figure you might generate for a manuscript and so I think that’s the only tricky thing trying to think if we have tips on that if not we could add an FAQ on how to import your SAS data into Excel you can you can you can just type a SAS command to save that file as an X I’ll file and it will be sitting on your harddrive waiting for you to open it up and make beautiful publication ready figures alternatively you can produce figures in SAS directly if you love status and you’re willing to you know manipulate figures in SAS Sarah vasilenko is my colleague one in particular who really likes to produce her figures in SAS and I can’t remember if she’s ever shared that code I’m publicly in an appendix but I know that we’d be happy to share it with you and but in general the output is the curve so you don’t get the figures of the overlaid specified in your exact effects of XY our x1 x2 x1 times x2 and so on intercept all of those are those functions are produced graphically for you the analyst to see for every model that you fit so you get an immediate in-depth picture of the questions that you’re trying to answer the question then is how do you get from there to publication ready conveyance of the these time varying effects and you might want to manipulate or interpret interactions and and so you you can calculate simple slopes age-specific simple slopes is a really good way to interpret interactions you can do all of that really easily in SAS or in excel so we don’t have a lot of syntax out with again we are we are putting together a book but if our level you won’t be out in until next year but that is the kind of tutorial information that will be in there but meanwhile you know we’d be happy to share pieces of code with with the users so Stephanie can you hear me really are getting down to the end across the great majority people are still around so I would like to ask how to use T them to ask proper falsifiable questions when using a hypothesis the effect of between-subjects factor acts will vary over time the capacity to reject that hypothesis approaches zero hence giving an opportunity to capitalize on chance alternatively an opportunity to invite ad-hoc hypotheses in excusing ad hoc hypothesis definition we expect a greater effect of acts at ages fourteen point zero to sixteen point five years what might be an appropriate way you use T then to provide a hard test of time during

effects hmm yeah so so I teach statistical inference and regression to graduate students in my department and we spend a lot of time talking about this kind of thing um not specific to team M these are bigger these are bigger questions I think and team M is vulnerable to potential criticisms of inflated air rates and things so so one thing that I should share with the group is that Team M relies on information criteria to to either automatically or the user can manually select their final model now what does model selection in Team M really mean it means that you’re selecting the complexity of each regression coefficient function so how many knots are required to capture the complexity in this trend over over time or age and so the model selection procedure is again it’s either automated or you can do it manually and rely on whatever information criterion you’re most comfortable with but so so one one maybe week answer to that question is that if if the if the coefficient function really was not substantially varying with time you would select curves because of the the way the information criteria work that protects you from / parameterizing models you would you would estimate basically flatlines they would be flatlines and so so that’s one potential empirical test if you have a question of whether whether a coefficient function increases linearly with time I would say you’re better off fitting uh you could use T them or multi-level modeling but fitting time-invariant coefficients and the coding a linear effect of time and testing whether that is significantly different from zero or not I think you know what T Bone seems to open the door to is more nuanced hypothesis testing and and I think there are a lot of opportunities for doing very systematic and rigorous hypothesis tests that are less fluid and loose than I kind of presented in the in the examples today for example one could code one could code zero for adolescence and one for young adults and that could be a you know that could be time varying or or not if it’s cross-sectional data and you can empirically test whether an association is statistically different during the adolescent period than it is during the adult period with one you know cleans very specific hypothesis test so but you know you’d have to have that question on priori and pose it so uh I think you know you know if we want to think about whether these cigarette users are more likely to be combustible cigarette users and we have developmental longitudinal data well II cigarette use is on or off at each time point and so that’s time varying variable and you could estimate the effect of it in a non varying way over time and get at marginally is the rate of cigarette use higher for e-cigarettes and that gets you a very again I mean I think you’ve got to get explicit about your research questions and and then think creatively about how to code variables and team em to test them though the way that I’ve been presenting it in the way that we do in our in our most of our papers is more on the exploratory side but a lot of specific hypotheses could be tested if you have the questions specifically laid out I’m not sure that was a very specific answer but it depends so two more questions okay how would you approach the nicotine data if you are interested in daily cycles waking random point one random point to bedtime that’s a great question I mean I actually have thought a lot about the fact that people are sleeping in between the bedtime and the wake up times and that’s something that we should be more thoughtful about in the future but regardless if you had an empirical question or or it or a theoretical reason to believe that

craving or the leak between mood and craving was going to be particularly important upon wakeup which I actually believe might be the case and less so in the other time points then let’s think about the data structure for a moment so you have a quick date and then people have for assessments at time one for assessments at time two and so on and so on well you could code an indicator for whether it was a wake-up assessment or not a dummy variable that’s time varying and it would be coded 100 100 100 and so on and so again this is a similar to the last response when I think if you have a specific research question like that that’s a really interesting research question and you can code it in to you I’m an estimated and so in that case you would estimate an on time varying effect of being a wake-up moment versus a non wakeup moment and and you could estimate and test so that’s basically you know an example of that and it might be it might make sense to allow that effects a very during the course of the study as well no no you’re not at all down to users you’ve got 30 people hanging in yeah that’s the number of their video totally general um so at any rate so the last question we have unless I missed one and someone can message me if I did though we are very short on time here should the tests of the multiple parameter estimates bisects the doctor Lanza just mentioned in her example in lieu of looking for non-overlapping confidence intervals be adjusted for multiple comparisons ah and so that was obviously earlier well so yes it’s yeah I think that the multiple comparisons issue is is significant in TM huh no pun intended um the the I don’t have great answers for this I think if you come to a study and you have a priori specific hypothesis generated hypothesis research questions then I think it’s it’s prudent to think very carefully about how you can specify your team m to test that hypothesis in in many cases we have never looked at time or age bearing effects in a sum in an area of study we just haven’t looked and in addition our behavioral theories are insufficient to generate hypotheses and so I would argue in those cases these are exploratory studies and we can interpret age or time periods of significance or significant moderation report or whatever and then those those should be put into the literature and put forward for future replication I feel very strongly that these kinds of models can advance our thinking and advance our theoretical developments our theory development but that they’re there in no way from one particular study going to give us the definitive answer that that sex significantly moderates the link between height and weight during ages 17.4 and 26.2 I mean it doesn’t work that way it’s statistics doesn’t work that way right it’s an incremental advancement in our knowledge I would caution I would caution all of us making myself included from over relying on the p-value tests again and again and again and you know and you know interpreting the p-value at this age and this age and this age and this age and so on I think if you approach it more from a broad brushstroke but this is the age range when we see significance in this study again it’s putting it forward for replication in future studies but there’s more more thinking to do absolutely about the the type 1 error rate in int event but I but I think it’s it’s not a simple answer it’s not a simple story the fact that information at time T is leveraging information in times prior to T in times post T in building the the constructing the confidence interval at that time point I think these are not independent hypothesis tests at every moment in real time it’s it’s not that egregious of a

problem but again I do think that we need to think more deeply about this ok well give me just a second here and you check in with the chat and is there anything else that I just missed please don’t raise your questions cuz she hasn’t answered questions for an hour anyhow yes thank you very very much for you our continued participation and if you have any questions or feedback afterward please send us an email and thank you very much have a good day thanks everybody