 Hello, everyone. I'm asking for taking an interest in my research. I was asked to come here today and present to help give an example of some of the work that CLSA data is being used for and also as an example for prospective students who are looking to work on this data set as some of the work that is possible. So I'm just going to jump right into it. This presentation is called measuring frailty in older Canadians. It's an analysis of the Canadian logitudinal study on aging and it was my master's thesis which was recently decided. So presenting part of the problem, people in Canada are living longer and the population as a whole is getting older. For the first time last year in this country's history, there were more people over the age of 65 than under the age of 15. That's nearly one in six and there are well documented high healthcare needs for this population. It's not just that people are getting older. It's there's a burden of providing care to these people. That being said, there's a variation in health status among people of the same age. Some people who are 65 are healthier than other people who are also 65. So age doesn't necessarily mean the be all and end all in determining health. As this sort of variation increases, there's a vulnerability to declining health and a risk for institutionalization and mortality and this this variation health status is commonly called frailty. Frailty as a concept is complex. It involves multiple health systems and has been linked to physical health, psychological health and social factors as well and frailty is dynamic. It changes over time. So a person's frailty level can increase or decrease. This concept is useful for a number of reasons. The ability to measure frailty would be good for planning health care interventions as an outcome measure and also for assessing the health resource needs of a population. Unfortunately, there's no consensus on how frailty comes about or how it should best be measured, but there have been several attempts to measure frailty in the past. There's two key main theories, the freed phenotype of frailty and the Rockwood cumulative deficit model. None of these methods has been established as the clear criteria on reference or gold standard for this is the be all and end all way to measure frailty in a population. So the just a little bit of background the freed phenotype of frailty involves five physical factors that can be measured in a in a clinical setting. Slow mobility or walking speed weakness, weight loss, decreased activities and exhaustion and a person is classified as frail or non frail if they have three or more or two or fewer of these criteria. The cumulative deficit model is a little bit different. A person is assessed using a list of health deficits that have to be associated with age. So gray hair would not count because it's well associated with age not the health deficit. So health related factors associated with age and within that category the selection of deficits included is very flexible. As long as you have 30 to 70 items the scales perform more or less the same. So your frailty index score is the proportion of the number of items in the index that person is deficit on or that is present. So the proportion indicates frailty. It's a continuous number. As opposed to the phenotype model, which is dichotomous. So our goal here was to evaluate the level of frailty of participants in the CLSA. We wanted to look at what is the best way we can measure frailty here in this data set and we looked at a bunch of different ways that had been done in the past to try to figure out what was the most appropriate for this setting. Secondary objectives for the research were to study the underlying construct of frailty. So what can we learn about frailty from this data and to identify some of the potentially key factors contributing to frailty and key factors for its measurement? We use structural equation modeling to compare some theoretical measurement models for frailty and we assess the validity of the measurement tools that we came up with. So the data source, just something of interest, is the Canadian longitudinal study in aging. As I'm sure many of you know, it's a large scale population based study covering across Canada. It includes community dwelling, Canadian adults aged 45 to 85 at times enrollment and collects a wide variety of physical, psychological and social health indicators, making it a well suited data set for studying frailty. Participants to be eligible for inclusion had to be able to complete telephone interviews, which is important later on. The tracking cohort specifically was what we used because when we started the study, it was the first data set available. It includes 21,241 participants and is cross sectional, even though CLSA is a longitudinal data set, we only had one time point for the study, so there were some limitations to this, but it was cross sectional data. Since then there has been more data available in the spring 2016 release, so if you're interested in using the data set, there is more than just what I talked about here. The data was collected, participants were recruited primarily through those who had completed the Canadian Community Health Survey and to fill out the rest of the sampling frame. Provincial health care registration databases and random digit dialing were also used. All of the participants completed a 60-minute telephone interview and all of the data that we looked at for this study were self-recorded and recorded by trained research staff using computer-assisted telephone interview software. To access the data, first we developed a research question and looked at which CLSA variables we needed and were relevant to the study of frailty. You can look at the data available online through the data preview portal and sort of look at what data is available and how does this fit my research question. After obtaining ethics approval, we submitted an application to the data and sample access committee and then once that was approved, we received our data set for this research. So the variables that we included were all health indicators that were theorized to contribute to frailty or be associated with frailty. They were chosen based on several systematic reviews that we looked at that were focused on primary research studies that measured frailty using original tools. So which variables had been previously used in an attempt to measure frailty? And we also consulted with some experts in the area of frailty measurement to make sure that we had good content coverage. We organized these variables because there were a lot of them based on the international classification of functioning disability and health that was put out by the World Health Organization. And we made sure we represented all of the ICF domains with the variables that we had. So as you can see, I've separated them into the five ICF categories. We had for health conditions, self-reported chronic conditions, there was an exhaustive list, as well as self-rated physical and mental health. Body mass index was included as was continent and self-reported sensory impairment. Depression was measured using the CESD-10 and satisfaction with life scale was also included, as were reported anxiety and mood disorders. Cognition was measured using telephone interview validated cognitive tests, the mental alternation tests, the rate auditory verbal learning, and then animal fluency tests were all included as measures of cognitive function. Functional status was measured including items from four different established scales. We sort of looked at which items could we include that would measure three key domains, upper body strength, lower body strength and dexterity. So all three of those are covered for various items. Activities of daily living were included as was social participation and the availability of social support as measured by the MOS Social Support Survey. So lots of validated tools and standardized measures for different indicators of health across several different domains was the key factors here. So we decided to construct a frailty index for a couple of different reasons. One, because it allowed us to represent the wide array of data available in the CLSA in a frailty measure and also because the phenotype of frailty index required physical function assessments which were not yet available in this dataset. We didn't have walking speed in grip strength so it was most appropriate to use the frailty index as the primary method of assessing frailty in this dataset. All of the variables that we included were transformed to a value from one being the maximum deficit and zero being no deficit so that they were all equally weighted when we summed them all together to give somebody a frailty index value. So as an example of the coding below you can see that ordinal variables, the worst option was coded as one and the best or least frail option was coded at zero as it was for dichotomous variables like chronic conditions. Continuous variables like the cognitive test score were recorded as a proportion of the maximum achievable score. So if you had the worst possible score anybody got you got one and if you had the best possible score anybody got you got a zero. We ended up including 90 health deficits in our frailty index for the CLSA and all the participants who had data missing on less than 5% of the available indicators were included. So we did a couple, we looked at these scales in a couple different ways. We did an exploratory factor analysis on all of the variables included in the frailty index to estimate latent factors that might be causing or driving the associations between these variables. We used polychoric and tetrachicoric correlations because there were several ordinal and dichotomous variables so this was the most appropriate way to study that. And we observed the factor pattern that accounted for the maximum variance among the observations. This method doesn't have any prior predictions as to how these relationships should play out. We also used structural equation modeling which conversely requires us to specify how we expect these variables to interrelate. So we specify a measurement model and then test the fit of that model with the observed data. And then we make stepwise modifications to that model to achieve the best fit. So our hypothesized based model is organized based on the domains outlined by the ICF and those were predicted associations. All of the latent domains should be interacting with each of the other latent domains. And we predicted that each of the variables would load only on one domain. We had a very, very large sample size with over 21,000 people. We had enough of a sample that we could split the data set into two halves. We used the first half to develop the measurement models and we used the second half to evaluate the fit in an independent sample so that we could test whether or not our results were simply due to chance. We also evaluated the construct validity of the frailty index by testing with predicted associations with some sociodemographic variables we had available in the data set. So our hypothesis for the construct validation was what were the associations we would expect for a good measurement of frailty. So we would expect a good measure of frailty to be positively associated with female gender, increasing age, home care received and assistive device use. And we would expect lower levels of frailty in those who had higher education and annual household income. So those were the hypothesis we made for a priori for the frailty index that we developed. So a descriptive summary of some of the data we had available. In the total sample we had slight overrepresentation of the 55 to 64 age group, but nonetheless each of our age groups were all represented. We had a slight female bias, but 51% to 49%, not too much of a difference to be worried about. Most of our data set had achieved a post-secondary degree or diploma. Around 12% reported informal home care services and, or sorry, around 30% reported informal home care services and assistive device use. Formal home care was somewhat less common given that these are all relatively healthy community dwelling adults. Most people didn't report any falls or injuries as well. So those were rare outcomes. The frailty index when constructed for the population could be calculated for almost 21,000 people. So there were 367 that were missing or had data such that it couldn't be calculated. The maximum value was 0.67. So in this again relatively healthy sample of older adults, the maximum frailty we saw in our sample was 2 out of 3 or 0.67%. The average was 0.14, presence of all of the deficits that we recorded. And if we wanted to try to decodamize frailty, the most common way to do that for the most established way is to use a cutoff of 0.25, and everybody above that being frail and everybody below that being non-frail, we saw a prevalence of 6.9%, and both the maximum frailty and the prevalence were consistent with other population-based studies that calculated frailty index. The results of our exploratory factor analysis, we found that based on the SCRE test, we should retain three key factors that should account for the majority of the variation in sample. And when we looked at those factors, we saw that the first one was dominated by the measures of functional status and self-care variables that we had available. So the physical functioning is what we called this factor. It also included a few chronic conditions as well as self-rated health. The second factor, counting for a smaller proportion, but still a significant one of the variants in the sample, included the satisfaction with life-scale items, self-rated mental health, and a few items from the CESD-10 scale. And these were sort of the positive depressive symptoms related to life satisfaction. So how often do you feel happy? How often do you feel hope for the future loaded on this factor? And the third factor included the rest of the items from the CESD scale, which included how often do you feel sad or lonely, how often do you feel depressed, reported mood disorders. This is more of the depressive symptoms sector. These three factors appeared to be the strongest contributors to the variants among the observations. And these patterns persisted when we stratified the sample by sex and age. So these three factors were present throughout the sample. This is our base model for the structural equation modeling, organized the way that I had discussed before. All of the variables were significant based on the WALT tests. So all of them contributed to the model in some way, and we couldn't take those out just based on the statistical tests. So we still made modifications to the model to improve fit. Highlighted in light blue are the changes that we made. The mental health item was moved to the impairments to body structure and function domain, which makes sense if you consider the association between depression and anxiety, whereas body mass index and consonants were moved up to the health conditions domain with the other chronic conditions that were reported. We also included error covariance terms, which it helps to account for two observations that are more heavily associated than would be explained by their common presence in the model. So an example of that is continence is associated heavily with activities of daily living, which makes sense because ADL includes an item that measures continence. So they were more closely associated than just the model would explain. So once we included those terms, the fit improved substantially, but the important finding from this is all of the variables we looked at to measure cruelty were significant contributors to the measurement of cruelty. As an exploratory analysis, we started removing factors just to see if the fit improved because it's possible that because the sample size was so large, any test for significant contribution would be positive. It's possible that all of the variables were significant because of the sample size and to our surprise, the fit improved up until the significantly simplified model, which included only the depression scale, lower body strength, upper body strength and dexterity. So the physical function items and depressive symptoms, when included with a single latent domain which we termed cruelty, had the best fit as any of the other models. So we looked at this a little further. As you can see by the fit statistics, the ones I want to draw your attention to, the comparative fit and non-known fit index, the closer those get to one, the better the fit of the data. Anything over 0.95 is considered a very good fit. So the modified model with the error covariance terms, but still all of the variables included, had very good fit of the data, but the simplified model was very close to one. This could be due to a number of different reasons. It could be to noise of many different measures measuring the same latent domain. And by simply removing other factors, we've removed some of the noise and we've reduced ourselves to some of the core domains of measuring frailty or the most important factors. And it's also important to note that physical functioning and depressive symptoms were the key domains measured by the freed phenotype of frailty. So we're seeing some commonalities between different measures of frailty that were included. We took this a little bit further to see what if we measured frailty using only those variables that were included in the simplified model. So we included all of the indicators of health included in those core domains and scored them by weighting the items equally based on recommendations from Norman and Strider. So we wanted to make sure that our measurement equally weighted all of those four domains because in the simplified model they more or less made an equal contribution to the model. So the score was weighted and one was considered maximum frailty and zero minimum frailty. So we transformed it to make it directly comparable to the frailty index. Looking at the distribution of both, the frailty index has much less skew and kurtosis, which makes sense because it includes many more items. So the frailty index, while more complicated, is more sensitive to smaller changes among participants, whereas the simplified frailty is much less sensitive. And a high proportion of the population had a very low or zero simplified frailty score. But when using the same cutoff, it identified roughly six percent, six and a half percent of participants as frail. So there was, we wanted to look at how closely do those two scales agree, both the simplified and the more complicated frailty index. Also, I should note that the simplified frailty scale did have some floor effects. So there were a lot of people who had a square of zero, which was not the case for the frailty index. When we looked at correlations with some of our construct validity outcomes, we saw all of the associations were in the direction that we predicted for both scales, and were of comparable magnitude, although the frailty index was a little more sensitive to age and injuries, but aside from that, or sorry, and also income, but all of these associations were still in the direction predicted. And this sensitivity to age is illustrated in this chart here, where you can see in blue the frailty index changes a little more with age, whereas the simplified score was less sensitive. The two scales agreed significantly beyond chance, but the CAPA was still less than 0.8, which suggests that it cannot be confidently said that they are measuring exactly the same thing, but the agreement was nonetheless significant. When we looked at a little bit further, which people do the two scales disagree on, the frailty index was again more sensitive to age, but also the participants who were frail on the frailty index, but not on the simplified scale, had many more chronic conditions than those in the vice versa category. So it appears that the key difference is that if you have many chronic conditions, the frailty index is more likely to classify you as frail, but that is not necessarily the case for the simplified scale. So just as a summary of some of the results, we explored the factor structure of the frailty index, and the structural equation models showed that all of the variables contributed significantly. And our theory of how they should relate had a good fit of the data, which persisted in the independent data set. So when we looked at that data set we had set aside beforehand, we saw a similar fit and similar results to our test data set. Both the frailty index and simplified frailty score showed good constrict validity and had acceptable agreement. Some of the limitations of the study is we didn't have all of the indicators of frailty that we would have liked to include, given that it was a secondary data set, but a lot of those variables like medication adherence are included in subsequent data releases. So that will be available in the future for us to look at. In a cross-sectional data set it's difficult to discern what things are affected by frailty and what things contribute to frailty. We can only look at associations. And the participants who could not be included in the analysis due to missing data were a little bit older and had some more chronic conditions than the people who were included. So we may be missing a vulnerable population that can't be measured using this method. And it's possible that the simplified frailty score might be measuring a component of frailty or a different but related concept and they might not be exactly agreeing. So it could be measuring something like disability. So what we are concluding from this is that the frailty index is simple to calculate and interpret. It was relatively straightforward to conduct one and it performed well in this data set. Also the frailty index method is commonly used and generalizable. So we can compare it to other population studies who are also measuring frailty using this method. We have a recommendation for measuring frailty in this CLSA. And we want to look at including a frailty derived variable future data sets. And for people who want to use the data to study other questions they can have frailty index value included in their data set. The simplified scale was a nice surprise. It had acceptable agreement with the more established frailty index and included similar domains to the phenotype of frailty model. So it still had some content validity. It may identify some important components to measuring frailty and had many fewer variables. So it might be useful as a screening tool or used in studies where the number of variables you can collect from your participants is very limited. But its potential as a screening tool should be definitely explored further. So what we plan to do next is look at the simplified frailty score using longitudinal data to see if it associates with outcomes. So all of the variables we included in the simplified score are available in the Canadian study of health and aging. So that might be the next step for evaluating that scale. And we want to look at the frailty index in relation to prospective outcomes like mortality and hospitalizations. And it's definitely good to compare the sensitivity to change and longitudinal data of the frailty index and of the simplified score. Thank you all for joining me today and listening to me talk about the work. I hope that you learned something and enjoyed it. I'm going to pass it over to Mark now just briefly before we respond to some of the questions. Great. Thank you, David, for this excellent presentation. Certainly it's elicited a lot of interest given the number of attendees. So we do have a couple of questions from the audience that I'll get right to you before I ask some of my own questions. Okay, so we have a question from Dr. Pichet. Was vitamin D intake or status considered for this study? Both the Canadian Community Health Survey and CHMS data for this group are available. And before David, you answered the question, I'll just let everyone know that biospecimen data from CLSA are not yet available. So if this question pertains to biospecimens, you were not able to look at vitamin D in the blunt. There is, to answer your question, there's self-reported data on vitamin D intake and status available in the maintaining contact data release, but it wasn't for the data release that I used. So it had been used to measure frailty in the past, absolutely, but it isn't something that I could include based on the data set I used, but it's something that we'll look at further in the next release. So from Alicia Rodriguez, did you consider to use the CSHA clinical frailty scale to conduct this study? Well, I'm not mistaken. The CSHA clinical frailty scale used a very similar method to construct a frailty index, and their data set as we did in this one, so we're hoping that the results are comparable. And we did see similar results in terms of the proportion of frail participants in the distribution frailty index scores. From Nader, did you calculate phenotype frailty, i.e. the freed frailty index as it seems your index is close to it? We looked at calculating phenotype frailty using the freed frailty index, but it requires functional assessment data, which we didn't have in the data set, so we thought it was better to look at freed frailty score when we have the physical assessment data, which I think is coming in one of the next releases. So we need things like gate speed and grip strength to measure that properly, and when that is available in the CSHA, that's something that we'll look at. Great. Comment from Larry Chambers. It would be helpful if at the end of the presentation a slide showed the items in each of the frailty indices. I can post that with the rest of the slides. I'll add a slide that includes that. It's just a very busy slide, so I didn't put it in, but I can definitely add that for people who are interested in the slide deck that eventually gets posted. Great. Fantastic. From Fatima, the simplified index is the sum of the, oops, I lost it here. The simplified index is the sum of the four items with equal weight. Is that so? That's correct, yes. Because each of the domains in the simplified model from the structural equation analysis had more or less an equal contribution, but each included a different number of items that measured that domain. We weighted the domains equally, so if the depression scale had 10 items, it would still make the same contribution as the dexterity domain, which only had three items. So, yeah, we weighted each of those equally so that we didn't unintentionally measure depression as more important to screening frailty than dexterity, for example. Great. Thanks. So, the next question has to do with missing data. And indeed, that was a question I had intended to ask you as well concerning the missing data. Let's find my question. Right. So, I was going to ask you, David, how you handled missing data? Because you said that you had included people with less than 5% missing data. So, I was wondering how you may have handled the missing data in your analysis. And from Kojo, the question is, why not consider multiple imputation to treat missing variables? So, we looked at doing multiple imputation for the missing data, for the missing variable, absolutely. The main reason we chose not to is the proportion of people with missing data was very small. So, it was unlikely to make a large effect on our results. The other reason being because the frailty index included so many different variables, there weren't too many other socio-demographic factors we could use to impute the data with. So, there was a couple of different reasons. But the most important being the proportion of people with missing data was very small. So, in those people with missing data that you included, how did you handle the missing data? Especially when it came to computing the frailty index, if they had missing data on one of the items that composed your index. So, we didn't exclude everybody with any missing data for the index. As long as they had 95% of the items reported, they could still have a frailty index number calculated. It's just if they were missing three, then their total sum of all deficits would be divided by 87 instead of the full 90s. So, they were removed from contribution to the index completely. They didn't bias the sample to the direction of being more frailty just because they were missing. But previous studies that had calculated frailty index for other data sets found that you lost integrity of the score if you included people with many fewer missing data than the 5%. And so, we followed that same convention. Great. Okay. And yeah, I've seen indices calculated using that exact type of process you described. From Meguna, it seems that your simplified frailty index covers only two domains, physical functioning and depression. It may miss other major domains such as weight loss, et cetera. Could this index be too simplified? I think that's an excellent question. I think if you want to be comprehensive and include everything that could possibly contribute to frailty and make your measurement as accurate as possible, it's been shown that including more items makes your index more accurate because the more you measure about a person, the more accurate your assumption of their health is going to be, or your measurement of their health is going to be. The purpose of the simplified scale seems to be most likely to be a screening tool. So, the goal wasn't to cover everything that could possibly contribute to frailty. It was to narrow it down to the most important or core thing so that we could have at least a crude measurement of frailty in as few items as possible. So, this didn't necessarily state that these are the only things that are important to frailty. Rather, that they have the potential to be the most important things to measure frailty or the most poor part of it. Great. So, a question from another mark, not me. Comorbidity burden is a predictor of outcome. Indices for this exist such as Charleston by including chronic conditions to frailty index, including concepts of coorbidity as well as frailty. Would it be better to exclude chronic conditions from a frailty index? It's possible. Part of the problems with studying frailty is there's so many interrelated concepts. And frailty could potentially include many interrelated things like comorbidity or disability or mental health or anything like that. And so, part of what we looked at and part of the research in looking at the simplified scale is taking out some of those other things that could be heavily related but creating noise in the measurement. And so, the simplified scale didn't include any measures of comorbidity but still seem to be relatively close to the frailty index. So, it is possible that we're including a separate heavily related concept like comorbidity. And so, that's part of what we looked at is taking out stuff like that. So, yes. Okay. From Emuna again, how did you set your frailty index cutoff at greater than or equal to .25? And the follow-up is your index distribution with SKU? Thank you. The SKU was expected for a couple different reasons. We had a relatively healthy sample and so, we're not going to get all components or everybody, we're not going to get an accurate sort of frailty estimate of the whole population. We have, these people have to be relatively healthy community-building adults and they have to be capable of completing a full 60-minute telephone interview on their own. So, we're going to exclude people who are sick, who are institutionalized, who have some cognitive decline to the point where they can't participate. And so, for lots of reasons for our sample, you'll see that the SKU is to the less frail. So, to the healthier end of the people in the, especially those who are, you know, 65, 75, 85. The people who can participate in this study are the healthier members of that age group. So, that helps to account for, or that's our hypothesis as to how it was skewed to that direction. The frailty index threshold of .25 comes from a paper that came out a couple years ago where somebody did a free, they collected a large sample and measured both the freed phenotype of frailty dichotomous measure. So, like a fail and non-fail and they also calculated a frailty index. And they saw that the curve of the frailty index value for the robust people crossed the curve of the frailty index values for the frailty people at .25. So, that was the point where it was most responsible to make that cutoff. But if you refer to the original Rockwood paper, he'll say that the frailty index was not designed to be a dichotomous measure and should be looked at continuously. I did that more for transparency. So, people who are used to the dichotomous classification can get an idea of the level of frailty of the participants. Great. From Laura Anderson, great presentation. Could you further describe how the results of the factor analysis informed the final indices that were developed? The factor analysis was a little bit more exploratory. We wanted to look at sort of what are the domains that come up and then compare those to the results of the equation modeling. The results of the factor analysis weren't used to inform the construction of the model. It was more of a sort of a confirmation that both identified similar domains. So, the factor analysis results, the domains identified were consistent with those we found present in the simplified model. But both of those analyses were conducted independently. Great. From Ahmed, okay. He says, good job. Have you thought about doing rush analysis on the index? I have not, but that's definitely something I will look into. Thank you for the comment. Great. So, we'll get to Satya's question in just a second. A couple of questions from me. The MOS Social Support Availability Index, the questioner is one of the items in your frailty index. So, we have about 19 questions. And I was just curious, did you, you gave your description of how you took each item and quantified it for inclusion in your index, 01, or if it was more than one response options, you gave 0.25 points for one answer, 0.5 points, etc. Is that how you took each of the 19 items on the Social Support Availability Questionnaire, scored them and put them into your index? Yes. Okay. So, you didn't give one value for the whole Social Support Availability Index. You gave one value for each of the questions. Each item for the MOS scale and for the other scales that were included was considered deficit on its own. Okay. So, then that's interesting because could you then, in total then you have 90 items. So, 19 of those 90 items came from the Social Support Availability Questionnaire. And you had, I don't know exactly how many, 15, 20 maybe more from the chronic conditions questionnaires. There are around 20 chronic conditions. There was about 16 reports of physical functioning. We didn't pay too, too much attention to worrying about there being an equal number of items in each of those domains. We just wanted to make sure that we had more than one item or representing each of the domains that we looked at. Right. So, then could it be possible that, and I don't know the answer. I'm just throwing this out to get your thought. Could it be possible that you may have in so doing given more weight to certain constructs than others? If you just treated each question, for example, as contributing one to the denominator of 90 in your index, could you not have given more weight then to certain constructs than others? It's possible, but unlikely and it can be something that we look at, weighting, like sort of assigning, look at the number of items in each of those domains and then weighting the domains equally to see if that affects the measurements at all. But in practice for developing it for the index, sort of the standard procedure is to not worry too, too much about domain weights and just include any sort of deficit. So, based on the theory, if you have deficits in any area, they all more or less contribute equally as long as you're measuring enough deficits. So, I feel like that would become more of a problem if you were collecting 20, 25 deficits, but when you get past 40, 50, 60, as long as you're representing a number of different domains, the number of items in each domain becomes less important. Great. And I noticed in your demographic table that the proportion of people with post-secondary education was quite high compared to the average that you would see in the Canadian population. And you usually get that in studies, even the CLSA tracking cohort was a random sample of the population, but you still see self-selection in most epi studies of individuals who are more educated than the norm. So, do you think that this could have created a bias in your results? And also, did you think about using the sampling weights to adjust for that? It's definitely possible that it contributed bias. As somebody else had noted, that the skew of frailty was towards the healthier age group. So, it's possible that there's a sampling bias which exists in any self-reported study that you're more likely, as you had mentioned, to catch people who are more educated because they're more likely to self-select first today. So, I hadn't done sampling, but it's definitely something that we can look at for sure. Okay. And so, another bias-related question from Satya. Thank you for the interesting talk. If frailty is to identify vulnerable older people, then excluding those with missing data could mean excluding more vulnerable participants. And you had mentioned that. How would that impact your findings? So, I'm thinking of potential selection bias here. It's possible that we're missing some people who are the most frailty because they... And we talked about that a little bit in the inclusion criteria for the study, is the people who are able to provide complete data might be a healthier group than the people who are unable to provide data or unable to be selected. And I think the people who couldn't report for a lot of questions, who maybe didn't complete cognitive function tests, who maybe over the course of the 60-minute interview didn't answer all of the questions. It's possible that the people with the missing data through the interview are a more vulnerable group. And so, the purpose of this study wasn't to assess the frailty of everybody. It was to assess the frailty of people who have complete data in the CLSA. It's important for, I think, from a generalizability standpoint to make sure that these results are reported in the context of the group that is sampled. And it's possible that there are many more frailty people, and it might be interesting to look at frailty in other contexts, such as people who are in nursing homes, and not healthier community-dwelling adults. Great. From Peter, thanks for an interesting presentation. I was interested that you included measures of social support availability, so we've just been talking about that, as part of the frailty index. While I understand the ICF model includes environmental factors, in the ICF model, these seem more as predictors of the components of disability, or factors that could potentially modify how disability components relate. Should the component of social support sit outside of the measure of frailty, or are you confident that including it within the frailty index was the right way to go? This is a phenomenal question. That's part of the reason that we did this study this way in the first place, because the original question was, what's the best way to measure the frailty of this population of these participants? And we said, well, how else has everybody measured frailty in the past? And there are people who include only physical function measures and chronic conditions. There are people who, there are scales that include social factors and environmental factors, and consider those just as important. There are social frailty scales that include only social environmental factors. So our goal here was to be comprehensive and to include everything that people have used as potential domains for inclusion in a frailty measurement scale, and to see their relative performance. So we wanted to see how the social support variables contributed in the context of the other ones. And this analysis showed that they did make a significant contribution. So based on these results and the model that we had, the social support survey items did contribute to the measure. And so based on these results, they should be included. But excluding those and other items for the more core items like the depression and the physical function ones that are included in the FREED index also showed promising performance. And so for a frailty index, I think they should be there because the results, that's what we got from this analysis. But if you want to have a simplified screening tool that looks at only physical function and depressive symptoms, then it doesn't need to be there. It depends on what your question is, I think. Great. Thanks, David. I don't see any other questions. So we'll give people a final minute or two to type any last questions if there are any. While we're waiting for that to happen, I should just note that Laura, our communications officer, if you scroll upwards from the last question in the chat, you scroll upwards, you'll see that she has given a link to where the webinar will be available. And all registered attendees will be notified by email when the webinar is posted. So David, you'll add that additional slide concerning the components of the frailty index and then it will be posted. So that was a response to Fatima's question, where can I find the slides? Just again, scroll up to see the link and you'll get email notification of when the slides are posted. From Ahmed, a final question. CLSA includes participants who are 45 years or older, which is unlikely to be included in a frailty study or to be frail. Do you think the index will be different if you include only people aged 65 and over? Thank you for the question and an interesting one that we thought of when we were designing the study. Most people who have looked at frailty restricted their analysis to people aged 65 and over. And if you are just interested in classifying people as frail or non-frail, you'll definitely find more people who are frail in the over 65 age group. I think one of the benefits of using this data set to study aging is we are including people who are in a younger age group of 45 and up. And it is interesting to see if frailty is detectable. And if that population is indeed less frail, then they're the people who are in the 65 and older age group. So it's an interesting way to look at, to see how the frailty of those participants changes as the study progresses. And I think measuring them at a younger age group is going to be interesting as a comparison point and also to see how that changes over time as the CLSA continues. I think there's a number of advantages of looking at the younger age group that isn't necessarily seen in previous research. I would agree with you. People don't suddenly become frail when they turn 65. So it's not like people who are less than 65 years of age cannot be frail. And I think you hit the nail on the head. One of the benefits of the longitudinal component of CLSA is that precisely you'll be able to do exactly what you just said is look at the development of frailty in middle age and that maps directly onto the life course perspective. May Muna's question, clinically speaking, do all chronic conditions have the same impact on potential frailty or can it depend on what chronic condition you may have? And if the answer is yes, the impact of different chronic conditions can have differential impact, coding each condition is 01 and having it contribute equally to the index could be too simplistic. Another excellent question. The way that we did it is weighting all of the items equally. So every item was coded from 01 whether it was including all the chronic conditions but also many other items. And so that's the standard procedure for creating a frailty index. We wanted to follow along with how it had been done in the past and this has shown good validity a number of times. This equal weighting of all deficits, the more common ones, the less common ones, the ones that are interpreted as having a very high impact on health and the ones that are interpreted as having a relatively low one. So the way that we handled chronic conditions in this analysis was the same as that. It was weighted just as equally as self-rated health. So I don't know about trying to assign different frailty weights for different chronic conditions but because we have enough items, weighting individual items is unlikely to change the overall index score of the individual. So if you had many fewer items that would make more of an impact. If you were collecting only 10 or 15 things, the relative contribution of each item would be, I think, more likely to affect the results. But if we're including 90 deficits, the individual weighting for each item I don't think is worth changing because it's very unlikely to affect their overall score. And I think the performance of our score in the population more than supports that. Right. And there's still a lot of disagreement over exactly how to define frailty or how to measure and insert studies. And I think that we're seeing that in this question and also in my earlier question. And you can certainly discuss different weightings for different questions and that will just generate another 20 years of debating discussion as to how to measure frailty. But it's still a valid point. So I think, sorry, go ahead. No, I agree that it's absolutely a valid point. And sort of the way we handled it was including many, many items so that weighting was less effective. Right. So at this juncture, we're going to end the webinar. David, it was a really interesting talk and it's great to see students using CLSA data for their work. You're not the first student to present. But you know, it's fantastic that we're getting students to present their work with the CLSA data. So it was a very interesting and clear, lucid presentation covering a very important topic. And on behalf of all of us at CLSA, thank you very much. And once again, congratulations on your MSc. Thank you everyone for participating and thank you for your excellent questions. Great. So I'd just like to take an opportunity before closing to mention our next webinar. So the theme is going to be primary prevention of cancer and chronic disease through improvements in unhealthy lifestyle behaviors. The presenter is going to be Katarina Maximova from the School of Public Health at the University of Alberta. We'll have more details on this webinar in the coming weeks and it's scheduled for November 1st at 12 p.m. eastern time. So we look forward to having all of you join us again for that webinar. And again, David, we wish you all the best in your future endeavors. Thank you very much everybody for joining us. Have a good day.