 Hello, everybody. It is noon here in Hamilton, Ontario, so we'll go ahead and get started. Welcome all. I'm Carol DeSine with the Canadian Logistical Study on Aging, our CLSA. Thank you for attending the May talk of our ongoing CLSA webinar series. Today, our webinar is Exploring the Geography of Cogs to Function and Social Support Availability, a spatial analysis of the CLSA, presented by Dr. Jane Law and Matthew Quick at the University of Waterloo. Before we begin, let's review a couple of housekeeping points. Everyone but the presenters will be muted throughout the seminar. At the end of the presentation, there will be a question and answer session. If you have a question at any time, you can post that question by typing it into the chat box located in the bottom right of the WebEx menu. These questions will then be talked through at the very end of the webinar. Ensure that you select all presenters from the chat drop-down menu before you send a question or a comment. Mobile users, please select chat with everyone. Questions will be visible to all attendees, and please remember to exit the WebEx session at the conclusion of the webinar. Now, today's webinar is Exploring the Geography of Cogs to Function and Social Support Availability, a spatial analysis of the CLSA. Let me introduce you to our distinguished speakers, Dr. Jane Law and Matthew Quick. Jane Law is an associate professor in the School of Public Health and Health Systems, jointly appointed to the School of Planning at the University of Waterloo. She is a professional member of the Geomatics Division of the Royal Institute of Chartered Travellers in the United Kingdom. She specializes in health, geomatics, geographic information systems, spatial statistical analysis, and spatial epidemiology. Dr. Law has conducted research involved in longitudinal studies such as European prospective investigation into cancer nutrition when she was a postdoctoral research associate at the University of Cambridge Medical Research Council by a statistical unit. She has held a natural science and engineering research council of Canada Discovery Grant on spatial statistical analysis since 2009. Matthew Quick, who I believe is presenting most of the webinar today, is a PhD candidate in the School of Planning at the University of Waterloo. His research uses spatial and spatial temporal statistical methods to investigate the relationships between geographical context and social demographic and health phenomenon. His work has been published in many prestigious geographical analysis journals. Again, before I begin, just remember that there will be a question and answer session moderated by me at the end of the webinar, but feel free at any time to write any questions or any comments during the webinar and to the chat box. So now I'd like to ask Matthew Quick to begin the webinar. Thanks for the introduction. My name is Matt Quick. I'm a PhD candidate at the University of Waterloo. And thanks for joining me and everybody today to talk about the Geography of Cognitive Function and Social Support using spatial analysis methods applied to the CLSA data. Now through the presentation today, I have two broad goals that I'd like for the participants to take home with them. And the first is that the presentation will illustrate the ways that spatial information can be incorporated into analysis of the CLSA data and broadly how to think about geography in the CLSA. And then second is specifically to focus on how geography helps us understand the relationship between cognitive functioning or the processes that through which one becomes aware or comprehends ideas and social support or the perception and actuality that one is cared for has assistance from other people. So but in particular, the objectives for today's presentation are first to understand the geography of the CLSA, what sort of geographical data and spatial information is incorporated into the CLSA, how we can incorporate this into analysis and then challenges associated with it. The second is exploratory and we're looking at is there overlap between clusters of high cognitive function and high social support or low cognitive function and low social support and how to what degree do these areas overlap. And the third is how does geography help us understand the relationship between cognitive function and social support, what covariates are associated with cognitive function and how much variability is explained by location. Broadly the outline today, first we're going to define spatial analysis and situate this methodological framework within the public health context. We're going to outline what type of geographical information is contained within the CLSA and how to use it in analysis. We'll define cognitive functioning in social support and briefly justify why a geographical lens is useful and helpful in understanding this relationship and then present two sets of analysis. The first is a cluster analysis, exploratory analysis looking at cognitive functioning in social support and overlapping clusters using the tracking assessment. And the second set of analysis will use the comprehensive assessment and fit a set of multi-level regression models that look at the relationship between cognitive function and social support and accounting for a number of risk factors. And finally, we'll conclude with some limitations of the research, some challenges and some future directions. So what is spatial analysis? Spatial analysis focuses on the use of spatial data, which is data where the attribute of interest and its location are recorded. And broadly, these can be considered two types. There's point data or instances on the earth with XY coordinates or latitude, longitude coordinates. For example, individuals, houses or buildings, street intersections, point sources or sensors for pollution, or perhaps sampling locations of fish or trees or something like that. And for point data, we're interested in the attributes of interest, such as where people live, their individual characteristics, their income and their education level, as well as their relationship to other people in space. For area data, we have polygons or areas bounded by lines, most commonly administrative boundaries like neighborhoods or census areas or zip codes. And in these contexts, we're interested in characteristics within these areas and how the areas relate to each other. One of the classic examples in spatial epidemiology is lip cancer risk within counties in Scotland. But we could also look at the distance between neighborhoods and the city center and cognitive function of the people within neighborhoods. And of course, when we're analyzing spatial data, we want to take into account geography or location when we're doing our analysis. And this involves focusing specifically on the relative relationship between data, so where points are in relation to other points. And this follows Tobler's first law of geography, that everything is related to everything else, but near things or closer things are more related than distant things. And we can incorporate these assumptions into analysis. For example, if we're looking at point data, we're looking at air pollution, we can assume that air pollution is a continuous surface over space, and that we incorporate geography by a distance decay function where the highest pollution is from the point source and it decreases as we move farther away. So if we're looking at aerial data or neighborhoods, we can assume that nearby areas have similar levels of income, but that that neighborhoods at a farther distance have less similar levels of income. Here's an example of spatial data from this was created in a GIS software. And sorry, there we go. And you can see here points are in light blue, and points are located within areas which are shaded in purple. So for example, we could use the point might be individuals and each individual has cognitive functioning associated with it. And then each individual is located within area and area has certain attributes as well, like income and like education level. So broadly, spatial analysis can be divided into two analytical approaches. The first is exploratory. And here we're looking at patterns of data, we look at a single variable. And one example is cluster detection and cluster detection is commonly used in public health application, including disease surveillance of where there's clusters of flu emerging. And the second is confirmatory spatial data analysis, which uses frequently regression models. And we're concerned about making inferences on covariates or relationships between outcomes and explanatory variables. And one example in the spatial context is spatial regression. So exploratory methods attempt to describe a phenomenon, describe its pattern, whereas confirmatory spatial methods try to explain it. In public health context, and I'll be using this sort of methods through the presentation, we often have multi level data where individuals are nested within neighborhoods. Historically, best of my knowledge, public health researchers often looked at individual level data. And looking at risk factors and outcomes at the individual level. But increasingly, the literature has identified that contextual group level or area level of risk factors also influence individual health outcomes above and beyond individual characteristics. So combining these two levels, the individual nested within areas naturally leads to a multi level design, compared to, for example, analyzing only individuals or analyzing only groups. So here I've illustrated a graphic. We have individuals. And each of these individuals has different levels of cognitive functioning. Each individual has a different level of social support. They're different ages, they have different levels of educational attainment. And they're all located within different neighborhoods. And these neighborhoods to be characterized by the housing type, whether or not they're located in the central city, their access to green space, population density, and so forth. And so the goal of multi level data analysis is to quantify how important these contextual or area level characteristics are to understanding individual health outcomes. So when we move multi level data into an analytical framework, we have multi level analysis, which is the study of the effects of collective or group characteristics on individual level outcomes. So this comes from a starting point that individual health health is shaped by individual characteristics and area level characteristics. And when we think about this, we can think about it broadly as place or space. So the classic geography talk, where place is group membership. We have level one people or individuals within shared geographical units. So myself and my family live within one neighborhood, and yourself and your family live within a different neighborhood. And these units are state and region, or neighborhoods or schools or families, they don't necessarily need to be geographical. For today, we consider them as zip codes. And then space is the relationship between these areas. So how neighborhoods are located next to each other or far away from each other, how they're located relative to the city center or in suburban areas. Broadly, multi level methods can be considered are often used in the neighborhood effects literature, or the literature that tries to develop theoretical models of disease that explain how group level and individual level variables interact in shaping health outcomes. And specifically in the cognitive functioning literature, few studies use multi level methods. Some of that might be a data availability issue. But typically, most of the variability has been found between individuals rather than between areas. And some characteristics just quite briefly. Some area level characteristics associated with cognitive functioning include educational payment on average, socioeconomic status or socioeconomic disadvantage, and poverty or income. So now that we've got a handle on spatial data, and we're going to think about how to incorporate geography and location within analysis, let's look at how geography is incorporated in the CLSA data. Broadly, the CLSA data has a handful of geographical information fields. One is province, the next is census subdivision, which are municipalities. The third is the data collection site. This is for the comprehensive assessment only. And there's 11 sites, including Victoria, Calgary, Hamilton and Montreal. And then the most precise or most granular geographic level is the forward rotation area. And that is the three character postal code. And this is included in both tracking and comprehensive assessments. So for example, forward rotation area would be K6V or N2H. And those are very in size throughout the country. And each unit of geography has attributes that you can assign to it. And you can look at in analysis. For example, at the province level, we could look at healthcare funding or certain programs. At the census subdivision or municipality level, we could look at air pollution for manufacturing industry. And at the forward rotation area, we could look at whether or not an area is located in in rural areas or urban areas, and as well as characteristics available from the census like income, ethnic composition or educational attainment. So the next slide shows an example of forward rotation areas in Kitchener Waterloo. So you can see, this is the University of Waterloo right here. And it is located within a forward rotation area. Conceptually, in this analysis that I'll present today, we have individuals nested within forward rotation areas. And because there's no more precise or more granular spatial information, we assume that each individual is located at the centroid or the geographical center of each area. Examples, this might be a centroid, that might be a centroid, and that might be a centroid. This doesn't really change our inferences at the area level, but it would have implications if we had more detailed information at the individual level. Forward rotation areas, as you can see from this map, are often contiguous areas. They cover the entirety of Canada, but they vary greatly in size, typically along an urban and rural divide where urban forward rotation areas are smaller, and rural areas are much larger. And postal codes or forward rotation areas are commonly used in public health research because this data is sensitive and reporting individual locations isn't wanted. An important thing to note when we're analyzing forward rotation areas, as we'll do in this study, is that they aren't necessarily representative of a neighborhood, or an activity space, or a meaningful unit of analysis at the individual level. This is a common limitation of geographical analysis at the area level, given data limitations and financial limitations on collecting data. So it's a question about which scale of analysis, if we have a choice, is most optimal to represent sort of exposure to different contexts and different geographical areas. Within the CLSA, there's two assessments, the tracking assessment and the comprehensive assessment. The tracking assessment was done by phone, collected for all the provinces. There's about 20,000 participants distributed over just over 1500 forward rotation areas. That's about, on average, 13 participants per forward rotation area, but this varies dramatically through the sample. The second assessment is the comprehensive assessment, and this is collected within 11 data collection sites in Canada. There's more participants in the sample, 30,000 participants, and this is distributed over a smaller number of forward rotation areas, about 500. This leads to about 60 participants per forward rotation area. So the implications for analysis when we're considering the tracking assessment and the comprehensive assessment is that the tracking assessment likely has more noise when we're analyzing at the area level because there are fewer participants, whereas the comprehensive assessment may have more signal coming out of the individuals up to the area level. The second thing to consider is that the tracking assessment, because it's collected across Canada, has very heterogeneous forward rotation area sizes, or area like kilometer squared of the forward rotation areas. We have very large rural areas that capture almost half of a province in some cases versus very small urban forward rotation areas in cities like Toronto or Vancouver. This raises a handful of challenges with geographical analysis, and I'll discuss these more later. But one of the assumptions we make is that internally within each forward rotation area, that the characteristics are relatively homogenous, that people have the same exposure within that area. Of course, when you have very large areas, that assumption might be violated. When we have large areas in the tracking assessment, we also question the assumption that nearby areas are more related to other areas because the internal homogeneity is very different. This has implications for how we interpret what neighborhoods mean, what forward rotation areas mean, and how we measure variables at that level. To give an example of the differences between the tracking data set and the comprehensive data set, I've extracted some sample sizes here, each assessment in Ottawa or Ottawa Gatineau, rather. You can see here, this is approximately the Ottawa River. I can delineate that because in the comprehensive data set, we have data collected for Ottawa, but not Gatineau, and in the tracking assessment we have data collected for Ottawa and Gatineau. It's collected across the country. You can see here, for example, this area has 58 participants and the same area in the tracking assessment has one participant. For reference, this area is about five kilometers squared or so in area. So it's clear that there are many more respondents per FSA in the comprehensive assessment. So now I've laid out some principles of geographical analysis of spatial data, and so far what I've talked about in terms of the comprehensive and the tracking assessment can be applied to any analysis of CLSA data. And now I'm going to specifically focus on cognitive functioning and social support. And first we're going to consider the rule that geography plays in understanding each of them separately. We'll separately analyze cognitive function and separately analyze social support and look for overlapping clusters. And then the second set of analysis will consider how geography and how spatial information helps us to understand the relationship between cognitive function as our outcome and social support as our explanatory variable or risk factor. So cognitive function is the process through which one becomes aware of or comprehensive ideas. And cognitive function is important because it's precursor to Alzheimer's and other dementias. It's a source of impairment in activities of daily living. It requires considerable informal caregiving or formal caregiving and leads to institutionalized care and that's costly in the long run. So if you can better understand how to reduce cognitive decline, perhaps via social support we can eliminate these problems. So broadly literature has found support that social support is associated with physical and mental health outcomes including cognitive function. And this is cross-sectional. So at one point in time high social support is associated with high cognitive functioning and longitudinal where high social support reduces the rate of decline of cognitive functioning. It's also associated with physical health outcomes like cardiovascular disease and hypertension. And social support comes in many forms. For example emotional support or the things that make us feel loved or cared for. Instrumental support or the tangible support to help with housekeeping or providing transportation and informational support or offering advice, providing information, perhaps that access healthcare. So some of the mechanisms through which social support influences cognitive functioning I've listed on the slide and often these are complex and they interact but for example engaging in social activities via your social support networks, phosphorus communication and interactions this stimulates your brain and increases synapse genesis. Your brain cells work more efficiently or take over the functions of affected cells. We also have positive emotions through social support. So talking to somebody or reconciling your emotions with somebody at least a more positive self-image protects against stress and anxiety and this reduces your cardiovascular reactivity your blood pressure and heart rate. We also have research showing that social support and social networks facilitate physical activity which also reduces stress and anxiety and benefits cognitive function as well as other risk factors associated with cognitive function. So situating this research geographically we consider how the characteristics of one's living environment influences cognitive functioning and recall that this comes from the neighborhood effects literature and multi-level models that individuals have exposures that are beyond themselves that behaviors and risk factors are due to individuals as well as their group setting and geography is particularly important for older adults because they often have mobility constraints such as limited capacity to drive they spend relatively more time in their local environments they're more dependent on local resources and services so they're more closely tied to the walkability or the food environment or the shops and banks nearby there's typically a fewer contacts with social network members so they have relatively stronger associations or they have relatively stronger exposures to the local environment but situating the method or the mechanisms through which social support influences cognitive functioning geographically so putting them through a spatial lens somehow neighborhoods or areas have to get into the body and influence the processes the the direct processes that influence cognitive functioning the first is engaging in social activities and past work has shown that areas with relatively higher proportions of adults have better cogn the adults sorry older adults that live in neighborhoods with higher proportions of older adults have better cognitive functioning perhaps due to better service availability policies or programs targeting areas with high concentrations of older adults but also based on social network performance or social network formation within these areas in certain areas they may have greater number and higher quality of resources and amenities that encourage physical activity and social activity and also the built environment the quality and the design of a living environment the safety and walkability associated with physical activity if it's an area is unsafe people may withdraw and experience social isolation and the built environment the walkability also facilitates access to services amenities like parks and libraries that serve as sites for cognitive stimulation and also physical activity and recognizing past work finding that land use mix and walkability are positively associated with cognitive functioning and memory recall the role of the built environment also shows how or provide some opportunities for interventions and this is perhaps most commonly in the planning literature referred to as H80 cities so designing cities and designing neighborhoods so that people aged eight people age 80 can can move and and and live in those areas well and changes to the physical environment and walkability are relatively long-lasting influence all individuals and and can be implemented graphically or targeting specific areas so focusing now on the data that we've analyzed and this is data that we analyzed through all the remainder of the session is global cognitive functioning that's our individual level outcome variable and it's a sum of memory and executive functioning and memory was operationalized using the ray auditory verbal learning and delayed recall and executive functioning is the mental alternation test so counting numbers and residing the alphabet and alternating between the two and animal fluency or is naming as many animals as you can and this was the average of these two Z scores so the variable centers about at zero and then our explanatory variable of interest we we do analyze others is overall social support and this is the combination of emotional tangible positive social interaction and affectionate levels of support so emotional is your sense of being looked after tangible is a concrete way of providing support positive social interaction is did you share or enjoy a good time with other people and affectionate support is feeling of being loved or receiving a section so it's the average of overall social support is the average of all four of these sub-scales and the variable center around four so now I'm going to provide or show the first set of analysis which is the space the cluster analysis it's exploratory we're not really making any inferences about the relationships between the two but just looking at patterns and similarities between cognitive function and social support and this uses the tracking assessment and we specifically focus on the forward sorciation area scale so for this level this set of analysis got three sort of guiding questions and the first is methodological how can we generate area level estimates of cognitive functioning and social support the second is how important is geography in explaining each of these outcomes is geography more important for cognitive functioning or is it more important for social support and why and question three is where are these clusters located and do areas with high cognitive functioning and high social support overlap what can this tell us about future analyses perhaps regression models and for this assessment we use the tracking or for this analysis we use the tracking assessment so here is the data this is all of the FSAs included in the tracking assessment there's about 1500 areas and about 16 000 participants using the complete cases data you can see the difference in sizes of areas and one of the limitations of doing analysis using the entire tracking data set geographical analysis using the entire tracking data set is we have very small forward rotation areas in the urban areas but very large in some of the rural areas I'll talk about more about that later so here's the data for global cognition overall social support see over global cognition mean it's about zero and standard deviation of three and overall social support closer between four and five so most people have relatively high social support mean of 4.3 and so our first question is how do we generate area level estimates and so we use a multi-level model and in this case we fit one individual level covariate using the tracking data set we have individual level cognitive functioning or social support they're analyzed separately and then here we have our individual level variance or the variability explained at the individual level and then each of our individual level cognitive functioning or social support is explained by an intercept or the average coefficient for virality whether or not person lives in a rural area or not and then we have the residual effect of being located within a forward rotation area and this was we don't impose any assumptions on this other than it follows a normal distribution and there's some common variance that explains the between area variability so I've provided the interpretation down here if you want to read that specifically for those interested we implement this in a Bayesian framework so there's priors and and this is just a straightforward way of implementing multi-level models and I'm happy to talk about this more in the questions or you can email me personally so we analyze virality and only virality in this context because rural areas are not terribly prevalent in the comprehensive assessment the comprehensive assessment if you recall are collected mostly for urban data collection centers so there are very few rural areas included in these samples past research has suggested although it's inconclusive that cognitive functioning is lower in rural areas compared to urban areas perhaps because urban areas have more complex environments offer more cognitive stimulation and a greater diversity of experiences it also could be due to the measurement tool or that people living in rural areas are less familiar with procedures used in cognitive testing so this shows the results of our analysis of the tracking assessment for across all of Canada and we analyze them separately so you can see that the intercepts broadly align with our data you can see cognition was centered at zero so that's about right and overall social support was average was about 4.3 so the intercepts align with the data and we can also see here the coefficient estimate for whether or not an individual lives in a rural area or not and we can see that virality is associated with global cognition so people who live in rural areas have lower global cognition slightly lower but there's no effect on social support and we can see this because this is the uncertainty interval 95% interval and this crosses zero whereas this interval is unambiguously negative so that's at the individual level at the forward rotation area level we have our area level estimate so how much of the individual clustering can be explained at the area level at the forward rotation area level and here we see the range of the forward rotation area estimates for global cognition is much greater than for overall social support this suggests and is confirmed by our variance partition coefficients that areas or geography is much more important for understanding cognition than it is for understanding overall social support and just to give you some context on these numbers 2.25% of the variability of cognition is explained by area past research has found that about 1% of mortality is explained across public health units 2.5% of respiratory function or blood pressures is explained between districts about 1% of drinking behaviors are explained between districts so we're about in line it's small but we're about in line with what people have found in past public health literature just showing the calculation of variance partition coefficient here and you can refer back to this slide to identify those parameters so here the map finally right so here we've got global cognitive functioning and we've classified them into turtiles so areas with the highest forward rotation area estimates or area level estimates and areas with the lowest one-third of area level estimates and the first thing that we can see is that when we're analyzing the tracking data set and we're visualizing these results is the map is dominated by these large rural areas the second thing we can see is that the pattern is heterogeneous there's a mixture of high estimates low estimates middle estimates throughout the country but at the regional level I've sort of identified a couple clusters that might be of interest the first is much of British Columbia you can see a lot of it is green or in the high turtile or positively associated with cognitive functioning you see much of sort of central Ontario has low levels cognitive functioning and much of Atlantic Canada also has low levels I'm not going to try and hypothesize why this might be this is just simply describing the patterns that are observed the second we show overall social support at the area level and the same interpretation where green is high levels of overall social support and orange is low but recall that there's a difference here in sort of the high estimates of cognitive functioning are of a greater magnitude than the estimates of social support and we can see that the pattern is much more heterogeneous the mixture of green and orange and gray in this map is much greater than in cognitive functioning so for example compared to cognitive functioning we have more positive effects throughout eastern Ontario we have more heterogeneous high low mixtures in Atlantic Canada and in British Columbia so putting these analyses together we can look at where areas overlap or where areas have high cognitive functioning and have high social support or areas that have low cognitive functioning and low social support so there's about 366 green areas or high high overlaps about 200 low low overlaps low cognitive functioning and low social support and the remainder are are non-overlapping are about 900 so some notable high high clusters we can see Victoria Vancouver throughout eastern Ontario and then there's some clustering here through Atlantic Canada and some low low clusters New Brunswick and Quebec and also these large areas but recall that these areas are only a handful one or two areas just they're very large so they dominate the map so giving just a more detailed look we can see these clusters in Vancouver and Victoria to also illustrate the size difference in forward rotation areas that's one of the broader themes running throughout this presentation is it with very small forward rotation areas in the urban areas but very large throughout the rural areas same in Ottawa and Kingston lots of high high overlaps throughout and not too much in the city and we've got Calgary we've got some low low low cognitive functioning and low social support throughout southern Alberta and our surrounding Quebec city okay so just to summarize that first set of analyses area level effects or sort of aggregating from individuals up to areas is more important for cognitive functioning than social support but two times more areas had different cluster classification so high cognitive functioning and low social support than they had overlapping cluster classifications and of course it's challenging to interpret these patterns using the tracking data set there's a lot of noise because there's not too many people within each area but there is some evidence of high high overlapping geographically okay so the next set of analyses we're going to look at multi-level spatial modeling and this sort of confirmatory we're going to quantify the relationship between social support and cognitive function and account for a handful of individual characteristics and also the residual variability that's due to being located within a certain board rotation area and we apply this in three ways we're going to look at the largest six data collection sites so that is Calgary, Halifax, Montreal, Vancouver, Winnipeg and Ottawa and we're also going to look specifically at Vancouver and we're going to look specifically at Montreal so guiding this sort of sub-analysis we've got three questions is social support associated with cognitive functioning in the comprehensive assessment and what does this relationship persist after we control for individual covariates like education or age or gender the second question is how important are forward rotation areas for understanding this relationship and more empirically does including neighborhood or context level effects improve our model fit and second or third where are areas with high cognition or low cognition located within data collection sites specifically Montreal and Vancouver and does this provide any insights into potential area level risk factors that we could analyze in future research so here's a map of the forward rotation areas included in the data collection sites for Vancouver and Montreal note that I've omitted a handful of rural areas with that were designated rural as per the forward rotation area census data just to make sure that our analysis is somewhat comparable between forward rotation areas such that they're all urban so in multi-level analysis and in geographical analysis there's a lot of questions around how do we operationalize neighborhoods and just to give you an example on the left we have forward rotation areas and on the right we have neighborhoods such as they're defined by Montreal planning so urban planning and these are relatively larger the neighborhoods defined by the planning department are relatively larger than forward rotation areas but forward rotation areas are relatively larger than census tracts and census tracts are the largest or the most common most commonly used level of analysis at that level two at the context level in multi-level modeling here we can see Vancouver and we have a similar insight that neighborhoods as defined by the planning department are slightly larger than forward rotation areas so we can't necessarily say that forward rotation areas are not indicative of neighborhoods but again this is a question that deserves further discussion and further analysis so here's the data for cognitive functioning in Vancouver and Montreal we've got each site is in purple and the overall for all of the the six largest data collection sites is in green so there's no real differences between between the sites and the overall sample but we can see that Montreal has more areas than Vancouver and I'll talk a bit about this later when we when we get to the model fit but this is important when we're considering the multi-level analysis and that more areas provides you know helps us improve our model fit so our covariates include overall social support that's our interest we've got a handful of risk factors there I'm not going to go over them now but these are all have been all associated with cognitive functioning in previous work so here we've got a multiple multi-level model similar to before but we include multiple risk factors at the individual level we've still got our between individual variants and our between area variants and comparing these two variance parameters helps us understand how important areas are note that I have model one which uses only individual level and we're going to compare this to model two which uses individuals and areas so comparing sort of a single-level analysis and a multi-level analysis so our results are here we have overall for the six data collection sites in red Montreal is in green in and blues in Vancouver and we can see that the relatively consistent results throughout all data collection sites that are comparable to overall the point estimates which are the dots are relatively similar and uncertainty intervals are smaller for the overall sample because there's more people and they're slightly larger for the Montreal Vancouver because there's less people so here we have social support positively associated with cognitive functioning so is being female and your education level that's achieving a high school education and we can also see that gender or age sorry as you get older your cognitive functioning clients gradually just give you some point estimates on these parameters we have overall social support all of these uncertainty intervals are overlapping so we have relatively similar effects between the overall sample or the six data collection sites and the specific Montreal Vancouver case studies and this confirms insisting research not going to really get into it too much more detail so I think I'm running out of time our second is age we have consistently declining cognitive functioning or negative relationships as age increases again consistent with past work and here we have our positive covariates associated with better cognitive functioning achieving a high school education and being female one interpretation of the gender or female male differences that men have less intensive social support networks but more widespread so perhaps it's the meaningful of the social connection that's driving cognitive functioning and not the number of social connections that's driving this relationship to our second question drive it in this set of analyses does it improve model set is there even a purpose of including this relationship including area level effects and we can see here that when we include area level effects in Montreal and in the overall sample we improve model fit and I'm not going to go into this in detail but basically we want to try and minimize the DIC or the deviance information criterion here we have D with a bar on top of it that's sort of the goodness of fit or how much difference there is between the real data that we observe in the in the survey versus the fitted data coming out of the model and PD is the number of effective parameters so the complexity of the model so smaller D bar which we have better matching with the real data and smaller PD we have simpler models and we can see that this improves adding area level improves model fit in the overall sample or the six data collection sites and in Montreal but not in Vancouver and one reason that this might be is that because Vancouver doesn't have as many areas as in the overall sample and is in Montreal so we're not capturing as much variability at that level but we can see consistently through all of the samples that our model fit or our goodness of fit sorry is smaller across all of the data all of the all of the data collection sites but that this is balanced out by increasing the number of increasing the complexity of the model or increasing the number of parameters by adding in that area level effect okay so quickly to summarize I'm showing the area level results here of Montreal or green is positive or positive cognitive functioning and purple is low negative cognitive functioning this is the entire study region on the left and a zoom up of sort of the downtown area this is Mont Royal the park I've highlighted that there so we can see here that that many areas in the downtown of Montreal are green so that's positive functioning so maybe there's some influence of density or being located close to the city center or resources that are located around there associated with cognitive functioning and here I've shown Vancouver the entire study region again on the left and a zoom up of sort of the city of Vancouver and the downtown area on the right and we can see that through Kislano point gray through the center area here we have high functioning and on the west side we have high functioning but we have low relatively lower cognitive functioning in Gaston and the downtown east side so maybe that points to some socioeconomic respectors influencing cognitive functioning at the area level so to conclude I'm just going to review some limitations some challenges of doing geographical analysis in the CLSA and then and then some directions for future work the first is the size and how FSAs are defined and the internal homogeneity of forward rotation areas they're constructed for coastal delivery not for representing health related processes or exposures um so uh this is an important consideration to take into account when we're doing geographical analysis in the CLS data the second thing is is that little research has specifically articulated the mechanisms through which areas or neighborhoods or geographies influence cognitive functioning the areas have to come into the body somehow but we don't really understand how that's an important direction of future research and the second thing the third limitation is how to consider sampling and the representativeness of each data collection site or even within forward rotation areas how does this compare to the entire sample I'm going to look at the using weights future research and comparing these sub-samples so challenges the first and most prominent challenge is that the most precise geographical information for participants is the forward rotation area and these are slightly larger in urban areas and much larger in rural areas than census tracts which are commonly used in neighborhood effects literature including in neighborhood effects literature focusing on cognitive functioning in particular the second thing is how do we interpret the variants explained at the area level they're not directly transferable to interventions because they're capturing residual variability and we haven't included any covariates yet so that's my next stage of future research is to add covariates to help us understand this and these might include education level or population density or walkability or access to libraries or public transit that'll help better understand how we can implement interventions and what area level factors are influencing cognitive functioning and the third thing is common in neighborhood effects literature is the broader social political forces influencing residential location and tenure and mobility and how people move how does that influence exposure so future research one of the real advantages I see in the CLSA data is this longitudinal in the follow-up studies so we can look at how change in cognitive function interacts with change in area areas areas gentrifying or we add new services how does that influence cognitive functioning that's a huge opportunity for this data set that hasn't been done in past research future research also look at the correlation structures between multiple indicators of cognitive function past work has found that memories associated with socioeconomic status but not attention or other things or other indicators so that's worth further examination and the second thing is modeling the complex variants so we assume a common variants but it could be hypothesized that socioeconomic status influences mean area level socioeconomic status influences the mean cognitive functioning within areas but that also influences the variability perhaps there's more people at the high level than at the low level versus in low cognitive functioning it's more the individual or more dispersed across the cognitive functioning spectrum and just to revisit the objectives we've understood the geography spatial analysis how can be incorporated we've looked at overlapping clusters and we've also looked at the relationship between cognitive functioning and social support I just want to thank Dr. Mark Ormus he's the leader of this subgroup of cognitive functioning Dr. Colleen Maxwell Dr. Student Tias Dr. Holly Twocco and Dr. Candice Connert they're all on the team as well Emily Rudder provided an excellent list of covariates that I did use so thank you Emily and Katherine Galli and Shirley Crocus for getting this webinar up and running I'm now ready to field any questions thanks Matt that was really really a great presentation I'll open it up now to questions for both Matthew Quick and Dr. Jane Law who's available for questions I'm pretty sure just a reminder muting remains on so I will be reading the questions but you can enter your questions into the chat box for the bottom right corner of the Web X window and I will read them out so we have a question but before we get to that and to allow other people to put some in I'll go ahead and start off with one so you talked about the scale of the analysis as your kind of background for using the forward rotation area that for most of your analysis for all of your analysis here can you can you tell us why you looked at the SA instead of maybe the census I mean not using something else sure so forward rotation areas are the smallest or most precise most granular geographical information and I thought that would be a good start past research has typically focused on census tracks or neighborhoods and forward rotation areas are the unit within the CLSA that best approximates this there's no reason why you couldn't do the same sort of analysis using the entire data collection site or entire census subdivisions although there are some limitations when you want to look across the tracking cohort for example because census subdivisions aren't they don't cover the entirety of Canada so you're going to get it's the most granular data available it is yes oh okay okay thank you I have a question from you long Lang Hi Matthew could you please elaborate a bit on the following how a sample weight used in the spatial analysis we haven't used sampling weights yet but that is the next step as I think I noted towards the end there maybe that question came in before that and looking at the you know how the sampling weights partition out between areas between data collection sites or forward rotation areas do you expect that to make a big impact on the analysis and the outcomes that's a good question and I'm glad you don't really have a very good answer for you right now and also question two could you please elaborate a bit on how the predictive variable social support was constructed from its four components sure concrete support diffraction etc sure I can flip back in my notes here but there is a handful of questions asked each individual along each domain and I think they were scaled on one to five and then like a Likert scale and then the overall social support was the average of those four components and each of those components was consistent of a set of questions okay thank you from Christina Wolfson there is some evidence of differences in some of the cognitive test results in the function of language such as English or French also in Montreal there is a very strong association between the location in the city or outside of the city and language so it might be useful to look at this when interpreting the results from Montreal it makes sense to maybe look at a language language effect okay yeah or did you consider that I didn't consider it to the best of my knowledge I think for cognitive functioning they were standardized separately for English and French respondents but we could include a language I suppose and also sort of a I've been thinking a lot about distance to the city so outside the city inside the city again there's a question like how do you measure distance until well but definitely that's something that really became evident from Colleen Maxwell Matt how do you describe relevance of this work to policymakers and charges dementia strategy for the provinces sure so I would say that so we've got the individual level analysis and I assume that the you know sort of translation of that into policies relatively straightforward but translating the area level results I would say that the first thing is I would say that we need to look at these areas that have relatively low cognitive functioning or negative associations cognitive functioning and trying to understand what it is about them maybe it has to do with access to services or physical activities or things like that and thinking about what we can do to improve improve that within the areas great we have a question from Brittany Scarfo which software was used for the spatial analysis so for these multi-level models I fit them in WinBugs which is just like Bayesian modeling software from the early 2000s I think you could also fit them in ML Win or or it's probably an R package to do multi-level models as well we have a question from Nazmou Sohal thanks Matthew for an excellent presentation for the regression model are you using OLS or spatial regression and what kinds of weight metrics are you considering here based on spatial weight metrics spatial dependencies may change for for conferencing we may expect more variability due to small FSA size an area but we will lose information on a reality can you how can't how do you address these sure so it's not fit in OLS but the results are comparable to if you would fit it using ordinary squares excuse me I did not include spatial component in particular so that would be a parameter that describes relationship like a leg or a spatial weight or something like that I did try I did use a spatial parameter but there was no it didn't improve model fit in some cases it made models at worse because you're adding an extra parameter but not gaining any more information from the model so I think one explanation for that is these forward transition areas are relatively large if we constraint it to only the smallest we might see some sort of spatial clustering but you know if we flip back to the maps if we included a spatial parameter we would expect to see clusters of green dark green leading to clusters of roughly the less green leading to white leading to light purple you know sort of a smooth map and we don't really see that so I think we might have to do a size of the FSA I hope that answers the question okay from Anne Tooney excellent analysis presentation of very interesting results I'm curious how well you expect to be able to add covariance to the FSAs as some may be quite homogeneous while others may be quite heterogeneous in terms of education income etc we'll see this in some Calgary neighborhoods though not in others thanks again sure yeah so we can analyze sort of like socioeconomic variables at the individual level and I think they're included in some of them aren't included in the field is that as well as at the neighborhood level so some will have more variability for example access to green space will be really high in some neighborhoods but really low in others perhaps those are located close to the city center so that variability would be entered in or the variable we entered in at the area level I suspect that some like median income for example or educational attainment will be not as strongly associated at the area level or those contextual effects not as large in magnitude as perhaps some of the variables that are more heterogeneous throughout the cities yes that makes sense we'll take one more question here from Brittany Scarfo great job that we didn't think the decrease in DIC was large when adding FFA to the model was it statistically significant sure so the typical threshold used is five points anything less than five points is sort of considered the same and anything less or greater than five points difference is is better so you can see here that the increase in one but we have six points and we have uh 60 points about I think more than that 90 points so yeah we consider these better models overall great and so Mike Ormister you uh you reference as being a collaborator for you has written a couple of explanatory notes that I'll just read through real quick FSA is the only data available in CSA and the CSA won't release full postal holds to preserve preserve confidentiality uh overall SSA is computed by taking the average score and all liking quests to the social support module so a little bit clarification in addition to the answer the question on that and these scores for cognition were computed separately for English and French speakers so maybe that that addresses the French speaking question a little bit better well I want to thank you again it was a really wonderful presentation I can I can see that you from all the questions and the interesting discussion we've been having how how well people have taken it so thank you again and we appreciate your participation in the CSA webinar series great thank you so I'd like to remind everyone that CSA data access request applications are ongoing the next deadline for applications is June 11th 2018 please visit the CSA website under data access to review available data for their information and details about the application process oh I'll skip to the very end and I'd like to remind everyone that our next webinar scheduled for next month we'll be welcoming Dr Chris Vershore an assistant professor in the department of pathology and molecular medicine at McMaster University to discuss age of menopause and to relation to frailty and biologic age and the CSA comprehensive cohort so please register soon and join us for next month webinar next month's webinar will be the end of our 2017-18 webinar series we'll be taking a break for the month of July and August and starting our next webinar series in September so please join us next month for the end of this year's webinar series thank you everybody