Social Network Factors Affect Nutrition Risk in Middle-Aged and Older Adults: Results from the Canadian Longitudinal Study on Aging

To determine which social network, demographic, and health-indicator variables are associated with SCREEN-8 (nutrition risk) scores at two time points, three years apart, using data from the Canadian Longitudinal Study on Aging. A retrospective cross-sectional study. 17051 Canadians aged 45 years and older with data from baseline and first follow-up of the Canadian Longitudinal Study on Aging. Nutrition risk was measured using SCREEN-8. Social network factors included social network size, frequency of contact with social network members, social participation, social support, self-rated social standing, and household income. Demographic variables included age, sex assigned at birth, marital status, educational attainment, and living situation (alone or with others). Health-indicator variables included depression, disability, and self-rated general health, mental health, healthy aging, and oral health. Multivariable linear regression was used to analyze the relationship between the social network, demographic, and health-indicator variables and SCREEN-8 scores at two time points, three years apart. Among the social network variables, individuals with higher social participation, self-rated social standing, and social support had higher SCREEN-8 scores at baseline and follow-up. Among the demographic variables, individuals who were single or widowed, compared to married or partnered, had lower SCREEN-8 scores at both time points. For the health-indicator variables, individuals who screened negative for depression, and those with higher self-rated general health, healthy aging, and oral health had higher SCREEN-8 scores at both time points. At baseline, as age increased, SCREEN-8 scores also increased. Individuals with low social participation, low social standing, and low social support may be at increased nutrition risk and should be proactively screened by healthcare professionals. Interventions and community programs designed to increase levels of social participation and foster social support may help to reduce the prevalence of nutrition risk.


Introduction
N utrition at midlife and beyond has an influence on health and well-being (1,2). However, as people age, there can be physiological, psychological, and social changes that affect food and nutrient intake (3,4).
Inadequate dietary intake can lead to nutrition risk, the risk of poor nutritional health or status (5). Unaddressed, nutrition risk can lead to malnutrition (6). In Canada, approximately one-third of community-dwelling adults 55 and older are at nutrition risk (7,8). The consequences of nutrition risk include frailty, hospitalization, mortality, and decreased quality of life (9)(10)(11).
Social networks also change with age (12), and influence health throughout the lifespan (13,14). We define social networks as "the web of social relationships that surround an individual and the characteristics of those ties" (15, p847). A variety of health outcomes, physical and psychological, are associated with social network characteristics (13)(14)(15)(16).

Conceptual Framework
Social network theory, as described by Berkman and colleagues, describes how social networks affect health (15). This framework posits that social and cultural contexts (macro level) influence social networks (mezzo level), which in turn affect social and interpersonal behaviour (micro level) (15). These psychosocial mechanisms impact health (15).
Characteristics of social networks include their size, range, density, boundedness, proximity, homogeneity, and reachability (15). Characteristics of the ties within these networks include frequency of contact, frequency of participation, reciprocity of ties, and duration of ties (15). Social network theory describes how these social networks influence behaviour through psychosocial mechanisms that include the provision of social support, social engagement and attachment, and access to resources and material goods (15).
Many social factors are associated with nutrition risk and food intake (16). For example, eating with others improves dietary intake and reduces nutrition risk (16). Some studies have found that social support helps reduce nutrition risk (16), likely because individuals with increased social support have greater assistance with food-related activities, such as meal preparation and grocery shopping (16). However, most studies have examined individual social factors in isolation, while our study seeks to examine many social factors through the lens of Berkman and colleagues' social network theory (15).
Focusing on the mezzo and micro levels of this social network theory (15), our study aims to determine what

Social Network Factors Affect Nutrition Risk in Middle-Aged and Older Adults: Results from the Canadian Longitudinal Study on Aging
social network characteristics and psychosocial mechanisms, collectively referred to as social network factors, predict nutrition risk scores at two time points, three years apart, using data from the Canadian Longitudinal Study on Aging (CLSA).

Methods
Using data from the CLSA we conducted a multivariable regression analysis to discover predictors of nutrition risk in Canadian adults aged 45 and older. CLSA participants were between 45 and 85 years old when recruited between 2010 and 2015 (17). Baseline data from participants was gathered at that time. The CLSA gathered the first wave of follow-up data between 2015 and 2018 (17,18). As we have data at two time points, we can conduct longitudinal analysis and examine which variables at baseline can predict nutrition risk at both baseline and at first follow-up, henceforth called follow-up.

Data source
The CLSA research team has reported the study design elsewhere (17). To summarize, there are two cohorts of participants, tracking and comprehensive. There are 21,241 tracking participants who are followed by telephone interview only and there are 30,097 comprehensive participants who are interviewed in person, undergo physical assessments, and provide urine and blood samples (17). The CLSA randomly selected participants in the tracking cohort within sex and age strata in each Canadian province. The proportion of individuals in the tracking cohort from each province is proportional to the province's population to allow CLSA data to be generalized to a given province's population and the Canadian population (17,18). Therefore, we used data from the tracking cohort for our study.
The CLSA research team has described the selection and recruitment process elsewhere (17). Briefly, there were three sampling frames used for the tracking cohort: a subset of participants from Statistics Canada's Canadian Community Health Survey -Healthy Aging (CCHS-HA), registries from provincial health care systems, and random digit dialing (RDD) of landline telephones (19). There were 3,923 participants recruited from the CCHS-HA, 3,810 through mail-outs from provincial health care ministries, and 13,508 through RDD (18). Since individuals with lower socioeconomic status and less education are often under-represented in populationbased studies, there was an attempt to over-sample certain regions identified from census data to ensure representation of these groups (19). The CLSA does not include residents of the Canadian territories and some remote regions, individuals living on First Nations reserves and settlements, regular force members of the Canadian Armed Forces, and individuals living in institutions, including long-term care homes (19). CLSA participants had to understand English or French and be cognitively and physically capable of answering the study questions by themselves (19). Trained interviewers identified individuals who were unable to provide informed consent and judged whether the individual was able to provide reliable information (20).
The core questionnaires in the CLSA are common across the tracking and comprehensive cohorts. There are questionnaires examining demographics, social measures, health status and function, lifestyle and behaviour, and psychological measures (19). When available in English and French, the CLSA used validated questionnaires. When there was no validated questionnaire available, the CLSA used established questionnaires from national surveys, such as Statistics Canada's CCHS (19).
We mapped CLSA data onto the variables in Berkman and colleagues' social network theory. Table 1 illustrates how we mapped CLSA measures onto the variables in this theory. The CLSA measures are described in more detail below.
The CLSA research team has reported the complete list of CLSA measures in the literature (17). We used data from baseline for social network, demographic, and health-related variables. We used data from both baseline and follow-up for nutrition risk. We describe these measures below.

Social Network Size
Participants indicated the number of people in each of these groups: children (biological, adopted, step), siblings, close friends, relatives, and neighbours. Participants also reported the number of people known through work or school, through community involvement, and through other activities.

Frequency of Contact with Network Members
Participants reported when they last met with members of each of the following groups: children, siblings, close friends, relatives, and neighbours. Responses included: more than one year ago, within the past year, within the past six months, within the past month, within the last week or two, within the last day or two, and live with me. These responses were collapsed into two categories: low contact for more than one year ago, within the past year, and within the past six months, and high contact for within the last month, within the last week or two, within the last day or two, and live with me.

Social Participation
Participants reported how often they took part in eight different types of activities over the past 12 months. These were: family/friend activities, religious activities, sports or physical activities with others, education or cultural activities, clubs or fraternal organizations, association activities, volunteer or charity work, and other recreational activities. Participants could respond never, at least once a year, at least once a month, at least once a week, and at least once a day. The responses from each category were summed to create a social participation variable that could range from 0 to 32, with higher numbers indicating increased frequency of participation.

Social Support
Social support was measured using the 19-item Medical Outcomes Study Social Support Survey (MOS). It measures multiple components of social support, including affection, emotional and informational support, tangible social support, and positive social interaction. The MOS has excellent internal consistency (overall and subscale Cronbach's alpha ranging from 0.91 to 0.97), and test-retest reliability (ICC = 0.78 after 1 year) (21).

Self-rated Social Standing
Participants were asked to think of a ladder with 10 steps as representing where people stand in their communities. At the top of the ladder (or step 10) are the people who have the highest standing in their community. At the bottom (or step 1) are the people who have the lowest standing in their community. On which step would you place yourself on this ladder?

Demographic Variables
Demographic measures in the CLSA include age, sex assigned at birth, living situation, marital status, education, and income. Participants reported their marital status: married/ common-law, single (never married), divorced, separated, or widowed. We combined these into three categories: married/ common law, single (including single, divorced, or separated), and widowed. Participants were asked about their highest level of education. We used the level of education (4 levels) variable. Possible responses were: less than secondary school graduation, secondary school graduation, some post-secondary, and post-secondary degree/diploma. Participants reported their household income from all sources using the following categories: less than $20,000, $20,000 -$49,999, $50,000 -$99,999, $100,000 or more. Participants were asked the number of people living in their household, other than the participant. We categorized individuals as living alone if they indicated no other person resided in their household.

Health Indicator Variables
Self-rated general health, mental health, healthy aging, and oral health: These were measured by asking participants: would you say your general health/mental health/healthy aging/oral health is excellent, very good, good, fair, or poor? We collapsed these into three categories: excellent or very good, good, and fair or poor.

Depression
Depression was measured using the short form of the Center for Epidemiologic Studies -Depression (CES-D10) Scale. The CES-D10 has excellent sensitivity and specificity (22). The CES-D10 has a positive correlation with poor health status and a negative correlation with positive affect (23).

Disability
Basic activities of daily living and instrumental activities of daily living were measured using modifications of the questions of the Older Americans Resources and Services (OARS) Multidimensional Assessment Questionnaire. The OARS scale is a valid and reliable tool (24). Participants were classified in the CLSA into 1 of 5 categories, ranging from no functional impairment to total impairment. We collapsed these into three categories: and mild or no impairment, moderate impairment, and total or severe impairment.

Nutrition Risk
Nutrition risk was measured using the abbreviated version of Seniors in the Community: Risk Evaluation for Eating and Nutrition II (rebranded as SCREEN-8) (25). Eight questions ask about typical daily eating habits, and include questions on weight change, meal skipping, appetite, swallowing, servings of fruit and vegetables, fluid intake, eating with others, and meal preparation (25). Scores range from 0 to 48 (26). When compared to registered dietitians' assessment of nutrition risk, SCREEN-8 has very good specificity and sensitivity with an AUC ≥ 78%. The test-retest reliability of SCREEN-8 is very good with ICC=0.84, as is the inter-rater reliability with ICC=0.79. A SCREEN-8 score less than 38 indicates that an individual is at high nutrition risk (26). SCREEN-8 was treated continuously using the full scale as the outcome for this study.

Data analysis
Our primary analytic strategy was multivariable linear regression. For all analyses, we applied weighting to the data, as recommended by the CLSA. We used inflation weights for descriptive statistics and analytic weights for all other analyses. Due to the large sample size, we chose α ≤ 0.001 as our significance level and reported effect sizes (27). We analyzed the data using IBM SPSS Statistics, Version 28 and RStudio Version 2022.02.3.
We started by running univariate analysis in SPSS to obtain descriptive statistics for each variable. For continuous variables, we reported means and standard deviations (SDs). For categorical variables, we reported percentages and the number of participants. We used the naniar (28) package in RStudio to examine the proportion of missing data for all variables, to explore the patterns of missingness, and to determine if cases were missing completely at random.
We ran two hierarchical multiple linear regression models in RStudio, one with SCREEN-8 score at baseline as the outcome variable and one with SCREEN-8 score at followup as the outcome variable. We used a stepwise approach, entering all the variables into the model simultaneously.
Step 1 included the social network variables of interest, including social participation, social support, self-rated social standing, household income, frequency of contact with network members, and number of network members. Step 2 also included the demographic variables of interest, which were age, sex assigned at birth, marital status, highest educational attainment, and living situation (alone or not alone). We chose these variables as they have been associated with nutrition risk in previous studies (8,29,30).
Step 3 added the health-related variables, including depression, self-rated general health, self-rated mental health, self-rated healthy aging, disability, and oral health. We chose these as previous research has indicated they are associated with nutrition risk (8,(29)(30)(31).
We repeated our regression analyses after multiple imputation of missing data to assess the potential for bias due to missing values (Supplementary Tables 3 and 4). We performed multiple imputation using RStudio and the package mice (32). The mice package uses fully conditional specification (FCS) to conduct multiple imputation. This method is also called multivariate imputation by chained equations (MICE) (32). MICE is practical and flexible and uses a series of regression models for each variable with missing data (33). This method can handle a variety of different variable types, including continuous and categorical variables (32,33). We included all variables in our regression analyses in the imputation model, as recommended (33). We chose 20 imputations as recommended in the literature (33). After multiple imputation we re-ran our multiple regression analyses. We used the package mice (32) to obtain the pooled regression results.

Sample Description
There were 17,051 individuals in the tracking sample for whom we had data at baseline and follow-up. There were approximately equal percentages of males and females when looking at sex assigned at birth (48.2% males). The mean age of the participants was 59.88 (SD=10.29) and most participants were married or partnered (75.0%). Mean SCREEN-8 scores at baseline were 38.66 (SD=6.40) and were 37.87 (SD=6.61) at follow-up. At baseline, 36.5% of participants were at nutrition risk compared to 42.2% at follow-up. Of those who took part at baseline, 80.3% of participants also provided data at follow-up. The CLSA confirmed death of 1165 participants (5.5%). We provide further details on the social network, demographic, and health-related variables in Table 2.
When we examined the proportion of missing data in our dataset, we were missing SCREEN-8 scores for 13.1% of the sample at baseline and 19.7% at follow-up (Supplemental Table  1). Out of 21,240 participants for whom we had data, 9,437 (44.4%) participants had complete data on all the variables of interest. Little's Missing Completely at Random test (34) was not statistically significant (χ2 = 29703.91, p=0.394), therefore we report the complete case analysis here.

SCREEN-8 scores at baseline
The unadjusted model, using social network variables only, was statistically significant, R2=0.081, p<0.001 (Table 3). The effect size (Cohen's f2) was 0.088, which is considered small (27). Among the social network variables, we found that social participation, self-rated social standing, social support, and household income were statistically significant predictors of SCREEN-8 scores at baseline. As social participation, selfrated social standing, social support, and household income increased, SCREEN-8 scores also increased, therefore lowering nutrition risk.
Adding demographic variables to the model increased R2 to 0.100 and the effect size to 0.111, also a small effect size (27) ( Table 3). Social participation, self-rated social standing, social support, and household income continued to be statistically significant predictors of SCREEN-8 scores at baseline. Among the demographic variables, as age increased, SCREEN-8 scores increased. Being single or widowed, compared to married or partnered, was associated with lower SCREEN-8 scores. Living with others, as opposed to living alone, was associated with higher SCREEN-8 scores. Having a post-secondary degree or diploma (compared to a less than secondary education) was associated with higher SCREEN-8 scores. The full model predicting SCREEN-8 scores at baseline, using social network variables, demographic variables, and health indicators, was statistically significant, R2=0.172, p<0.001. The effect size (Cohen's f2) was 0.208, which is considered medium (27) (Table 3). In the full model, among the social network variables, we found that social participation, self-rated social standing, and social support were statistically significant predictors of SCREEN-8 scores. Household income was no longer a significant predictor. Higher social participation, higher self-rated social standing, and increased social support were associated with higher SCREEN-8 scores and therefore lower nutrition risk. Regarding the demographic variables, we found that age, marital status, having a postsecondary degree or diploma, and living situation were statistically significant predictors of SCREEN-8 scores. As age increased, SCREEN-8 scores also increased, showing lower nutrition risk. Being single or widowed, compared to being married or partnered, was associated with lower SCREEN-8 scores and therefore higher nutrition risk. Having an educational attainment of post-secondary degree or diploma (compared to less than secondary) was associated with higher SCREEN-8 scores and therefore lower nutrition risk. Among the health-related variables, screening negative for depression, self-rated general health, a self-rated healthy aging of very good or excellent (compared to fair or poor), and oral health were statistically significant predictors of SCREEN-8 scores. Screening negative for depression, having a self-rated general health and self-rated oral health of good, very good, or excellent (compared to fair or poor), and having self-rated healthy aging or very good or excellent (compared to fair or poor) were associated with higher SCREEN-8 scores and therefore lower nutrition risk.

SCREEN-8 scores at follow-up
The unadjusted model, using social network variables, was statistically significant, R2=0.075, p<0.001. The effect size (Cohen's f2) was 0.081, which is considered small (27) (Table  4). We found that social participation, social support, and household income were statistically significant predictors of SCREEN-8 scores at follow-up. As each of these increased, SCREEN-8 scores also increased.
Adding demographic variables to the model increased R2 to 0.087 and the effect size to 0.095, which is small (27) (Table 4). Social participation, self-rated social standing, social support, and household income continued to be statistically significant predictors of SCREEN-8 scores at follow-up. Among the demographic variables, as age increased, SCREEN-8 scores increased. Being single or widowed (compared to being married or partnered) was associated with lower SCREEN-8 scores.
Having an educational attainment of post-secondary degree or diploma (compared to less than secondary) was associated with higher SCREEN-8 scores.
The full model predicting SCREEN-8 scores at follow-up, using social network variables, demographic variables, and health indicators, was statistically significant R2 = 0.152 (p <0.001) ( Table 4). The effect size (Cohen's f2) was 0.179, a medium effect size (27). Among the social network variables, baseline social participation, self-rated social standing, and social support were statistically significant predictors of SCREEN-8 scores at follow-up. As social participation, selfrated social standing, and social support increased, so did SCREEN-8 scores, resulting in reduced nutrition risk. Having a household income of $50,000 or more was associated with higher SCREEN-8 scores and lower nutrition risk. Regarding the demographic variables, being single or widowed at baseline were statistically significant predictors of SCREEN-8 scores at follow-up. Being single or widowed, compared to being married or partnered, was associated with lower SCREEN-8 scores and therefore higher nutrition risk.
Among the health-related variables, depression, self-rated general health, self-rated healthy aging, and having a self-rated oral health of very good or excellent (compared to fair or poor) were statistically significant predictors of SCREEN-8 scores at follow-up. Screening negative for depression was associated with higher SCREEN-8 scores and, therefore, lower nutrition risk. As self-rated general health and healthy aging increased, so did SCREEN-8 scores, resulting in lower nutrition risk. Having a self-rated oral health of very good or excellent (compared to poor or fair) was associated with higher SCREEN-8 scores and therefore lower nutrition risk.

Sensitivity analysis
Compared to the complete cases, those with missing data differed on all the social network variables, except for the frequency of contact with neighbours, however the effect sizes were all small or trivial (35). Among the demographic variables, age, marital status, living situation, and education attainment differed between the two groups, but again, the effect sizes were small or trivial (35). All the health-related variables also differed between the two groups, however, the effect sizes were all trivial (35). SCREEN-8 scores at baseline and at follow-up also differed between the complete cases and those with missing data, however, the effect size was trivial (35) (Supplemental Table 2).
The regression model predicting SCREEN-8 scores at baseline after performing multiple imputation was similar to the complete case analysis, with the only difference being that self-rated mental health became a significant predictor of SCREEN-8 scores (Supplemental Table 3). For the regression model predicting SCREEN-8 scores at follow-up that included SCREEN-8 score at baseline, again the model after imputation was similar to the complete case analysis (Supplemental Table  4). However, the number of living children, living alone, age, self-rated mental health, and having a self-rated oral health of good (compared to fair or poor) became significant predictors of SCREEN-8 scores at follow-up.

Discussion
To our knowledge, our study is the first to examine nutrition risk using a reliable, validated tool, SCREEN-8, at two time points, in a large, nationally representative sample of community-dwelling adults that includes Canadians aged 45 to 64 as well as those 65 and older. Our models, based on Berkman and colleagues' social network theory (15), could explain 22% of the variance in SCREEN-8 scores at baseline and 15% of the variance in SCREEN-8 scores at follow-up. Using social network theory (15), we found that variables representing psychosocial mechanisms predicted SCREEN-8, and therefore nutrition risk scores, at both time points.
While we did not find that social network factors (at the mezzo level) were predictors of SCREEN-8 scores, our study found that psychosocial mechanisms (micro level) predicted SCREEN-8 scores at both baseline and follow-up. In Berkman and colleagues' social network theory, the social network factors at the mezzo level influence psychosocial mechanisms at the micro level, which in turn influence health outcomes (15), therefore this result is not surprising. The effects of the mezzo level factors are therefore likely captured by the micro level factors in our study, lending further support to Berkman and colleagues' (15) social network theory.
In our full models, at both baseline and follow-up, higher levels of social participation, higher self-rated social standing, and higher levels of social support were associated with higher SCREEN-8 scores and lower nutrition risk. Previous Canadian studies have found that low social support (8,36), and infrequent social participation (8) were associated with nutrition risk in adults aged 65 and older. Other studies have also found that having low levels of social support is associated with increased nutrition risk (37,38). Social support may affect nutrition risk in several ways. An individual's social support system may encourage healthy behaviours, such as consuming adequate amounts of nutrient-rich foods (38). If an individual requires assistance with food-related activities, such as grocery shopping or meal preparation, adequate social support can ensure they meet these needs (39). Higher levels of social support may also provide an individual with more opportunities to eat with others. Eating with others improves food intake (40), whereas eating alone is a well-known risk factor for poor nutrition and food intake (16,41,42). Even after adjusting for demographic and health-related variables that may influence nutrition risk, social participation, self-rated social standing, and social support remained significant predictors of SCREEN-8 scores and therefore nutrition risk, lending further support to the idea that social relationships play a key role in food and eating behaviours. In our full models, among our demographic variables, being single or widowed, compared to being married or partnered, was associated with lower SCREEN-8 scores and increased nutrition risk at baseline and at follow-up. Other studies have also found an association between marital status and nutrition risk (43). Additionally, previous Canadian work has found that widowhood leads to fewer regular meals and less interest in meal preparation, as eating alone is less appealing than eating with a partner (44). Eating with others can also reinforce social norms around food and eating, for example, by providing cues for mealtimes and appropriate food intake (16).
We also found that, in our full model at baseline, not living alone was associated with increased SCREEN-8 scores and therefore decreased nutrition risk, similar to a previous Canadian study that found increased nutrition risk in adults aged 65 and older who lived alone (8). Those who live alone may eat alone more often compared to those who live with others. As previously mentioned, eating alone is associated with reduced food intake and increased nutrition risk. Other studies have also found an association between increased nutrition risk and living alone (43,45). However, we did not find that living alone at baseline was a significant predictor of SCREEN-8 scores at follow-up. It is possible that participants' living situation may have changed between baseline and follow-up, and they could now live with others.
Having an educational attainment of a post-secondary degree or diploma, compared to less than secondary, was associated with higher SCREEN-8 scores at baseline in our full model. Previous Canadian research has found that having a less than secondary school graduation, compared to a secondary school graduation, was associated with increased nutrition risk (8). Higher education may influence nutrition risk in several ways. Those with higher levels of education may have higher incomes and greater access to other resources (8,37). Higher education may also lead to better food and nutrition choices and improved quality and quantity of food consumed (8,37).
In our full models, we found that many health-indicator variables were associated with SCREEN-8 scores. We found that higher self-rated health status was associated with higher SCREEN-8 scores at both time points, similar to previous studies (26,31,46). It is well known that nutrition influences health and chronic disease (1,2). Additionally, those in better health may have an easier time with food-related activities such as meal preparation and grocery shopping or may find it easier to eat a healthy diet. We also found that higher self-rated healthy aging was associated with higher SCREEN-8 scores at both baseline and follow-up. This is similar to a previous Canadian study that found an association between increased nutrition risk and lower self-rated successful aging (46). As with self-rated health, those with higher self-rated healthy aging may find it easier to complete food-related tasks. Alternatively, the consumption of a healthy diet may lead an individual to rate their healthy aging as higher. We also found that higher self-rated oral health was associated with higher SCREEN-8 scores at baseline, and that having a self-rated oral health of very good or excellent (compared to fair or poor) was also associated with higher SCREEN-8 scores at follow-up. Other studies have found an association between oral health and nutrition risk (8,47,48). Poor oral health can lead to difficulties with chewing and swallowing (49), which in turn affects food and nutrient intake. It is possible that individuals' oral health changed (improved or worsened) between baseline and follow-up, which is why having a good oral health (compared to fair or poor) is no longer a predictor at follow-up.
Not screening positive for depression was associated with higher SCREEN-8 scores and lower nutrition risk at both time points. This is not surprising, as a previous study using data from the CLSA found that depression was associated with high nutrition risk (50), as have other studies (36,43,45). Depression can lead to a loss of appetite and reduced food intake (4), leading to nutrition risk.
Most of the previous research into nutrition risk has only included adults aged 65 and older, although some studies do include adults aged 50 and older (7,57). As our findings have similarities with previous research conducted with older age groups, the factors discussed above appear to be associated with nutrition risk through midlife and into older adulthood.
In contrast to other studies (8, 51), we found that increasing age was associated with higher SCREEN-8 scores at baseline and therefore reduced nutrition risk, however, the coefficient was small (B=0.051). Our result could be due to several factors. First, participants in the CLSA are community-dwelling. Adults in older age groups with increased nutrition risk may no longer be living in the community, as nutrition risk is associated with hospitalization, mortality, and institutionalization (9,10,52). We have previously seen this same survivor effect (53) in a study of residents of naturally occurring retirement communities (54).
Second, SCREEN-8 scores at baseline were highest for those aged 65-74 (39.1), followed by those aged 55-64 (38.7), comprising in total 50.4% of the sample. While those aged 75 and older had the lowest SCREEN-8 scores (38.4) they were only 11.5% of the sample. Those aged 45-54 had SCREEN-8 scores similar to those 75 and older (38.5) and comprised 38.1% of the sample. We should note that the SCREEN-8 tool has only been validated for use in adults aged 50 and older (58). It is possible that SCREEN-8 does not measure nutrition risk accurately in those younger than 50, which could also explain why we found that SCREEN-8 score increased with age.
Third, we purposefully included this younger age group so that in the future we can examine how their nutrition risk scores change, particularly as they enter older adulthood. Participants in this younger age group may still have children living at home or they may have children moving out of the home for the first time, affecting their food choices (55,56). Similarly, these may be working adults, which may affect their food choices during the workday (55,56). Future research will explore how the determinants of nutrition risk vary for different age groups.
One strength of our study is the use of CLSA data, which is a large, representative sample of the Canadian population aged 45 and older (17). As there are currently two waves of CLSA data available, we could also examine what factors at baseline predicted nutrition risk scores at both baseline and at follow-up three years later. As the CLSA will follow participants every three years for 20 years or until participant death, in the future, we will be able to examine the factors that predict changes in nutrition risk over time. We will also be able to examine trajectories of nutrition risk.
While we were able to examine the baseline factors that predicted SCREEN-8 scores at two time points, we cannot imply causation. There could be other factors that affect both our predictors and SCREEN-8 scores (and therefore nutrition risk). Additionally, our models only explained between 15 and 22 percent of the variance in SCREEN-8 scores, indicating that there are additional factors that predict SCREEN-8 scores and nutrition risk.
While the CLSA is a large, nationally-representative sample of Canadians living in the ten Canadian provinces, it does exclude several groups, including full-time members of the Canadian Armed Forces, those living in the territories, in some remote areas, and those living on First Nations reserves and settlements (19). It also only includes those who can speak English or French, and those with the capability to answer the questions themselves (19). Therefore, groups who may be more likely to have lower SCREEN-8 scores and be at higher nutrition risk may not be well represented in the CLSA.
While we were able to map several CLSA measures onto the factors in Berkman and colleagues' social network theory (15), there are other social network factors that were not represented in the CLSA data. For example, the CLSA only measured the frequency of face-to-face contact with friends, siblings, relatives, friends, neighbours, and children, and did not include other forms of contact. While the CLSA did ask about social media use, there were large numbers of missing data for those questions, so we did not include those variables in our analyses. At the mezzo level, other characteristics of network ties, such as the reciprocity of ties, their duration, and their intimacy (15), were not captured in the CLSA data. Similar, other social network structures, such as density, boundedness, proximity, homogeneity, and reachability (15), were not available in the CLSA data. At the micro level, social influence factors, such as constraining and enabling influences on health behaviours, norms towards health seeking and adherence, and peer pressure (15), were similarly not captured in the CLSA.

Conclusions
Using Berkman and colleagues' (15) social network theory, we found that psychosocial mechanisms, as well as some demographic and health-indicator measures, were predictors of SCREEN-8 scores in community-dwelling Canadians aged 45 and older. Health care professionals should know that middle-aged and older adults with low social participation and low social support may be at increased nutrition risk, and therefore should be screened proactively. Additionally, individuals who are single or widowed, those who screen positive for depression, as well as those with low self-rated health, low self-rated mental health, poor healthy aging, and poor oral health may be at increased nutrition risk and should be proactively screened and managed with a variety of community-based programs. Nutrition risk occurs before there are overt signs of malnutrition (6). For that reason, and because it is easier to treat nutrition risk compared to malnutrition, it is important to identify those who are at increased nutrition risk (6). Interventions targeting nutrition risk should explore ways to increase both social support and social participation.