An Investigation of Activity Profiles of Older Adults

Abstract

Objectives.

In this study, we advance knowledge about activity engagement by considering many activities simultaneously to identify profiles of activity among older adults. Further, we use cross-sectional data to explore factors associated with activity profiles and prospective data to explore activity profiles and well-being outcomes.

Method.

We used the core survey data from the years 2008 and 2010, as well as the 2009 Health and Retirement Study Consumption and Activities Mail Survey (HRS CAMS). The HRS CAMS includes information on types and amounts of activities. We used factor analysis and latent class analysis to identify activity profiles and regression analyses to assess antecedents and outcomes associated with activity profiles.

Results.

We identified 5 activity profiles: Low Activity, Moderate Activity, High Activity, Working, and Physically Active. These profiles varied in amount and type of activities. Demographic and health factors were related to profiles. Activity profiles were subsequently associated with self-rated health and depression symptoms.

Discussion.

The use of a 5-level categorical activity profile variable may allow more complex analyses of activity that capture the “whole person.” There is clearly a vulnerable group of low-activity individuals as well as a High Activity group that may represent the “active ageing” vision.

Over many years of gerontological scholarship, theoretical and empirical work has identified activity engagement as a major determinant of well-being in later life. The current literature has been generally limited by focusing on one or two activities at a time to explain outcomes. In this study, we advance knowledge about activity engagement by considering many activities simultaneously to identify profiles of activity among older adults. Further, we use cross-sectional data to explore factors associated with activity profiles and prospective data to explore activity profiles and well-being outcomes.

Background and Related Literature

Several established bodies of literature inform this work. First, there is a diverse literature on the effects of activity on well-being outcomes. A wide range of types of activities, including working, caregiving, volunteering, grandparenting, leisure, and physical activity, have been studied in regard to well-being outcomes, including health, cognitive function, functional status, and mortality (Buchman, Wilson, & Bennett, 2008; Glass, Mendes de Leon, Bassuk, & Berkman, 2006; Hsu, 2007; Janke, Payne, & Van Puymbroeck, 2008; Karp et al., 2006; Lennartsson & Silverstein, 2001). This research has generally supported the idea that activity participation leads to positive outcomes for the individual older adult (Chipperfield, 2008; Lampinen, Heikkinen, Kauppinen, & Heikkinen, 2006; McDonnall, 2011). Yet, certain activities, like caregiving, custodial grandparenting, or employment under certain conditions, have been associated with negative outcomes (Adams, McClendon, & Smyth, 2008; Son et al., 2007).

Second, there are numerous studies seeking to increase understanding about factors associated with various types of activity engagement. The literature is too large and diverse to review here, but several examples are presented to illustrate the wide range of associative factors that have been examined. At the level of the individual older adult, demographics like ethnicity, gender, and age, as well as measures of human and social capital, are commonly employed. When studying work and volunteering, these personal characteristics are important predictors of activity engagement. Measures of socioeconomic status (SES), family structure, and living arrangements have been related to the activities of caregiving and grandparenting (National Alliance for Caregiving & AARP, 2009). In regard to social activities, there has been attention to the predictive ability of health and functioning as well as mental health and personality (Jang, Mortimer, Haley, & Borenstein Graves, 2004; Krueger et al., 2009). At the community level, measures of built environment and neighborhood characteristics have been examined in regard to physical and social activities (Glass & Balfour, 2003; Wen, Hawkley, & Cacioppo, 2006). Driving cessation and lack of transportation are associated with reduced participation in a range of activities (Marottoli et al., 2000). Although far from exhaustive, these examples demonstrate the wide range of factors, from psychological to environmental, that have been studied in relation to activity engagement.

This literature, however, largely focuses on one activity at a time, making it difficult to understand how the range of activities in which an older adult engages relates to health and wellness outcomes. A trend toward investigating multiple activities is emerging and represents a step toward this goal. Some studies have considered two activities simultaneously, like volunteering and working (Hao, 2008; Luoh & Herzog, 2002), caregiving and working (Fredriksen-Goldsen & Scharlach, 2006), and informal helping and volunteering (Choi, Burr, Mutchler, & Caro, 2007; Jegermalm & Grassman, 2009). A few researchers have simultaneously analyzed three or more activities, often focusing on the productive activities of working, volunteering, informal helping, caregiving, and grandparenting (Baker & Silverstein, 2008; Hinterlong, 2008; Sugihara, Sugisawa, Shibata, & Harada, 2008). In general, findings from these studies documented that most older adults were involved in more than one productive activity and that the well-being outcomes varied in response to combination of activities; further, these effects were modified by demographic characteristics. Arai and coworkers (2007) used canonical correlations to associate depressive symptoms with 18 lifestyle activities, from gardening to housekeeping to religious activities. They documented that less interaction with neighbors, society, and friends correlated with depressive mood for men, whereas less interaction with family members related to depressive mood in women.

A more limited number of studies have gone beyond studying several activities simultaneously to consolidating multiple activities into broader categories, and these studies are most relevant to the work presented here. Burr, Mutchler, and Caro (2007) analyzed five productive activities (formal volunteer work, informal help to others, unpaid domestic work, caregiving, and paid work) and identified four classes of older adults: helpers, home maintainers, worker/volunteers, and super helpers. Those older adults who were younger, white, and had more income were more likely to be worker/volunteers or super helpers. Bielak, Hughes, Small, and Dixon (2007) created seven subscales from 67 separate survey items: self-maintenance, travel, social, passive, integrative, physical, and novel activities. Although activity engagement across all domains was related to better cognitive performance, involvement in novel activities (completing income tax forms, playing bridge) had the strongest relationship. Paillard-Borg, Hui-Xin, Winblad, and Fratiglioni (2009) reduced 31 leisure activities into five domains: mental, social, physical, productive, and recreation. Factors related to decreased engagement, including low education, limited social networks, and mental disorder, were identified. Croezen, Haveman-Nies, Alvarado, Van’t Veer, and De Groot (2009) studied 17 social activities of older people and identified five subgroups. About half of the sample was classified as less socially engaged, and these older adults were older, living alone, less well educated, and in poorer physical and mental health. Jopp and Herzog (2010) started with 70 leisure activities and identified 11 activity categories: activities with close social partners, group-centered public activity, religious activities, physical activities, developmental activities, experiential activities, crafts, game playing, TV watching, travel, and technology use. They conducted analyses to demonstrate that the proposed factor structure fit the data well and advocated this measurement approach in subsequent work on leisure activity engagement. Janke, Davey, and Kleiber (2006) created three domains of leisure (informal leisure, formal leisure, and physical leisure) by adding related items and used growth curve analysis to assess change in leisure over an 8-year period. In general, leisure activity decreased over time in all three domains, although informal leisure stayed more stable for men, and health had the strongest relationship to change profiles. Lennartsson and Silverstein (2001) classified 10 leisure activities along two continuums, solitary-social and sedentary-active, and they found that men who engaged in more solitary activities were likely to live longer than those engaged in more social activities.

As seen in this overview, there is precedent for consolidating information about engagement in discrete activities into summary measures. Most often, these studies focus on a restricted range of activities (productive, leisure, social, etc.). Given the central role of activity engagement in our thinking about late-life health and well-being, we should advance methodologies that consider a fuller range of activities as well as apply current knowledge about antecedents and outcomes of single activities or smaller sets of activities to a more comprehensive picture of activity patterns.

Conceptual Framework

The conceptual framework used in this study is presented in Figure 1. This framework builds on the World Health Organization’s Active Ageing Framework (World Health Organization, 2007), which highlights the role of activity participation in quality of life and uses the word “active” to refer to participation on social, economic, cultural, spiritual, and civic affairs. Further, the framework posits that “active ageing” is determined by a multitude of factors (personal, behavioral, health and social services, economic, social, and physical environment), some of which are mutable through policies and programs. For this study, we utilize this basic conceptualization: that activity participation is determined by a wide range of factors and that activity participation is associated with well-being outcomes. The literature reviewed earlier and a long history of social science theory, from activity theory to the successful and productive aging paradigms, supports this conceptualization. We use a broad definition of activity as participation in a wide range of behaviors—a description of what people do (Putnam et al., 2013). Figure 1 is more specific than the WHO model in which “active ageing” is posited as the outcome of interest. We view activity patterns as an intermediate outcome, leading ultimately to quality of life or well-being outcomes. We acknowledge unidirectional linearity is a limitation in this conceptual framework, but we suggest it has utility in advancing knowledge about activity engagement.

Figure 1.

Conceptual framework for the study of antecedents and outcomes of activity profiles.

Building on the literature reviewed earlier and guided by this conceptual framework, we posed two research questions: (a) what activity profiles occur among older adults? and (b) what antecedents and well-being outcomes are associated with these profiles? The primary objective of the work was to identify activity profiles from numerous activity items, allowing the simultaneous consideration of many activities that reflect the reality of daily life for older individuals. Given that there is not much precedent in the literature on activity profiles, we took an exploratory approach to analyzing antecedents and outcomes to, in a sense, validate these profiles. Based on the previous work on single activities or smaller sets of activity items or domains, we expected to see factors at the various levels (personal, social, physical environment, etc.) related to the patterns. Further, based on theory and past findings on activity, we expected to see patterns related to subsequent well-being outcomes, and in general, with higher activity engagement associated with better outcomes. However, pending more understanding of activity patterns, we did not pose hypotheses. To our knowledge, the work we present here is unique in that it considers 36 activity items in the creation of activity profiles, and it assesses antecedents and outcomes of these profiles. We believe that this work advances the study of activity, both methodologically and through its substantive findings, permitting greater understanding of how engagement patterns across a broad range of activities relate to healthy aging.

Method

Data

This study used data from the Health and Retirement Survey (HRS), perhaps the leading source of data for studies of older adults in the United States (National Institute on Aging, 2011). The original HRS cohort is a nationally representative sample of individuals born from 1931 to 1941, with oversampling for African Americans, Latinos, and residents of the state of Florida (Heeringa & Connor, 1995). Surviving respondents have been surveyed every 2 years since 1992. The HRS has since expanded to include additional cohorts of older adults, such that it now provides statistically representative samples of all U.S. households that include adults aged 51 and older (Hauser & Willis, 2005). In each wave, approximately 20,000 individuals were interviewed. Data were collected by both person-to-person and mail surveys, and response rates from wave to wave range from 85% to 89% (HRS, 2011).

In this study, we used the 2008 and 2010 core survey data from the RAND HRS data files (version L), as well as the 2009 Health and Retirement Study Consumption and Activities Mail Survey (HRS CAMS). The HRS CAMS includes questionnaires assessing individual activities, measured by hours per week or hours per month. For the 2009 HRS CAMS, 7,231 questionnaires were mailed to the random subsample of the HRS, and 5,530 questionnaires were returned with a response rate of 74%. Six questionnaires had missing observations across all activities. Therefore, the number of observations used for the next step of sample identification in this study was 5,324.

Because the HRS CAMS only includes the subsample of the HRS main data sets, this subsample was retained in the merged data set. We identified 4,832 respondents aged 55 or older included in both the HRS 2008 and 2009 HRS CAMS. Among those 4,832 respondents, 158 did not respond to the 2010 HRS and 147 died between the waves. Additionally, missing data across study variables ranged from 0% to 7%. Although the proportion of sample attrition and missing data was relatively low, these were imputed using Markov chain Monte Carlo multiple imputation. This method has been widely reported to be effective in handling missing data due to sample attrition and non-response and to produce unbiased estimates even with a large fraction of missing data (Goldstein, 2009; Schafer, 1997; Shaffer & Chinchilli, 2007). Five imputed data sets were created and all the estimates were combined across these data sets. We also conducted the analysis both with and without imputation, and findings were not substantially different.

The final sample size was 4,593 after excluding 92 participants who had zero or missing observations in the HRS CAMS weight variable, which is a product of the prior core HRS weight and a non-response adjustment factor (Hurd, Rohwedder, & Carroll, 2011). To reduce potential bias from initial non-response in the 2009 HRS CAMS, we used the respondent weight variable for statistical analysis.

Measures

Activity measures.

We started this analysis with a thorough review of the questions in the 2009 HRS CAMS to identify activity measures. We determined whether the intent of a survey item was to inquire about “doing” something, as opposed to inquiring about feeling, thinking, believing, having, getting help with, and so forth. If it seemed a survey item was about “doing,” we tested our assessment by rephrasing the survey question to see if it was possible to reword it as “do/did you do X” or “how much/often do you do X?”. This approach yielded 36 activity measures from the 2009 HRS CAMS (blinded for review).

Original activity items were measured continuously (e.g., hours per week or hours per month). Two activity items, reading newspapers/magazines and reading books, were added together to create one item: reading newspapers/magazines/books. We excluded three items from the analyses—sleep/nap, personal grooming, and eating meals—because about 99% of the sample was engaged in these activities, indicating that the measures would produce no meaningful variation in our analysis.

To reduce the number of activity items for use in subsequent analyses, we conducted exploratory factor analysis (EFA) and confirmatory factor analysis (CFA) to identify activity domains underlying the discrete activities. Details of these analyses are available from the authors (blinded for review). In general, we took an empirical approach to preserve the use of as many activity items as possible to determine if this consolidation into profiles can be supported. Final results indicated nine factors among the 36 activity items. Activity items in each factor were summed. Given the wide variation in level of engagement among the various activities, we coded the scores into a three-level ordinal measures representing low, medium, and high levels of engagement in each domain. We studied the distribution of the amount of each activity and divided the distribution into three groups. When the proportion of those who reported no activity was small (5% or less), the sample was split evenly into one-thirds (low, medium, and high), with those not participating in the activity included in the low group. When there was a larger group reporting no activity, that group was retained and the remaining sample was split evenly, yielding three groups (none, medium, and high). These ordinal measures of none/low, medium, and high activity levels in nine activity domains were utilized in latent class analysis (LCA).

Table 1 presents a description of each activity item, grouped by activity domains, as well as the unstandardized factor loadings and model fit statistics from the CFA. We also report the percent of people engaging in each activity.

Table 1.

Activity Items, Participation Rates, and Domains From Confirmatory Factor Analysis

Activity domain (factor)	Activity item	Factor loading (unstandardized)	Participation rates N (%)
Personal leisure	Watch TV	1.00	5,132 (97%)
	Read papers/magazines/books	5.28***	4,780 (91%)
	Play cards/games/puzzles	4.33***	2,509 (48%)
Civic/religious activity	Pray/meditate	1.00	4,134 (79%)
	Volunteering	2.40***	1,622 (31%)
	Religious attendance	1.61***	2,864 (55%)
	Attend meetings	2.09***	1,628 (31%)
	Sing/play instruments	1.77***	1,069 (20%)
Physical exercise	Listen music	1.00	4,076 (78%)
	Walk	1.15***	4,501 (86%)
	Sport/exercise	1.03***	2,096 (40%)
Interior household chores	House cleaning	1.00	4,211 (80%)
	Wash/iron/mend	0.84***	3,648 (69%)
	Shop/run errands	1.44***	4,597 (87%)
	Meal preparation/clean-up	1.16***	4,454 (84%)
	Money management	1.12***	4,235 (80%)
	Arts and crafts	0.80***	1,084 (21%)
Exterior household chores	Yard work/garden	1.00	2,757 (52%)
	Pet care	0.55***	2,135 (40%)
	Home improvement	1.49***	2,148 (41%)
	Vehicle maintenance	1.14***	2,534 (48%)
Managing medical conditions	Managing medical conditions	1.00	3,252 (62%)
	Seeing a physician/nurse/and so on	0.98***	1,197 (22%)
	Managing medical bills	1.25***	900 (17%)
Employment/computer use	Work for pay	1.00	1,625 (31%)
	Use computer	3.64***	2,757 (52%)
Interpersonal exchange/helping others	Visit in person	1.00	4,510 (85%)
	Phone/letter/E-mail	0.94***	4,874 (92%)
	Show affection	0.89***	4,210 (81%)
	Help others	1.31***	3,062 (58%)
	Treating others’ medical condition	0.67***	1,001 (20%)
Community leisure	Concert/movies/lectures	1.00	1.325 (25%)
	Leisure dining/eat out	0.99***	4,083 (78%)
Model fit statistics from CFA	χ2(df = 199) = 14259.84***; CFI = 0.90; TLI = 0.92; RMSEA = 0.04

Note. Negative variance in “use computer” item is set to zero, and error terms are allowed to be correlated in confirmatory factor analysis (CFA). CFI = comparative fit index; RMSEA = root mean square error of approximation. TLI = Tucker−Lewis index.

***p < .001.

Antecedents and well-being outcomes.

Antecedents are grouped into factors included in the conceptual framework. Personal factors included functioning in mobility/activities of daily living, age, gender, race, and education. Economic factors included income and assets. Social factors included marital status, number of people in the household, number of children, number of close friends, and social support. Physical environment factors include urban–rural residence, neighborhood cohesion, and neighborhood disorder.

We used two measures to capture well-being: self-rated health and depressive symptoms. These well-being indicators derived from two waves, 2008 and 2010. The 2008 assessments were used as associative factors to 2009 activity profiles, and 2010 assessments were used as outcomes of 2009 activity profiles. More detailed descriptions on measuring antecedents and outcomes are presented in Table 2 and descriptive statistics are presented in Table 3.

Table 2.

Description of Measures of Antecedents and Well-Being Outcomes

	Variables	Measure
Antecedents
Personal factors	Age at interview	Continuous years
	Gender	1 = female; 0 = male
	Race	1 = white; 2 = African American; 3 = Hispanic; 4 = other; A series of dummy variables was created for each category of race and included in the model
	Mobility/ADL	Summed scores based on five dichotomous items (1 = yes; 0 = no) measuring functional limitation (difficulty in walking, getting in and out of bed, dressing, bathing, and eating)
	Education (years)	Continuous years
Social factors	Marital status	1 = married; 2 = currently not married; 3 = never married; A series of dummy variables was created for each category of race and included in the model
	Number of people in household	Continuous numbers
	Number of children	Continuous numbers
	Number of close friends	Continuous numbers
	Positive social support	Summed score based on three items; Each item was measured by a 4-level ordinal scale (from 1 = a lot to 4 = not at all). Items were reverse coded so that higher scores indicate higher positive social support; positive social support from spouse/partner, child, family, and friend were summed to create overall positive social support scores; a single factor solution was preferred in EFA (validity), and Cronbach’s alphas were .82 for positive support from spouse/partner, .83 from child, .86 from family, and .83 from friends a. How much do they really understand the way you feel about things? b. How much can you rely on them if you have a serious problem? c. How much can you open up to them if you need to talk about your worries?
	Negative social support	Summed score based on three items; each item was measured by a 4-level ordinal scale (from 1 = a lot to 4 = not at all); items were reverse coded so that higher scores indicate more negative social support; negative social support from spouse/partner, child, family, and friend were summed to create overall negative social support scores; a single factor solution was preferred in EFA (validity), and Cronbach’s alphas were .78 for negative support from spouse/partner, .78 from child, .77 from family, and .75 from friends d. How often do they make too many demands on you? e. How much do they criticize you? f. How much do they let you down when you are counting on them? g. How much do they get on your nerves?
Economic factors	Total household income	Continuous U.S. dollars; Given the high skewness, original measures were transformed using natural logarithm
	Total household wealth (excluding housing)	Continuous U.S. dollars; Given the high skewness, many zero and negative values, original measures were transformed using the inverse hyperbolic sine function
Physical environment factors	Urban/rural	Ordinal measures ranging from 1 through 9. Higher values indicate more urban
	Neighborhood safety	A single question about neighborhood safety measured by a 5-level ordinal scale ranging from 1 = excellent to 5 = poor; Original responses were reverse coded so that higher scores indicated better safety a. Would you say the safety of your neighborhood is excellent, very good, good, fair, or poor?
	Neighborhood cohesion	Summed scores from eight items; Each item was measured by a 7-level ordinal scale; Original responses were reverse coded as necessary so that higher scores indicate higher cohesion; A single factor solution was preferred (validity); Cronbach’s alpha of .91 (reliability) Example of items: I really feel part of this area/I feel that I don’t belong in this area; Most people in this area are friendly/Most people in this area are unfriendly; There are many vacant or deserted houses or storefronts in this area/There are no vacant or deserted houses or storefronts in this area
Well-being outcomes	Self-reported health	A 5-level ordinal measure ranging from 1 = excellent to 5 = poor; Reverse coded so that higher scores indicated better health
	CES-D	Summed scores based on eight dichotomous items (1 = yes; 0 = no) measuring depressive symptoms

Note. ADL = activities of daily living; CES-D = Center for Epidemiologic Studies Depression; EFA = exploratory factor analysis.

Table 3.

Descriptive Statistics of Antecedents and Well-Being Outcomes

	M (SD)/N (%)
Antecedents
Personal factors
Age	69.45 (8.91)
Gender
Male	1,942 (41%)
Female	2,743 (59%)
Race
White	3,694 (79%)
African American	553 (12%)
Hispanic	344 (7%)
Other	93 (2%)
Mobility/ADL 2008	0.26 (0.80)
Education	12.78 (2.98)
Social factors
Marital status
Married	2,960 (63%)
Not married	1,563 (33%)
Never married	162 (4%)
Number of people in household	2.08 (1.02)
Number of children	3.19 (2.03)
Number of close friends	4.57 (5.80)
Positive social support	32.48 (8.91)
Negative social support	22.04 (7.83)
Economic factors
Household income	76,641 (880,966)
Household wealth	351,041 (1,108,344)
Physical environment
Urban/rural	8.09 (0.93)
Neighborhood safety	4.01 (0.99)
Neighborhood cohesion	44.28 (10.82)
	M (SD)
Well-being outcomes
Self-rated health 2008	3.19 (1.06)
Self-rated health 2010	3.17 (1.07)
CES-D 2008	1.34 (1.91)
CES-D 2010	1.32 (1.88)

Note. ADL = activities of daily living; CES-D = Center for Epidemiologic Studies Depression.

Statistical Procedures

We used LCA to identify activity profiles of the older adults in the study sample. LCA is a widely used statistical method for identifying subtypes of cases (or latent classes) from multivariate categorical data and has several advantages over other clustering methods (Eshghi, Haughton, Legrand, Skaletsky, & Woolford, 2011; Hagenaars & McCutcheon, 2002; Magidson & Vermunt, 2002). Unlike other traditional clustering methods (e.g., k-means clustering, hierarchical cluster analysis), LCA uses a model-based approach; thus, maximum likelihood estimates can be used to classify individuals based upon their posterior probability of class membership. LCA further provides several model fit statistics that can be used to determine the optimal number of classes. LCA is also less restrictive than other clustering methods in that it does not depend on statistical assumptions such as linearity, normally distributed data, or homogeneity of variance.

In this study, we utilized the 9 activity domains derived from the 36 activity measures for LCA, and LCA empirically determined discrete latent classes from these observed activity measures and created subgroups of older adults that share similar engagement profiles. After extracting the smallest number of classes that fit the data, LCA provided probabilities of being in each class for each observation. To characterize each class, LCA produced probabilities of the class engaging in each activity (Muthén & Muthén, 2000). Using this information, we classified the entire sample into classes that shared similar engagement profiles and, for each profile, produced probabilities of that group engaging in high, medium, or low levels of activity in each domain.

We used three different model fit statistics to determine the number of classes. A significant result on the Lo–Mendell–Rubin (LMR) test indicates that there is a significant improvement in model fit between k-class and (k − 1)-class models (Lo, Mendell, & Rubin, 2001). Second, lower values of the Bayesian Information Criterion (BIC) indicate better model fit, and the BIC is one of the best indicators for class enumeration (Hagenaars & McCutcheon, 2002; Nylund, Asparouhov, & Muthén, 2007). Lastly, we reported entropy, which is a measure of uncertainty in classification. Entropy ranges from 0 to 1, with a higher value indicating high certainty.

To assess factors associated with activity profiles and outcomes associated with activity profiles, two waves of HRS data were merged. Antecedents were derived from the 2008 data set, activity measures from the 2009, and well-being outcomes from 2010. This permitted the time ordering required to support a causal argument.

We ran generalized linear modeling to examine the relationships between antecedents and activity profiles as well as the relationships between activity profiles and well-being outcomes. Separate analyses were conducted for each relationship. For the relationships between antecedents and activity profiles, we used multinomial logistic regression because the dependent variable was a five-level nominal measure. To examine the relationships between activity profiles and well-being outcomes, dummy variables representing each latent class were created and included as main independent variables. Multivariate linear regression was conducted in an structural equation modeling framework where correlations among well-being outcomes were allowed.

It would be best to analyze relationships among antecedents, activity engagement profiles, and well-being indicators using a General Mixture Model so that all estimates could be made simultaneously. However, this one-step approach has certain disadvantages relevant to the current study. When a series of antecedents was included while estimating latent activity classes (activity profiles), activity classes varied according to other variables in the model. Because our purpose is to identify activity classes (profiles), and then to explore the relationships among antecedents, activities, and outcomes longitudinally, the one-step approach was less applicable (Vermunt, 2010). Further, model convergence was not obtained in the presence of numerous antecedents and well-being indicators in the model because the one-step approach significantly increases computational complexity and time (Clark & Muthén, 2009). Therefore, in this study, latent classes were estimated first based on activity domains, and after creating class membership variables, separate analyses were conducted using multinomial logistic regression and multivariate linear regression. Although this two-step approach has been used widely in the literature, it may produce incorrect estimates and standard errors because the variability in estimating latent classes cannot be considered in the subsequent regression analysis (Clark & Muthén, 2009; Croon & Bolck, 1997). This limitation should be considered when interpreting estimates and their statistical significance (Clark & Muthén, 2009).

Individuals in the merged data set were clustered within the households, and thus there is some chance that within-cluster correlations could produce incorrect standard errors. Taking this into consideration, we utilized cluster-robust standard errors in our analysis described subsequently. Multiple imputation was conducted using SAS version 9, and all other statistical analyses, including EFA, CFA, LCA, and others, were conducted using Mplus version 5.

Findings

Activity Profiles

As seen in Table 4, the LCA produced various model fit statistics. As suggested in the literature, we relied on the BIC, where a change of 10 points can be considered meaningful (Muthén, 2001). The BIC decreased until the 5-class model and then increased for the 6-class model, indicating that the 5-class model fit the data best. The entropy values were not very useful in determining the best fit in this sample as they did not vary much (and it has been noted that entropy measures often do not work as well as other fit measures). The LMR test comparing the 4-class and 5-class model was non-significant, suggesting that the 5-class model was not better than the 4-class model. We selected the 5-class model for subsequent analysis because it was supported by the BIC and because it allows for maximum variation in activity profiles.

Table 4.

Latent Class Analysis Model Fit Statistics

	Chi square	BIC	Entropy	LMR test
Class 2	31,495*** (df = 19,302)	98,642	0.60	2757.70***
Class 3	30,319*** (df = 19,298)	98,007	0.60	732.86***
Class 4	28,988*** (df = 19,277)	97,832	0.55	276.15*
Class 5	28,978*** (df = 19,262)	97,728	0.54	205.12
Class 6	28,875*** (df = 19,243)	98,028	0.53	161.39

Note. BIC = Bayesian Information Criterion; LMR = Lo–Mendell–Rubin.

*p < .05; ***p < .001.

Next, each of the five classes was labeled and interpreted based on the visual interpretation of Figure 2 and the conditional probabilities in Table 5 that characterize the level of engagement of the group in each activity domain. The study sample was distributed reasonably across the five classes, from the smallest class of 13% to the largest of 28%. More than a quarter of the sample (28%) was assigned to Class 3, where activity level across all types of activity was largely in the medium range, compared with other groups. As seen in Table 5, the conditional probabilities of personal leisure, civic/religious activity, and managing medical conditions were the highest within this group (e.g., 0.45, 0.41, and 0.40 for the “High” engagement category, respectively), whereas conditional probabilities for employment/computer use were lowest (0.12 for “High” engagement). The next largest class, Class 2, included 22% of the sample and was characterized by high levels of physical exercise, interior and exterior household chores, and interpersonal exchange/helping others, with moderate levels of all other activities. Class 1, comprised of 20% of the sample, was characterized by high activity in all domains except employment/computer use. Seventy percent of respondents in this group were medium or low activity on employment/computer use. We know from other analyses that 63% of this group reported being retired. Class 5, comprised of 17% of the sample, mirrored this pattern in the other direction: low on most activities compared with other classes, but high in employment/computer use. Eighty-one percent of this class reported being employed. The last class, Class 4, was the smallest group at 13% and was less likely to be engaged in most activities. In this class, conditional probabilities of engaging in personal leisure and managing medical conditions were the highest for this group.

Figure 2.

Five activity profiles. Note. Probabilities of engaging in “high” activities for each domain are plotted.

Table 5.

Conditional Probabilities of Engaging in Activity Domains (the 5-class Model)

	High Activity (n = 919)	Physically Active (n = 1,010)	Moderate Activity (n = 1,286)	Low Activity (n = 597)	Working (n = 781)
Personal leisure
Low	0.11	0.32	0.22	0.60	0.60
Medium	0.28	0.48	0.34	0.16	0.34
High	0.61	0.20	0.45	0.24	0.06
Civic/religious
Low	0.02	0.36	0.21	0.67	0.61
Medium	0.14	0.50	0.38	0.29	0.29
High	0.84	0.14	0.41	0.04	0.10
Physical exercise
None	0.13	0.13	0.44	0.69	0.35
Low	0.35	0.35	0.39	0.21	0.37
High	0.52	0.52	0.18	0.10	0.28
Interior household chores
None	0.09	0.16	0.36	0.77	0.53
Low	0.23	0.35	0.37	0.15	0.43
High	0.68	0.49	0.28	0.08	0.04
Exterior household chores
None	0.19	0.06	0.45	0.64	0.19
Low	0.27	0.35	0.32	0.24	0.51
High	0.53	0.59	0.22	0.13	0.29
Managing medical conditions
None	0.22	0.31	0.25	0.36	0.49
Low	0.26	0.36	0.35	0.37	0.40
High	0.53	0.33	0.40	0.28	0.11
Employment/computer use
None	0.32	0.27	0.56	0.81	0.10
Low	0.38	0.37	0.32	0.12	0.18
High	0.31	0.37	0.12	0.07	0.73
Interpersonal exchange/helping others
None	0.04	0.11	0.35	0.78	0.53
Low	0.26	0.41	0.41	0.16	0.32
High	0.70	0.48	0.24	0.07	0.15
Community leisure
None	0.07	0.17	0.19	0.56	0.19
Low	0.23	0.44	0.47	0.38	0.55
High	0.70	0.40	0.34	0.06	0.26

Note. Numbers represent conditional probabilities of engaging in each level of activity for each class. For example, the conditional probability of engaging in “high” levels of personal leisure for the “High Activity” group is 0.61.

It is not possible to characterize these profiles with simple labels. However, for convenience and discussion purposes, we use the following labels. Class 4 is identified as Low Activity, Class 3 as Moderate Activity, and Class 1 as High Activity. The two other groups had defining characteristics that also led to labels. We labeled Class 5 as Working and Class 2 as Physically Active.

Antecedents and Activity Profiles

Table 6 shows the independent variables related to these five profiles, with the Low Activity class being the constant comparison. In three of the four models, those with low self-rated health were more likely to be Low Activity. More functional limitations distinguished the Low Activity group from the Physically Active and Working groups, but functional limitations were not different between the Low Activity, High Activity, and Moderate Activity groups. More depressive symptoms distinguished the Low Activity group from the High Activity and Working groups. It is notable that health, functioning, and depression measures did not differentiate the Low Activity and the Moderate Activity profile.

Table 6.

Multinomial Logistic Regression: Antecedents and Activity Profiles

	High Activity vs Low	Physically Active vs Low	Moderate vs Low	Working vs Low
Self-rated health	0.34 (0.08)***	0.40 (0.08)***	0.10 (0.07)	0.47 (0.09)***
CES-D	−0.11 (0.04)**	−0.08 (0.04)*	−0.04 (0.03)	−0.09 (0.04)*
Mobility/ADL	−0.09 (0.08)	−0.55 (0.12)***	0.01 (0.07)	−0.28 (0.13)*
Age	−0.06 (0.01)***	−0.12 (0.01)***	−0.02 (0.01)*	−0.17 (0.01)***
Female	0.78 (0.15)***	0.25 (0.14)	0.50 (0.13)***	−0.44 (0.16)**
Education	0.10 (0.03)***	0.07 (0.03)*	0.09 (0.02)***	0.13 (0.03)***
Number of people in household	−0.12 (0.08)	−0.07 (0.07)	−0.11 (0.06)	0.02 (0.07)
Number of living child	0.02 (0.04)	0.05 (0.04)	0.002 (0.03)	0.004 (0.04)
Race (white)
African American	−0.37 (0.23)	−1.32 (0.25)***	−0.33 (0.20)	−0.98 (0.25)***
Hispanic	0.11 (0.30)	−0.18 (0.28)	−0.36 (0.26)	0.19 (0.31)
Other	−0.58 (0.46)	−0.28 (0.44)	−0.44 (0.45)	−0.19 (0.46)
Marital (married)
Not married	0.28 (0.20)	0.38 (0.20)	0.06 (0.17)	0.65 (0.22)**
Never married	0.41 (0.43)	0.43 (0.43)	0.26 (0.37)	0.71 (0.45)
Family assets	0.04 (0.02)**	0.05 (0.01)***	0.03 (0.01)*	0.03 (0.01)*
Family income	0.08 (0.05)	0.10 (0.06)	0.02 (0.04)	0.28 (0.10)**
Positive social support	0.03 (0.01)*	0.01 (0.01)	0.03 (0.01)**	0.03 (0.01)
Negative social support	0.01 (0.01)*	0.03 (0.01)**	−0.01 (0.01)	−0.001 (0.01)
Number of close friends	0.04 (0.01)**	0.03 (0.01)*	0.01 (0.01)	0.01 (0.01)
Urban/rural	0.23 (0.07)***	0.25 (0.07)***	0.11 (0.06)	0.25 (0.08)***
Neighborhood safety	0.11 (0.09)	0.04 (0.08)	0.10 (0.07)	0.13 (0.09)
Neighborhood cohesion	0.01 (0.01)	0.03 (0.01)**	0.02 (0.01)**	0.01 (0.01)

Note. ADL = activities of daily living; CES-D = Center for Epidemiologic Studies Depression.

*p < .05. **p < .01. ***p < .001.

When considering education and assets, two measures of SES, there was a consistent pattern: lower SES individuals were more likely to be Low Activity compared with any other class, even when controlling for the measures of health and functioning. Further, men were more likely to be Low Activity compared with High and Moderate Activity groups, but women were less likely to be in the Working group. Younger older adults were less likely to be Low Activity. Compared with whites, African Americans were less likely to be in the Physically Active and Working groups; there were no differences between whites, Hispanic, and other ethnic groups in regard to activity profiles.

Measures of household size and number of children were not related to activity profiles, although married individuals were less likely to be in the Working group. Social support measures as well as number of close friends distinguished the High Activity group from the Low Activity group. Social support measures were not related to the Working versus Low Activity groups. Those individuals in the Physically Active profile reported having more friends and more negative social support, whereas more positive social support distinguished Moderate from Low Activity groups.

Finally, in regard to the physical environment, more urban living characterized the High Activity, Physically Active, and Working profiles from the Low Activity profile. Neighborhood safety was not significant in any model, but neighborhood cohesion was higher for the Physically Active and Moderate Activity groups compared with the Low Activity group.

Activity Profiles and Well-Being Outcomes

Table 7 presents the relationship of activity profiles to well-being outcomes at a subsequent observation period. As would be expected, baseline measures of well-being in 2008 were strongly associated with well-being status 2 years later. After controlling for baseline well-being and other independent variables, activity profiles did retain predictive ability in regard to self-rated health and depression symptoms. In regard to self-rated health, the High Activity, Physically Active, and Working groups had better outcomes than Low Activity. The Moderate Activity class did not relate to health differently than Low Activity. In regard to depressive symptoms, the Low Activity group had poorer outcomes compared with all other groups.

Table 7.

Activity Profiles and Well-Being Outcomes

	Health	CES-D
Health (2008)	0.61 (0.02)***	−0.20 (0.03)***
CES-D (2008)	−0.03 (0.01)***	0.45 (0.02)***
Mobility/ADL (2008)	−0.10 (0.02)***	0.14 (0.05)**
Activity pattern (low activity)
High activity	0.12 (0.05)*	−0.44 (0.12)***
Physically active	0.10 (0.05)*	−0.37 (0.12)**
Moderate activity	0.04 (0.05)	−0.36 (0.12)**
Working	0.13 (0.06)*	−0.42 (0.12)***
Age	−0.01 (0.002)**	0.00 (0.004)
Female	0.03 (0.03)	0.26 (0.06)***
Education	0.02 (0.01)***	−0.03 (0.01)*
Number of people in household	−0.04 (0.01)*	−0.02 (0.03)
Number of living child	0.00 (0.01)	0.01 (0.01)
Race (white)
African American	0.05 (0.05)	−0.04 (0.11)
Hispanic	0.11 (0.06)	−0.09 (0.15)
Other	−0.16 (0.09)	0.07 (0.25)
Marital (married)
Not married	0.00 (0.04)	−0.22 (0.09)*
Never married	0.09 (0.08)	−0.33 (0.18)
Family assets	0.001 (0.003)	−0.01 (0.01)
Family income	−0.004 (0.01)	−0.01 (0.03)
Positive social support	0.005 (0.003)	−0.03 (0.01)***
Negative social support	−0.003 (0.002)	0.02 (0.01)**
Number of close friends	0.00 (0.003)	0.01 (0.01)
Urban/rural	0.01 (0.02)	0.03 (0.03)
Neighborhood safety	0.05 (0.02)**	−0.03 (0.04)
Neighborhood cohesion	0.001 (0.002)	−0.001 (0.01)

Note. Correlations among well-being outcomes are allowed in the model. ADL = activities of daily living; CES-D = Center for Epidemiologic Studies Depression.

*p < .05. **p < .01. ***p < .001.

Discussion

Using 36 activity items factored into nine domains, we were able to identify five distinct profiles of activity engagement among older adults. These profiles can be described in terms of types and amount of activity, and some signature features make them interpretable as unique profiles. Of course, fewer categories could be described, as indicted by the model fit statistics for the LCA, but maintaining the maximum number of categories supported by the statistics was instructive at this point for advancing the study of activity profiles.

Roughly speaking, there are high, medium, and low activity profiles, with the lines representing activity levels not exactly parallel, but never crossing. In other words, these classes included individuals who did less or more of all the activities, not just certain of the activities. However, the lack of parallel form suggests that some features characterize the groups. For example, the High Activity group was lower on employment/computer use compared with other activities those individuals engaged in. The Low Activity group was particularly low on community leisure and high on personal leisure, perhaps indicating that these individuals were more housebound.

The Low Activity group represented the older adults most vulnerable to poor outcomes. Members of this group were more likely to be older, to have fewer assets, and to have less education. Associative factors suggested that they were in poorer health than the other groups and their activity pattern was prospectively related to worse outcomes, in general. This is consistent with the finding that the activity they engaged in the most was managing medical conditions. It could be said that this profile did not achieve the “active ageing” imagined by the WHO. This group likely represents the biggest challenge in increasing activity engagement, given the lack of health and personal resources.

On other hand, the High Activity group may represent the stereotypic “busy” retiree and may be the model of “active retirement.” It is notable that antecedents included higher levels of health, lower levels of depressive symptoms, and higher ratings of life satisfaction, but not higher level of mobility/activities of daily living functioning. This may imply that despite some physical limitations, these individuals had other personal and social resources that enabled engagement. Correspondingly, findings do suggest that this group had high levels of social support, whether positive, negative, or number of friends. Findings also suggest that, although the individuals with this activity profile might have had some physical limitations, they had the best well-being outcomes 1 year later.

The Working group spent more time engaged in employment activities, and it is notable that the level of engagement in the other activities was similar to the Low or Moderate Activity group. The members of the Working group were low on household chores and managing medical conditions, indicating that they spent less time managing aspects of their personal lives and were perhaps able to devote more time to work. It is important to understand how work cessation will affect activity engagement for these individuals. These data could suggest that, unless these individuals increase activities in other domains, they may become low activity. This pattern would be consistent with results of other studies, such as Agahi, Ahacic, and Parker (2006), which found that, in general, there is continuity of activity engagement across the life course and that most people do not adopt new activities later in life. Shaw, Liang, Krause, Gallant, and McGeever (2010) suggest that leisure-time physical activity levels tend to decrease rather than increase in middle and later life, supporting the speculation that those in the Working group may not increase their activity engagement in retirement. The findings of this study suggest that programs and policies in the preretirement and transition period that focus on activity engagement in retirement may be important.

In future studies of changes in activity pattern over time, gender and social support will be important. In this analysis, men were more likely to be in either the Working group or the Low Activity group. One interpretation of this finding is that men are not engaged in other activities while working (demonstrated by the activity levels in the Working profile) and after they stop working, these levels of activity stay low. In other words, they may not start new activities or increase activity engagement in other arenas in retirement. Given the wide influence of social connections on human behavior, it is expected that social support is related to activity profiles. Yet the relationships may be complex, as suggested by the findings that the three measures of social support in this study (positive and negative social support and number of close friends) related differently to profiles.

Findings regarding two different types of well-being indicators offer some interesting ideas. Low Activity individuals experienced worse mental health outcomes over time compared with all other activity groups. Level of activity engagement may affect mental health more than physical health. The Moderate Activity individuals did not differ from the Low Activity group in terms of self-reported health in the prospective analyses, whereas the High Activity, Working and Physically Active groups reported better outcomes on these two assessments. This offers general support for the fundamental idea that higher activity engagement maintains or produces health for older adults.

There are several limitations in this work worth noting. First, although there is prospective assessment of the effects of activity on subsequent well-being outcomes, this study remained observational, limiting a causal argument. Longitudinal analyses involving multiple waves of activity data over time is feasible and will provide stronger arguments for causes and effects of activity patterns. Despite the fact that the HRS has more activity items than most publicly available nationally representative data sets, an even wider range of activities could be included in the survey. Further, the 36 activities used in the study clearly represent things that people do, but these items are ambiguous. Items are not consistent in the range of behaviors captured. For example, listening to music involves a narrower focus than volunteering or managing medical conditions, both of which may encompass a wide range of behaviors. Additionally, the contexts of the activities are not considered, like the extent to which other people are involved or the subjective aspects of the activity. For example, the levels of discretionary involvement, enjoyment, stress, and so forth remain unspecified, and there is evidence that activity involvement in and of itself is not adequate to understand activity outcomes (Matz-Costa, Besen, James, & Pitt-Catsouphes, 2012). Lastly, our analysis spans a short period in time, 2008–2010, which limits our ability to understand activity trends and their relationship to health and wellness outcomes over time.

In sum, we believe that, foremost, this study demonstrates that a fuller range of activities can be studied simultaneously, allowing for more complex analyses that capture the “whole person.” The findings themselves support various conceptualizations, like the WHO Active Ageing Framework or the successful aging paradigm, that older adults are more or less active, in that we clearly document a group of Low Activity individuals as well as a High Activity group. This study also supports the WHO Active Ageing framework that activity is determined by a broad range of factors, from personal to environmental. It is especially noteworthy that physical limitations did not always differentiate activity patterns, suggesting the importance of social and physical contexts in enabling engagement. Some of these determinants are amenable to policy and program interventions, and such interventions have not received systematic attention to increase the activity participation by older adults. Finally, empirical work supports the relationship between single activities or a smaller set of activities and well-being outcomes, but the effect of activity profiles that capture a wider range of activities has not been considered. This study offers evidence that activity profiles affect well-being and that these profiles might relate differently to physical and mental health outcomes. This study raises many important questions about activity engagement in later life and offers researchers a more comprehensive way to study how activities relate to healthy aging.

Funding

This work is supported by a grant from the National Institute on Aging (1R21AG038868-01A1).