Skip to main content
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2012 Dec 1.
Published in final edited form as: Psychol Aging. 2011 May 23;26(4):830–843. doi: 10.1037/a0023180

What contributes to perceived stress in later life? A recursive partitioning approach

Stacey B Scott 1, Brenda R Jackson 2, C S Bergeman 2
PMCID: PMC3177031  NIHMSID: NIHMS299090  PMID: 21604885

Abstract

One possible explanation for the individual differences in outcomes of stress is the diversity of inputs that produce perceptions of being stressed. The current study examines how combinations of contextual features (e.g., social isolation, neighborhood quality, health problems, age discrimination, financial concerns, and recent life events) of later life contribute to overall feelings of stress. Recursive partitioning techniques (regression trees and random forests) were used to examine unique interrelations between predictors of perceived stress in a sample of 282 community-dwelling adults. Trees provided possible examples of equifinality (i.e., subsets of people with similar levels of perceived stress but different predictors) as well as for the identification both of contextual combinations that separated participants with very high and very low perceived stress. Random forest analyses aggregated across many trees based on permuted versions of the data and predictors; loneliness, financial strain, neighborhood strain, ageism, and to some extent life events emerged as important predictors. Interviews with a subsample of participants provided both thick description of the complex relationships identified in the trees, as well as additional risks not appearing in the survey results. Together, the analyses highlight what may be missed when stress is used as a simple unidimensional construct and can guide differential intervention efforts.

Keywords: Perceived Stress, Loneliness, Regression Trees, Random Forests, Aging


Stress has implications for mental and physical health in later life. Although some research has focused on exposure to specific life events (e.g., loss of a spouse) or chronic conditions (e.g., pain), most of this work conceptualizes stress as one's perception of difficulty adjusting to or coping with everything from major trauma to daily hassles. For example, Lazarus and Folkman (1984) emphasize the demands versus resources aspect of psychological stress and describe it as the “particular relationship between the person and the environment that is appraised…as taxing or exceeding his or her resources and endangering his or her well-being” (p. 19). The aim of this project is to better understand the combinations of strain in diverse life domains that relate to perceived stress in a later life sample, as well as to illuminate the possible pathways by which people appraise the stressfulness of their lives.

Although perceived stress is a commonly used predictor in the literature, the stress-outcome link is not a simple unidirectional effect, in which stress produces negative health and well-being outcomes. This relationship could also work in the opposite way, with health problems or depression contributing to the appraisal that life is stressful (Mitsonis et al., 2009). Resilience research (see Ong, Bergeman, & Boker, 2009 for review) has focused on possible protective factors (e.g., social support, personality) that may buffer the effect of stress. Although much research has been directed to these questions, less attention has focused on what factors produce this stressed state. Specifically, prior to determining which characteristics of an individual may modify the effect of perceived stress, it is useful to explore which features of person's life contributed to him or her appraising life as stressful in the first place. A variety of contextual and personal risk factors have separately been linked to negative outcomes using traditional regression techniques. Beyond examining which domain of risk is most important, it is also important to consider for whom certain circumstances are particularly risky; more exploratory methods specialize in identifying these nuanced relationships (Gruenewald, Mroczek, Ryff, & Singer, 2008). By examining multiple domains in the same analysis, it is possible to examine the contexts in which people live and the unique combination of factors that put them in jeopardy.

Certain aspects of later life may be especially noxious. For example, regardless of how well or poorly the individual fares in other life domains, the presence of chronic health problems may distinguish those who are highly stressed from those who are not. There may also be specific combinations of strain from various life domains (e.g., financial concerns, health problems, isolation) that interact with one another to impact the extent to which people feel overwhelmed in later life. In an effort to address these questions, the present study explores the patterns by which life events, neighborhood strain, age-related discrimination (i.e., ageism), social isolation (i.e., loneliness), financial strain, and physical health affect the extent to which older adults perceived themselves as generally “stressed.” A brief review of the relevant research on each of these six domains is presented here to support their inclusion in the analysis.

Life Events

Theorists have long recognized that experiences such as widowhood and retirement are major life events (Palmore, Cleveland, Nowlin, Ramm, & Siegler, 1979; Holmes & Rahe, 1967, Neugarten, 1979), and as such are important influences on perceptions of stress. Older adults nominate these experiences and others such as personal illness, death of a loved one, and loved one's illness as being particularly pernicious (Hardy, Concato, & Gill, 2002; Sematter-Bagnoud, Karmaniola, Santos-Eggimann, 2008). Studies examining the impact of negative events in later life have focused on mental health outcomes, such as depression or anxiety (De Beurs, Beekman, Geerlings, Van Dyck, Van Tilburg, 2001; Kessing, Agerbo, & Mortensen, 2003), on indices of physical health and mortality (Clémence, Karmaniola, Green, & Spini, 2007;) and disease precursors and symptoms (Golden-Kruetz et al., 2005; Mitsonis et al., 2009; Rafanelli et al., 2005; Rosengren et al., 2004; Wigers, 1996).

Financial Strain

Financial strain is another factor that is likely to impact the degree to which people feel “stressed.” Lower socioeconomic status and circumstances of economic disadvantage have been associated with increased rates of depression (Pinquart, Silbereisen, & Korner, 2010), as well as detrimental health behaviors such as smoking, inactivity, and poor diet, which in turn lead to increased morbidity, particularly cardiovascular disease (Kershaw, Mezuk, Abdou, Rafferty, & Jackson, 2010; Rosengren, et al., 2004). There is also evidence that financial strain has an impact on well-being over and above current socioeconomic status or economic circumstances (Kahn & Pearlin, 2006).

Neighborhood Strain

Neighborhoods provide both physical (i.e., upkeep of houses in area, access to stores and religious outlets) and social (i.e., long-standing relationships with neighbors, safety) contexts and have implications for both health and well-being in later life (see Yen, Michael, & Perdue 2009 for review). Pearlin and Skaff (1996) refer to concerns about the quality of neighborhood as ambient strain, which are conditions in a person's environment that may increase feelings of vulnerability through threats to security and safety or limitations on access to services. In older adults, such neighborhood strain has been linked to reports of more physical health problems, as well as poorer mental health functioning (Krause, 1998), and is a potentially powerful predictor of overall perceptions of stress.

Ageism

Discrimination is consistently associated with negative outcomes across a variety of demographic characteristics, but discrimination on the basis of age—or ageism—is especially threatening to the health and well-being of older adults (Levy, 2009). For example, being a victim of negative stereotypes of aging have been linked with elevated cardiovascular response to stress (Levy, Hausdorff, Hencke, & Wei, 2000; Levy, et al., 2008, Pascoe & Richmond, 2010), increased mortality risk (Barnes, et al., 2008), the will to live (Levy, Ashman, & Dror, 2000), greater psychological distress and lower positive well being (Yuan 2007), and a variety of psychosocial domains such as social involvement and responsibility, physical activity, self-rated health, and subjective age (Sanchez, Palacios, Trianes, Torres, & Blanca Mena, 2008).

Loneliness

Feelings of loneliness are associated with a myriad of negative outcomes. These include psychosocial factors such as depressive symptoms (Aanes, Middlemark, & Hetland, 2010; Steptoe, Owen, Kunz-Ebrecht, & Brydan, 2004), poorer health-related quality of life (Maxwell, et al., 2009), and greater perceptions of distress (Cacioppo, Hawkley, & Berntson, 2003; Kiecolt-Glaser, et al., & Glaser, 1984); behavioral factors such as physical inactivity (Hawkley, Thisted, Masi, & Cacioppo, 2010) and sleep problems (Cacioppo et al., 2003; Steptoe, Owen, Kunz-Ebrecht, & Brydan, 2004); physiological factors including increased levels of cortisol (Doane & Adam, 2010, Steptoe et al., 2004), markers of poor cardiovascular health (Hawkley, Burleson, Berntson, & Cacioppo, 2003; Hawkley, Masi, Berry, & Cacioppo, 2006), more pronounced physiological reactions to stress (Steptoe et al., 2004), and poorer immune system functioning (Glaser, Kiecolt-Glaser, Speicher, Holliday, 1985; Kiecolt-Glaser, et al., 1984). Loneliness is consistently associated with an increased risk of morbidity and mortality across demographic characteristics (Berkman, Leo-Summers, & Horwitz, 1992; House, Landis, & Umberson, 1988; Penninx, et al., 1997).

Chronic and Somatic Health Problems

Although the association between chronic health issues and perceptions of stress makes intuitive sense, there is a large body of empirical support for this relationship as well (Lagana & Reger, 2009; Linn, Sandifer, Stein, 1985; McDade et al., 2006; Strodl et al., 2003). Chronic pain from arthritis, for example, may itself be a stressor; in this overtaxed state, individuals may narrow their focus to negative experiences rather than the more demanding processing of both positive and negative information which has been linked with well-being (Davis, Zautra, & Smith, 2004). For example, among individuals with rheumatoid arthritis and fibromyalgia, more severe and/or chronic symptoms were related to elevated levels of stress (Murray, Daniels, & Murray, 2006; Tak, Hong, & Kenneday, 2007).

Present Study

The focus of the current study is to elucidate the contexts (e.g., financial strain, neighborhood, social isolation, ageism, health problems) and life events that may converge to make older adults generally perceive their lives as stressful. The first aim was to use exploratory recursive partitioning to analyze survey data of contextual stress indices across a variety of domains. These analyses allow us to examine how combinations of the contextual features of individuals' lives may be related to overall feelings of stress. Specifically, these analyses allow for an examination of equifinality, meaning multiple different pathways to the same outcome (Cummings, Davies, & Campbell, 2000; Gruenewald et al., 2008). In other words, individuals who have comparable levels of perceived stress may have similar responses for very different reasons. Intervention or prevention strategies designed to diminish levels of stress need to be sensitive to these different pathways. The second aim is to use interview data to profile individuals who have these unique combinations of life contexts to further probe our quantitative findings. Although subjective in nature, it is hoped that the life stories of older adults will aid our understanding of the processes through which individuals view their lives as challenging, and better articulate what it means to feel stressed.

Regression trees and random forests, described in the method section, are particularly appropriate because they allow for different weightings and contributions of the risk factors across different portions of the sample. In traditional regression, the influence of a predictor such as number of life events experienced is constant across all levels of the predictor itself (e.g., the same regression coefficient is applied to both 1 and 10 life events). In contrast, the recursive partitioning procedures employed here not only allow for the identification of the domain that best divides the sample on the outcome at each split, but also determines the level at which this predictor best splits the sample (e.g., people who report experiencing life events one standard deviation above the sample average are at particular risk for high perceived stress). Additionally, unique interactions and nonlinear relationships may be identified that are especially useful in profiling one subset of participants whereas other predictors and interactions may be identified to describe another. The selected risk domains are not typically examined together, thus it is difficult to specify a priori the numerous interactions that may appear or the predictors that may best “split” the subjects into the various pathways. These techniques have been applied in prior aging research (see Gruenewald et al., 2008; Gruenewald, Seeman, Ryff, Karlamangla, & Singer, 2006; Wallace, Bergeman, & Maxwell, 2002), however, the present study is the first article to appear in this journal using the method and serves as an introduction to the framework, terminology, and application. Exploratory techniques such as this one provide an opportunity to describe the relationships in a particular dataset, which can then be used to develop specific confirmatory hypotheses in other samples and to link separate domain-specific stress literatures.

Method

Participants

Participants included 282 adults ranging in age from 54 to 91 years (M=68.79, SD=5.10) who completed the second yearly survey of the longitudinal Notre Dame Study of Health & Well-Being (NDHWB; Bergeman, et al., 2010). Demographic information is provided in Table 1. Year 2 data was used because it coincided with the interview portion of the study. Slightly more than half the sample was female and it reflects the cultural diversity of the Northern Indiana area. The interview participants were selected from the pool of participants who endorsed at least two events on the life events checklist in their survey. From this larger group, 37 NDHWB participants were selected for possible interviews; 26 interviews were completed. The remaining 11 participants were not interviewed for various reasons (e.g., unable to be reached by phone, lived too far away, not interested). Although not selected to be directly representative of the larger sample, they reflect similar demographic characteristics.

Table 1.

Demographic Characteristics of the Full and Interview Samples

Full Sample
Interview Sample
N % N %
Total N 282 -- 26 --
Gender (Missing = 4)
 Female 169 61 16 62
 Male 109 39 10 38
Marital Status (Missing = 3)
 Married 129 46 9 35
 Divorced 65 23 8 31
 Widowed 61 22 4 15
 Single 21 8 4 15
 Separated 3 1 1 4
Race (Missing = 3)
 Caucasian 234 84 18 69
 Black/African American 29 10 7 27
 Hispanic/Latin American 9 3 0 --
 Asian/Pacific Islander 2 1 1 4
 Native American/Aleutian Islander 1 <1 0 --
 Other 4 1 0 --
Education (Missing = 2)
 Grade School (grades 1–6) 1 <1 0 --
 Middle School (grades 7–9) 5 2 0 --
 High School (grades 10–12) 110 39 11 42
 Vocational Education 17 6 1 4
 Some College Classes 61 22 7 27
 College Degree 46 16 4 15
 Post College Professional Degree 23 8 2 8
 Graduate, Medical, or Law Degree 18 6 1 4
Annual Household Income (Missing = 7)
 Less than $7,500 3 1 0 --
 $7,500 to $14,999 49 18 7 27
 $15,000 to $24,999 50 18 7 27
 $25,000 to $39,999 72 26 8 31
 $40,000 to $74,999 71 26 2 8
 $75,000 to $99,999 18 7 1 4
 $100,000 or more 12 4 1 4

Note. Number of people missing refers to the Full Sample; demographic data for the Interview Sample is complete.

Survey measures

Global perceptions of stress

Overall perceptions of stress were assessed using the 14-item Perceived Stress Scale (PSS; Cohen et al., 1983). The PSS was designed to assess the degree to which individuals appraise the situations in their lives to be unpredictable, uncontrollable, and overloaded. Participants are asked to reflect over the last month and report the frequency of agreement on a four-point scale (never, sometimes, often, always). A higher overall score represents more perceived stress. Cronbach's alpha in this sample is 0.89.

Life events

Life events in the past year were assessed using the Elders Life Stress Inventory (ELSI; Aldwin, 1990). Participants endorsed the occurrence of 31 events across a variety of domains including health, finance, family, and work. A simple count of events is typically used with this instrument (i.e., Isaacowitz & Seligman, 2001; von Dras, Powless, Olson, Wheeler, & Snudden, 2005).

Chronic and somatic health problems

Respondents reported on a variety of medical conditions and physical health problems using a checklist following Belloc, Breslow, and Hochstim (1971). Fourteen items assessed chronic health problems such as high blood pressure, arthritis, cancer, and diabetes. Fifteen items assessed somatic health complaints such as joint stiffness, headaches, tiring easily, and shortness of breath.

Financial strain

Four items, modeled after those contained in the MIDUS (Brim et al., 1995–1996) were used to assess financial strain. Using an 11-point scale, participants rated their current financial situation, the amount of control over their current financial situation, and the amount of thought or effort they put into their current financial situation. Additionally, participants selected one category from three to describe their current finances (i.e., “We do not have enough money to meet our needs,” “We have just enough money to meet our needs,” “We have more money than we need”). Responses to the four items were standardized and summed; α in this sample was 0.74.

Neighborhood strain

Participants' ratings of their neighborhood were assessed using a 12-item scale developed based on research by Keyes (1998) and Ryff, Magee, Kling, and Wing (1999). Items tapped a variety of features of neighborhood, including cleanliness, safety, and extent to which participants could call upon neighbors for help. Participants responded on a 4-point scale; α was 0.91.

Loneliness

Social isolation was measured using the 20-item Revised UCLA Loneliness Scale (Russell, Peplau, & Curtona, 1980). Respondents are presented with a statement about perceived social connectedness and are instructed to rate each on a scale of 1, indicating never, to 4 indicating often. One item was changed from the original source (i.e., from “People are around me but not with me” to “I often feel alone even when in a crowd”) for clarification; α in this sample was 0.92.

Age discrimination

Experiences of discrimination were assessed using a 20-item checklist of experiences following work by Palmore (2001). Respondents were asked to endorse events that they experienced because of their age. Higher scores were indicative of more discrimination experiences.

Interview

The interview component of the larger project was not designed specifically for this study; the data, however, allow for a more rich profile of the lives of individuals in the terminal nodes produced by the recursive partitioning analysis. The focus of the interviews, which included a description of the life events and difficulties in the last year fits well with the components assessed in the survey data. It allows not only for a confirmation of the recursive partitioning results, but also extends this approach to provide a more detailed examination of these pathways to perceived stress.

The interviews were semi-structured, following the format of the Life Events and Difficulties Schedule (Brown & Harris, 1978) a semi-structured interview with a series of questions regarding life events in the previous year in a variety of domains (e.g., work, family, health). In addition, interviewees were asked more broadly about their life experiences and were encouraged to discuss important turning points and to contextualize the joys and sorrows of their lives. The interview was designed to be flexible and have a conversational flow, rather than a clinical interview or standard survey (Wethington et al., 1997). Interviews were audio recorded and transcribed verbatim.

Procedure

Participants were mailed a packet of questionnaires which included a variety of paper-and-pencil stress measures. Upon returning their packets, participants received a $20 gift card. Those participants who met the sampling criteria were contacted via telephone and asked if they would participate in a follow-up interview. Interviews took place either at participant homes or at research offices. Interviews generally lasted about 1.5 hours, but ranged from 1 to 3 hours. Participants received a $30 gift card for their participation.

Analyses

Recursive partitioning: Regression trees of perceived stress

A goal of the present study was to examine the interrelations between various life events and difficulties in putting people at-risk for generally perceiving their lives as stressful. Regression trees were used to conduct a tree-structured, non-parametric data analysis of predictors of perceived stress in a later life sample. The description of regression trees below draws from several sources (Berk, 2008; Brieman, Friedman, Olshen, & Stone, 1984; Hastie, Tibshirani, & Friedman, 2001; Merkle & Shaffer, in press; Strobl, Malley & Tutz, 2009). Perceived stress was used as the outcome in this analysis and treated as a continuous variable. Regression trees fall under the broad category of a binary recursive partitioning techniques. Recursive partitioning has been used in the aging literature to identify interrelations between protective factors (Wallace et al., 2002), biomarkers predicting mortality (Gruenewald et al., 2006), and pathways to positive and negative affect (Gruenewald et al., 2008).

In this study, we used the software program CART Pro Version 6.0 (Steinberg & Colla, 1997) to conduct the tree analyses. The procedure begins with a root node, containing the entire sample. In the case of regression, the software uses a least squares splitting rule to compare all possible cutoff points on all the predictors. The optimal predictor and its cutoff point, which minimizes the within-group variance on perceived stress, are identified; the two groups produced by these binary splits are called child nodes. The same procedure is repeated for each of the child nodes, recursively identifying the predictor that optimally separates the participants in that node into two subsequent child nodes. This process is often displayed in an inverted-tree structure in which the multiple splits resemble branches that end in terminal nodes, groupings that cannot reliably be further divided.

CART Pro first grows a large tree because it is impossible to identify a good stopping point a priori. If the procedure stopped too early, key splits could be missed. It should be noted, however, that overly large trees are difficult to interpret and their structures are unlikely to be replicated. One way of limiting tree complexity is by imposing a constraint on the minimum number of cases necessary for a terminal node. In this study, the minimum terminal node size was set at 10. This large tree is then pruned, removing branching splits that may capitalize on chance relationships.

With a large enough sample, an ideal method for assessing the extent to which a particular tree overfit the idiosyncrasies of a dataset is to split the data into a learning sample and an independent test sample. A large tree is grown on the learning sample, then the test sample is used to examine the error the sequence of trees produced that are special cases of the large tree, and a best tree is selected by its cost. In smaller samples for which dividing the participants into learning and test samples is not feasible, as in this study, cross-validation methods are used. In cross-validation, the software determines how much to prune the large tree by constructing a series of trees that are grown on subsamples of the entire dataset. In this study, 10-fold cross-validation was chosen. In this procedure, the sample is randomly divided into 10 equal-sized subsets. A tree is then grown on nine of those subsets (90% of the overall sample) and then the remaining 10% of the sample is used as a pseudo-test sample, and the sum of squared error for the sequence of trees that are special cases of the large tree is calculated. This was repeated 10 times, resulting in all subsets serving as both learning sample and pseudo-test sample. Those branches or splits that do not improve the accuracy in cross-validation are pruned. The error information is combined across trees and used to prune the large tree to identify the optimal tree. The optimal tree has the lowest cross-validated relative error rate for trees grown, providing the smallest tree with error comparable to the large tree.

Missing data and data transformations

Of the 297 available participants at this wave, 15 (approximately 5% of the sample overall) were missing on at least one of the domains of interest. These participants were excluded from the dataset1. Regression tree results are unaffected by monotonic transformations of the predictors (Wallace et al., 2002). As neither the perceived stress outcome nor the predictors have a common metric or clinically-relevant-cutoff values, all survey scales were standardized prior to analysis to aid in interpretation of the cut-points

Results

Descriptive Statistics

Descriptive statistics and correlations are displayed in Table 2. As expected, all the contextual risk domains were positively related to the Perceived Stress Scale. Age was unrelated to PSS and all predictors except for chronic health problems. Gender, race, and marital status were not related to PSS, however, both SES-related indicators (e.g., education: F=2.65, p < 0.02; income: F= 2.42, p < 0.03) were, with higher SES participants reporting lower overall perceived stress.

Table 2.

Means, Standard Deviations, and Correlations of Variables in the Full Sample (N=282)

M SD 1. 2. 3. 4. 5. 6. 7. 8. 9.
1. Perceived Stress 29.12 5.84 1.00
2. Life Events 2.28 2.58 0.26 1.00
3. Neighborhood Strain 20.77 5.48 0.35 0.19 1.00
4. Financial Strain 0.003 3.02 0.36 0.24 0.35 1.00
5. Ageism 2.53 2.71 0.31 0.32 0.23 0.16 1.00
6. Loneliness 34.69 9.24 0.55 0.23 0.38 0.28 0.27 1.00
7. Chronic Health 2.02 1.53 0.18 0.21 0.17 0.22 0.28 0.24 1.00
8. Somatic Health 2.48 2.12 0.26 0.19 0.17 0.25 0.29 0.29 0.57 1.00
9. Age 68.84 5.20 −0.09 0.08 −0.07 0.02 0.04 −0.09 0.16 0.02 1.00

Note. Financial Strain is comprised of standardized items. The scales for Life Events, Ageism, Chronic Health, and Somatic Health reflect counts of exposure or symptom reports. Raw (unstandardized) scores are reported here, with the exception of Financial Strain; Analyses use standardized scores. Significant correlations are bold.

Regression Tree Results

The optimal tree, which is the tree with the minimum cost regardless of size, identified is displayed in Figure 1; all predictors and the outcome were z-transformed to aid interpretation. Additional descriptive data by node are provided in Tables 3 and 4. It is important to remember that although this tree was identified as the smallest with an error comparable to the large tree, the analyses produced many trees including 11 of which were within 1 standard error of the large tree. Each of the trees identified within the 1 SE rule (Breiman et al., 1984) could be nominated as the best tree; the selected tree, however, was chosen for its low relative error and relatively even distribution of participants and interviews across terminal nodes. The tree includes six pairs of branches (e.g., splits) and seven terminal nodes describing the relationships between contextual risk domains and overall perceived stress in this sample. Following other studies using recursive partitioning (Gruenewald et al., 2008; Seroczynski, Cole, & Maxwell, 1997), we calculated the percent variance explained by the model as one minus the ratio of average within node error variance of the optimal tree to the total sum of squared errors of the baseline model with only the root node. The optimal tree selected explained 45% of the variance in perceived stress. Upon recommendation of a reviewer, we examined the R-square of a traditional linear regression model for perceived stress using all the contextual risks and their two-way interactions as predictors. This resulted in an R-square of .42 which is quite close to the pseudo-R-square value calculated for the selected tree. Given that previous work tended to examine individual contextual domains and perceived stress separately, we did not have specific hypotheses about how these may interact. The tree approach employed has the benefit of being exploratory from the outset.

Figure 1.

Figure 1

Regression tree diagram: Regression of perceived stress scores onto contextual risk domains. PSS: Perceived Stress Scale; Neighbr: neighborhood strain; Fin. Str.: financial strain; Life Ev.: life events. Standardized scores are reported. Nodes display level of splitter variable, node size, and average perceived stress relative to entire sample. Terminal nodes are indicated by circles and include number of interviewed participants which fell into that node.

Table 3.

Demographic Characteristics by Terminal Node; Values are Percentages.

Node 8 12 13 10 11 6 7

N=24 N=76 N=41 N=27 N=67 N=21 N=26
Gender (Missing = 4) (0) (0) (1) (0) (3) (0) (0)
 Female 63 55 60 56 61 48 92
Marital Status (Missing = 3) (0) (0) (0) (0) (3) (0) (0)
 Married 41 53 47 59 41 53 27
 Divorced 23 24 29 11 27 24 19
 Widowed 24 22 12 22 23 19 31
 Single 12 1 12 8 6 4 19
 Separated 0 0 0 0 3 0 4
Race (Missing = 3) (1) (0) (1) (0) (0) (0) (1)
 Caucasian 83 83 80 93 87 86 76
 Black/African American 4 8 15 7 12 5 20
 Hispanic/Latin American 9 5 3 0 1 5 0
 Asian/Pacific Islander 0 3 0 0 0 0 0
 Native American/Aleutian Islander 0 1 0 0 0 0 0
 Other 4 0 2 0 0 4 4
Education (Missing =1) (0) (0) (0) (0) (1) (0) (0)
 Grade School (grades 1–6) 0 1 0 0 0 0 0
 Middle School (grades 7–9) 4 1 0 4 1 0 4
 High School (grades 10–12) 25 26 37 44 55 38 50
 Vocational Education 8 4 5 0 14 0 4
 Some College Classes 13 28 27 19 14 29 23
 College Degree 21 20 15 18 12 14 15
 Post College Professional Degree 12 12 11 11 3 5 0
 Graduate, Medical, or Law Degree 17 8 5 4 1 14 4
Annual Household Income (Missing =7) (1) (2) (0) (1) (2) (0) (1)
 Less than $7,500 0 0 0 0 2 5 4
 $7,500 to $14,999 13 4 17 4 31 10 52
 $15,000 to $24,999 9 9 32 8 29 14 16
 $25,000 to $39,999 35 30 29 19 25 19 20
 $40,000 to $74,999 17 45 17 42 12 33 4
 $75,000 to $99,999 9 8 3 19 1 14 0
 $100,000 or more 17 4 2 8 0 5 4

Note. Number of people missing refers to the full sample. Values in parentheses indicate the number of people missing on that characteristic in each node.

Table 4.

Descriptive Statistics by Terminal Node

Node 8 12 13 10 11 6 7

N=24 N=76 N=41 N=27 N=67 N=21 N=26
Perceived Stress 23.6 26.2 28.8 26.1 31.3 32.9 37.7
(4.3) (4.7) (4.3) (3.6) (4.5) (5.3) (4.2)
Loneliness 27.3 27.7 28.3 36.2 37.0 50.3 51.9
(3.9) (3.5) (3.1) (2.7) (3.1) (5.4) (6.2)
Financial Strain −1.3 −2.0 2.4 −3.2 1.7 −0.1 2.6
(2.7) (1.5) (1.7) (1.0) (2.4) (2.6) (3.6)
Life Events 0 2.1 2.7 1.9 2.4 3.1 3.7
-- (1.2) (1.8) (1.5) (3.5) (3.0) (3.9)
Neighborhood Stress 17.4 19.2 20.6 17.6 22.7 19.3 28.3
(4.4) (4.8) (4.8) (5.1) (4.6) (2.6) (5.0)
Ageism 1.4 2.1 2.2 1.8 2.9 3.5 4.2
(1.2) (2.3) (2.1) (2.2) (2.6) (3.0) (4.6)
Chronic Health 1.3 1.7 2.0 1.5 2.4 2.8 2.5
(1.2) (1.4) (1.3) (1.4) (1.4) (2.0) (1.8)
Somatic Health 1.3 2.0 2.5 2.2 2.8 3.2 3.8
(1.7) (1.7) (1.9) (1.6) (2.6) (2.0) (2.2)

Note. Means and standard deviations (in parentheses) of raw scores, with the exception of Financial Strain which is standardized, on each variable by node.

In the resulting tree, loneliness, financial strain, neighborhood, and life events were identified as splitters. Loneliness appears twice in the tree. It is the primary splitter at the root node (Node 1), which separates those who report very high loneliness (Node 3) from the rest of the sample. Participants in this high-loneliness pathway can be further described and separated by their level of neighborhood strain; this is an example of an interaction and is discussed below. Loneliness appears again as a splitter in the path to the left of the root node (Node 2), an example of the intricate relationships that trees and other recursive partitioning techniques can identify. That is, among those participants who do not report extremely high levels of loneliness, different profiles emerge. For those who report moderate to low levels of loneliness (Node 5), financial strain further explains differences in overall level of perceived stress. For those who report very low levels of loneliness (Node 4; loneliness 0.20 standard deviations [SDs] below the sample average), recent life events further subdivides this group, differentiating those who report extremely low perceived stress (Node 8; PSS scores that are nearly one SD below the mean) from those who report more low to average levels (Node 9; PSS 0.34 SD below average). Financial strain also appears multiple times within the tree, both in the low-to-medium loneliness path (Node 5) as well as for those who report very low loneliness, but some life events (Node 9).

Terminal nodes (Nodes 6, 7, 8, 10, 11, 12, and 13) are indicated by circles in Figure 1. Multiple pathways to perceived stress are apparent in the terminal nodes. Those persons in Node 8, who reported the lowest loneliness and life events scores (0.20 and 0.69 SDs below average, respectively), reported perceived stress scores nearly one SD below the average person in the study. (Note: in this sample, 1 SD reflects a difference of almost 6 points on the raw scores of the PSS which ranges from 15 to 47). Node 9 participants had similarly low levels of loneliness, but reported more life events, and these individuals' perceived stress was further differentiated by their financial strain. Participants in the Node 12 pathway, had average (or lower) levels of financial strain and reported perceived stress about ½ SD below the sample average; whereas those participants in Node 13 reported higher levels of financial strain and had scores at about the average of perceived stress.

Moving across the figure from left to right, we can examine the role of more moderate levels of loneliness and its interactions with other domains in predicting perceived stress. Interestingly, the adults in terminal Node 10 have similar levels of perceived stress as terminal Node 12 (approximately ½ SD below average), but it is the interaction between low-medium loneliness and very low financial strain that is most related to their low overall level of perceived stress. Their pathway neighbors in terminal Node 11 report higher levels of financial strain, and perceived stress scores 0.4 SDs above the sample average.

For those participants who reported very high values of loneliness (Node 3 pathway), neighborhood strain was a useful predictor in distinguishing levels of perceived stress. Participants in terminal Node 6 report high levels of loneliness, but relatively low-to-average neighborhood strain. Average perceived stress in this group is 0.6 SDs above average, not that dissimilar to Node 11. In addition to the very high loneliness characteristic of this pathway, those in terminal Node 7 reported higher levels of neighborhood strain. These participants appear to be at particular risk for feeling that life is overwhelming or unpredictable; they report perceived stress nearly 1½ SDs above the rest of the sample.

In summary, the highest levels of global perceived stress tended to be related to higher levels of loneliness. Different predictors, however, were helpful in separating participants with moderate and high levels of loneliness. For example, it may be that their lower levels of strain in other domains (i.e., finances and neighborhood) provided offsetting influences for Nodes 10 and 6 compared to their counterparts in nodes 11 and 7 who perceived their lives to be much more stressful. We use random forests to further examine the contributions of predictors.

Follow-up: Random forests

It is important to keep in mind that trees are very sensitive to the particularities of the sample, meaning that “the entire tree structure could be altered if the first splitting variable, or only the first cut point was chosen differently because of a small change in the…data” (Strobl, Malley, & Tutz, 2009, p. 330). Cross-validation is one way to prune a tree and limit the tendency to overfit the data. A recent development in recursive partitioning techniques is to take advantage of this variation in individual trees by aggregating over multiple trees to obtain more stable results, a procedure called random forests (Breiman, 2001). Random forests incorporates additional variability through two methods – varying the sample and varying the predictors. Similar to the resampling of the dataset in the cross-validation procedure above, the random forest procedure either bootstraps or subsamples the participants to vary the sample; this step is sometimes referred to as bagging. In addition to varying the participants by bagging, in random forests the predictors are sampled with replacement in order to produce an even more diverse set of trees. When a particularly strong predictor (e.g., loneliness), is not sampled, another predictor may be selected and interactions that may have been missed can be identified.

Although random forests produce too many trees to make sense of visually, they provide indices of variable importance across the “forest” of trees (Strobl et al., 2009a). In this study, we use the cforests procedure in the party package in R (Strobl, Boulesteix, Zeileis, & Hothorn, 2007; Strobl, Boulesteix, Kneib, Augustin, & Zeileis, 2008)2. To address sample size and correlated predictors, we follow recommendations to vary the starting seed and to examine the conditional permutation importance (Strobl, Hothorn, & Zeileis, 2009; Strobl et al., 2009a). If we were to randomly permute (i.e., shuffle or reorder) the values on a predictor such as financial strain, the original relationship between financial concerns and perceived stress would be broken. If the original version of financial strain was related to perceived stress, using this permuted-version along with the other original (i.e., non-permuted) contextual predictors3 would result in much worse prediction. Poor predictors, on the other hand, would only be subject to a small random decrease and in some cases a small increase in prediction after permutation. This difference in prediction before and after permuting a predictor is referred to as permutation importance. The relative ranking of predictors in terms of permutation importance should be interpreted rather than their absolute values (Strobl et al., 2009a).

Random forests results provide support for the domains selected as splitters (and their relative importance when both participants and predictors available are varied) as well as a reminder to use caution when interpreting a single tree. Across starting values, loneliness was by far the strongest predictor (Figure 2), meaning that permuting loneliness greatly decreases its predictive utility. It is followed by financial strain, neighborhood strain, and ageism. Chronic and somatic health problems were poor predictors; negative permutation importance values indicate that randomly shuffling responses on these health domains actually increases prediction of perceived stress. Life events also had a small permutation importance value, which means that the life events split in the tree in Figure 1 should be interpreted cautiously as it may be the product of random variation in the dataset. Depending on the sample, this domain may not appear as an important predictor.

Figure 2.

Figure 2

Example conditional permutation importance table. LifeEv: life events, Neighbr: neighborhood strain, FinStr: financial strain, Lonlns: loneliness, ChrHlth: chronic health problems, SomHlth: somatic health problems.

Together, the selected tree and the random forests indicate that loneliness is a particularly important risk domain, but that the impact of loneliness on perceived stress may depend on other, different contextual domains for different people. Although permutation importance suggests that the impact of life events may not replicate, the tree displayed in Figure 1 still provides one possible depiction of the multidimensional relationships between life contexts that may underlie responses – even similar levels of response – on the commonly used predictor perceived stress. One goal of this study was to not only identify which domains are especially related to overall perceived stress, which the random forests results speak to, but to use the trees to profile the people for whom certain situations are particularly risky. Pursuing this idiographic goal, we also examined the interviews of participants in selected nodes to gain another perspective on the contexts of these individuals' lives.

Follow-up: Interviews

Transcribed interviews were sorted based on terminal node membership derived from the optimal tree. A formal qualitative analysis was not conducted for this study; instead, the authors listened to and read the interviews multiple times and used the stories to add additional contextual detail to the other findings. The distribution of interviews across terminal nodes is indicated in Figure 1. Selection criteria for the interview project required endorsing at least 2 life events in the last year; therefore, no interviews appear in Node 8. Of those interviewed, terminal nodes 12 and 7 represent maximal diversity, both in terms of PSS and interacting domains of risk.

Node 12 participants reported low loneliness and financial strain and were nearly 1/2 SD below average on perceived stress. In their interviews, these individuals had social networks of varying sizes, which appeared to fit their needs. These participants described being financially secure and those who were still working were doing so mostly because they enjoyed it or liked the benefits (e.g., discounts at a retail store). An absence of pressure and confidence in being able to manage was audible in some responses, “my goal is to be where I need to be when I need to be there,” and “don't take bad events seriously. Things are going to happen, [either] there's nothing you can do or if [there is something] you can do, you can overcome.” Interview participants reported events ranging from retirement, an incident at work involving police, best friend's death, and surgeries.

Node 7 represents the highest risk group in the sample. In the interviews, social isolation is apparent – the loss of two close friends in the last month, a major break in friendship, considering ending relationship with boyfriend, and reluctance to ask children for help in order not to be a burden were all mentioned. Additionally, one person explained her social isolation, “I have a tendency to hibernate…and because I hibernate, I don't do so good on making friends.” Neighborhood strain, however, is not as clear in the interviews. Although financial strain did not appear as a predictor in this branch of the tree, their financial resources were likely quite limited. More than half of the participants in this node reported an annual income less than $15,000 (Table 3). In some ways it is surprising that more risk domains do not appear in the tree, given the contexts described in the interviews. One woman related the story of a difficult year with many events and strains. Isolated from her children since her divorce over 10 years ago, she fell at home and hurt her wrist which resulted in multiple surgeries and complications. She had to take unpaid leave from work and did not have insurance. Bills accumulated, including taxes which she could not pay and she faced losing her house. Before she could get medical clearance to return to work, she was let go from her position. She said, “I really wanted to commit suicide…I've never had it so bad and been alone so totally…I lost my job, I lost my insurance…my family wasn't there to help me out…losing my house was the final blow.” She recounted that on the night she was going to commit suicide by leaving the car running in the garage, her car was stolen. This insight, albeit extreme, gives us opportunity to better understand the contexts of the lives of highly stressed individuals.

These interviews provide support for the role of chronic difficulties, like financial strain and loneliness, in putting people at-risk for feeling stressed. Further, the role of appraisal may be important to the understanding the difference between objectively identifiable stressors such as those detailed in the interviews and the risk domains appearing in the perceived stress tree. It is not necessarily the case that participants in Nodes 6 and 7 did not have any other strains in their lives beyond loneliness and financial strain, clearly some of them did. Instead, as suggested by permutation importance, the effect of loneliness may be particularly overwhelming and related to how they generally perceive the stressfulness in their lives.

Discussion

The current study provided both idiographic and nomethetic information regarding the life domains that are particularly salient to perceptions of stress in this later life sample. First, at the idiographic level, regression trees were used to identify the contextual features that contributed to perceptions of being stressed and reflected the pathway of splits that led to node membership. The interview data helped to better flesh out these important relations. Second, the use of random forests, a nomethetic approach to understanding contexts important to perceptions of stress, identified risk domains in the analysis that were of greatest importance—in this case, loneliness, followed by financial strain, neighborhood stress and age discrimination. These findings are discussed in turn.

Regression Trees: Idiographic Understanding of Perceived Stress

One way to consider the results of the tree analysis is to consider cumulative and compensatory pathways to perceptions of stress (Seroczynski, et al., 1997; Wallace et al., 2002). Cumulative pathways indicate a compounding of risk or, alternatively, protection through the interactions of multiple domains; whereas compensatory pathways display possible ways in which being of low risk in one domain may offset high risk status on another. For example, the split that lead to terminal Nodes 7 and 8 are good examples of cumulative pathways, in which the experience of each of the domains result in splits that further differentiate the most (Node 7) and the least (Node 8) stressed groups. For the high-stressed group there may be a cumulative negative effect of high loneliness and high neighborhood strain, in addition to the many insights gleaned from the interview data described above. On the least-stressed side (Node 8), low loneliness or high social integration and support, in addition to fewer major life events, related to low levels of perceived stress. It is not surprising that greater severity of stressors accompany higher perceptions of being stressed and vice versa. Of greater interest are the compensatory pathways that reside in the center nodes of Figure 1. In Node 6, participants reported very high loneliness relative to the rest of the sample, but there may have been something protective about their neighborhood situation that resulted in above average levels of perceived stress, but ones that were much lower when compared with Node 7. In other words, something about the neighborhood characteristics (or contexts related to it) were protective in the relationship between loneliness and perceived stress.

Regression tree results also provided interesting examples of equifinality (Nodes 10 and 12; and to a lesser degree Nodes 6 and 11). For the participants in Nodes 10 and 12, perceptions of stress are about ½ SD below the mean of the sample, but the pathways are different. That is, even though individuals report similar levels of perceived stress, the contexts of their lives contribute differently to their feelings of stress. Efforts to intervene at the level of social isolation would likely benefit Node 10 participants more than those in Node 12 who appraise their social opportunities as adequate. Both sets report relatively low financial strain. What might be contributing to adults in Node 12 feeling similarly overwhelmed is the experience of life events in the last year. For example, one interviewee in Node 12 described multiple deaths in his extended family and social network, his wife experienced moderate health concerns, and his sister was in the later stages of Alzheimer's disease. On the other hand, he had good relationships with family and was active with friends (e.g., volunteering, senior men's club), had been pleasantly surprised about how well his investments had done, and felt comfortable in his financial situation. Thus, social support and a lack of financial hardship may help to compensate for life stressors, and shows how individuals with similar levels of stress have different predictive risks.

Also interesting are the somewhat similar mean scores of Nodes 6 and 11. Although Node 6 is clearly higher (average perceived stress 0.64 vs. 0.38 SDs above the mean), both are above the sample average and participants reside in these nodes through a different set of predictive associations—Node 6 through high loneliness and low neighborhood strain and Node 11 through moderate loneliness and high financial strain. For example, interviewees in Node 6 reflect on social losses (death of a spouse or fiancé), whereas subjects in Node 11 reported “losing their nest egg,” and having constant “struggles” or “worries” about finances. Thus in one group, low neighborhood strain may have buffered the negative effects of high loneliness, whereas high financial stress undermined the positive effects of social integration. The tree analysis allows for the identification of these differing reasons for the perceptions of stress and provides valuable insights for prevention and intervention efforts.

Random Forests: Nomethetic Understanding of Perceived Stress

Although the identification of the course by which individuals (or groups of individuals) perceive their lives as stressful is important, there is a benefit to understanding the particular attributes that are generally the best predictors of perceived stress. By varying the sample across many trees in random forests, it is possible to identify predictors that may generalize to other samples. In other words, the conditional permutation importance provides important information for understanding the factors that are most salient to perceived stress, in this case, loneliness. Later life may be a segment of the lifespan that is particularly ripe for changes in social relationships. For example, death of family members or friends limits the number of people in our social networks; changes in health restrict involvement in activities that promote social engagement; and a move to a new residence can make gathering with others more difficult. It should be noted, however, that loneliness is not simply physical isolation; it is more than just time spent alone or the number of contacts that people have with others (Fischer & Phillips, 1982; Hawkley, Burleson, Bernston, & Cacioppo, 2003; Jones, 1982; Wheeler, Reis, & Nezlek, 1983; for review see Heinrich & Guillone, 2006). “Loneliness, … may be best described as an internal assessment of the meaning derived from social exchanges” (p. 5; Russell & Bergeman, 2010), and feelings of loneliness reflect the quantity and/or quality of unmet social needs (Peplau & Perlman, 1982). Much research has shown that social dynamics can serve as both a source of the stress and as a coping or resilience mechanism (Bergeman et al., 2010). It is clear from both the regression tree and random forest analyses, however, that loneliness is especially detrimental, and is related to elevated perceptions of stress.

Limitations

Much discussion in the stress literature has centered on the notions of objective and subjective measures of stress (Monroe, 2008). Perceived stress, the outcome here, is certainly a subjective measure. On the surface, life events checklists and counts of physical symptoms appear to be objective measures of exposure, however, the reporting of these indicators are intertwined with appraisal processes as well (Dohrenwend, 2006; Monroe, 2008) and so may best be described as subjectively-identified stress exposure (Duggal et al., 2000). From this perspective, the measures utilized in the current study document both the self-nominated exposure as well as the perceived severity of conditions across social, health, financial, and neighborhood domains. In addition, the analyses were limited to the stress measures available in the NDHWB, and different trees might result if additional measures were included. On the positive side, this multi-domain subjective data available for these analyses provides important information on the contexts of individual lives and how these components may put people at differential risk for perceiving their lives as stressful.

Recursive partitioning analyses are very sensitive to the characteristics of the sample, both a drawback and benefit of these methods. The random forests analyses produced 500 trees, which displayed many different patterns depending on which participants and predictors were sampled. It is impossible to interpret the specific patterns in this many trees, however, the random forests provide further evidence for the importance of loneliness, financial strain, and neighborhood strain as consistent predictors of perceived stress. The optimal tree examined is one of many possibilities produced by these analyses, but the one with the lowest cross-validated relative error. As a visual depiction, it underscores the intricate and node-specific relationships in the present sample and directs our attention towards the factors that may underlie perceived stress. More importantly, it highlights the varied pathways, delineating possible, but different, contributors to perceived stress. Understanding these differences is fundamental to targeting appropriate interventions. That is, for people like X, we should pay special attention to changing financial circumstances; whereas for person Y, it might be most effective to enhance social integration. It is important to note, however, that although it is tempting to read trees as flow charts that describe processes in action, the causal processes underlying the stress appraisal are complicated and necessitate longitudinal data to examine empirically.

Future Directions & Conclusions

Although perceived stress is commonly employed as a predictor, little research has focused on the contextual stress factors that are related to these perceptions. The current study identified loneliness, neighborhood, and financial strain as particularly important correlates of this often used construct. Beyond simply assessing correlational relationships among these indicators or their individual correlations with perceived stress, the tree analyses provide intricate combinations of contexts that better contribute to our understanding of individual differences. Although someone's level of perceived stress gives information on the extent to which he or she generally feels overwhelmed, the contexts underlying this appraisal may not be the same across people even those with similar levels of perceived stress. In other words, people perceive their lives as stressful for different reasons. Public health interventions often use demographic characteristics to target their interventions. The highest perceived stress in this sample was found among those who reported some of the lowest income, however, some participants in this high stress node also reported mid-range and high incomes. These people, who also report feeling extremely stressed compared to the rest of the sample, would have been missed in an intervention using solely socio-economic status as a targeting tool. Intervention and prevention strategies need to be designed to take this equifinality of pathways into account. Finally, the mixing of quantitative and qualitative data can allow insights into complex developmental phenomena (Yoshikawa, Weisner, Kalil, & Way, 2008). Regression trees, random forests, and the interviews each provided a different vantage point for understanding stress in later life. The result is a rich, albeit exploratory, examination of the antecedents and concomitants of stress.

Acknowledgments

The Notre Dame Study of Health & Well-being is supported by a grant from the National Institute of Aging, 1 R01 AG023571-A1-01 to C. S. Bergeman. Earlier versions of this work were presented at the annual meeting of the Gerontological Society of America, the International Congress of Qualitative Inquiry, and the International Colloquium on Promises and Perils of Exploratory Data Mining in Predicting Risk. We thank Scott Maxwell, Jack McArdle, and the members of the colloquium for their helpful feedback.

Footnotes

1

As a rule of thumb, listwise deletion is an option for handling missing data in cases where fewer than 5% of the total observations would be excluded (Berk, 2008). CART Pro software manages missing data by using surrogate predictors as splitters (Steinberg & Golovyna, 2006). For a given node, surrogates closely mimic the primary splitter (i.e., node size, composition, which cases are assigned to left and right child nodes) and function as back-up splitters in cases for which the primary splitter is missing. This can be a useful option in some applications, but because the focus of the current study was in determining the relationships between specific contextual risk variables and perceived stress in this sample, we chose not to rely on surrogates and thus limited the available data to complete cases.

2

A constraint of at least 10 participants per terminal node, the default setting for bagging (subsampling .632*n), number of trees (ntree=500), and sampling from the predictors (mtry=5) were specified.

3

For technical details on permutation in the case of continuous and correlated predictors, see Strobl et al., 2009b.

Publisher's Disclaimer: The following manuscript is the final accepted manuscript. It has not been subjected to the final copyediting, fact-checking, and proofreading required for formal publication. It is not the definitive, publisher-authenticated version. The American Psychological Association and its Council of Editors disclaim any responsibility or liabilities for errors or omissions of this manuscript version, any version derived from this manuscript by NIH, or other third parties. The published version is available at www.apa.org/pubs/journals/pag

References

  1. Aanes MM, Middlemark MB, Hetland J. Interpersonal stress and poor health: The mediating role of loneliness. European Psychologist. 2010;15(1):3–11. [Google Scholar]
  2. Aldwin CM. The Elders Life Stress Inventory: Egocentric and non-egocentric stress. In: Stephens AMP, Crowther JH, Hobfoll SE, Tennenbaum DL, editors. Stress and coping late life families. Hemisphere; New York: 1990. pp. 49–69. [Google Scholar]
  3. Barnes LL, Mendes de Leon CM, Lewis TT, Bienias JL, Wilson RS, Evans DA. Perceived discrimination and mortality in a population-based study of older adults. American Journal of Public Health. 2008;98(7):1241–1247. doi: 10.2105/AJPH.2007.114397. [DOI] [PMC free article] [PubMed] [Google Scholar]
  4. Belloc NB, Breslow L, Hochstim JR. Measurement of physical health in a general populations survey. American Journal of Epidemiology. 1971;93:328–336. doi: 10.1093/oxfordjournals.aje.a121265. [DOI] [PubMed] [Google Scholar]
  5. Bergeman CS, Braun ME, Scott SB, Baird BM, Montpetit MA, Ong AD. Effects of daily stress on negative affect: Do social interactions help or hinder? 2010 under review. [Google Scholar]
  6. Berk RA. Statistical learning from a regression perspective. Springer; New York: 2008. [Google Scholar]
  7. Berkman LF, Leo-Summers L, Horwitz RI. Emotional support and survival after myocardial infarction. Annals of Internal Medicine. 1992;117:1003–1009. doi: 10.7326/0003-4819-117-12-1003. [DOI] [PubMed] [Google Scholar]
  8. Brim OG, Baltes PB, Bumpass LL, Cleary PD, Featherman DL, Hazzard WR, Shweder RA. National survey of midlife development in the united states (MIDUS), 1995–1996. Inter-university Consortium for Political and Social Research; Ann Arbor, MI: 2007. [Google Scholar]
  9. Breiman L, Friedman JH, Olshen RA, Stone CJ. Classification and regression trees. Wadsworth; Belmont, CA: 1984. [Google Scholar]
  10. Breiman L. Random forests. Machine Learning. 2001;45:5–32. [Google Scholar]
  11. Brown GW, Harris TO. Social origins of depression: A study of psychiatric disorder in women. Free Press; New York: 1978. [Google Scholar]
  12. Cacioppo JT, Hawkley LC, Berntson GG. The anatomy of loneliness. Current Directions in Psychological Science. 2003;12(3):71–74. [Google Scholar]
  13. Clémence A, Karmaniola A, Green EG, Spini D. Disturbing life events and wellbeing after 80 years of age: A longitudinal comparison of survivors and the deceased over five years. Ageing & Society. 2007;27:195–213. [Google Scholar]
  14. Cohen S. Contrasting the Hassles Scale and the Perceived Stress Scale: Who's really measuring appraised stress? American Psychologist. 1986;41(6):716–718. [Google Scholar]
  15. Cohen S. Aftereffects of stress on human performance and social behavior: A review of research and theory. Psychological Bulletin. 1980;88(1):82–108. [PubMed] [Google Scholar]
  16. Cohen S, Kamarck T, Mermelstein R. A global measure of perceived stress. Journal of Health and Social Behavior. 1983;24(4):385–396. [PubMed] [Google Scholar]
  17. Cohen S, Kessler RC, Gordon LU. Strategies for measuring stress in studies of psychiatric and physical disorders. In: Cohen S, Kessler RC, Gordon LU, editors. Measuring stress: A guide for health and social scientists. Oxford University Press; New York: 1997. pp. 3–26. [Google Scholar]
  18. Cohen S, Williamson GM. Perceived stress in a probability sample of the United States. In: Spacapan S, Oskamp S, editors. The Social Psychology of Health. Sage; Newbury Park, CA: 1988. [Google Scholar]
  19. Cummings EM, Davies PT, Campbell SB. Developmental psychopathology and family process: Theory, research, and clinical implications. The Guilford Press; New York: 2000. [Google Scholar]
  20. Davis MC, Zautra AJ, Smith BW. Chronic pain, stress, and the dynamics of affective differentiation. Journal of Personality. 2004;72(6):1133–1160. doi: 10.1111/j.1467-6494.2004.00293.x. [DOI] [PMC free article] [PubMed] [Google Scholar]
  21. De Beurs E, Beekman A, Geerlings S, Deeg D, Van Dyck R, Van Tilburg W. On becoming depressed or anxious in late life: Similar vulnerability factors but different effects of stressful life events. The British Journal of Psychiatry. 2001;179:426–431. doi: 10.1192/bjp.179.5.426. [DOI] [PubMed] [Google Scholar]
  22. Doane LD, Adam EK. Loneliness and cortisol: Momentary, day-to-day, and trait associations. Psychoneuroendocrinology. 2010;35:430–441. doi: 10.1016/j.psyneuen.2009.08.005. [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. Dohrenwend BP. Inventorying stressful life events as risk factors for psychopathology: Toward resolution of the problem of intracategory variability. Psychological Bulletin. 2006;132(3):477–495. doi: 10.1037/0033-2909.132.3.477. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Duggal S, Malkoff-Schwartz S. Assessment of life stress in adolescents: Self-report versus interview methods. Journal of the American Academy of Child & Adolescent Psychiatry. 2000;39(4):445–453. doi: 10.1097/00004583-200004000-00013. [DOI] [PubMed] [Google Scholar]
  25. Fischer CS, Phillips SL. Who is alone? Social characteristics of people with small networks. In: Peplau LA, Perlman D, editors. Loneliness: A sourcebook of current theory, research and therapy. Wiley; New York: 1982. pp. 21–39. [Google Scholar]
  26. Glaser R, Kiecolt-Glaser JK, Speicher CE, Holliday JE. Stress, loneliness, and changes in herpesvirus latency. Journal of Behavioral Medicine. 1985;8(3):249–260. doi: 10.1007/BF00870312. [DOI] [PubMed] [Google Scholar]
  27. Golden-Kruetz DM, Thornton LM, Wells-Di Gregorio S, Frierson GM, Jim HS, Carpenter KM, Anderson BL. Traumatic stress, perceived global stress, and life events: prospectively predicting quality of life in breast cancer patients. Health Psychology. 2005;24(3):288–296. doi: 10.1037/0278-6133.24.3.288. [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. Gruenewald TL, Mroczek DK, Ryff CD, Singer BH. Diverse pathways to positive and negative affect in adulthood and later life: An integrative approach using recursive partitioning. Developmental Psychology. 2008;44(2):330–343. doi: 10.1037/0012-1649.44.2.330. [DOI] [PubMed] [Google Scholar]
  29. Gruenewald TL, Seeman TE, Ryff CD, Karlamangla AS, Singer BH. Combinations of biomarkers predictive of later life mortality. Proceedings of the National Academy of Sciences of the United States of America. 2006;103(38):14158–14163. doi: 10.1073/pnas.0606215103. [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Hastie T, Tibshirani R, Friedman J. Elements of statistical learning: Data mining, inference, and prediction. Springer; New York: [Google Scholar]
  31. Hardy SE, Concato J, Gill TM. Stressful life events among community-living older persons. Journal of General Internal Medicine. 2002;17(11):841–847. doi: 10.1046/j.1525-1497.2002.20105.x. [DOI] [PMC free article] [PubMed] [Google Scholar]
  32. Hawkley LC, Burleson MH, Berntson GG, Cacioppo JT. Loneliness in everyday life: Cardiovascular activity, psychosocial context, and health behaviors. Journal of Personality and Social Psychology. 2003;85(1):105–120. doi: 10.1037/0022-3514.85.1.105. [DOI] [PubMed] [Google Scholar]
  33. Hawkley LC, Masi CM, Berry JD, Cacioppo JT. Loneliness is a unique predictor of age-related differences in systolic blood pressure. Psychology and Aging. 2006;21(1):152–164. doi: 10.1037/0882-7974.21.1.152. [DOI] [PubMed] [Google Scholar]
  34. Hawkley LC, Thisted RA, Masi CM, Cacioppo JT. Loneliness predicts increased blood pressure: 5-year cross-lagged analyses in middle-aged and older adults. Psychology and Aging. 2010;25(1):132–141. doi: 10.1037/a0017805. [DOI] [PMC free article] [PubMed] [Google Scholar]
  35. Heinrick LM, Guillone E. The clinical significance of loneliness: A literature review. Clinical Psychology Review. 2006;26:695–718. doi: 10.1016/j.cpr.2006.04.002. [DOI] [PubMed] [Google Scholar]
  36. Holmes TH, Rahe TH. The social readjustment rating scale. Journal of Psychosomatic Research. 1967;11(2):213–218. doi: 10.1016/0022-3999(67)90010-4. [DOI] [PubMed] [Google Scholar]
  37. House JS, Landis KR, Umberson D. Social relationships and health. Science. 1988;241:540–545. doi: 10.1126/science.3399889. [DOI] [PubMed] [Google Scholar]
  38. Isaacowitz DM, Seligman ME. Is pessimism a risk factor for depressive mood among community-dwelling older adults? Behaviour Research and Therapy. 2001;39(3):255–272. doi: 10.1016/s0005-7967(99)00178-3. [DOI] [PubMed] [Google Scholar]
  39. Jones WH. Loneliness and social behavior. In: Peplau LA, Perlman D, editors. Loneliness: A sourcebook of current theory, research and therapy. Wiley; New York: 1982. pp. 238–252. [Google Scholar]
  40. Kahn JR, Pearlin LI. Financial strain over the life course and health among older adults. Journal of Health and Social Behavior. 2006;47(1):17–31. doi: 10.1177/002214650604700102. [DOI] [PubMed] [Google Scholar]
  41. Kershaw KN, Mezuk B, Abdou CM, Rafferty JA, Jackson JS. Socioeconomic position, health behaviors, and c-reactive protein: A moderated-mediation analysis. Health Psychology. 2010;29(3):307–316. doi: 10.1037/a0019286. [DOI] [PMC free article] [PubMed] [Google Scholar]
  42. Kessing LV, Agerbo E, Mortensen PB. Does the impact of major stressful life events on the risk of developing depression change throughout life? Psychological Medicine. 2003;33(7):1177–1184. doi: 10.1017/s0033291703007852. [DOI] [PubMed] [Google Scholar]
  43. Keyes CLM. Social well-being. Social Psychology Quarterly. 1998;61(2):121–140. [Google Scholar]
  44. Kiecolt-Glaser JK, Garner W, Speicher C, Penn GM, Holliday J, Glaser R. Psychosocial modifiers of immunocompetence in medical students. Psychosomatic Medicine. 1984;46(1):7–14. doi: 10.1097/00006842-198401000-00003. [DOI] [PubMed] [Google Scholar]
  45. Kiecolt-Glaser JK, Ricker D, George J, Messick G, Speicher CE, Garner W, Glaser R. Urinary cortisol levels, cellular immunocompetency, and loneliness in psychiatric inpatients. Psychosomatic Medicine. 1984;46(1):15–23. doi: 10.1097/00006842-198401000-00004. [DOI] [PubMed] [Google Scholar]
  46. Krause N. Neighborhood deterioration, religious coping, and changes in health during late life. The Gerontologist. 1998;38(6):653–664. doi: 10.1093/geront/38.6.653. [DOI] [PubMed] [Google Scholar]
  47. Lagana L, Reger SL. A pilot study on perceived stress and PTSD symptomatology in relation to four dimensions of older women's physical health. Aging and Mental Health. 2009;13(6):885–893. doi: 10.1080/13607860903004023. [DOI] [PMC free article] [PubMed] [Google Scholar]
  48. Lazarus RS. Psychological stress and the coping process. McGraw-Hill; New York: 1966. [Google Scholar]
  49. Lazarus RS, Folkman S. Stress, appraisal, and coping. Springer; New York: 1984. [Google Scholar]
  50. Levy B. Stereotype embodiment: A psychosocial approach to aging. Current Directions in Psychological Science. 2009;18(6):332–336. doi: 10.1111/j.1467-8721.2009.01662.x. [DOI] [PMC free article] [PubMed] [Google Scholar]
  51. Levy B, Ashman O, Dror I. To be or not to be: The effects of aging stereotypes on the will to live. Omega. 2000;40(3):409–420. doi: 10.2190/y2ge-bvyq-nf0e-83vr. [DOI] [PubMed] [Google Scholar]
  52. Levy BR, Hausdorff JM, Hencke R, Wei JY. Reducing cardiovascular stress with positive self-stereotypes of aging. Journal of Gerontology: Psychological Sciences. 2000;55B(4):P205–P203. doi: 10.1093/geronb/55.4.p205. [DOI] [PubMed] [Google Scholar]
  53. Levy BR, Ryall AL, Pilver CE, Sheridan PL, Wei JY, Hausdorff JM. Influence of African American elders' age stereotypes on their cardiovascular response to stress. Anxiety, Stress, & Coping. 2008;21(1):85–93. doi: 10.1080/10615800701727793. [DOI] [PubMed] [Google Scholar]
  54. Levy BR, Slade MD, Kunkel SR, Kasl SV. Longevity increased by positive self-perceptions of aging. Journal of Personality and Social Psychology. 2002;83(2):261–270. doi: 10.1037//0022-3514.83.2.261. [DOI] [PubMed] [Google Scholar]
  55. Linn MW, Sandifer R, Stein S. Effects of unemployment on mental and physical health. American Journal of Public Health. 1985;75(5):502–506. doi: 10.2105/ajph.75.5.502. [DOI] [PMC free article] [PubMed] [Google Scholar]
  56. Maxwell CJ, Kang J, Walker JD, Zhang JX, Hogan DB, Feeny DH, Wodchis WP. Sex differences in the contribution of social and clinical factors to the Health Utilities Index Mark 2 measure of health-related quality of life in older home care clients. Health and Quality of Life Outcomes. 2009;7(80):1–11. doi: 10.1186/1477-7525-7-80. [DOI] [PMC free article] [PubMed] [Google Scholar]
  57. Merkle EC, Schaffer VA. Binary recursive partitioning as an alternative to regression in Psychology. British Journal of Mathematical and Statistical Psychology. doi: 10.1348/000711010X503129. in press. [DOI] [PubMed] [Google Scholar]
  58. McDade TW, Hawkley LC, Cacioppo JT. Psychosocial and behavioral predictors of inflammation in middle-aged and older adults: The Chicago health, aging, and social relations study. Psychosomatic Medicine. 2006;68:376–381. doi: 10.1097/01.psy.0000221371.43607.64. [DOI] [PubMed] [Google Scholar]
  59. Mitsonis CI, Potagas C, Zervas I, Sfagos K. The effects of stressful life events on the course of multiple sclerosis: A review. International Journal of Neuroscience. 2009;119(3):315–335. doi: 10.1080/00207450802480192. [DOI] [PubMed] [Google Scholar]
  60. Monroe SM. Modern approaches to conceptualizing and measuring human life stress. Annual Review of Clinical Psychology. 2008;4:33–52. doi: 10.1146/annurev.clinpsy.4.022007.141207. [DOI] [PubMed] [Google Scholar]
  61. Murray TL, Daniels MH, Murray CE. Differentiation of self, perceived stress, and symptom severity among patients with fibromyalgia syndrome. Families, Systems, & Health. 2006;24(2):147–159. [Google Scholar]
  62. Neugarten BL. Time, age, and the life cycle. The American Journal of Psychiatry. 1979;136(7):887–894. doi: 10.1176/ajp.136.7.887. [DOI] [PubMed] [Google Scholar]
  63. Oh J, Kim S. Aging, neighborhood attachment, and fear of crime: Testing reciprocal effects. Journal of Community Psychology. 2009;37(1):21–40. [Google Scholar]
  64. Ong AD, Bergeman CS, Boker SM. Resilience comes of age: Defining features in later adulthood. Journal of Personality. 2009;6:1777–1804. doi: 10.1111/j.1467-6494.2009.00600.x. [DOI] [PMC free article] [PubMed] [Google Scholar]
  65. Palmore E. The ageism survey: First findings. Gerontologist. 2001;41(5):572–575. doi: 10.1093/geront/41.5.572. [DOI] [PubMed] [Google Scholar]
  66. Palmore E, Cleveland WP, Nowlin JB, Ramm D, Siegler IC. Stress and adaptation in later life. Journals of Gerontology. 1979;34(6):841–851. doi: 10.1093/geronj/34.6.841. [DOI] [PubMed] [Google Scholar]
  67. Pearlin LI, Skaff MM. Stress and the life course. The Gerontologist. 1996;36(2):239–247. doi: 10.1093/geront/36.2.239. [DOI] [PubMed] [Google Scholar]
  68. Penninx BW, van Tilburg T, Kriegsman DM, Deeg DJ, Boeke AJ, an Eijk J. Effects of social support and personal coping resources on mortality in older age: The longitudinal aging study Amsterdam. American Journal of Epidmiology. 1997;146(6):510–519. doi: 10.1093/oxfordjournals.aje.a009305. [DOI] [PubMed] [Google Scholar]
  69. Peplau LA, Perlman D, editors. Loneliness: A sourcebook of current theory, research, and therapy. John Wiley & Sons; Hoboken, NJ: 1982. [Google Scholar]
  70. Pinquart M, Silbereisen RK, Korner A. Coping with family demands under difficult economic conditions: Associations with depressive symptoms. Swiss Journal of Psychology. 2010;69(1):53–63. [Google Scholar]
  71. Rafanelli C, Roncuzzi R, Milaneschi Y, Tomba E, Colistro MC, Pancaldi LG, Di Pasquale G. Stressful life events, depression, and demoralization as risk factors for coronary heart disease. Psychotherapy & Psychosomatics. 2005;74(3):179–184. doi: 10.1159/000084003. [DOI] [PubMed] [Google Scholar]
  72. Rosengren A, Hawken S, Ounpuu S, Sliwa K, Zubaid M, Blackett KN, Yusuf S. Association of psychosocial risk factors with risk of acute myocardial infarction in 11 119 cases and 13 648 controls from 52 countries (the INTERHEART study): Case-control study. Lancet. 2004;364:953–962. doi: 10.1016/S0140-6736(04)17019-0. [DOI] [PubMed] [Google Scholar]
  73. Russell A, Bergeman CS. Daily social exchanges and affect in middle and later adulthood: The impact of loneliness and age. Psychology and Aging. 2010 doi: 10.2190/AG.74.4.c. [DOI] [PMC free article] [PubMed] [Google Scholar]
  74. Russell D, Peplau LA, Curtona C. The revised UCLA loneliness scale: Concurrent and discriminant validity evidence. Journal of Personality and Social Psychology. 1980;39(3):472–480. doi: 10.1037//0022-3514.39.3.472. [DOI] [PubMed] [Google Scholar]
  75. Ryff CD, Magee WJ, Kling KC, Wing EH. Forging macro-micro linkages in the study of psychological well-being. In: Ryff CD, Marshall VW, editors. The self and society in aging process. Springer-Verlang; New York: 1999. pp. 247–278. [Google Scholar]
  76. Sanchez Palacios C, Trianes Torres MV, Blanca Mena MJ. Negative aging stereotypes and their relation with psychosocial variables in the elderly population. Archives of Gerontology and Geriatrics. 2008;48:385–390. doi: 10.1016/j.archger.2008.03.007. [DOI] [PubMed] [Google Scholar]
  77. Segerstrom SC, Miller GE. Psychological stress and the human immune system: A meta-analytic study of 30 years of inquiry. Psychological Bulletin. 2004;130(4):601–630. doi: 10.1037/0033-2909.130.4.601. [DOI] [PMC free article] [PubMed] [Google Scholar]
  78. Sematter-Bagnoud L, Karmaniola A, Santos-Eggimann B. Adverse life events among community-dwelling persons aged 65–70 years: Gender differences in occurrence and perceived psychological consequences. Social Psychiatry and Psychiatric Epidemiology. 2010;45(1):9–16. doi: 10.1007/s00127-009-0035-3. [DOI] [PubMed] [Google Scholar]
  79. Seroczynski AD, Cole DA, Maxwell SE. Cumulative and compensatory effects of competence and incompetence on depressive symptoms in children. Journal of Abnormal Psychology. 1997;106(4):586–597. doi: 10.1037//0021-843x.106.4.586. [DOI] [PubMed] [Google Scholar]
  80. Steinberg D, Colla P. CART: Classification and regression trees. Salford Systems; San Diego, CA: 1997. [Google Scholar]
  81. Steinberg D, Golovyna M. CART 6.0 User's Manual. Salford Systems; San Diego, CA: 2006. [Google Scholar]
  82. Steptoe A, Owen N, Kunz-Ebrecht SR, Brydan L. Loneliness and neuroendocrine, cardiovascular, and inflammatory stress responses in middle-aged men and women. Psychoneuroendocrinology. 2004;29:593–611. doi: 10.1016/S0306-4530(03)00086-6. [DOI] [PubMed] [Google Scholar]
  83. Strobl C, Boulesteix A, Kneib T, Augustin T, Zeileis A. Conditional variable importance for random forests. BMC Bioinformatics. 2008;9(307) doi: 10.1186/1471-2105-9-307. Retrieved from http://www.biomedcentral.com/1471-2105/9/307. [DOI] [PMC free article] [PubMed] [Google Scholar]
  84. Strobl C, Boulesteix A, Zeileis A, Hothorn T. Bias in random forest variable importance measures: Illustrations, sources and a solution. BMC Bioinformatics. 2007;8(25) doi: 10.1186/1471-2105-8-25. Retrieved from http://www.biomedcentral.com/1471-2105/8/25. [DOI] [PMC free article] [PubMed] [Google Scholar]
  85. Strobl C, Hothorn T, Zeileis A. Party on! A new, conditional variable-importance measure for Random Forests available in the party package. R Journal. 2009a;1/2:14–17. [Google Scholar]
  86. Strobl C, Malley J, Tutz G. An introduction to recursive partitioning: Rationale, application, and characteristics of classification and regression trees, bagging, and random forests. 2009b;14:323–348. doi: 10.1037/a0016973. [DOI] [PMC free article] [PubMed] [Google Scholar]
  87. Strodl E, Kenardy J, Aroney C. Perceived stress as a predictor of the self-reported new diagnosis of symptomatic CHD in older women. International Journal of Behavioral Medicine. 2003;10(3):205–220. doi: 10.1207/s15327558ijbm1003_02. [DOI] [PubMed] [Google Scholar]
  88. Tak SH, Hong SH, Kenneday R. Daily stress in elders with arthritis. Nursing and Health Sciences. 2007;9:29–33. doi: 10.1111/j.1442-2018.2007.00301.x. [DOI] [PubMed] [Google Scholar]
  89. Tucker-Seeley RD, Li Y, Subramanian SV, Sorensen G. Financial hardship and mortality among older adults using the 1996–2004 Health and Retirement Study. Annals of Epidemiology. 2009;19(12):850–857. doi: 10.1016/j.annepidem.2009.08.003. [DOI] [PMC free article] [PubMed] [Google Scholar]
  90. VonDras DD, Powless MR, Olson AK, Wheeler D, Snudden AL. Differential effects of everyday stress on the episodic memory test performances of young, mid-life, and older adults. Aging & Mental Health. 2005;9(1):60–70. doi: 10.1080/13607860412331323782. [DOI] [PubMed] [Google Scholar]
  91. Wallace KA, Bergeman CS, Maxwell SE. Predicting well-being outcomes in later life: An application of classification and regression tree (CART) analysis. In: Shohov SP, editor. Advances in psychology research. Nova Science Publishers; Huntington, NY: 2002. pp. 71–92. [Google Scholar]
  92. Wethington E, Brown GW, Kessler RC. Interview measurement of stressful life events. In: Cohen S, Kessler R, Gordon LU, editors. Measuring stress: A guide for health and social scientists. Oxford University Press; New York: 1997. pp. 59–97. [Google Scholar]
  93. Wigers SH. Figromyalgia outcome: The predictive values of symptom duration, physical activity, disability pension, and critical life events—A 4.5 year prospective study. Journal of Psychosomatic Research. 1996;41(3):235–243. doi: 10.1016/0022-3999(96)00070-0. [DOI] [PubMed] [Google Scholar]
  94. Wheeler L, Reis H, Nezlek JB. Loneliness, social interaction, and sex roles. Journal of Personality and Social Psychology. 1983;45(4):943–953. doi: 10.1037//0022-3514.45.4.943. [DOI] [PubMed] [Google Scholar]
  95. Yen IH, Michael YL, Perdue L. Neighborhood environment in studies of health of older adults: A systematic review. American Journal of Preventive Medicine. 2009;37(5):455–463. doi: 10.1016/j.amepre.2009.06.022. [DOI] [PMC free article] [PubMed] [Google Scholar]
  96. Yoshikawa H, Weisner TS, Kalil A, Way N. Mixing qualitative and quantitative research in developmental science: Uses and methodological choices. Developmental Psychology. 2008;44(2):344–354. doi: 10.1037/0012-1649.44.2.344. [DOI] [PubMed] [Google Scholar]

RESOURCES