UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Loneliness as a risk factor for mortality and morbidity Patterson, Andrew C 2008

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
24-ubc_2008_fall_patterson_andrew.pdf [ 499.68kB ]
Metadata
JSON: 24-1.0066582.json
JSON-LD: 24-1.0066582-ld.json
RDF/XML (Pretty): 24-1.0066582-rdf.xml
RDF/JSON: 24-1.0066582-rdf.json
Turtle: 24-1.0066582-turtle.txt
N-Triples: 24-1.0066582-rdf-ntriples.txt
Original Record: 24-1.0066582-source.json
Full Text
24-1.0066582-fulltext.txt
Citation
24-1.0066582.ris

Full Text

LONELINESS AS A RISK FACTOR FOR MORTALITY AND MORBIDITY  by ANDREW C. PATTERSON  A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF ARTS in THE FACULTY OF GRADUATE STUDIES (Family Studies)  THE UNIVERSITY OF BRITISH COLUMBIA (Vancouver)  August 2008  © Andrew C. Patterson, 2008  Abstract Studies over the past couple of decades have depicted loneliness as a significant concern to physical health, although its meaning for overall health outcomes is still unclear. The precise impact of loneliness on life expectancy and on specific disease processes remains unknown. With regression modeling techniques, this thesis uses data from the Alameda County Health and Ways of Living Study to characterize the impact of loneliness on self-rated health, mortality, and fatalities from specific diseases. A key hypothesis is that loneliness as a health problem hinges on its persistence over time. This hypothesis is also tested by examining the reliability of the loneliness measure across the full 34 years of the survey. A second test is to examine its interplay with marital status as a mutable social circumstance. Results show that loneliness is a risk factor for poor selfrated health, non-ischemic cardiovascular diseases, cerebrovascular diseases, infections, and overall mortality. Results also show that loneliness need not be a stable problem across the life span in order to pose health risks. The reliability of the loneliness measure fades across time and levels of loneliness also vary with changes in marital status. Loneliness did not clearly mediate the impact of marital status on self-rated health, mortality, or specific causes of death.  ii  TABLE OF CONTENTS Abstract .......................................................................................................................... ii Table of Contents ........................................................................................................... iii List of Tables ................................................................................................................. iv List of Figures ................................................................................................................ v Acknowledgements ........................................................................................................ vi Dedication ...................................................................................................................... vii CHAPTER I Introduction ............................................................................................. 1 CHAPTER II Review of the Literature ........................................................................ 1 2.1. Mortality, Morbidity, and Loneliness ........................................................ 3 2.2. The Stability of Loneliness ......................................................................... 8 2.3. Marital status ............................................................................................... 12 2.4. Summary of Research Questions and Hypotheses ...................................... 15 2.5. Analytical Strategy ...................................................................................... 17 CHAPTER III Data ...................................................................................................... 19 3.1. Validation of the loneliness measure .......................................................... 23 3.2. Data Preparation and Exploration ............................................................... 28 CHAPTER IV Methods ............................................................................................... 38 CHAPTER V Results ................................................................................................... 43 5.1. Phase 1: The Stability of Loneliness ...................................................................... 5.2. Phase 2: Self-reported Health in 1965 .................................................................... 5.3. Phase 3: Loneliness and Mortality .......................................................................... 5.4. Phase 4: Exploring Particular Fatalities ..................................................................  43 50 55 64  CHAPTER VI Discussion ............................................................................................ 68 6.1. Review ..................................................................................................................... 68 6.2. Reflections ............................................................................................................... 73 References ...................................................................................................................... 91  iii  List of Tables Table 1. Responses for Loneliness and Marital Status .................................................. 81 Table 2. Sociodemographic Information for Alameda County Residents in 1965 ........ 82 Table 3. Kendall's Tau-b Coefficients for Reports of Loneliness .................................. 84 Table 4. Change in Loneliness Over Time ..................................................................... 84 Table 5. Negative-Log Ordinal Regression Models Predicting Loneliness at Time 2 .......................................................................................................... 85 Table 6. Logistic Regression Models Predicting Status as Married at Time 2 .............. 85 Table 7. Multinomial Regression Models Predicting Marital Change .......................... 86 Table 8. Sample Sizes for Models Testing Social Causation Versus Social Selection ................................................................................................ 87 Table 9. Coefficients for Ordinal Probit Regression Models Gauging Self-Rated Health in 1965 ................................................................................ 87 Table 10. Odds Ratios for Logistic Regression Models Predicting All-Cause Mortality ....................................................................................... 88 Table 11. Odds Ratios for Models Restricted to Person Years Within 5 Years of the Most Recent Wave ............................................................................... 88 Table 12. Odds Ratios for Models Testing the Persistence of Loneliness ..................... 89 Table 13. Odds Ratios for Models Examining Married Status, Gender, and Loneliness ................................................................................................ 89 Table 14. Odds Ratios for Models Predicting Particular Causes of Death .................... 90  iv  List of Figures Figure 1. Loneliness by Age .......................................................................................... 80  v  Acknowledgements I give my enthusiastic gratitude first to Gerry Veenstra for his mentorship during the last year. I feel I have accomplished much under his guidance and that working with Dr. Veenstra was an important step towards fashioning myself into a capable researcher. My thanks also go to Dan Perlman, who offered me the opportunity to pursue this work at the University of British Columbia and who opened many doors with his patient guidance. Dr. Perlman is a world-renowned scholar and I feel quite privileged to have had the experience of working with him. I thank Nathan Lauster as well. His input was not only valuable to this thesis but also to informing and elaborating my interests in sophisticated modeling techniques. Finally, as this area of research was challenging in the emotional sense as well as the intellectual, I express my gratitude to my family for their love and support while I grappled with this thesis. For even the researcher (or at least, this researcher) cannot adequately study social isolation while condemned to this very affliction.  vi  For my beloved twin brother Ben, who was with me even in the very, very beginning.  vii  I. Introduction Within the past two decades, researchers have come to regard loneliness as a serious health issue. Loneliness appears to alter immune and cardiovascular functioning in direct ways and may even accelerate the bodily deterioration that accompanies age. The exact nature and degree with which loneliness compromises health remains unclear, however. Loneliness is associated with a very long list of mental and physical ailments, but it is altogether unclear that it is the deciding factor leading to those ailments. On the other hand, significant progress has been made. Hawkley and Caccioppo (2003; 2006) have offered a hypothesis that explains possible reasons for the association between loneliness and specific causes of disease and death. They argued that loneliness is a subtle stressor that contributes to the wear and tear of the body, accelerating the ageing process. An important component to this argument is that loneliness is also a persistent problem across the life span. Few studies have used prospective data, however, which has hindered efforts to determine whether loneliness has a causal effect on health. Using both cross-sectional and longitudinal data, the study discussed herein is an effort to amend these gaps in the literature. Featured below is a thorough analysis that attempts to better characterize the position of loneliness as a health risk by bringing into focus mortality and specific morbidities. The important contention that loneliness is a health risk by way of being persistent problem over time is also examined. II. Review of the Literature The first known appearance of the word “lonely” was in Shakespeare’s play, Coriolanus, in which the eponymous character compared himself to a dragon that was “feared and talked of more than seen” (Simpson & Weiner, 1989, p. 1122). For the next  1  two centuries, the word merely referred to objective solitude and did not necessarily carry negative implications. By the early nineteenth century, however, “lonely” had assumed a new meaning. It is as though English speakers came to appreciate and relate to the perspective of Coriolanus’s dragon, and the word expressed empathy for this perspective. To be feared, shunned, alienated, or otherwise isolated from others carried saddening ramifications. This rings true today according to findings in social science research. As scholars in the area would agree, feeling “lonely” has troublesome implications. Loneliness is believed to affect millions of North Americans each month (Rubenstein & Shaver, 1982). Surveys conclude that about one in four people have recently felt lonely (e.g., Weiss, 1973; Bradburn, 1969; Patterson, 2007). As such, loneliness is a phenomenon to which many of us can relate, often following the death of a loved one, dissolution of an intimate relationship, or relocation to a far place as for a career change or for matriculation into an educational program. On a societal level, though, the picture is rather gloomy. Demographic trends appear to favor future increases in the incidence of loneliness. Family members represent a critical source of social support, yet family size and cohesion have been in decline over the past two decades (Ernst & Caccioppo, 1999). This has regrettable implications for the elderly of tomorrow, many of whom will have no children or spouses to depend upon. Friendship networks have also shown decline (McPherson, Smith-Lovin, & Brashears, 2006). Loneliness, although a common malady that most of us manage to survive, is far from a trivial problem. Literature reviews routinely associate it with a long list of detrimental effects that extend into both mental and physical health domains (e.g., Miller, Perlman, & Brehm, 2007; Jones & Hebb, 2003; Ernst & Cacioppo, 1999). The breadth of  2  the problems that have been implicated is sobering. Miller et al. (2007), for example, notes that loneliness is associated with juvenile delinquency, poor academic performance, faulty memory, general maladjustment, bulimia nervosa, alcohol abuse, personality disorders, schizophrenia, poor sleep, and stress. Heinrich and Gullone (2006) conclude that loneliness is serious enough to be considered a problem of clinical significance akin to depression, and that satisfaction with social relationships should be routinely explored in clinical settings. Moreover, the troublesome effects of loneliness may sometimes even reach beyond the individual who is lonely. Associations have been found with phenomena as grave as sexually motivated murder (Milsom, Beech, & Webster, 2003) and thoughts of rape and pedophilia (Proulx, McKibben, & Lusignan, 1996). Combined with all other considerations, then, loneliness is a public policy issue as well as one of a personal nature. 2.1. Mortality, Morbidity, and Loneliness Social epidemiologists concerned with mortality have understood the importance of personal relationships for some time now. In a classic study, Berkman and Syme (1979) established a correspondence between all-cause mortality and a lack of close personal relationships. Using an index gauging size and composition of the social networks of respondents, they found higher mortality rates among those who were less connected to others. This association was graded: generally, the fewer the kith and kin, the higher the likelihood of death. In another oft-cited article that followed up on Berkman and Syme’s work, House, Landis, and Umbarson (1988) argued that social isolation is such a significant risk factor that it could be commensurate with health hazards like smoking and obesity.  3  Still, Berkman and Syme (1979) and others (e.g., Uchino, 2004) have noted that the precise mechanisms linking social isolation to mortality and disease remain poorly understood. Some scholars have proffered social support as a phenomenon that may fill this explanatory gap, wherein significant others presumably interact in a way that protects against illness, for example, by encouraging health-promoting behaviors. The concept of social support has frustrated researchers, however. Callaghan and Morrissey (1993) impugned social support as a vague construct with little consensus among scholars on its meaning and possessing substantial methodological problems. Wethington and Kessler (1986) made some headway by offering a distinction between perceived support and actual support received, arguing that the former is more important than the latter when predicting adjustment to stressful life events. Although this development did not seem to pave the way toward a clearer conception of social support1, the distinction has been useful. The precise means by which the objective presence or absence of close relationships affects health remain enigmatic (e.g., Uchino, 2004), but scholars have created abundant literature connecting subjective aspects of personal relationships to health. Of these, a particularly large amount of attention has been devoted to studies of loneliness. The literature on loneliness has framed it as a significant complication for both mental and physical health. Loneliness is believed to correlate with poor self-rated health (Mullins, Smith, Colquitt, & Mushel, 1996), but it may even impact life expectancy. An association with suicide has been documented (Conroy & Smith, 1983; Peck, 1983; Kidd,  1  Measurement could easily be a determining factor here. Wethington and Kessler (1986) assessed perceived social support through agreement with the single statement, “These days I really don't know who I can count on for help,” and acknowledged that this was a weak measure of social support. Moreover, objective support can manifest itself in many ways throughout a given day.  4  2004) while less obvious pathways to mortality are apparent via other health measures as well, all making loneliness an important epidemiological concern. The most recent standing hypothesis has been that loneliness impacts health over the long term by affecting autonomic, endocrine, and immune functioning (Hawkley & Cacioppo, 2002) in a way that leads to faster aging and bodily deterioration over time (Hawkley & Cacioppo, 2007). Age, then, is presumed to moderate the effect of loneliness on health, with older adults who are lonely bearing the brunt with respect to their health. This seems concordant with other studies that have targeted older people. Cutrona, Cutrona, Russell, de la Mora, and Wallace (1997) used survival analysis to examine survival in the community versus entry into retirement homes among rural Iowans over a four-year period. Although they did not take specific health outcomes into account, they did find that loneliness coincided with entry into retirement homes, even after controlling for age. This complements the notion that loneliness accelerates the bodily deterioration that comes with age. More specific age-related problems are also implicated, though. Wilson et al. (2007) observed in a longitudinal study that loneliness was predictive of deteriorating cognitive functioning akin to dementia in old age. They observed the symptoms of Alzheimer’s disease in particular, although they did not see more cerebral plaques among the lonely subjects as would be expected in this case. Parallel to the apparent effects of other distressful emotions (Kubzansky & Kawachi, 2000), accumulating evidence shows detrimental impacts of loneliness on both the cardiovascular and the immune system. Studies have demonstrated associations with weakened immune defense against viruses (Dixon, Cruess, Kilbourn, Klimas, & Fletcher, 2001; Pressman et al., 2005), higher systolic blood pressure (Hawkley, Masi, Berry, &  5  Cacioppo, 2006), and lower cardiac output (Hawkley, Burleson, Berntson, & Cacioppo, 2003). Loneliness also seems to be a problem of its own with respect to health. In contrast with a unidimensional model of stress, Hawkley et al. (2006) found that loneliness impacted cardiovascular functioning in unique ways when compared with other troubling emotional states. Studies show that these effects may carry over into chronic illnesses and mortality via those illnesses. At least one study (Sorkin, Rook, & Lu, 2002) demonstrated that loneliness corresponds with an increased risk of having a heart condition, and another (Herlitz et al., 1998) showed a negative association with recovery from coronary artery bypass surgery. The case for loneliness as a risk factor for mortality via a compromised immune system does not seem quite as consistent. In their earlier review, Hawkley and Cacioppo (2003) regarded as tenuous the evidence linking loneliness to changes in levels of stress hormones. However in their later review (Hawkley & Caccioppo, 2007) they cited more recent studies supporting this link, especially among older adults. They postulated that loneliness may encourage a consistently overactive hormonal response to stress, which could either accelerate wear and tear of the body directly or alter the regulation of inflammation and thus impair physiological resilience. On the other hand, immune hormones may not be the issue of interest. Cole et al. (2007) found that gene expression from white blood cells was different among a small sample of chronically lonely subjects when compared to a group of non-lonely subjects. These differences occurred irrespective of cortisol levels, which was a counterintuitive result for the research team. Many of the genes performed functions specifically germane to immune reactions while others played roles in the life and death of those blood cells.  6  Loneliness may therefore change the way white blood cells behave and thus their adequacy in protecting against foreign bodies, irrespective of stress response. Granted, linking behavioral variables to a disease via immunity has been difficult in general (Brannon & Feist, 2007), and this seems to be true in the case of loneliness. Yet as Cole et al. (2007) found loneliness may compromise immune functioning in a very direct fashion. That said, changes in antibody activity for the worse could result in any number of diseases. It is therefore plausible that mortality in general could be significantly impacted even though no statistically significant results may be seen for any one immune-related disease. According to Hawkley and Cacioppo’s hypothesis (2002; 2007), then, loneliness is a subtle but persistent stressor across the life span that accelerates the wear and tear of the body by taxing the cardiovascular and immune systems. Loneliness may also impact health by impacting health-related behaviors, such as smoking and sleep, the latter of which is important to bodily recuperation. Still, scholars have found the most supportive evidence for poor health stemming from loneliness through changes in immune and especially cardiovascular functioning. Despite their prolific work in this area, Hawkley, Cacioppo, and other authors have often emphasized the need for and lack of prospective designs. Longitudinal methodology would allow pursuit of the gold standard in studies of behavior and health: corroboration of a pathway from the phenomena of interest to time-ordered changes in variables used to assess health (e.g., blood pressure), and then from those to morbidity and mortality (see Forlenza & Baum, 2004). However, the overwhelming majority of studies examining the health effects of loneliness have used cross-sectional designs,  7  leaving very few in the research literature that are prospective in nature. Determining causal pathways has therefore been extremely difficult. This thesis presents an effort to address this concern, starting with a key supposition: loneliness is a problem that is resistant to change. 2.2. The Stability of Loneliness The linchpin to Hawkley and Cacioppo’s model of loneliness as a risk factor for disease is their argument that loneliness is stable over time. Loneliness is perhaps more a question of disposition and less one of life’s routine vicissitudes, at least with respect to health outcomes. Borrowing a term from the literature on obesity, Boomsma, Willemse, Dolan, Hawkley, & Cacioppo (2005) suggested that everyone may have a “setpoint” for loneliness. That is, people may presumably overcome it, but must do so with effort and otherwise tend towards a particular level or setpoint that is unique to each individual. Given no such effort and no extraneous influences, people would therefore experience levels of loneliness that are largely predetermined by genes or some other internal, unchanging quality. Genes are believed to be an important part of this picture. Boomsma, Cacioppo, Slagboom, and Posthuma (2006) went so far as to describe a link between reports of loneliness and a specific area of the human genome, namely, chromosome 12q23–24. On the other hand, they noted that a range of other psychiatric disorders such as schizophrenia may also originate from this same chromosome and possibly even the same alleles. Taking another tack, Boomsma, Cacioppo, Muthén, Asparouhov, and Clark (2007) examined the concordance of reports to the statement “I feel lonely” among a sample of identical and fraternal Dutch twins. They estimated the heritability of  8  responses to this item at 77%. With more longitudinal data at their fingertips, this number was a revision to a previous study estimating heritability at 48% (Boomsma et al., 2005). Boomsma et al. (2007) emphasized, however, that environmental factors are also important. Rather than arguing strict genetic determinism, they painted a conception of loneliness consistent with the diathesis-stress model, namely, that vulnerability to loneliness may be a genetically endowed trait, and that expression of that genetic endowment may be moderated by environmental influences. Central to Hawkley and Caccioppo’s conception of loneliness as a risk factor for disease is that it is also a stable problem, although the evidence here is not crystal-clear. In an article that clearly showed the development of Cacioppo’s thought process, Cacioppo et al. (2002) cited some studies showing test-retest reliabilities of loneliness measures that were between 62% and 73%. These, however, covered time spans of a year or less. While arguing that the heritability is around 48%, Boomsma et al. (2005) showed a correlation matrix for responses to the item “I feel lonely” across five waves of a longitudinal survey. This chart indicated that correlations generally weakened across time. Among responses for waves that were 2-3 years apart, correlations ranged between 50% and 66%, but those between the 1991 and 2002/2003 waves, the full length of the survey, were only 27% to 30%. One possible explanation is that existing loneliness remains across the life span but that more people become lonely as they get older. Wenger and Burholt (2004) found that, among those who had reported loneliness at some point throughout a longitudinal study spanning twenty years, less than 10% had overcome it by the most recent wave. Many more people became lonelier. This was often in the context of various factors  9  resulting in less social interaction, such as retirement, loss of a marital partner, or poorer health and thus poorer mobility, vision, or hearing. Similarly, Jylhä (2004) contested that older people become lonelier as they continue to age, and that this coincides with functional disability as well as the loss of a spouse and a diminution in social participation. In contrast, Dykstra, van Tilburg, and de Jong Gierveld (2005) clarified in their study that not all groups of older people become lonelier. Exceptions were those whose health improved and those whose social network sizes increased between waves. Jylhä (2004) also noted that some people in her study appeared to have recovered from widowhood, although given the large gaps between waves this may have taken quite some time. These studies are not inconsistent with the contention that loneliness experienced persistently across time is a conduit to chronic disease later in life. They suggest that there may be some movement between the lonely and non-lonely groups, but that more people become lonely than not due to changing circumstances in their social lives as they age. All the same, very few studies if any have examined loneliness across the life span while bringing health outcomes into focus. Even longitudinal studies focusing on older adults all too often do not confront attrition in their samples due to mortality. Two important tasks would be to gauge the extent to which loneliness is felt persistently across time and then to determine whether persistence in reports of loneliness corresponds to worse health outcomes. The first of these tasks could be done directly using longitudinal data, for example, by creating a simple correlation matrix for reports of loneliness across waves. Another way to test whether loneliness is a persistent problem is to investigate its interplay with the social circumstances of respondents. If it were  10  found to change in tandem with changes in social life then we would be in a position to suggest that loneliness is not strictly a manifestation of personality but also a routine vicissitude. Boomsma et al. (2007) concede that environmental influences are important in contributing to the incidence of loneliness, however very few studies have examined the role that the persistence of loneliness may play in creating poor health. This brings to the fore the second task of examining health outcomes against longitudinal data. To test the argument that loneliness is a stable problem, an examination of changes in social circumstances and how those correlate with loneliness may help to elucidate the issue. One worthy pursuit may be to examine changes in the characteristics of personal social networks as they relate to loneliness and health. This however entails some methodological problems. Many surveys, such as the Alameda County Health and Ways of Living Study (discussed in more detail later), asked respondents how many “close friends” they had. Yet loneliness, the key variable under consideration, could easily be a confounding variable here. Duck, Pond, and Leatham (1994) painted a portrait of the lonely person as having a cynical view of existing personal relationships. This cynicism is pervasive, creating an altogether gloomy view toward social life. The integrity of questions attempting to gauge the number of “friends” one has, not to mention the number of “close friends,” is questionable because lonely people may be less (or perhaps more) inclined to regard another individual as such. A more serious problem may be that the objective size and composition of social networks (even if measured adequately) do not correspond with loneliness in a clear-cut way. Perlman and Peplau (1981) formulated their well-respected conception of loneliness by describing it as “the unpleasant experience that occurs when a person’s network of  11  social relations is deficient in some important way, either quantitatively or qualitatively” (p. 31). The quantitative dimension, then, captures only part of the puzzle. People become lonely when they see a disparity between the quantity or quality of social relations they have and that which they desire, which also makes the criteria subjective to the individual in question. Some may just as easily feel discontent when surrounded by friends while others feel completely happy by themselves, making the actual number of friends less meaningful. Indeed, this problem seems to emerge in the literature. While predicting loneliness, Hughes, Waite, Hawkley, and Cacioppo (2004) found only a modest correspondence with their index of social network size and composition. While studying immune response, Pressman et al. (2005) found a synergistic effect between objective social isolation and loneliness in predicting immune response, but failed to find that either mediated the other. More investigation may be merited in this area, but I would not anticipate much progress. In a contemplative article Perlman (2004) explained that, being a question of discrepancy between wants and haves, loneliness is necessarily a subjective experience. A question of whether one has four or ten friends must be followed up with a question of how many the individual wants as well as the quality of those friendships and whether that quality conforms to expectation. The objective nature of kith and kin networks does not seem to carry much currency when attempting to describe the relationship between loneliness and health. 2.3. Marital status Far from posing methodological concerns, marital status is one of the most objective and reliable variables available. It also has long been regarded as an important  12  predictor of both loneliness and health. On the one hand, House et al. (1988) in their literature review speak of marital status as a predictor of such outcomes as mortality, tuberculosis, and even accidents. On the other hand, Stack (1998) demonstrates that people all over the world tend to be lonelier if they are not married. Putting these pieces together in the knowledge that loneliness is a health risk may show that it also mediates the relationship between marital status and health. Granted, some qualitative dimensions are worthy of consideration. Marriage may be beneficial to health only for those married people who are satisfied with their relationships (Troxel, Matthews, Gallo, & Kuller, 2005). It is conceivable that health effects could be lost if a spouse is altogether unsupportive. The corresponding lack in the desired emotional intimacy may in turn be conducive to loneliness. This is illustrated well by older spousal caretakers. Dykstra et al. (2005) remarked that having to care for a housebound spouse in old age may increase loneliness due to the various stresses involved. Versus the view that marriage is ubiquitously protective, all of this is consistent with the notion that loneliness is a discrepancy between what one wants and what is lacking (as per Perlman & Peplau, 1981; Perlman, 2004). Consideration of the dynamics of marital satisfaction is beyond the purview of this thesis, but marital status as an objective predictor of both loneliness and health will be useful for my purposes. There is one suspicious parallel between the loneliness literature and the literature describing social network deficiencies and mortality. Citing a number of studies, House et al. (1988) observe that the relationship between marital status and poor health is stronger for men than for women. Men, on average, incur more  13  damage than women regarding life expectancy for the same deficiencies in their social lives. Similarly, there is a gender effect for loneliness. Wheeler, Reis, and Nezlek (1983) found that on average, attached men were the least lonely, followed by attached women, then unattached women, and finally unattached men. Married men may enjoy more benefits with respect to both health and loneliness because women, through their free expression of affect, seem to provide better emotional support (Miller, et al., 2007). Men may more often depend on romantic partners for emotional intimacy, and if they do not manage to procure one such partner, or if they lose the one they have, then they may suffer greater loneliness than would women in the same situation. This pattern could also underlie the gender interaction with respect to social network characteristics and mortality, especially given that loneliness ties to such a wide variety of troublesome outcomes as discussed above. That is to say, since men have been found to be more sensitive to attachment status than women while predicting either loneliness or mortality, it is plausible that these variables relate in some way. The parallel is rather striking. Somehow gender interacts with both poor health outcomes and loneliness in the exact same fashion, with married men benefiting the most, followed by married women, then unmarried women, and unmarried men least of all. If loneliness also predicts poor health, then it may mediate the gender X marital status effect. Stack (1998) discussed two possible explanations for why married people are less lonely. The social causation hypothesis describes marital status as buffering against loneliness by providing physical and emotional intimacy as well as financial and other kinds of support. The social selection hypothesis, on the other hand, describes marital  14  status as the outcome of interest: married people are likely to have already suffered less loneliness in advance of their selection of a mate, and this may have made them all the more attractive to future partners. Literature reviews describe lonely people as suffering poorer social skills and poor regard for their social environs (e.g., Ernst & Cacioppo, 1999), which may hamper their pursuit of (or desire for) a marital partner. Miller et al. (2007) go so far as to describe loneliness as a recurring cycle by which lonely people reinforce negative regard from others through poorer social skills, and this poor regard from others only further reinforces their negative self-concept and cynicism about their social prospects. It is plausible, then, that people who are chronically lonely are less adept in managing the skills needed to attract a partner. If the first hypothesis holds true and indeed marriage is in some way protective against loneliness, then that would support the importance of environmental factors (in this case the absence of a marital partner) as playing a role in the genesis of loneliness. This in turn would bring some contrast to the argument that loneliness is felt persistently across the life span. If, on the other hand, lonely people are simply less successful in finding and keeping a marital partner, then that would support the view that loneliness is more a dimension of personality than a routine vicissitude. 2.4. Summary of Research Questions and Hypotheses The situation presented by the literature suggests that several hypotheses and points of inquiry about loneliness can be raised in terms of the way it relates to health. First, loneliness is seen as a risk factor for poor health and can thus be expected to correlate with some health measures. The dependent variables chosen for this thesis are self-rated health and mortality. Second, loneliness may play a role in the association  15  between marital status and poor health. This may be a mediating influence of some kind, but gender may also be a moderator since men and women appear to have their own unique experience of marriage and the way it relates to both loneliness and health. Third, loneliness is understood to contribute to poor health by being stable across time and thereby accelerating age. This means that either loneliness is a stable condition and hence a slowly developing health risk or that only those who experience it persistently across time are impacted by poor health while those who recover from it quickly are not. The stability of loneliness, a key aspect of how loneliness has been framed in the literature as a health risk, may be tested either through test-retest reliability correlations or via its relationship with marital status. If the social causation hypothesis is supported, namely, that marriage is protective against feelings of isolation, then loneliness can be seen as amenable to social circumstances and therefore need not be a persistent condition. If on the other hand the social selection hypothesis is supported, namely, that lonely people are less likely to marry, then loneliness probably structures social circumstances rather than vice-versa and may be a more persistent condition. Finally, loneliness has been presented as a subtle stressor that alters an individual’s circulatory functioning. It may therefore more specifically pose risks for fatalities involving the cardiovascular system. Literature has argued as well that loneliness may affect the immune system, but fewer studies have consistently supported this possibility, and none that I know of have supported a causal link between loneliness and specific infectious diseases. In the case of cancer, negative emotions in general have been regarded as a weak risk if it even exists at all (Brannon & Feist, 2007). Still, with  16  comprehensive health and mortality data, all of these possibilities can be explored quite easily. 2.5. Analytical Strategy In summary, Hawkley and Cacioppo’s hypothesis (2002; 2007) is presently the dominant one linking loneliness to poor health. Namely, loneliness is a subtle stressor that accelerates the wear and tear of the body over time, resulting in poorer cardiovascular functioning and greater susceptibility to immune-related diseases. These, in turn, are presumed to contribute to a shorter life expectancy for lonely adults, particularly those who are middle-aged or older. A key assumption in this picture is that loneliness is also a persistent problem across the life span. Given available longitudinal data, this paradigm of loneliness as a risk factor for disease presents the opportunity for a four-pronged analytical strategy. First is to explore the nature of loneliness as it is manifested in the available data. After first checking that the chosen measure conforms to expectation on its own right and in comparison to known correlates, a second component to this stage is to examine whether loneliness is a persistent problem. This can be done by checking its test-retest reliability with the longitudinal aspect of the data and also by seeing whether it is amenable to changing social circumstances. In this case, loneliness would not be expected to fluctuate with marital status if it is a stable problem. Should it indeed be seen as stable, then that would open the possibility that loneliness is a gradual health risk by nature of its consistency across time. The next three stages involve health-related inquiries. The second stage would be to begin exploring whether loneliness may impose any health risks by analyzing the data  17  cross-sectionally as it was reported in 1965. This facilitates familiarity with the data and the modeling process but also helps establish an understanding of the starting circumstances of the sample population, which would in turn provide grounds from which to frame the research questions. Using variables that can only be fashioned with longitudinal data, the third stage is to put the pieces together and see if loneliness poses a health risk via its stability across time. Parallel to this question is whether loneliness mediates the association between marital status and health. If it does, then it may be seen as a health risk that is amenable to circumstance. The same may be true if marital status is seen as moderating the association between loneliness and health. Finally, comprehensive longitudinal health data provide ample opportunity to explore exactly how loneliness may affect physical health. Therein would lie possible answers to the final question of whether loneliness is a precursor to certain particular fatalities.  18  III. Data The Alameda County Health and Ways of Living Study, headed by George A. Kaplan, was a longitudinal survey that focused on mortality and other health outcomes as they related to various dimensions of daily life, most notably health-related habits and personal relationships. Questions relating to habits covered smoking, diet, exercise, and a number of other variables believed to impact health. Respondents were asked a broad spectrum of questions pertaining to both mental and physical health, including the presence of chronic illnesses such as cardiovascular disease and cancer. Respondents were also asked about various dimensions of their social lives, including participation in voluntary organizations, the number of friends and relatives with whom they regularly communicate, and marital satisfaction. Questions relating to emotional health pertained to feelings of depression, anxiety, general happiness, and many other outcomes. One of these questions specifically asked about loneliness. The initial survey was deployed in 1965 using a stratified random sample of 8,074 non-institutionalized residents of Alameda County, California. Preliminary information about the targeted respondent was first requested from each household. Those responding to this initial request were not necessarily the targeted respondents, however these questions pertained to very basic, objective information, such as marital status, presence of a telephone in the household, and similar details that were ostensibly obvious to anyone responding on behalf of the targeted respondent. Questions covering even standard sociodemographic details that may be subject to interpretation in some way, such as income and racial background, were not asked at this stage.  19  Subjects were then mailed a paper questionnaire and were asked to mark their responses by hand and then return it via post to the researchers. Among the total of 8,074 initially contacted, 6,928 residents responded, yielding a response rate of 85.8% for the first wave of the survey. Follow-up questions were asked among survivors in 1974, 1994, 1995, and 1999, and the respective sample sizes for these waves were 4,864, 2,729, 2,569, and 2,123. In 1995, subjects were given an abbreviated version of the questionnaire that largely covered psychosocial variables, including the loneliness question but excluding marital status. Among 3,180 residents confirmed to have died throughout the course of the study, cause of death was documented in a mortality dataset, along with the month, day, and year of death. The codebook accompanying the data explained that the initial sample included residents aged 21 and older, or aged 16 through 20 if married. Since marital status was a focal point of this thesis, and due to concern about the possible introduction of selection bias influencing the outcomes being tested, the group of young persons aged 16 to 19 (N = 49) was deemed non-representative and excluded from the models. In contrast with the information declared by the researchers in the codebook, about half of young adults aged 20 reported themselves as single rather than married. Due to the ensuing uncertainty about the criteria with which this group was selected, and given the other concerns stated here, adults aged 20 were also excluded (N = 90). The final sample sizes for the various waves were 6,789 (1965), 4,765 (1974), 2,654 (1994), 2,503 (1995), and 2,061 (1999). The last mortalities were recorded in 2000, with 3,168 individuals confirmed to have died by that time. Contact was presumably lost for 1,626 people (23% of the total) who responded in 1965. These were the individuals in the abbreviated sample described here  20  who were not confirmed to have died at any point but who also did not submit data for themselves in 1999. One part of the survey addressed emotional health and began, "Here is a list that describes some of the ways people feel at different times. How often do you feel each of these ways?” Listed were several items gauging anxiety, depression, and other dimensions of emotional well-being that allowed respondents to circle “Never”, “Sometimes”, or “Often”. An item reading “Very depressed or unhappy”, for example, may be used to gauge depression, and perhaps could be combined with reverse-coded answers to another that reads “On top of the world”. The item chosen to measure loneliness reads, “Very lonely or remote from other people”. Percentages of respondents answering “Never”, “Sometimes”, and “Often” in 1965 were 51.6%, 43.3%, and 5.1%, respectively. The percentages were similar across waves, although some variation was seen. Table 1 shows this breakdown, along with reports of marital status across waves. Several other items gauge agreement with statements relating to emotional health utilizing dichotomous responses (i.e., “True” / “False” or “Agree” / “Disagree”). These include the statements, “What is lacking in the world today is the old kind of friendship that lasted for a lifetime”, “It's hard for me to feel close to others”, and “Often, when I'm with a group of people, I feel left out – even if they are friends of mine.” The creation of a composite scale was considered, combining the item that explicitly asked about loneliness with items that seemed to reflect similar feelings. After testing the reliability of various combinations, however, they yielded at best a modest Cronbach's α of 0.661. Only the one question explicitly asking about loneliness was therefore used throughout  21  the analyses. Answers to this item were reported for all five waves. Values were missing for 83 individuals, or 1.2% of the population, in 1965. In all analyses using loneliness as an independent variable cases were deleted listwise. Income was assessed with the question, “About how much was the total income, before taxes, of your family from any source last year?” and the comment in parentheses, “Do include the income of all other members of your immediate family who received income last year. If you live alone, consider yourself a one-member family.” Respondents were allowed to mark “Under $1,000”, “$1,000 – 1,999”, “$2,000 – 2,999”, “$3,000 – 3,999”, etc., up to $9,999, followed by “$10,000 - 14,999”, “$15,000 - 24,999”, and “$25,000 and over”. About 5% of the values for income were missing in 1965. Education was assessed with a question asking, "How many grades did you finish in school?" followed by the instructions, "Circle last grade completed". The numbers 1 through 8 were listed next to "Grade school", 9 through 12 next to "High school", and next to "College" were the numbers 1 through 4, "5+", and "Advanced degree". Although the 1965 codebook did not show how gender was assessed, later waves asked the question, “What is your sex?” and respondents were able to mark “Male” or “Female”. Table 2 shows the prevalence of men and women, as well as educational attainment, income, race, and age in ten-year increments.2 One of the questions began, “All in all, would you say that your health is generally…” and respondents were allowed to indicate “Excellent”, “Good”, “Fair”, or “Poor”. This was regarded as a measure of self-rated health and was used during initial stages of the analyses as detailed below. For those subjects who had died by 2000, cause of death was represented in the mortality dataset with ICD-9 codes. Codes in the ICD-9 scheme, also known as the Ninth 2  Apparently, the values for race were updated after 1965 but were still incorporated into the first wave.  22  Revision of the International Classification of Diseases, comprise an internationally recognized system of disease classification used for reporting cause of death and the presence of disease among the living. It was published by the World Health Organization in follow-up to an international conference that convened in 1975 with the purpose of updating the system that existed at that time. These codes are typically reflected in death certificates, hospital records, and other official documentation about the progress of diseases among individuals. They were used to determine cause of death among respondents in some of the analyses discussed below. 3.1. Validation of the loneliness measure Special care was taken to ensure that this project hinged upon a valid measure of loneliness. As a first step toward this end, literature covering the use of similar metrics was consulted. Steed, Boldy, Grenade, and Iredell (2007) compared responses to a similar self-rated item with two established scales, the UCLA Loneliness Scale (as per Russell, 1996) and the scale created by de Jong Gierveld and van Tilburg (as cited in Steed et al., 2007). They compared how well each of these associated with a variety of sociodemographic measures understood to bear some impact on loneliness, such as marital status, self-rated health, and a solitary living situation. More specifically, they tested differences across various sociodemographic strata in the level of loneliness as per each metric. In terms of either confirming or failing to confirm a statistically significant association, the unidimensional measure agreed with the other two scales for 16 out of 24 sociodemographic variables examined by Steed et al. (2007). The unidimensional measure agreed with at least one of the scales for five additional variables, but disagreed with both for three others: those assessing gender, whether children lived within close  23  range, and whether the respondent had regular contact with neighbors. Although differences in these results may not be easily explained for the latter of these two, due to greater stigmatization men tend to be somewhat more hesitant about declaring themselves as lonely when explicitly asked (Miller et al., 2007). Because the word “lonely” was explicitly mentioned, and uniquely so, a gender difference for the unidimensional item is therefore not surprising. More important is that agreement among scales was observed for all of those variables that are particularly relevant to the current study, namely, age, marital status, and self-rated health. As mentioned above, the proportions of Alameda County residents reporting that they “Never”, “Sometimes”, and “Often” felt “Lonely or remote from other people” in 1965 were 51.6%, 43.3%, and 5.1%, respectively. Jylhä (2004) reported that 65% of a sample of Tampere, Finland residents indicated “Never”, 27% “Sometimes”, and 8% “Often” in response to the question, “Do you feel lonely?” Among a sample of older adults living in Great Britain (Victor, Scambler, Bowling, & Bond, 2005), 61% reported “Never”, 31% “Sometimes”, and 7% either “Often” or “Always”. Steed et al. (2007) found similar proportions as well (61.5%, 31.5%, and 7%, respectively). In sum, the proportion reporting “Often” in Alameda County resembled that found in the other studies, but the proportion reporting “Sometimes” versus “Never” varied somewhat from other studies, with higher proportions indicating “Sometimes”. Further tests of validity were therefore pursued. Using data from the 1965 wave, the loneliness measure was validated by examining its association with variables that are each understood to correspond with loneliness in some way. The degree of reported loneliness was first compared across  24  cohorts. Age was categorized into seven levels that each covered ten-year increments, starting with ages 21-30, followed by 31-40, and so on. Using the loneliness measure as a continuous variable (1= “Never”, 2 = “Sometimes”, 3 = “Often”), a line was then plotted showing average loneliness by age category (Figure 1). This resulted in a trend that remarkably resembled the one reported by Perlman (1991) in his meta-analysis of loneliness as it related to age. He likened the curve to a reverse check shape, and indeed, loneliness levels among Alameda County residents were comparably high among young adults, trended downward and reached a low point for those aged 61 – 70, then trended back upward. Notably, there was another downturn for those aged 81 and older. Although this trend did not quite reach the lowest point again, the pattern contrasts somewhat with other literatures. Still, results on the prevalence of loneliness among the oldest old have been mixed and thus remain a subject of some controversy. Since the sample size for those aged 81 and older was comparably small (N=116, or 1.7% of the total), and since elaboration on the prevalence of loneliness among the oldest old falls beyond the purview of the current study, discussion of this finding is omitted here. Due caution is encouraged while interpreting the results. The presence of a relationship between marital status and loneliness was tested next, along with an interaction for gender. As expected, mean loneliness was higher for those who were single (1.71), divorced (1.72), separated (1.78), or widowed (1.64) than for those who were married (1.48). A binary variable was created that distinguished the “Sometimes” and “Often” responses from “Never” for loneliness. Pearson’s Chi-squares were calculated using this indicator against marital status in a 2X5 cross-tabulation, and  25  indeed this test was statistically significant (χ2 = 140.7, df = 4, p < .001). Using a 2X2 table a test for gender was also significant (χ2 = 72.4, df = 1, p < .001). The possibility of an interaction between gender and marital status was also explored. As expected, mean loneliness scores were lowest among married men (1.41), followed by married women (1.55), unmarried women (1.68), and finally unmarried men who were loneliest of all (1.73). This conforms quite well to the literature. Using the dichotomous version of the loneliness metric in a 2X4 table, Pearson’s Chi-squares were calculated among these four groups to test how they compared in their reports of loneliness. The test was statistically significant (χ2 = 204.0, df = 3, p < .001), although this did not necessarily demonstrate an interaction affect since both gender and marital status were already seen to be significant. The possibility of such an interaction is explored with more sophisticated methods later in this thesis. Also tested were associations with 30 measures that bore face validity for loneliness (e.g., “It's hard for me to feel close to others”), hostility (e.g., “I am sometimes cross and grouchy without any good reason”), depression (e.g., “It often seems that my life has no meaning”), social anxiety (“It is hard for me to start a conversation with strangers”), hopelessness (“I get pretty discouraged sometimes”), overall happiness (“All in all, how happy are you these days?”), and the self-reported number of close friends. Depending on the shape of the cross-tabulation (i.e., square or rectangular), either the tau-b or the tau-c was examined. All 30 tests were statistically significant (p < 0.001) and all 30 occurred in the expected direction. Absolute values for the tau coefficients ranged between 0.052 and 0.466, with an average of 0.205.  26  The question that seemed to resonate most with the loneliness metric was one gauging depression (“[How often have you felt] depressed or very unhappy?”; tau-b = .466), followed by boredom (tau-b = .370), hopelessness (tau-c = .360), and overall happiness (tau-b =.313). That being the case these are potential confounders. Still, various aspects of social life were also clearly implicated, as ordinal associations between loneliness and items gauging shyness, number of close friends, and possible alternative measures were all statistically significant and occurred in the expected direction. Importantly, self-rated health was found to correlate with loneliness in the expected direction (tau-c = 0.125; p < 0.001), further confirming the validity of the loneliness measure. Victor et al. (2005) commented that unidimensional measures of loneliness offer some advantages, such as ease of inclusion in a large survey, but are limited because the conceptual interpretation of loneliness may change over time and across age groups. Such changes were not evident, however, for Alameda County residents. The above exercises were repeated for the subgroups aged 21-40, 41-60, and 61 and older. All associations again occurred in the expected directions across age groups and were statistically significant (p < 0.001). Although the strength of the associations varied somewhat between age groups, these differences were marginal. It must be conceded that only a single, simple measure was used here to assess loneliness. The face validity of this measure is clear, however, since the word “lonely” is explicitly mentioned. Moreover, despite the vague, multidimensional, and even philosophical implications of the word “lonely”, and despite a possibly artificial gender difference in responses, researchers studying loneliness (e.g., Perlman, 1989; Routasalo  27  & Pitkala, 2003) have generally expressed comfort in trusting the validity of a rough, unidimensional measure. With respect to associations with a large array of variables, and in terms of those associations being extant and in the expected direction, the simple measure for loneliness used in this study performed flawlessly. It was therefore presumed that there was enough evidence showing that the metric used throughout this thesis is valid for loneliness. Lacking the more sophisticated scales, a greater threat would be the loss of reliability and power. With only a single response item occurring in three levels, the measure is admittedly weak in comparison to other scales. Still, the trade-off was that the prospective nature of responses on this measure could be observed easily among a representative population sample, a feat that has been rare in the loneliness literature. 3.2. Data Preparation and Exploration In a similar spirit, independent variables were also checked for validity. Marital status was cross-tabbed with number of children and home ownership to ensure that they conformed to expectation. A higher number of children was found living in the home among married people, fewer among widowed, separated, and divorced people, and fewest among singles. Similarly, home ownership was highest among married people. Age also seemed to be represented well in the survey. Initial exploration of the data showed that the poorer the health, the higher were respondents in age. For each of the seven age categories representing 10-year increments, a pie graph was created showing the proportion of those rating their health as “Poor”, “Fair”, etc. A clear pattern was seen in that, as they aged, more people fell out of the “Excellent” categories and into “Fair” and “Poor”. The one exception to this was the category for those 81 years of age or older,  28  which showed the opposite effect. It is possible that people in this group lived long enough to feel optimistic about their health, having “earned” a high rating. Perhaps also they had enjoyed good long-term health, which would have helped them survive to that age. In any case this issue was deemed to be of less concern because those in this group composed only 1.7% (N=116) of the population as a whole. With the exception of those of 81 years and older, average self-reported health followed a roughly linear trend with a negative slope when plotted against age. This showed deteriorating health as may be expected. For the cross-sectional phase it was therefore deemed appropriate to use raw age without any transformation. Moreover, preliminary models of self-rated health using both age and age squared showed only the former to be statistically significant. Self-reported race / ethnicity, education, and income were also checked by crosstabulating them with each other. Indeed, respondents reporting higher education also reported higher income. By far the most people reported themselves as having completed high school as opposed to any other educational level. Still, education roughly followed a normal curve, with the supposed mean falling close to the middle of the graph. The distribution for income, on the other hand, was skewed to the left, with a large number of respondents indicating total incomes of $8,000 – $9,999 and $10,000 – $14,999. A potential problem is that the latter of these is the mode and yet covers a comparatively large range. Actual mean and median income could therefore fall anywhere in this rather substantial range and may even fall outside of that range. All the same, a graph depicting income against average self-rated health, using this ordinal response, showed a relatively linear trend. The same was also true for education. Since this feature was deemed more important, income and education were left untreated as ordinal variables and were later  29  used as interval variables while building the models gauged self-rated health against the cross-sectional aspect of the data. Seventy-eight percent of the population was White, followed by 12.5% Black, 3.8% Hispanic, and the rest of various other ethnicities in proportions of less than 2%. Blacks and Hispanics had less education and lower income than Whites, and Hispanics less education but higher income than Blacks. Proportions for Hispanics and each of the other ethnicities were very small in comparison to Whites and Blacks. Income and education for these groups also varied greatly (e.g., Japanese had both higher income and education than Whites). Black, White, and other non-White ethnicities were therefore the only categories deemed analytically useful. An indicator was created identifying Black people and another identifying those of other non-White ethnicities. These were used as control variables. The idea here was to represent the three categories as a control variable for ethnicity with White as the reference category, but the extra step of creating an actual categorical variable was omitted. A similar (albeit somewhat abbreviated) process was used to ensure the quality of the data reflected in later waves. The loneliness metric was validated in a similar way as for the 1965 data. For the most part, no inconsistencies or anomalies were seen that posed any particular concern, and the data were understood to bear out the expected patterns in each of the waves in which they were reflected. There were, however, two exceptions to this: education and income. Education. Self-reports for educational achievement were found to be somewhat questionable across waves. Of those reporting their education in the respective waves, 10.6% reported lower educational achievement in 1974 when compared with 1965, 8.5%  30  in 1994 when compared with 1974, and 7.2% in 1994 when compared with 1965. Since past educational achievement cannot be reduced, these discrepancies may have resulted from inconsistencies in the interpretation of the question. The occurrence of reductions in education was much more frequent among those reporting lower educational levels. Larger proportions were seen across waves of those reporting increases in education. Comparing 1965 with 1974, 1974 with 1994, and 1965 with 1994, the percentages reporting increases in educational achievement were 23.1%, 22.7%, and 33.0%, respectively. It is plausible that many of these individuals returned to school in order to improve their credentials. However it is also plausible that much of this is also measurement error, considering the surprising number reporting decreases in education as discussed above. Also worthy of mention is that the minimum age in the population was 30 as of 1974, and yet 22.3% supposedly returned to school through 1994, which may seem somewhat high for this age group. All the same, those erroneously reporting increases in education could not be distinguished from those who had legitimately returned to school. Generally, no pattern could be seen that suggested which discrepancies appeared erroneous. The option of removing all reductions in education and using the prior report was considered, however this seemed unjustified for two reasons. First, since there were so many mysterious reductions in education, it seemed likely that at least some of the reported increases in education were also erroneous. Yet no concessions could be made to correct those cases that may have reflected erroneous increases. Correcting only the decreases could have introduced some bias. Second, lacking additional information on the nature of these discrepancies, it was impossible to know which of two divergent  31  reports was the correct one in the first place. Any attempt to do so would have entailed some assumptions that may have been difficult to justify. The second report reflecting lower educational achievement may have merely been a more honest response. Nevertheless, for those cases where data from three waves were available it was possible to identify the likely erroneous report when the other two were identical. For example, if one reported 10 years of education in 1965, 12 in 1974, and 10 again in 1994, it was likely that the individual erroneously reported his or her educational achievement in 1974. Changing the one discrepant report, however, would likely have biased the sample, creating a problem similar to left-censoring. That is to say, this type of correction would have been possible only among those reporting data through 1994. While the erroneous report would have been corrected for these individuals, they could not have been corrected for those who had died or lost contact with the researchers by 1994. Proceeding in such a fashion may result in unknown and uncontrolled bias. Attempts to correct the discrepancies among waves were therefore abandoned, and raw categories for education were used as they were originally represented. The one exception to this is that, for the 1965 and 1974 waves, the last two categories were combined. This was done so as to align them with the 1994 wave, in which those reporting advanced degrees and those reporting five or more years of college (as opposed four or less) appeared combined under one category. This left the total number of categories as a continuous variable representing the number of years spent in school, ranging from zero to 17 for all waves, the last of which represented 17 or more years of schooling. The Cronbach’s alpha for the resulting variable across all three waves was 0.966 (p < 0.001), which was equivalent to the Cronbach’s alpha seen prior to the change  32  made to the final category. With such a high reliability estimate, the variable was deemed appropriate for my purposes. Income. As another important control variable, income required some special considerations when contemplating the longitudinal aspect of the data. The increases and decreases in income that were seen across time would understandably coincide with changes in occupation and other circumstances, so this was not seen as a problem. However the categories were relatively incommensurable when comparing the different waves. Intermediate categories were added in 1974 that provided better definition of the higher ranges, and the first two categories seen in the 1965 wave were combined. By 1994 the problem of inflation became obvious. Income was capped with a category describing earnings of “$150,000 or over”. This is quite a contrast with the maximum of $25,000 seen for the 1974 and 1965 waves, which meant that the scales were quite different from each other. The option was considered to use the ordinal values in a continuous fashion, but the number of categories changed across waves, and a given category did not carry the same meaning at a later time point due to inflation. The option was considered to use as the variable of interest cumulative percentiles that described increasing income levels. This would have solved the problem of the differences in categories by taking into account inequalities within the population in a way that did not depend on any particular income cut-off or on an equivalent number of categories among waves. Should income be protective against disease, those in the lowest 5% would presumably be impacted quite differently compared to those earning the better of 95% of the population. This would  33  have also provided an interesting point of comment on income inequality as opposed to objective income. On the other hand, it was supposed that as people age they may earn job promotions, are more adept in their various investments, or in some other way become more experienced in terms of earning money. If indeed the population as a whole in Alameda County enjoyed growth in their income over time, and if indeed income was protective of health, it would make more sense to allow that growth of income to serve its own protective effects against advancing age. The more plausible solution, then, was to use the natural log of the median values of the ranges described by each income category.3 This strategy was borrowed from Cacioppo, Hughes, Waite, Hawkley, & Thisted (2006), although an adjustment for inflation was also included so that income would have the same scale among the different waves. It was assumed that debt need not be a subject of this analysis and thus that income could only go as low as zero. On the other end of the spectrum, however, the highest categories presented a complication. The true ranges for these categories were unknown, and therefore the medians for those ranges were unavailable. Only known was the number of people in 1965, for example, reporting income of “$25,000 or more”. If the best solution was to use the median values of the ranges in order to mimic actual dollar amounts, then adjust for inflation and take the natural log, then in all fairness a best guess was needed for the median of the highest category. A best guess would be one that agrees with the rest of the data. In other words, if in 1965 a rather humble distribution of wages was seen, a very large value for the last category would not make sense. Neither would 3  For example, the median value for a range of $1,000 - $1,999 would be $1,499.5. In many cases this was the same as the mean average between the two numbers, but in those cases when this was not the case the middle number within the range was selected.  34  one wish to have the distribution fall abruptly by setting the median at barely above the minimum of $25,000. A value that seemed to follow the distribution smoothly would be most appropriate. A special procedure was undertaken to address this problem. The idea here is that, when one has the mean and standard deviation for a distribution and can also assume that the data are normally distributed, one can estimate an expected value if the percentile at which that value falls is known. That is to say, when applying the normal distribution to a problem, for each value seen in the data there also exists the theoretical value at that percentile. If the data were to perfectly follow the normal distribution, the mean would be the same as the median, otherwise known as the 50th percentile. At the extreme end, though, the value at the 95th percentile would fall at 1.96 standard deviations from the mean. My task was to find the values for the highest category that best agreed with the respective theoretical (expected) value. For the 95th percentile this could be found by multiplying 1.96 (the z-score for this percentile in the normal distribution) times the standard deviation and then adding the result to the mean. I began by assuming that the natural log of income follows a normal distribution. Using as an example the 1965 data, the proportion of people in the highest category was 2.8% of those reporting income. From this we know the percentile at which the median within this range fell. When 2.8% is divided by half and then subtracted from 100%, the resulting cumulative percentile is 98.6%. If logged income follows a normal distribution, then, those earning at the median for the highest category should have earned the better of 98.6% of the population in 1965. The z-score at this percentile is 2.197. Multiplying that amount by the standard deviation and adding it to the mean would produce the expected  35  logged income level at that percentile. Taking the antilog of this value would then show the expected median income among those reporting $25,000 or higher. Missing at this point, however, are the mean and standard deviation. In order to calculate those, the whole distribution is needed but of course the value for the top category is still lacking. Taking sample values (in other words, guessing) allows us to calculate the supposed mean and standard deviation and to see how they play out in the above procedure. Upon calculating our guesses for those, this leaves two values: the natural log of the assumed median for the highest category and the expected value for the log of this same median, presuming that the data are normally distributed. Whenever these two happen to converge, we have our best guess for logged median of the highest category. In order to gauge which values best fit the data, given again that the data were lognormal, sample median incomes for the final category were entered in increments of $2,500, all values for income were then logged, and descriptive statistics of these now complete distributions were calculated. The resulting standard deviations were multiplied by the z-score for the percentile at which the median of the highest category stood and then this was added to the resulting mean. The expected value that best aligned with the chosen logged median was selected. For the 1965 population, taking the natural log of income after assuming a value of $35,000 in the highest category resulted in a standard deviation of 0.713 and mean of 8.897 for the entire distribution. Given a z-score of 2.197 for the 98.6th percentile, solving the equation EXP(M + Z * σ) resulted in a value of 35027.02, which falls quite close to 35,000. As the median for the final category, $35,000 was therefore presumed to best  36  align with the rest of the data. The process was repeated for all other waves in which income was reported. Chosen median values of the highest income categories for 1965, 1974, 1994, and 1999 were $35,000, $45,000, $230,000, and $217,500, respectively. Ordinal values for the income variable were replaced with the median values for the ranges that they described, including the above for the highest categories. Values for later waves were then adjusted for inflation so as to be comparable to 1965. Finally, the natural log was once again taken of the dollar income levels, and the resulting variables were used in the longitudinal analyses as income measures. Further tests were carried out to see if the correct approach may have been to use raw median incomes throughout the process described above, without taking the natural log at any point. Using the same procedure described above for 1965, this yielded a median income of less than $25,000 for the final category, which would have been impossible. It was therefore deemed suitable to have started the process with the natural log for income as had been done. Simple preliminary models predicting deceased status4 showed that the revised variable for 1965 income (Wald = 211.6; p < 0.001; Nagelkerke R2 = 0.009) performed similarly to the original ordinal categories (Wald = 224.9; p < 0.001; Nagelkerke R2 = 0.010) in separate models. This was in contrast to the use of the median incomes as represented by dollar amounts (Wald = 55.8; p < 0.001; Nagelkerke R2 = 0.003). Logged income was therefore deemed the best metric to use in all longitudinal models. Because the raw ordinal variable appeared slightly superior, and because income posed special  4  Deceased status was used as the dependent variable here. The context within which it was used, i.e., in longitudinal models calculated against person-years, is described presently in the Methods section.  37  problems only when considered longitudinally, the original ordinal variable for income instead was used during the cross-sectional analyses of the 1965 population. IV. Methods The analysis for this thesis occurred in four stages. The main idea was to first explore the nature of loneliness as it was represented in the survey data and then to see whether it posed any substantial health risk. Important questions pertained to whether loneliness may be problematic by nature of it being a persistent problem over time, whether it was related to substantial health measures in any important way, and what particular fatalities it may cause. The methods entertained and discussed as follows were expected to help elucidate exactly how it as a psychological phenomenon carries through to bodily harm. The aim of the first stage was to piece together the nature of loneliness as it was represented in the 1965 wave of the Alameda County Health and Ways of Living Study. Discussed earlier, this involved rigorous validation of the loneliness measure and investigation of various constructs understood to covary with reports of loneliness. All correlates gleaned from the literature that were represented in the survey were tested for association with the loneliness measure. These included marital status, depression, and social anxiety. The trend in loneliness by age was also observed and compared with findings in the literature. In all regards, as noted above, the chosen measure conformed to expectation and was deemed acceptable. More germane to the nature of loneliness in general, also necessary was to explore whether it was more or less a persistent problem across the life span for Alameda County residents, an important assumption in Hawkley and Caccioppo’s model of disease  38  outcomes. Through the calculation of Kendall’s tau-b and tau-c correlation coefficients and the accompanying tests of significance, this involved gauging the reliability of the loneliness metric over time. If loneliness is a stable problem, then one may expect the correlations between reports to be robust among the five waves. Granted, correlations may still weaken simply because more people become lonely across time while those who are already lonely stay that way, but whether this would explain weakening correlations can be checked very easily. The prospective nature of the association between marital status and loneliness was then examined so as to investigate whether levels of loneliness change alongside changing social circumstances. That is to say, if marital status (e.g., single or divorced) is understood to play some sort of causal role in the onset of loneliness, then loneliness may be presumed to be amenable to circumstance and not just a static feature of personality or genetic makeup. Through the use of logistic regression, the social selection hypothesis was tested against the social causation hypothesis5 by examining how well marital status predicted loneliness and vice-versa. Results that may support the social selection hypothesis, namely, that loneliness somehow hinders progress toward marriage, would also support Hawkley and Cacioppo’s hypothesis by showing that loneliness is less amenable to the presence or absence of a significant personal relationship and thus may be a stable phenomenon. The second stage was a preliminary attempt to explore the nature of loneliness as a health threat through cross-sectional analysis. This exercise served as a first test of the main hypotheses under investigation. Following its initial use in the validation of the  5  Namely, whether loneliness “causes” a person to end up unmarried in some way, or whether marriage “causes” a person to feel less lonely.  39  loneliness measure, self-rated health was used here as the dependent measure. Using ordinal regression, loneliness in 1965 and several competing measures were included as independent variables in models that predicted self-rated health. The third stage involved testing the impacts of loneliness on health by combining data from the 1965 survey and subsequent waves with the mortality file. Through the use of logistic regression models, loneliness was thoroughly explored as a specific health risk by using deceased status as the dependent variable in regression models. This started with the simple introduction of the loneliness metric into the models. Several steps were then undertaken in order to examine the role that the stability of loneliness may have in structuring health outcomes. Some models were created that tested whether the recency of a given report had an impact. Others examined whether loneliness interacted with the aging process in some special way. Finally, the interplay among loneliness, gender, and marital status as predictors of mortality was thoroughly explored. All of these modeling exercises included the addition of socioeconomic variables as controls. This stage of the project was carried out through the creation of person-years. A new row was created for each year that the individual was known to have survived. For those study participants for whom contact was lost yet whose death was not recorded, the last row represented the last wave for which they offered responses. For those who had died, a new row was created for each year that they had survived, and the last row represented the year that they had died. Among those who had survived through all points of data collection, the last row represented the year 2000, the final year in which mortality data were collected.  40  For all person-years, a binary variable was created that indicated whether the individual had died during that year. This served as the dependent variable in logistic regression models describing mortality. Age was updated so as to represent actual age during each person-year rather than age as of the latest report. The option was considered to use squared or cubed age in the analyses, but in a simple logistic regression model with deceased status as the dependent variable and raw age, age squared, and age cubed as independent variables, only raw age was a significant predictor of mortality. Age was therefore used in all modeling exercises as it was originally represented in the data. Whenever person-years reached the year of a new wave, the values for loneliness, marital status, and all variables derived from these were updated with information from the new wave. Similarly, education and income were updated with data from each new wave (income using the process described in detail above), but only the 1965 data represented gender and race since these were presumed to be constant. In this fashion, several sets of models were created that showed how well loneliness predicted mortality amidst competing explanations. Age, race, education, and income were each deemed possible confounders and were pitted against the loneliness metric in separate models. Additional tests were carried out to demonstrate how the effect that loneliness had on mortality could change when considering marital status, gender, and an interaction between age and loneliness. Finally, various analyses were carried out to test whether loneliness is a health risk by nature of being a persistent problem. For the final stage of the study, exploratory analyses were carried out. Explanations for any emerging patterns were sought and preconceived notions about the particular manner with which loneliness affects health were tested. To this end,  41  cardiovascular and other circulatory diseases as reasons for death were of special interest. In logistic regression models similar to those described above, the indicator for mortality in each person-year was replaced with another indicator showing cardiovascular disease as cause of death. This new variable was assigned a value of ‘1’ only if the individual had died of cardiovascular disease. All other person-years were given a value of ‘0’ on this measure, and for those who had died of other causes, the last row reflected the year of death. This process was repeated for several other fatalities of interest.  42  V. Results All cross-sectional analyses, all basic tabulations, and the large majority of longitudinal analyses were carried out in SPSS 16.2. Some analyses were carried out in Stata 7.0 Intercooled. This was done so as to ensure that the process using the General Estimating Equations (discussed below) approach was carried out correctly in SPSS, and indeed results were identical. On some occasions, SPSS failed to output results. The reason for this was unknown, but appeared to be due to complications involving complex filtering structures that were used in some way to restrict the cases being observed. On those occasions models were created using Stata. 5.1. Phase 1: The Stability of Loneliness The stability of loneliness was tested by examining Kendall’s tau-b correlations between self-reports for each wave with those of each other wave. These results are shown in Table 3. Because the experience of loneliness may differ by gender, or at least the reporting thereof, correlations for men and women are separated. As Table 3 shows, however, results were not substantially different between genders. Correlations generally faded as time passed, but they never disappeared completely. Correlations also did not drop below 0.20, and all were statistically significant (p < 0.01). As may be expected the strongest correlations for both genders were seen when comparing reports from years 1994 and 1995, the closest time frame between waves. These correlations hovered around 0.50 for men and for women. Bearing in mind the weakness of the measures that were used, this does demonstrate some stability in the loneliness metric but may indicate change over time for a substantial number of people.  43  In order to check what patterns were responsible for these correlations, crosstabulations were performed. These are shown in Table 4. A key question in this exercise was whether the correlations faded over time simply because more people joined the group reporting feelings of loneliness while those who were already lonely did not manage to recover. For each consecutive pair of waves, the report for loneliness at time 1 was compared with that from time 2. Because the ordinal nature of the loneliness measure was used in all models, it made sense to regard the state of loneliness for any given individual as having improved or worsened versus simply “lonely” or “not”. One may experience an improvement in loneliness, for example, if they reported themselves as lonely “Often” at time 1 and “Sometimes” at time 2, or “Sometimes and time 1 and “Never” at time 2. In Table 4, it is not evident that the correlations faded because more people joined the group of those reporting loneliness. On the contrary, there seemed to be a fair amount of movement in both directions. It is evident that some people may experience levels of loneliness that fluctuate rapidly, as 29% either improved or worsened between 1994 and 1995. It is also evident that the change may take quite some time for other people. The period that showed the largest amount of movement between categories (42% of the population) was between 1974 and 1994, the longest span of time among the waves. In contrast, the largest proportion identifying themselves as “Never” lonely was between 1995 and 1999 (41%), followed by the comparison between 1994 and 1995 (40%). The fact that these two are the shortest time frames between waves suggests that change for the better (or for the worse) may take several years for some. To the extent that a threelevel self-report of loneliness may be regarded as a valid measure, these data present  44  some contrast to the view that the genetic contribution to loneliness is as high as 77%, per Boomsma et al. (2007). Loneliness and marriage. The prospective relationship between loneliness and marital status was examined by comparing competing hypotheses about the causal nature of this association. First, the social causation hypothesis was tested. Ordinal regression models were created with the negative log-log link using married status as the independent variable and loneliness from the subsequent wave as the dependent variable. Each model was then controlled for married status as of the second of the two waves. The logic here was that, if indeed loneliness was amenable to change via success or failure to find and keep a mate, then current marital status should wash out the effect of the report from the prior wave. That is, the protective effect of marriage against loneliness would be repeated for those who were still married, whereas those who were newly divorced or separated would experience greater loneliness. Those who moved into marriage may also experience a decrease in loneliness. Either way, current marital status would be a much better predictor of current loneliness. Table 5 shows the outcome of this exercise, with results for men and women presented separately.6 Four of the six pairs of models show that marital status at time 2 overtakes the effect of prior marital status in predicting loneliness at time 2. There were two exceptions, however. For men in 1999, current marital status appeared to be protective against loneliness (β = -0.941, p < 0.001), however prior marital status changed signs and also acquired statistical significance in the second model. A somewhat  6  Due to the nature of this exercise, inclusion of Ns in the table would have been complicated. Sample sizes in each set of models during this and subsequent exercises are therefore presented separately in Table 8.  45  similar pattern appeared for women in 1974, although the p-values for the coefficients in those models never dropped below 0.05. Next, the social selection hypothesis was tested using the converse of the same process. For each wave, a logistic regression model was created using loneliness for that wave as the independent variable and an indicator for married status from the subsequent wave as the dependent variable. Loneliness from the second of the two waves was then inserted into the model as a control. The logic here was that, if loneliness is a stable problem and predicted marital success, rather than vice-versa, then results should not change substantially when loneliness from the subsequent wave was inserted. Those results are shown in Table 6. Again, in most cases, current loneliness was a better predictor of current marital status when pitted against past loneliness. This was not true, however, for both men and women between 1965 and 1974. For men, both current (β = -.470, p < 0.001) and past loneliness (β = -.371, p < 0.01) correlated simultaneously with marital status in 1974. For women, only prior loneliness (β = -0.271, p < 0.001) correlated with current marital status in 1974. This suggests support for the social selection hypothesis only for women and partially for men between 1965 and 1974. This hypothesis was not supported by any of the other four models. To further illuminate the relationship between loneliness and future marital status, this process was repeated, except with change in marital status as the dependent variable. At the end of each two consecutive waves, individuals were characterized as newly married, still married, or newly divorced, widowed, or separated at time 2. Multinomial regression models were then created that predicted these categories based on past versus current loneliness (Table 7).  46  While predicting divorced, separated, and widowed status among men and women, loneliness was significant only at time 2 (T2) in most cases. The exceptions occurred in 1999 for men, when both T2 (β = 1.441, p < 0.001) and time 1 (T1) loneliness (β = -0.815, p < 0.05) were significant, and in 1974 for women, when both T2 (β = 0.232, p < 0.05) and T1 loneliness (β = 0.302, p < 0.01) were significant. Additionally, for men only, coefficients in all cases were smaller in the second model. This situation contrasted with the results among those who were newly married. In most cases, only loneliness at T1 predicted the attainment of marriage. Men in 1999 were the only exception, for whom loneliness at T2 instead was significant (β = -1.039, p < 0.05). Meanwhile, among those who were still unmarried at T2, the large majority of coefficients were statistically significant in all models. Women in 1994 were the only exception, among whom T1 loneliness was insignificant but approached significance in the second model (β = 0.231, p < 0.10). Summary. In conclusion to the first stage of the project, these results do not clearly depict loneliness as a stable problem and hence one that is robust against changing social circumstances. First, the test-retest reliability of the loneliness metric used throughout this study appears to fade across time. Further exploration revealed that the pattern was not simply because more people joined the population of lonely adults while those who already experienced loneliness continued to do so. Furthermore, contrasting with the view that it is a reliable individual difference (e.g., a dimension of personality), the present study demonstrates that loneliness is amenable to social circumstances. Models testing the social selection versus the social causation hypotheses lent more support to the latter. With some exceptions, coefficients  47  for past married status predicting present loneliness were washed out by present married status. In many cases coefficients for past loneliness predicting present marital status were also washed out by current loneliness. These findings suggest that marriage may buffer against loneliness more so than loneliness hampers progress toward marriage. The multinomial models presented in Table 7 show an even clearer picture. For many people loneliness appears as an outcome rather than a predictor of change in marital status. Among divorced, separated, and widowed people, past loneliness was nonsignificant, indicating that it was not a precursor to the ending of the relationship. Current loneliness was instead significant, seeming to reflect the trauma incurred by the loss of a partner. For those whose fortunes changed in the reverse, findings with respect to loneliness were also reversed. In most of these models past loneliness was predictive of success in attaining a partner while current loneliness was non-significant. This result is rather striking both in its consistency and in the predictive power of loneliness from the many years prior. One tempting proposition is that loneliness is a motivational force that prompts single adults to seek a marital partner. Further analysis on this point is necessary, though, because prior loneliness did not distinguish those who were newly married from those who ended up still unmarried. For this latter group, past and current loneliness were both significant. While this accommodates the view that singlehood may not be an especially pleasant state and that failure to procure a partner contributes to the persistence of loneliness, it is also possible that those ending up still unmarried simply have more trouble finding marital partners. Nevertheless, taking all groups together the multinomial models more strongly support the social causation than the social selection hypothesis. In  48  most of the multinomial models, loneliness (or the lack thereof) appears as a likely repercussion of marital status.  49  5.2. Phase 2: Self-reported Health in 1965 In the cross-sectional phase of the study, ordinal regression models (see Agresti & Finlay, 1997), shown in Table 9, were created to test whether loneliness mediated the relationship between marital status and health.7 Self-rated health was used as the dependent variable in these models. Because this measure appeared to follow the normal curve, probit was chosen as the link function.8 By assuming a normally distributed latent variable for the dependent variable, the probit link was particularly handy because it would presumably accommodate those reporting their health as “Poor”, a category that plausibly holds important information but which only 2% of the sample used to describe themselves. To represent gender, a binary variable was created to identify an individual as male. Another indicator represented status as married versus unmarried. Those who were single, divorced, widowed or separated thus fell into the latter group. These two variables were then multiplied so as to represent an interaction effect. Income, education, and age were entered without manipulation as they were originally manifested in the data. As such income, an ordinal variable with values ranging from 1 to 12, was treated as continuous in the models along with education. As a reminder, only for the longitudinal  7  One can interpret an ordinal regression model created by SPSS by seeing positive coefficients as lending to an increased likelihood of answering in the higher categories. In this case, with positive coefficients the respective term correlates with better self-reported health, while negative coefficients suggest a risk for poorer health with each increment in the respective term. The particular manner with which SPSS calibrates the general linear models so as to allow for this interpretation is somewhat unique, but is evidently done this way so as to make ordinal regression more accessible to the user. 8 In contrast, the more popular logit link function better accommodates a dependent variable that is equally distributed among the categories. Clearly that was not the case here. Unfortunately, because a completely different link function is used one cannot as easily speak of odds ratios associated with the beta coefficients. However, a value that is roughly equivalent to the odds ratios found using the logit link can be obtained by multiplying a given beta coefficient by 1.8 and then taking the antilog of the result.  50  analyses discussed later was income specially treated using the process described in detail above. The first model consisted of only the intercept and age. Successive models added the marital status X gender interaction variables, loneliness, a loneliness X age interaction, race, education, and income. Self-rated health was reverse-coded so as to range from “Poor” to “Excellent” in ascending order so that SPSS would regard the latter as the reference category. In the particular manner in which SPSS handles ordinal regression, coefficients that were negative conveyed poorer health for each increase in a given random variable. Those cases that had missing values for any of the variables used in the full model were deleted listwise. This left 462 individuals or 6.8% of the total out of the models, 77% of whom were missing income. A comparison of those who were taken out with those remaining in the analysis showed no substantial differences, either in self-rated health or loneliness levels. Because age is understood to be a primary factor causing health decline, it was the only independent variable included in the first model, the idea being that subsequent models could then be compared so as to gauge how different risk factors may accelerate health decline ahead of age. Indeed, the first model shows that age is a risk factor for poor self-rated health (β = -0.016, p < 0.001). The second model incorporated married status. The effect size was relatively small for a binary variable (β = 0.058) and was insignificant but approached significance (p < 0.10). This was followed by gender in the next model, which was statistically significant (β = 0.112; p < 0.001). Married status also no longer approached significance (β = 0.042; p > 0.217). The third model added a married status X gender interaction term  51  (β = 0.107; p > 0.124), which was not only non-significant but also appeared to cancel out the statistical meaning of both marital status and gender. Both of the latter two appeared with very large p-values and small effect sizes. In fact, the married status X gender interaction term proved unhelpful in all other exploratory models and therefore is not treated as a focus in the longitudinal analyses that follow. Loneliness appeared as a significant risk factor (β = -0.381, p < 0.001) in the fifth model. Compared to either gender or marital status its coefficient was large. This is true even when comparing to age. Dividing the coefficient for loneliness by the coefficient for age yields 23.8. This means that one increment in the level of loneliness (i.e., “Sometimes” to “Often” or “Never” to “Sometimes”) increased the risk of poor health almost to the same extent as an increase of a quarter century in age. I would, however, stop short of saying that one equates to the other in this way (see discussion by Hoetker, 2007, pp. 334-335). The loneliness X age interaction term was introduced in the sixth model, but it was not statistically significant (β = -0.002, p > 0.216). The small coefficient for this term and its non-significance were found among the remaining models as well. Other exploratory analyses similarly failed to find any reasonable interaction between loneliness and age in predicting poor self-rated health. Hence no support was found at this stage of the thesis that there is some interacting synergy between loneliness and age that hastens the advance of poor health. In the next set of models, control variables were added separately. These are listed together in the full model in Table 9 for simplicity but are discussed here as they were successively added. Race presented as a significant risk factor (Black: β = -0.498, p <  52  0.001; Other non-white: β = -0.403, p < 0.001), although its coefficients were reduced (0.371 and -0.263, respectively; ps < 0.001) when education was included (β = 0.066, p < 0.001). Income also appeared as germane to self-rated health in the last model (β = 0.026, p < 0.001). None of these control variables seemed to mediate the impact of loneliness. Even in the last model the coefficient for loneliness was not dramatically different from the one shown in the first and remained statistically significant (β = -0.330, p < 0.001). Married status was not a particularly reliable predictor of health outcomes in these models. An important premise to this thesis, though, was that loneliness in some way explains the relationship between marital status and health. This possibility was further explored with simpler models gauging self-rated health, this time introducing age and married status but then followed by either gender or loneliness as main effects in separate models. Married status appeared as statistically significant in the latter of these two models (β = 0.063, p < 0.05) and then lost statistical significance and had a reduction in its effect size (β = -0.029, p > 0.369) when loneliness was introduced (β = -0.396, p < 0.001). However, similar results were seen after gender was introduced separately, with loneliness also appearing to mediate the association between gender and self-rated health. It is interesting that married status reemerged as significant in the full model shown in Table 9. Although these results are not presented here, the possible interplay among loneliness, gender, and married status was further explored after first entering the various control variables. A model was created that included age, married, gender, race, education, and income. This showed gender (β = -0.095, p < 0.001) as statistically significant but not married status (β = -0.019, p > 0.584). When loneliness was entered, however, the coefficient for married status was significant in the full model and was even  53  negative, suggesting that the interplay among these variables may be quite complex. Contemplation of this unexpected outcome is omitted here, although the pattern is further explored during the longitudinal component of this thesis. Worth noting is that these models did not pass the test of parallel lines, an important assumption for ordinal regression. Unknown was how serious this problem was, how to amend it, and to what extent and in what way it threatened the conclusions drawn from these results. Logistic regression models were therefore created against an indicator identifying those who regarded their health as “Fair” or “Poor” as opposed to “Good” or “Excellent”. Similar results were found. Summary. In conclusion, three important results were observed during the crosssectional portion of this thesis. First, loneliness was seen to be a risk factor for poor selfrated health in all models. The coefficient for loneliness was negative and statistically significant in each model, even after controlling for income, education, race, and most importantly age. Increases in levels of loneliness therefore corresponded with risk of poorer self-rated health. Second, age did not appear to be a moderator between loneliness and health. The age X loneliness interaction term consistently failed to yield useful results in several exploratory models. Finally, loneliness did not clearly mediate the relationship between married status and self-rated health. However, the conclusion that marriage and loneliness have little to do with each other in terms of health would be premature. In simpler models loneliness appeared to be a mediator for marital status, but the same was true for gender. On the contrary, in the full model loneliness was a suppressor that allowed married status to be meaningful. These results are both mysterious and difficult to explain, but suggest some sort of complex interplay.  54  5.3. Phase 3: Loneliness and Mortality The next phase involved taking full advantage of the longitudinal aspect of the Alameda County Health and Ways of Living data. Using person-years as the unit of analysis, the relationship between loneliness and mortality was explored through the use of logistic regression models with deceased status as the dependent variable. In these models and in all others cited below, all censored observations were removed. That is, those person-years in which the individual was not confirmed to have died and yet also did not provide responses for the most recent wave were removed from the dataset. This affected 34,621 data rows, or 17.9% of the total person-years. With the General Estimating Equation (GEE) approach in SPSS, models were created using the AR1 working correlation matrix. To ensure that this was done correctly, models represented in Table 10 were repeated in Stata with the “cluster” subcommand. This yielded identical coefficients and standard errors indicating that the procedure was carried out correctly. The logic of using this type of technique is that observations reflected in person-years rather than persons are not independent from one to the next, which makes regular logistic regression modeling unjustified in this case. More simply put, while there were over 158,764 rows of data, the fact that those rows described only 6,789 people rather than 158,764 needed to be accommodated. By clustering the error terms within individuals the assumption of independent observations could be relaxed. In this way influences involving individual differences beyond the control or knowledge of the researcher, such as genetic contributions to longevity, were captured within individuals and thereby separated from calculations of the error terms for the model coefficients.  55  Those person-years that were missing values for any of the theoretical and control variables were deleted listwise from all models. Variables with some missing values were loneliness (1813 observations), marital status and thus the indicator for married status derived from it (178), education (753), race (18), and income (7568). As in the crosssectional analysis, gender and age did not have any missing values. Out of the 158,764 total person-years, this resulted in the removal of 9,382 or 5.9% of the total. Given the nature of the unit of analysis, information for one individual could be excluded for years associated with a wave that had missing values, and yet the person could reenter a model for years in which all the necessary fields were populated for that individual. Several models were built in order to examine the association between loneliness and mortality. The first included only age as the preeminent risk factor, the idea being to start with the natural aging process and then see how other phenomena factored in. Additional terms were then added separately, beginning with loneliness and continuing with each of the control variables. Table 10 shows the odds ratios found by taking the antilog of the coefficients in these models. Married status (0 = not married; 1 = married) was included in the exercise, but at this phase of the analysis was only regarded as an important control variable. As shown in Table 10, inclusion of this term did not alter results with respect to loneliness and mortality, nor did any of the other control variables. Loneliness was found to be significant in all models predicting deceased status. In the one that included terms for only age and loneliness, the coefficient for the latter was 0.087 (p < 0.05), comparing to 0.092 (p < 0.001) for age. In the full model the effect size did not change substantially for loneliness (β = 0.084; p < 0.05) but dropped slightly for age (β = 0.087; p < 0.001). Looking at these results, it is tempting to suggest that each  56  increment in the level reported for loneliness (i.e., “Sometimes” versus “Never” or “Often” versus “Sometimes”) meant an increase in risk of death that was roughly equivalent to an increment in one year of age. This may entail some questionable mathematical assumptions, however (see Hoetker, 2007). As expected, men incurred the brunt of health risks in terms of mortality at 62.8% increased risk versus women in the full model. Married status was also found to be protective in the fourth model (β = -0.242; p < 0.001), but its influence diminished somewhat as control variables were added. Risk associated with Black ethnicity was significant in the fifth model (β = 0.238; p < 0.01), however it lost significance and its coefficient was halved (β = 0.134; p > 0.088) in the sixth model when education was introduced (β = -0.037; p < 0.001). Education and income both appeared to be important in the final models. They also appeared to exert their influences independently of each other, which agrees with the argument that different dimensions of socioeconomic status are unique to each other rather than fungible control variables in research (see Adler et al., 1994). The risk associated with loneliness seemed somewhat small in these models, although it was obviously larger for those who were “Often” compared to “Never” lonely, which translates to an odds ratio of 1.182 for the full model.9 On the other hand, when person-years occurring beyond five years of the latest wave were filtered out,10 the effect for loneliness more than doubled (β = 0.184; p < 0.001). Worth noting also is that the p-value decreased substantially for the loneliness term. Compared to the equivalent  9  This is calculated by multiplying the coefficient by two and taking the antilog of the result. As a reminder, person-years rather than persons are restricted here. Since no one was added to the sample population in later waves, no one could have been removed by filtering out any person-years occurring after 1965. 10  57  model shown in Table 10, the risk associated with age was also slightly reduced (β = 0.086; p < 0.001). When the restriction was loosened to include person-years occurring within six years of the most recent wave, the coefficient for loneliness increased slightly to 0.190 (p < 0.001). The coefficient declined, however, when this time frame was either reduced to less than five years or increased beyond six, suggesting that the association between loneliness and mortality may be optimal within five or six years from the time loneliness is reported. Table 11 shows odds ratios for two models that were created using the 5-year restriction, one using only age and loneliness and the other using all of the control variables plus married status. The second of the two models may be compared to the full model shown in Table 10. Clearly, among all the variables the largest change in the coefficients was seen for the loneliness metric. As a special note, for the first of these models I did not remove any cases other than those missing values for loneliness. This was done so that the effect that loneliness had on mortality could be seen when using all possible observations. Although the p-value diminished somewhat between the two models, the size of the coefficient for loneliness did not change substantially. The Stability of Loneliness. Consideration of the stability of loneliness over time imposed some logistical problems. Bias could be introduced if models are created irrespective of the time frames imposed by the survey waves, for instance, if one were to include loneliness in 1965 as a separate variable from loneliness in 1974. There was suspicion that the models could unintentionally “see into the future” and distinguish those reporting anything in later waves because they had survived to that wave. Loneliness in 1994, for instance, could come up as having a positive effect since one would have had to  58  have survived to 1994 in order to report for that wave. The effect of any term showing persistent loneliness between 1974 and 1994 would therefore have been suspect. The influence of the persistence of loneliness over time was therefore tested with GEE logistic regression models that were restricted to those still reporting data in 1974 and to person-years occurring in 1974 and beyond. Two indicators were created, one representing those regarding themselves as “Sometimes” lonely in both 1965 and 1974 and another for those regarding themselves as “Often” lonely in at least one wave and either “Sometimes” or “Often” in the other. Logistic models were then created testing the effects of these indicators, both alongside the most recent measure for loneliness and without. These results are shown in Table 12. The coefficient for the indicator representing those who were “Often” lonely in at least one wave was not significant in any of the models. It also changed signs when loneliness from the last wave was removed. The coefficient for the indicator representing those who were “Sometimes” lonely in both 1965 and 1974 was significant only in the models that included loneliness from the most recent wave and was also negative, appearing to be protective against mortality. One exception to these circumstances was found, however. An interesting pattern was seen during the exploratory exercises discussed below in the fourth phase of the thesis. Those under the age of 55 who reported feeling lonely “Sometimes” in both 1965 and 1974 were found to be at greater risk of death from injury (β = 1.432; p < 0.05). This may have corresponded to suicides, although the number of total deaths from injury in this age group was already quite small (N = 21) and therefore conclusions stemming from elaboration of this pattern might be suspect. Still, this pattern held despite controlling for  59  gender, education, and income. These models were carried out in Stata due to the complex filtering scheme used here.11 Loneliness and Age. Although the results are not recorded here, other GEE models were created that tested the effect of an interaction between loneliness and age. When loneliness from the last wave was also included (β = 0.417; p < 0.10), the loneliness X age interaction term was non-significant in both a preliminary model that used age and loneliness (β = -0.005; p > 0.122) and the full model that also included married status plus all of the control variables (β = -0.006; p > 0.055). When the loneliness term was removed, the interaction term switched signs and again was not significant (β = 0.001; p > 0.065). As an extra step, the interaction between age and persistent loneliness was tested. Person-years were again restricted to those occurring in 1974 or thereafter and among those for whom data were collected in 1974. Interaction terms were created between age and each of the two indicators measuring the stability of loneliness, i.e., those confirming “Sometimes” on two occasions and “Often” in at least one of the two. GEE models were then created, one with age and the two new interaction terms for age X persistent loneliness, and the other adding in the loneliness metric and the original indicators for persistent loneliness as well. For each of these cases married status and gender were entered in separate models as control variables. In none of these models were the new interaction terms significant.  11  Interestingly, Stata removed marital status in these models and remarked that it predicted death from injury “perfectly”. This was mysterious because 3 of the 21 individuals who had died of injury were not married. It could have been due to correlation with other variables, but since it was a low priority I declined to seek resolution of this problem.  60  Loneliness and Marriage. The influence of marital status was thoroughly explored, but mediating and moderating relationships were not seen while prospectively examining mortality. Moderating associations were examined by breaking the gendermarital status interaction into categories referring to unmarried men, unmarried women, etc., and then examining each group separately with GEE models. The logic here was to see if the odds ratios for loneliness varied by group. This yielded non-significant results for the loneliness metric. Since the effect was significant when the data was not spliced in any way, explanation of the loss in significance beyond the loss of power proved difficult. Similarly, using a four-level nominal variable describing people as unmarried men, unmarried women, married women, and married men with the latter of these as the reference group did not prove additionally beneficial versus using separate indicators for married status and gender. Nor did loneliness appear to mediate the relationship between mortality and either gender or married status in GEE logistic regression models (see Table 13). The coefficient for married status was significant both before (β = -0.101; p < 0.05) and after (β = -0.103; p < 0.05) the term for loneliness was introduced. The coefficient itself also did not change substantially. Gender, however, seemed to be something of a suppressor variable. P-values for all terms were noticeably smaller when gender was included. In particular the effect of married status intensified, with the odds of death decreasing by 12.7 points. Efforts to further illuminate this pattern proved difficult, however. No substantial differences in coefficients were seen between models created separately for men and for women. Nor was any distinct pattern seen when separate models were created for married  61  and for unmarried people, and nor was loneliness found to be a mediator when status as widowed, divorced, separated, or single was added. Generally, any attempt to focus upon particular segments of the population that were theoretically relevant (e.g., men vs. women, unmarried people, divorced or separated people) failed to shed any light by turning up significant results or by otherwise showing any interesting patterns. As a final attempt to examine how marital status may pattern health outcomes via loneliness, models were created that gauged the effect of change in marital status. Personyears were once again restricted to those occurring in 1974 and beyond among those who reported data in 1974. Marital change was used as the primary independent variable of interest, in the categories of “Divorced, Separated, or Widowed”, “Still unmarried”, and “Newly married”, and “Still married”, with the last of these as the reference category. Loneliness emerged as non-significant and again was not seen as a mediator, however status as “Still unmarried” appeared as its own risk factor against mortality (β = 0.137; p < 0.05) compared to being “Still married”. Summary. In the longitudinal phase of the study that encompassed all 34 years of the survey, loneliness was a specific risk factor for mortality (Tables 9 and 10). Loneliness remained significant even after all the control variables were entered. The effect size more than doubled, however, when person-years were restricted to those occurring within five years of the data collection. This suggests that the recency of those feelings matters. Loneliness may be a more imminent concern than recent literature has argued, but it also does have some stability in terms of health risks. The sensitive time frame may be more like five or six years than, say, twenty.  62  Parallel to this finding, broad support was not found for the Hawkley and Cacioppo’s hypothesis about how loneliness poses health risks. Loneliness is presumed to be a stable problem across the life span that more heavily impacts older adults in terms of health, but this was not corroborated by the results. For one, age was not found to be a substantial moderating factor when predicting mortality. Moreover, the coefficient for the age X loneliness interaction term occurred in a direction that was counterintuitive. Similarly, persistent loneliness if anything was shown to be protective against mortality. Statistical significance for its effect also depended on the coincident inclusion of shortterm fluctuations in loneliness. Loneliness that was persistent across time therefore did not seem to carry much comparative meaning in terms of mortality for Alameda County residents. Short-term fluctuations in loneliness appeared as more important in the models. One may then concede that loneliness is amenable to social circumstances as a health risk, but again explanations here proved elusive. There was not any clear interplay with marital status as might be expected with this view. In the main model predicting mortality, loneliness did not appear to mediate the effect for married status. Nor did it prove particularly beneficial to separate the various groups of married and unmarried people and compare their model coefficients. Additional analyses were undertaken in attempt to find some interplay among these variables but they failed to yield meaningful results.  63  5.4. Phase 4: Exploring Particular Fatalities Taking advantage of the ICD-9 codes, indicators were created identifying individuals who had died of various particular causes. These indicators were then multiplied by the indicator for deceased status, and the resulting value appeared as a 1 for those person-years in which an individual died of the given cause and 0 otherwise. Death from a cause other than the one of interest was coded as 0 and constituted the last row for that individual. Various GEE models were then created with these new indicators as binary dependent variables. Table 14 shows odds ratios belonging to the full models that describe each cause of death, while the modeling steps are described in entirety below. The following ICD-9 codes were used in order to configure the cause of death: 140 – 239 for cancer, 430 – 438 for cerebrovascular disease, 290 and 331 for Alzheimer’s disease / dementia, 390 – 459 for general circulatory system diseases, 410 – 414 for ischemic heart disease, and 001 – 139 for infectious and parasitic diseases. The indicator for ischemic heart disease death was subtracted from the indicator for general circulatory system deaths so as to create an indicator for general circulatory fatalities other than ischemic heart disease. Coded in this way, 200 deaths from cerebrovascular diseases were observed, 23 from infectious and parasitic diseases, 599 from cancer, 21 from dementia or Alzheimer’s disease, 61 from injury, and 1,242 from general circulatory diseases, 596 of which were ischemic heart disease and 646 all other problems of the circulatory system. As a variable intended to capture circulatory problems more generally, the last of these categories also included cerebrovascular diseases. In terms of the immediate effects on the body, of particular interest are the findings pertaining to altered circulatory functioning. Hawkley et al. (2003) argued that  64  cardiovascular profiles are different for lonely adults compared to non-lonely adults, characterized by higher basal total peripheral resistance and lower cardiac output. This creates a situation of chronically less blood flow, even after controlling for exercise activity and other behaviors understood to relate to circulatory functioning. It is plausible that this pattern in some way places an individual at risk of cardiovascular death, but exactly how this may be so has not yet been established. Fatality from ischemic heart disease and other circulatory diseases was therefore carefully tested. Models for ischemic heart disease did not yield significant results with respect to loneliness, but other circulatory diseases did. GEE models were created that separately added age, loneliness, and the various control variables. Loneliness (β = 0.185; p < 0.01) was significant when added to age (β = 0.112; p < 0.05). As terms were added the effect size diminished somewhat but retained significance. The odds ratio associated with the loneliness term in the full model came to 1.16 (β = 0.148; p < 0.01). As a subset of this category, models predicting cerebrovascular fatality were also examined. In the initial model containing only age and loneliness, the latter turned out non-significant (β = 0.199; p > 0.114). Regardless, the coefficient for loneliness began to approach significance after gender and married status were added (β = 0.236; p < 0.10). By the time all control variables were included, it almost passed the α = 0.05 threshold (β = 0.240; p > 0.056). Most interestingly, despite that only 23 people had died of infectious and parasitic diseases according to the chosen ICD-9 codes, loneliness appeared as a significant risk factor for this cause of death (β = 0.736; p < 0.01). This was true even after having controlled for gender, marriage, and all the other control variables, of which only income  65  approached significance as a protective factor (β = -.554 ; p > 0.072). None of these variables notably altered the coefficient for loneliness or its p-value. Since the literature suggests that loneliness may imply reduced immunity, death from cancer was tested as a dependent variable since some agents of the immune system (natural-killer cells) may be responsible for eliminating malignant tumors. This yielded non-significant results, however. Other authors (e.g., Wilson et al., 2007) have implicated Alzheimer’s disease and dementia as possible outcomes, but this also yielded nonsignificant results. Nor did loneliness altogether predict injury deaths, although as discussed above people under the age of 55 who reported being lonely “Sometimes” in both 1965 and 1974 were an exception to this rule. Suicide seemed plausible as a more precise reason for death, but for this age group there were only 4 confirmed suicides out of the total 21 injury deaths, which makes further explanation problematic. In summary, exploratory models demonstrated that loneliness was a precursor to certain specific causes of death. Recent literature contends that loneliness alters circulatory functioning and immune functioning, and indeed loneliness predicted circulatory diseases that were other than ischemic cardiac deaths. Tests predicting cerebrovascular deaths also approached significance. Furthermore, despite a small number of deaths due to infectious and parasitic diseases, loneliness was reliably predictive there as well. The argument that loneliness is a risk factor for specific causes of morbidity and mortality by changing immune and circulatory functioning therefore found support in the current study. Mentioned earlier was also a possible effect for death from injury among those younger than 55, which is clearly a psychosocial issue germane to health, although this  66  pattern could not be seen among the population at large. Similarly, clear patterns were not seen for dementia and Alzheimer’s disease, cancer, strokes, or ischemic heart disease. The reason for this may be a loss of power in predicting only a small number of deaths due to some of these causes, particularly in the case of Alzheimer’s disease. These have not been central concerns in the literature as health risks with respect to loneliness, however. No association was particularly expected for cancer, and one may argue that ischemic heart disease has more to do with the development of arterial plaques than to wear and tear of the body or accelerated age.  67  VI. Discussion This thesis has been my first attempt at professional scholarship in the social sciences and is also the product of two years of research aimed at completion of a Master of Arts degree in Family Studies. As there is more than one way with which I can ponder these findings, the final section of the thesis is organized into two parts. First I review the broad picture presented by the results, conceptualize what they may mean to scholars working in the area, discuss the limitations, and provide suggestions for further research. I then talk more candidly about the position that loneliness may have in the broader picture of social interaction in North America today. More specifically, I speculate what this subject area may mean to sociological thought. I do so partly in light of my own changing circumstances at the University of British Columbia. Broaching the discussion in this way, I hope to entertain what may be my more immediate steps as well as possible long-term career paths while at the same time offering a context with which to evaluate my academic achievement thus far. 6.1. Review For Alameda County residents, loneliness presented itself as a significant health risk. Cross-sectional tests using the 1965 data showed that loneliness was a specific risk factor for poor self-rated health. Indeed, loneliness was highly significant even after controlling for gender, married status, race, income, and education (Table 9). Control variables did not substantially weaken the p-value for the loneliness coefficient, nor did they do so for the effect itself. Throughout the longitudinal analyses as well, loneliness consistently emerged as a significant risk factor for mortality and for poor self-rated health. These results also help shed light upon the possible mechanisms linking loneliness  68  to poor health outcomes. Loneliness emerged as a risk factor for infectious and parasitic diseases as well as circulatory diseases are other than myocardial infarction. Strokes were additionally implicated. All of this agrees with the literature that has defined loneliness as a psychosocial factor that alters cardiovascular and immune processes. One important limitation to this study was the nature of the measure used to gauge loneliness, a simple question asking respondents how often they felt lonely. They answered “Never”, “Sometimes”, or “Often”, which created only a three-level ordinal response. This imposed some important problems. First, if the findings of this study are valid, then much of the variation in response to this item was lost. Respondents had the opportunity to answer this question only five times throughout the thirty-four years of the survey. A special arrangement was necessary in order to test the impact of the persistence of loneliness, and this process involved removing a sizeable portion of the population who had died (N = 649) or were censored between 1965 and 1974. Efforts to gauge the impact of the stability of loneliness over time were therefore seriously hampered, and this may have been why some results were unclear. There was even a twenty-year gap between two of the waves, 1974 and 1994, during which a lot may be assumed to have happened in the lives of the individuals studied. With the exception of mortality, much information about the changes in circumstances among Alameda County residents was lost. Clearly, this was due to limited research funds. An attempt was made in 1986 to re-contact respondents for another wave, but funds unexpectedly ran out. That said, the creation of more longitudinal designs involving more time points for data collection is much easier said than done, and to offer such a suggestion would be facile. Still, it is fair to say that conclusions of this study are  69  limited due to the infrequency and irregular intervals with which the data were gleaned. The unclear findings on the relationship between loneliness and marriage, as well as the stability of loneliness, may have been an artifact of these difficulties. Furthermore, if the sensitive time frame with which loneliness affects health is, say, five years, and given that the first three waves involved a 9-year and a 20-year gap, exactly how sensitive health may to feelings of loneliness remains unknown. Regarding the self-reported measure on loneliness, perhaps a larger problem is that women and men are understood to answer this type of question differently. Due to the special stigma implied by feelings of loneliness, men tend to underreport those feelings when asked in an explicit manner, while women over-report (Miller et al., 2007). This may mean that the gender differences seen throughout this study were artificial. Thus the suggestion that gender may be important when gauging the relationship among loneliness, marital status, and health may be in error. Other authors have argued that gender is important in predicting health (as was found in this study) and in the interplay with both loneliness and marriage. Still, results may have been markedly different were a special scale used to measure loneliness such as the UCLA Loneliness Scale. Another problem with the measure is that it was subject to interpretation. No time frame was specified with the question, which was simply “How often” respondents felt “Very lonely or remote from other people.” It seems somewhat unlikely that such a large proportion of the audience would “never” feel lonely. Moreover, any number of problems may have underlain responses of “often”, such as depression. This is further complicated by the fact that “or remote from other people” was tacked onto the question, and the impact of this part of the item on responses is unknown. Given that individuals may  70  construe this question however they may wish, variability in responses over time could be artificial: the mere interpretation of the question may have changed. Conclusions about the mutability of loneliness across time in relation to health are therefore subject to question. Despite all these concerns, though, there was strong evidence that the measure for loneliness was valid. Rigorous procedures were undertaken to validate the measure at the beginning of the study, and all of the expected patterns were seen. There is therefore opportunity to defend the findings of this study. While the more formal measures such as the UCLA Loneliness Scale may have been more powerful and more easily trusted, there is some benefit to asking about loneliness with the simple question used in this study. For one, the face validity was clear. Routasalo and Pitkala (2003) remarked that the UCLA Loneliness Scale and similar endeavors, on the other hand, may have biased the recent several years of literature toward regarding loneliness as a dimension of personality. Indeed, in the full 20-item scale, one question asks how often respondents feel “shy”, while another asks how often “outgoing and friendly”. Contrary to its possible aspect as a personality trait, this thesis presents some tension with the suggestion that loneliness stems from one’s genetic endowment, or that it is otherwise resistant to change. While loneliness may have some stability across time, even spanning several years, the argument that it is an altogether persistent problem was not supported in these results. Tau-b correlations showed that its reliability faded across time. Further exploration also showed that this was not simply because more people joined the population of lonely adults. Rather, there seemed to be movement in either direction. Some people became lonelier over the course of the survey, while a seemingly  71  equivalent group became less so, and others stayed the same, suggesting that it is more of a routine vicissitude than has been argued in some recent literature. Among the most interesting outcomes of this thesis were the models that explored loneliness in relation to marital status as shown in Tables 5 and 6 and especially the multinomial models predicting change in marital status in Table 7. These findings indicate that loneliness is not obstinate to change but instead seems to occur in flux with changing circumstances. While interpreting results for those who were newly married, one may even suppose that loneliness is a motivating factor prompting an individual to improve his or her social situation, although further tests would need to be carried out to corroborate this kind of claim. Namely, the group of respondents who remained unmarried during follow-up would need to be more clearly distinguished from those who were newly married. While on the one hand these findings indicate that escape from lonely feelings is possible, either on one’s own or perhaps through an improvement in social circumstances, on the other hand loneliness appears to be a more imminent health concern than may have been supposed in recent literature. The sensitive time frame may be closer to 5 or 6 years than to 20, as the odds of death were substantially larger when examining shorter time frames. Similarly, no moderating effect was found for age, suggesting that people at all stages of life who are lonely are vulnerable to significant health risks associated with loneliness. Due to the irregularity and infrequency of the loneliness measure, however, zeroing in on how imminent a threat against health loneliness may be was not possible in this study.  72  More difficult, however, would be the topic of divergent validity. Depression is understood to correlate well with loneliness, and indeed when the loneliness measure was validated, the strongest correlate was the measure asking how often respondents felt “depressed or very unhappy” (tau-b = .466). Time did not permit a fair treatment of depression as a possible confounding factor, in part because this would have been difficult with only another three-level measure. Cacioppo et al. (2006) remarked that divergent validity is more difficult to establish with such simple metrics, and that formal scales measuring both loneliness and depression are better suited for the purpose. Future research on this subject is clearly merited. This is the only study of its kind that I am aware of, suggesting that many of its conclusions could be premature. The irregular pattern and infrequency with which the survey was conducted proved problematic, as did the simplicity of the metric for loneliness upon which this thesis hinged. Moreover, there may be subtler ways to capture the effect of the stability of loneliness as accelerating bodily wear and tear that escaped my imagination. Also possible are confounding factors, especially given that such a simple measure was used in this thesis. Indeed, much remains to be done in order to characterize the precise nature and gravity of the health risks that loneliness imposes. 6.2. Reflections Since starting this project, my circumstances have changed. I have accepted the opportunity to pursue a doctoral degree in sociology at the University of British Columbia. In deference to all of those who have me helped me thus far towards this next goal, I will now speak candidly about my experience while contemplating the results presented in this thesis. I will also attempt to marshal the insights offered by the  73  sociological perspective. In so doing, my purpose is to allow the reader to evaluate the nature and quality of my academic achievement up to this point while at the same time enjoying a conversational tone. Perhaps more important is that by speaking frankly I hope to facilitate my own inductive reasoning. Is loneliness relevant to sociology? I can start by acknowledging that social isolation is being researched by sociologists today, even as a specific health concern. Putnam’s popular book Bowling Alone (2001) presents personal alienation as the antithesis of social capital. Pages upon pages of his book are devoted to the argument that social capital is slowly disintegrating throughout North America and thus social isolation is becoming more and more widespread. McPherson et al. (2006) contend that part of this picture is a diminishing number of personal ties, and that people are more and more turning to their spouses and families rather than friends to share their more intimate thoughts. This may put new pressure on North Americans to find spouses and create families on their own, which is further complicated by rising divorce rates and shrinking marriage rates. Berkman, Glass, Brissette, and Seeman (2000) in turn discuss how social isolation can compromise personal health. Yet they only mention loneliness in passing, regarding it as a “less devastating” problem (p. 853). All the same, if sociologists are largely concerned with ameliorating human suffering, if loneliness is an outcome on its own right in that regard and one that may be a growing problem, if it is a plausible explanation of how social isolation affects health, and if there is already evidence that it may bear its own direct impact, then by all accounts loneliness has much to do with sociology. This seems true all the way back to Durkheim’s studies of suicide.  74  How might loneliness be theoretically relevant? In a section entitled On Reason and Freedom, Mills (1959) offered the intriguing argument that people can be bereft of freedom without knowing it. Quite blissfully ignorant, these are people whose circumstances are unfairly delimited yet remain complacent. Mills further argued that to ameliorate the position of these individuals, we must help them think for themselves so that they can more properly evaluate the meaning of their roles as citizens, although better yet of course would be to prevent their problem of blissful ignorance from coming into being in the first place through public education. While arguing this point, Mills made a number of allusions to the book 1984 by George Orwell, which was a bleak portrait of a future world in which people are brutally subjugated but come to be happy in their extremely constrained spaces. In this picture I think Mills is missing a crucial aspect of the human experience, and one that can easily compromise both reason and freedom: emotion. Loneliness was the particular emotion (or constellation of emotions) studied in this thesis, and I believe that it is one reason the world depicted by George Orwell may never happen. In 1984 the main character lived alone, worked alone, had no spouse, had no choice but to associate with vapid people, and was forbidden any romantic engagement. The findings presented in my thesis show this to be quite a lonely existence. Yet the main character miraculously walked away happy at the end of the story. If textbooks describe human beings as inherently social in nature (Miller et al., 2007), then loneliness is tenably among the natural outcomes when social life is constrained. Other authors describe loneliness as stemming from a paucity of meaningful exchanges with others (Wheeler et al., 1983). At least in the context of a totalitarian regime that  75  constrains social interaction, the “cheerful robot” as Mills (1959) describes him would not be possible. Still, on close inspection even our conception of loneliness is a surprisingly elusive idea. There are some curiosities about the word “lonely” worth mentioning here. First, it is a relatively recent addition to the English lexicon. Although he may not have been the first to use it, Shakespeare is credited with contributing the first known usage in a seventeenth-century play. Yet “lonely” carried a neutral meaning at that time, not the largely negative, unpleasant connotations that it carries today. This sense of the word came into English much later, in nineteenth century. Second, the concept of loneliness does not always translate well into other languages. This is at least true of Spanish. In order to capture the unpleasant sense as opposed to mere solitude, one must say “alone and sad” (solo y triste). There also does not exist an equivalent noun for loneliness. Whenever I talk about the subject with native Spanish speakers, they consistently have great difficulty understanding what “lonely” means in English, tending to confuse it with mere solitude. Interestingly, to my memory the only Spanish-language articles discussing loneliness that I have seen have come from researchers in Spain. Consideration of these points prompts the question of whether this is a phenomenon of the developed world. Is loneliness the price of economic success? Putnam (2000) would seem to agree. Pivotal to his arguments about the diminishment of social capital is the technology that is made easily available to consumers. With the advent of the television, people could be entirely entertained in the comfort of their own homes. Today we must add Netflix to this picture, along with online purchases, iTunes with earplugs, and many other inventions meant to make life more  76  convenient that also make social intercourse more unnecessary. We also have highly sophisticated computer and video games. Unlike a century ago, one no longer needs to leave the home in order to enjoy an evening of entertainment. In large cities, there is the additional problem of anonymity. With transport available over a large, densely populated area, it is quite easy to walk into a store, purchase an item, leave, and then never see anyone in that store again. This is an advantage for someone who is shy or who has misanthropic tendencies, but is quite a disadvantage for someone who wishes to make a connection. Anonymity is a doubleedged sword. If a young man is upset with a purchase or with a fellow customer who cuts in line in front of him, he can issue a stream of invective and then leave, never fearing a reprisal. Yet other people can do the same to him as they see fit because they are equally unlikely to cross paths with him in the foreseeable future. The problem, then, is “other people”. A new resident of a suburban community may find it quite difficult to carry on in that community because her neighbors have already found ways to entertain themselves in their own homes, making her unnecessary. Surround her with city dwellers who quickly disappear from her life as she carries out her daily tasks, then give the more familiar people iTunes with earplugs or a laptop from which to watch movies, and she may find few people to talk to. Another problem is the commuter lifestyle and the ease of transportation that coincides with it. Miller et al. (2007) argue that repeated contact with known persons is important to forming friendships and other relationships, yet this may not be easy in large cities whose denizens have a habit of quickly disappearing. Indeed, North American culture is a culture of mobility, quite hospitable to one wishing to move to a distant place  77  for school or for a job. Yet Miller et al. also regard long distances as troubling for couples, potentially threatening the stability of their more intimate relationships. Clearly, culture is part of this discussion as well. That Latin American culture can exude a sense of welcome to strangers and among its own members may explain why the word for loneliness does not translate well into Spanish. This line of thought could help shed light on why some societies may be lonelier than others, although cultural differences may also play a role in structuring social isolation. Integration into a social clique likely entails some familiarity with the culture known to those people. The less common culture is shared, the less may be the likelihood of two groups integrating. This is yet one more barrier to social cohesion in such multicultural, individualized, and highly mobile societies as the United States and Canada. If sociology has tended to shy away from psychosocial explanations of human outcomes, then I would offer encouragement against this pattern. Berkman and Syme (1979) first sought to explain why social isolation structures poor health outcomes a quarter century ago. An important part of how social scientists and epidemiologists describe causality is through a plausible explanation of why one thing may lead to another. Although loneliness has not yet efficiently explained why objective social isolation structures poor health, it does appear to be one thread through which poor outcomes happen. As discussed above it may also have bearing on how people are adjusting in post-industrial societies. Loneliness even seems to offer substance to the discussion of cultural consumption and consumerism, for instance when arguing that feelings of alienation stem from a paucity of meaningful social exchanges. One may even venture to argue that consumerism could lend to a finicky attitude toward the selection of  78  friends and romantic partners, which would further diminish the opportunities available to those who are looking for fulfillment in their social lives. I could go on, but this kind of reflection already points to a number of possible research directions. I would for example be curious to see how loneliness fluctuates with population size and density. In the meantime, were I to further pursue this kind of inquiry, I would bear in mind that in addition to wealth and education people may have such a thing as richness or poverty in their social lives. The poorest of the poor not only have empty wallets but also an empty house with few visitors.  79  Figure 1 Loneliness by Age  1.65 1.6  Loneliness  1.55 1.5 1.45 1.4 1.35 1.3 21 - 30  31 - 40  41 - 50  51 - 60  61 - 70  71 - 80  81+  Age  80  Table 1 Responses for Loneliness and Marital Status 1965  1974  1994  Lonely or remote from others? Never Sometimes Often  51.6% (3460) 43.3% (2902) 5.1% (344)  46.5% (2176) 46.5% (2175) 7.0% (326)  50.7% (1328) 44.6% (1169) 4.7% (122)  Marital Status Married Separated Divorced Widowed Single  74.7% (5070) 2.6% (176) 6.0% (404) 7.1% (484) 9.6% (655)  75.1% (3571) 2.9% (138) 8.3% (397) 9.2% (437) 4.5% (212)  69.2% (1831) 1.9% (49) 12.4% (329) 13.7% (362) 2.8% (74)  1995 53.1% (1321) 42.8% (1066) 4.1% (103)  1999 54.3% (1112) 41.8% (857) 3.9% (79)  68.5% (1410) 1.0% (20) 12.3% (254) 15.9% (328) 2.2% (45)  81  Table 2 Sociodemographic Information for Alameda County Residents in 1965 Gender  N  %  Male Female Total  3110 3679 6789  45.8 54.2 100.0  Missing  0  Age 21 - 30 31 - 40 41 - 50 51 - 60 61 - 70 71 - 80 81+ Total Missing  Race White Black American Indian Chinese Japanese Filipino Asian, Not elsewhere classified Hispanic Other Total Missing  N  %  1645 1422 1491 1030 692 393 116 6789  24.2 20.9 22.0 15.2 10.2 5.8 1.7 100.0  Cumulative % 45.8 100.0  Cumulative % 24.2 45.2 67.1 82.3 92.5 98.3 100.0  0  N  %  5351 851 85 107 75 40 6 264 8 6787  78.8 12.5 1.3 1.6 1.1 0.6 0.1 3.9 0.1 100.0  Cumulative % 78.8 91.4 92.6 94.3 95.4 95.9 96.0 99.9 100.0  2  82  Table 2 (Cont’d)  Total income before taxes Under $1000 $1000-1999 $2000-2999 $3000-3999 $4000-4999 $5000-5999 $6000-6999 $7000-7999 $8000-9999 $10000-14999 $15000-24999 $25000 and over Total Missing  Education No school Completed 1 year grade school Completed 2 years grade school Completed 3 years grade school Completed 4 years grade school Completed 5 years grade school Completed 6 years grade school Completed 7 years grade school Completed grade school Completed 1 year high school Completed 2 years high school Completed 3 years high school Graduated high school Completed 1 year college Completed 2 years college Completed 3 years college Graduated college Completed 5+ years of college Completed advanced college degree Total Missing  N  %  108 248 275 332 424 509 648 653 1121 1500 436 177 6431  1.7 3.9 4.3 5.2 6.6 7.9 10.1 10.2 17.4 23.3 6.8 2.8 100.0  Cumulative % 1.7 5.5 9.8 15.0 21.6 29.5 39.6 49.7 67.1 90.5 97.2 100.0  358  N  %  51 17 23 69 80 94 144 168 595 273 510 409 2040 462 515 252 464 313 280 6759  0.8 0.3 0.3 1.0 1.2 1.4 2.1 2.5 8.8 4.0 7.5 6.1 30.2 6.8 7.6 3.7 6.9 4.6 4.1 100.0  Cumulative % 0.8 1.0 1.3 2.4 3.6 4.9 7.1 9.6 18.4 22.4 29.9 36.0 66.2 73.0 80.6 84.4 91.2 95.9 100.0  30  83  Table 3 Kendall's Tau-b Coefficients for Reports of Loneliness 1965 1965  1974  1994  1995  1999  0.333 (2070)  0.270 (1128)  0.251 (1072)  0.221 (889)  0.305 (1130)  0.241 (1074)  0.270 (890)  0.504 (1068)  0.423 (886)  1974  0.374 (2575)  1994  0.244 (1481)  0.308 (1479)  1995  0.277 (1409)  0.316 (1408)  0.492 (1397)  1999  0.215 (1152)  0.257 (1152)  0.414 (1145)  0.448 (888) 0.482 (1150)  Note. Correlations for women are on the bottom-left and for men on the top-right. All correlations were significant at p < 0.01 (2-tailed). Ns are listed in parentheses.  Table 4 Change in Loneliness Over Time 1965 - 1974 1974 - 1994 1994 - 1995 1995 - 1999 Worse 22.2% 16.5% 13.5% 14.9% Same 28.8% 28.2% 31.0% 28.8% Better 15.7% 25.6% 15.9% 15.3% Never lonely 33.3% 29.7% 39.5% 41.0% N  4645  2609  2465  2038  84  Table 5 Negative-Log Ordinal Regression Models Predicting Loneliness at Time 2 MEN  WOMEN Model 1  β  Model 2  Model 1  β  β  Model 2  β  1965 - 1974 Married T1 Married T2  -0.294 ***  -0.097 -0.443 ***  1965 - 1974 Married T1 Married T2  -0.002  0.078 -0.128 ?  1974 - 1994 Married T1 Married T2  -0.382 **  -0.192 -0.409 ***  1974 - 1994 Married T1 Married T2  -0.251 **  -0.111 -0.270 **  1994 - 1999 Married T1 Married T2  -0.226 ?  0.498 * -0.941 ***  1994 - 1999 Married T1 Married T2  -0.247 **  0.148 -0.467 **  *** p < 0.001 ** p < 0.01 * p < 0.05 ? p < 0.10  Table 6 Logistic Regression Models Predicting Status as Married at Time 2  MEN  WOMEN Model 1  β  Model 2  Model 1  β  β  Model 2  β  1965 - 1974 Loneliness T1 Loneliness T2  -0.531 ***  -0.371 ** -0.470 ***  1965 - 1974 Loneliness T1 Loneliness T2  -0.296 ***  -0.271 *** -0.063  1974 - 1994 Loneliness T1 Loneliness T2  -0.280 *  -0.116 -0.547 ***  1974 - 1994 Loneliness T1 Loneliness T2  -0.168 ?  -0.048 -0.391 ***  1994 - 1999 Loneliness T1 Loneliness T2  -0.470 **  -0.162 -0.696 ***  1994 - 1999 Loneliness T1 Loneliness T2  -0.364 ***  -0.204 ? -0.388 **  *** p < 0.001 ** p < 0.01 * p < 0.05 ? p < 0.10  85  Table 7 Multinomial Regression Models Predicting Marital Change Status as Divorced, Separated, or Widowed MEN  Model 1  β 1965 - 1974 Loneliness T1 Loneliness T2  0.190  1974 - 1994 Loneliness T1 Loneliness T2 1994 - 1999 Loneliness T1 Loneliness T2  Model 2  WOMEN  β  Model 1  β  -0.018 0.570 ***  1965 - 1974 Loneliness T1 Loneliness T2  0.395 ***  0.176  0.010 0.562 **  1974 - 1994 Loneliness T1 Loneliness T2  0.104  -0.106  -0.815 * 1.441 ***  1994 - 1999 Loneliness T1 Loneliness T2  0.349 ?  Model 2  β 0.302 ** 0.232 *  -0.014 0.386 ***  0.074 0.659 **  Status as Still Unmarried MEN  Model 1  β 1965 - 1974 Loneliness T1 Loneliness T2  0.941 ***  1974 - 1994 Loneliness T1 Loneliness T2 1994 - 1999 Loneliness T1 Loneliness T2  Model 2  WOMEN  β  Model 1  β  Model 2  β  0.823 *** 0.353 **  1965 - 1974 Loneliness T1 Loneliness T2  0.348 ***  0.373 *** -0.063  0.594 **  0.430 * 0.542 **  1974 - 1994 Loneliness T1 Loneliness T2  0.357 **  0.231 ? 0.413 **  0.658 ***  0.493 * 0.382 *  1994 - 1999 Loneliness T1 Loneliness T2  0.397 ***  0.270 * 0.307 *  Status as Newly Married MEN  Model 1  β 1965 - 1974 Loneliness T1 Loneliness T2  0.819 ***  1974 - 1994 Loneliness T1 Loneliness T2  0.920 ***  1994 - 1999 Loneliness T1 Loneliness T2  -0.403  Model 2  WOMEN  β  Model 1  β  Model 2  β  0.875 *** -0.190  1965 - 1974 Loneliness T1 Loneliness T2  0.637 ***  0.677 *** -0.102  0.894 *** 0.095  1974 - 1994 Loneliness T1 Loneliness T2  0.593 **  0.567 * 0.086  1994 - 1999 Loneliness T1 Loneliness T2  1.017 *  1.232 ** -0.530  -0.048 -1.039 *  *** p < 0.001 ** p < 0.01 * p < 0.05 ? p < 0.10  86  Table 8 Sample Sizes for Models Testing Social Causation Versus Social Selection Predicting Loneliness at Time 2 (Table 5)  1965 - 1974 1974 - 1994 1994 - 1999  Men 2067 1130 883  Women 2592 1481 1155  Predicting Married Status at Time 2 (Table 6)  1965 - 1974 1974 - 1994 1994 - 1999  Men 2079 1134 889  Women 2572 1478 1145  Predicting Marital Change at Time 2 (Table 7)  1965 - 1974 1974 - 1994 1994 - 1999  Men 2067 1130 883  Women 2572 1476 1144  Table 9 Coefficients for Ordinal Probit Regression Models Gauging Self-Rated Health in 1965 (N = 6327) Model 1 Self-rated health: Poor Fair Good Excellent Age Married Gender Married X Gender Loneliness Age X Loneliness Race / Ethnicity: Black Other Non-White Education Income  Model 2  Model 3  Model 4  Model 5  Model 6  Model 7  -2.882 *** -1.655 *** -0.013  -2.831 *** -1.602 *** 0.040  -2.795 *** -1.566 *** 0.079  -2.835 *** -1.605 *** 0.040  -3.640 *** -2.371 *** -0.686 ***  -3.513 *** -2.241 *** -0.557 ***  -2.632 *** -1.303 *** 0.448 **  -0.016 ***  -0.016 *** 0.058 ?  -0.016 *** 0.042 0.112 ***  -0.016 *** 0.000 0.029 0.107  -0.018 *** -0.063 0.034 0.054 -0.381 ***  -0.015 *** -0.063 0.034 0.054 -0.299 *** -0.002  -0.011 *** -0.113 * 0.031 0.037 -0.330 *** -0.001 -0.348 -0.250 0.061 0.026  *** p < .001. ** p < .01. * p < .05. ? p < 0.10  87  *** *** *** ***  Table 10 Odds Ratios for Logistic Regression Models Predicting All-Cause Mortality (N = 149382) Model 1 Age 1.096 *** Loneliness Gender (M=1; F=0) Married Race = Black Race = Other non-White Education Income  Model 2 1.096 *** 1.091 *  Model 3 1.099 *** 1.126 ** 1.489 ***  Model 4 1.096 *** 1.102 ** 1.595 *** 0.785 ***  Model 5 1.096 *** 1.105 ** 1.590 *** 0.798 *** 1.268 ** 0.971  Model 6 1.094 *** 1.094 * 1.595 *** 0.808 *** 1.144 ? 0.902 0.964 ***  Model 7 1.091 *** 1.087 * 1.628 *** 0.876 * 1.110 0.897 0.974 *** 0.871 ***  *** p < .001. ** p < .01. * p < .05. ? p < 0.10  Table 11 Odds Ratios for Models Restricted to Person Years Within 5 Years of the Most Recent Wave  Age Loneliness Married Gender (M=1; F=0) Race = Black Race = Other non-White Education Income  Loneliness and Age (N = 83260) 1.090 *** 1.202 ***  Full model (N = 79043) 1.085 *** 1.200 ** 0.858 ? 1.547 1.057 0.982 0.972 * 0.890 *  *** p < 0.001 ** p < 0.01 * p < 0.05 ? p < 0.10  88  Table 12 Odds Ratios for Models Testing the Persistence of Loneliness  Age Loneliness Persistently lonely - Sometimes Persistently lonely - Sometimes, Often Gender (M=1; F=0) Married  (N = 98036) Model 1 Model 2 1.102 *** 1.102 *** 1.137 ** 1.142 ** 0.851 * 0.881 ? 0.932 0.954 1.575 *** 0.797 ***  (N = 98137) Model 3 Model 4 1.101 *** 1.101 *** 0.914 1.082  0.949 1.111 1.569 *** 0.789 ***  *** p < 0.001 ** p < 0.01 * p < 0.05 ? p < 0.10  Table 13 Odds Ratios for Models Examining Married Status, Gender, and Loneliness (N = 156828)  Age Married Gender (M=1; F=0) Loneliness  Model 1 1.095 ** 0.891 *  Model 2 1.095 ** 0.903 1.075 *  Model 3 1.095 *** 0.763 *** 1.600 ***  Model 4 1.096 0.776 1.610 1.097  *** *** ** ***  *** p < 0.001 ** p < 0.01 * p < 0.05 ? p < 0.10  89  1.092 0.986 2.179 0.934 0.754 0.827 0.959 0.813 *** ***  ?  ***  ***  1.110 1.160 1.326 0.739 1.294 0.860 0.980 0.899 ?  *** * ** ** ?  1.130 1.271 1.336 1.134 1.502 1.015 1.006 1.022 ?  *** ? ?  Other general circulatory Cerebrovascular (567) (180) 1.078 *** 2.088 ** 1.669 2.094 0.829 1.758 0.985 0.575  Infection / Parasites (21) 1.074 *** 0.941 1.425 *** 1.067 1.208 1.026 0.969 * 0.954  Cancer (535)  *** p < 0.001 ** p < 0.01 * p < 0.05 ? p < 0.10 Note. Total number of deaths are included in parentheses beneath the name of the respective disease.  Age Loneliness Gender (M=1; F=0) Married Race = Black Race = Other non-White Education Income  Ischemic heart disease (521)  Odds Ratios for Models Predicting Particular Causes of Death (N = 149,382)  Table 14  1.188 *** 1.237 0.467 2.423 * 1.768 1.752 0.937 0.935  Dementia / Alzheimer's (20)  1.041 *** 1.112 3.304 *** 0.357 *** 1.160 0.860 0.960 1.372  Injury (57)  90  References Adler, N. E., Boyce, T., Chesney, M. A., Cohen, S., Folkman, S., Kahn, R. L., & Syme, S. L. (1994). Socioeconomic status and health: The challenge of the gradient. American Psychologist, 49, 1, 15-24. Agresti, A., & Finlay, B. (1997). Statistical methods for the social sciences (3rd ed.). Upper Saddle River, NJ: Prentice Hall, Inc. Anderson, C. A., & Harvey, R. J. (1988). Discriminating between problems in living: An examination of depression, loneliness, shyness, and social anxiety. Journal of Social and Clinical Psychology, 6, 482-491. Berkman, L. F., Glass, T., Brissette, I., & Seeman, T. E. (2000). From social integration to health: Durkheim in the new millennium. Social Science and Medicine, 51, 843-857. Berkman, L. F., & Syme, S. L. (1979). Social networks, host resistance, and mortality: A nine-year follow-up study of Alameda County residents. American Journal of Epidemiology, 109, 186-204. Bradburn, N. (1969). The structure of psychological well-being. Chicago: Aldine. Brannon, L., & Feist, J. (2007). Health psychology: An introduction to behavior and health (6th Ed.). Belmont, CA: Thomson Wadsworth. Boomsma, D. I., Cacioppo, J. T., Muthén, B., Asparouhov, T., & Clark, S. (2007). Lognitudinal Genetic Analysis for Loneliness in Dutch Twins. Twin Research and Human Genetics, 10, 2, 267-273. Boomsma, D. I., Cacioppo, J. T., Slagboom, P. L., & Posthuma, D. (2006). Genetic linkage and association analysis for loneliness in Dutch twin and sibling pairs points to a region on chromosome 12q23-24. Behavior Genetics, 36, 1, 137-146. Boomsma, D. I., Willemse, G., Dolan, C. V., Hawkley, L. C., & Cacioppo, J. T. (2005). Genetic and environmental contributions to loneliness in adults: The Netherlands Twin Register Study. Behavior Genetics, 35, 745-752. Cacioppo, J. T., Hawkley, L. C., Crawford, L. E., Ernst, J. M., Burleson, M. H., Kowalewski, R. B., Malarkey, W. B., Cauter, E. V., & Berntson, G. G. (2002). Loneliness and health: Potential mechanisms. Psychosomatic Medicine, 64, 407417. Cacioppo, J. T., Hughes, M. E., Waite, L. J., Hawkley, L. C., & Thisted, R. A. (2006). Loneliness as a specific risk factor for depressive symptoms: Cross-sectional and longitudinal analyses. Psychology and Aging, 21, 1, 140-151. Callaghan, P., & Morrissey, J. (1993). Social support and health: A review. Journal of Advanced Nursing, 18, 203-210. Cole, S. W., Hawkley, L. C., Arevalo, J. M., Sung, C. Y., Rose, R. M., & Cacioppo, J. T. (2007). Social regulation of gene expression in human leukocytes. Genome Biology, 8, 9, R189. Conroy, R. W., & Smith, K. (1983). Family loss and hospital suicide. Suicide and LifeThreatening Behavior, 13, 179-194. Cutrona, C. E., Russell, D. W., de la Mora, A., & Wallace, R. B. (1997). Loneliness and nursing home admissions among rural older adults. Psychology and Aging, 12, 574-589.  91  Dixon, D., Cruess, S., Kilbourn, K., Klimas, N., & Fletcher, M. A. (2001). Social support mediates loneliness ad human herpes virus Type 7 (HHV-6) antibody titers. Journal of Applied Social Psychology, 31, 1111-1132. Duck, S., Pond, K., & Leatham, G. (1994). Loneliness and the evaluation of relational events. Journal of Social and Personal Relationshps, 11, 253-276. Dykstra, P. A., van Tilburg, T. G., & de Jong Gierveld, J. (2005). Chanes in older adult loneliness: Results from a seven-year longitudinal study. Research on Aging, 27, 6, 725-747. Ernst, J. M., & Cacioppo, J. T. (1999). Lonely hearts: Psychological perspectives on loneliness. Applied & Preventive Psychology, 8, 1-22. Forlenza, J. J., & Baum, A. (2004). Psychoneuroimmunology. In R. G. Frank, A. Baum, & J. L. Wallander (Eds.), Handbook of clinical health psychology (Vol. 3, pp. 81114). Washington, DC: American Psychological Association. Hawkley, L. C., Burleson, M. H., Berntson, G. G., & Cacioppo, J. T. (2003). Loneliness in everyday life: Cardiovascular activity, psychosocial context, and health behaviors. Journal of Personality and Social Psychology, 85, 1, 105-120. Hawkley, L. C., & Cacioppo, J. T. (2003). Loneliness and pathways to disease. Brain, Behavior, and Immunity, 17, S98-S105. Hawkley, L. C., & Cacioppo, J. T. (2007). Aging and loneliness: Downhill quickly? Current Directions in Psychological Science, 16, 4, 187-191. Hawkley, L. C., Masi, C. M., Berry, J. D., & Cacioppo, J. T. (2006). Loneliness is a unique predictor of age-related differences in systolic blood pressure. Psychology and Aging, 21, 1, 152-164. Heinrich, L. M., & Gullone, E. (2006). The clinical significance of loneliness: A literature review. Clinical Psychology Review, 26, 695-718. Herlitz, J., Wiklund, I., Caidahl, K., Hartford, M., Haglid, M., Karlsson, B. W., Sjöland, H., & Karlsson, T. (1998). The feeling of loneliness prior to coronary artery bypass grafting might be a predictor of short- and long-term postoperative mortality. European Journal of Vascular and Endovascular Surgery, 16, 120-125. Hoetker, G. (2007). The use of logit and probit models in strategic management research: Critical issues. Strategic Management Journal, 28, 331 - 343. House, J. S., Landis, K. R., & Umbarson, D. (1988). Social relationships and health. Science, 241, 4865, 540-545. Hughes, M. E., Waite, L. J., Hawkley, L. C., & Cacioppo, J. T. (2004). A short scale for measuring loneliness in large surveys: Results from two population-based studies. Research on Aging, 26, 6, 655-672. Kubzansky, L. D., & Kawachi, I. (2000). Affective states and health. In L. F. Berkman & I. Kawachi (Eds.), Social epidemiology (pp. 213-241). New York: Oxford University Press. McPherson, M., Smith-Lovin, L., & Brashears, M. E. (2006). Social isolation in America: Changes in core discussion networks over two decades. American Sociological Review, 71, 353-357. Miller, R. S., Perlman, D., & Brehm, S. S. (2007). Intimate relationships. New York: McGraw-Hill. Mills, C. R. (1959). The sociological imagination. New York: Oxford University Press.  92  Milsom, J., Beech, A. R., & Webster, S. D. (2003). Emotional loneliness in sexual murderers: A quantitative analysis. Sexual Abuse: Journal of Research and Treatment, 15, 285-296. Mullins, L. C., Smith, R., Colquitt, R., & Mushel, M. (1996). An examination of the effects of self-rated and objective indicators of health condition and economic condition on the loneliness of older persons. Journal of Applied Gerontology, 15, 1, 23-37. Patterson, A. C. (2007). The gloom-tinted lens: Hopelessness and boredom as mediators of loneliness. Unpublished manuscript, University of British Columbia, Vancouver, Canada. Peck, D. L. (1983). The last moments of life: Learning to cope. Deviant Behavior, 4, 313332. Perlman, D. (1991). Age differences in loneliness: A meta-analysis. Vancouver, Canada: University of British Columbia. (ERIC Document Reproduction Service No. ED 326767). Perlman, D. (2004). European and Canadian studies of loneliness among seniors. Canadian Journal on Aging, 23, 2, 181-188. Perlman, D., & Peplau, L. A. (1981). Toward a social psychology of loneliness. In S. Duck & R. Gilmour (Eds.), Personal relationships. 3: Personal relationships in disorder (pp. 31-56). New York: Academic Press. Pressman, S. D., Cohen, S., Miller, G. E., Barkin, A., Rabin, B., & Treanor, J. J. (2005). Health Psychology, 24, 3, 297-306. Proulx, J., McKibben, A., & Lusignan, R. (1996). Relationships between affective components and sexual behaviors in sexual aggressors. Sexual Abuse: Journal of Research and Treatment, 8, 4, 279-289. Putnam, R. (2000). Bowling alone: The collapse and revival of American community. New York: Simon & Schuster. Routasalo, P., & Pitkala, K. H. (2003). Loneliness among older people. Reviews in Clinical Gerontology, 13, 303-311. Russell, D. W. (1996). The UCLA Loneliness Scale (Version 3): Reliability, validity, and factorial structure. Journal of Personality Assessment, 66, 20-40. Simpson, J. A., & Weiner, E. S. C. (Eds.). (1989). The Oxford English Dictionary (2nd ed., Vol. 8). Oxford: Clarendon Press. Sorkin, D., Rook, K. S., & Lu, J. L. (2002). Loneliness, lack of emotional support, lack of companionship, and the likelihood of having a heart condition in an elderly sample. Annals of Behavioral Medicine, 24, 4, 290-298. Steed, L., Boldy, D., Grenade, L., & Iredell, H. (2007). The demographics of loneliness among older people in Perth, Western Australia. Australasian Journal on Ageing, 26, 2, 81-86. Troxel, W. M., Matthews, K. A., Gallo, L. C., & Kuller, L. H. (2005). Marital quality and occurrence of the metabolic syndrome in women. Archives of Internal Medicine, 165, 1022-1027. Uchino, B. N. (2004). Social support and physical health: Understanding the health consequences of relationships. New Haven, CT: Yale University Press. Weiss, R. S. (1973). Loneliness. Cambridge, MA: MIT Press.  93  Wenger, G. C., & Burholt, V. (2004). Changes in levels of social isolation and loneliness among older people in a rural area: A twenty-year longitudinal study. Canadian Journal on Aging, 23, 2, 115-127. Wethington, E., & Kessler, R. C. (1986). Perceived support, received support and adjustment to stressful life events. Journal of Health and Social Behavior, 28, 306-319. Wheeler, L., Reis, H., & Nezlek, J. (1983). Loneliness, social interaction, and sex roles. Journal of Personality and Social Psychology, 45, 4, 943-953. Wilson, R. S., Krueger, K. R., Arnold, S. E., Schneider, J. A., Kelly, J. F., Barnes, L. L., Tang, Y., & Bennett, D. A. (2007). Loneliness and risk of Alzheimer disease. Archives of General Psychiatry, 64, 234-240.  94  

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.24.1-0066582/manifest

Comment

Related Items