This pattern is reversed for appraisals of health status, for which physical functioning is more important ...... American Life: Perceptions, Evaluations, and Satis-.
Quality of Life Research 8: 447±459, 1999. Ó 1999 Kluwer Academic Publishers. Printed in the Netherlands.
Distinguishing between quality of life and health status in quality of life research: A meta-analysis Kevin W. Smith, Nancy E. Avis & Susan F. Assmann New England Research Institutes
Accepted in revised form 7 April 1999
Abstract. Despite the increasing acceptance of quality of life (QOL) as a critical endpoint in medical research, there is little consensus regarding the de®nition of this construct or how it diers from perceived health status. The objective of this analysis was to understand how patients make determinations of QOL and whether QOL can be dierentiated from health status. We conducted a meta-analysis of the relationships among two constructs (QOL and perceived health status) and three functioning domains (mental, physical, and social functioning) in 12 chronic disease studies. Instruments used in these studies included the RAND-36, MOS SF-20, EORTC QLQ-30, MILQ and MQOL-HIV. A single, synthesized correlation matrix combining the data from all 12 studies was estimated by generalized least squares. The synthesized matrix was then used to
estimate structural equation models. The meta-analysis results indicate that, from the perspective of patients, QOL and health status are distinct constructs. When rating QOL, patients give greater emphasis to mental health than to physical functioning. This pattern is reversed for appraisals of health status, for which physical functioning is more important than mental health. Social functioning did not have a major impact on either construct. We conclude that quality of life and health status are distinct constructs, and that the two terms should not be used interchangeably. Many prominent health status instruments, including utility-based questionnaires and health perception indexes, may be inappropriate for measuring QOL. Evaluations of the eectiveness of medical treatment may dier depending on whether QOL or health status is the study outcome.
Key words: Health status, Meta-analysis, Quality of life Introduction Quality of life (QOL) has now become ®rmly established as an important endpoint in medical care [1]. This is especially true of chronic diseases for which a cure is unlikely. However, despite its burgeoning popularity as an outcome measure, research continues to be hampered by a lack of conceptual clarity regarding precisely what is meant by QOL [2±5]. In a review of 75 articles purporting to evaluate QOL, Gill and Feinstein [6] found that only 15% of these papers provided conceptual de®nitions. All too often, QOL is used as a generic label for an assortment of physical functioning and psychosocial variables. Improvements over time in any of these variables may be proclaimed to be evidence of improved `quality of life'. Investigators frequently use the constructs `quality of life' and `health status' interchangeably [7, 8]. The World Health Organization's de®nition of health as ``a state of complete physical, mental and social wellbeing, and not merely the absence of disease or in®rmity'' also serves as a starting point for de®ning QOL. Any distinction between those two constructs is further obscured by references to `health-related' QOL. This term originated to distinguish outcomes relevant to health research from earlier sociological
research on subjective well-being and life satisfaction in healthy, general populations [9]. None of the articles reviewed by Gill and Feinstein [6] attempted to clarify the distinction between overall and health-related QOL. Some instruments go so far as to merge the two constructs. The European Organization for Research and Treatment of Cancer-Core Quality of Life Questionnaire (EORTC QLQ-30) [10], for example, asks patients to make separate assessments of QOL and physical condition (using a 7-point scale ranging from very poor to excellent) and then combines the two ratings into a single overall QOL score. Are quality of life and health status essentially the same construct or do they represent dierent constructs? In the absence of a conceptual consensus, one way of approaching this issue is to determine whether QOL and health status mean the same thing to patients with chronic diseases. Our objective in this paper is to re-analyze existing QOL studies in an effort to understand how patients make determinations of QOL and to determine whether QOL can be differentiated from health status. We begin by presenting a structural model for QOL and health status. The relationships between individual life domains and the two constructs in this model are then analyzed separately for 12 studies. Finally, we use
448 meta-analysis to merge the results of these dierent studies and estimate structural models of QOL and health status. Structural equation model for quality of life Structural equation models of health status or quality of life are a comparatively recent development in the research literature. Hays and Stewart have explored the structure of self-reported health in the Medical Outcomes Study [11]. Johnson and Wolinsky developed a model linking disease, disability and functional limitations to perceived health status in older adults [12]. Molzahn et al. [13] examined the impact of support, outlook, health status and functional status on the QOL of patients with end stage renal disease. Fayers et al. [14] describe a causal modeling approach to QOL, stressing the importance of distinguishing between causal indicators and eect indicators. Wilson and Cleary [15] described a comprehensive conceptual model tracing the relationships between physiologic variables, symptom status, functional status, general health perceptions and overall quality of life. However, the Wilson and Cleary model has two important shortcomings. First, their model does not specify how speci®c life domains are related to overall QOL. Second, the authors continue to use the terms health status and quality of life interchangeably when describing the model. Nearly all authorities agree that QOL is a multidimensional construct [16, 17]. We hypothesize that perceptions of life quality are based on a cognitive process similar to that used to formulate attitudes [18] and judgments [19]. This process involves: (1) identifying the relevant domains comprising QOL, (2) determining where one stands on each domain, and (3) integrating the separate domain judgments into an overall QOL assessment. Thus, QOL is multidimensional in the sense that subjects may simultaneously evaluate several dimensions to arrive at an overall judgment. Figure 1 displays a structural model of this judgment process in which QOL is an unobserved, latent construct. In this ®gure, we follow the usual causal modeling conventions of representing unobserved variables by circles and observed (measured) variables by rectangles. QOL is determined jointly by assessments of several domains. For convenience, three separate domains are depicted in the model, although the number of relevant domains may be larger or smaller than this. Domain scores computed from a QOL questionnaire would ordinarily serve as measures of the unobserved domain variables. It should be noted that the hypothesized relationship between domains and QOL is not a factor model. Rather, in Bollen's [20] terminology, the domains are `cause indicators'; that is, the combination of the individual domain assessments produces the QOL construct.
Figure 1. Structural model of the determinants of quality of life.
The right side of the model speci®es a factor model relating QOL to self-reported global ratings of quality of life. These global ratings may be elicited by life satisfaction scales, measures of well-being, or by items referring speci®cally to QOL. The global ratings are correlated through their common dependence on the underlying QOL factor. A major determinant of domain scores is presumed to be symptom severity, which is in turn in¯uenced by a patient's biologic and physiologic status. These two variables also appear in the Wilson and Clearly scheme. In general, our model posits that chronic diseases produce symptoms of varying severity which aect patients' assessments of the speci®c life domains that determine quality of life. The same general structural equation model may also be used for health status by substituting health status for QOL. A widely used global indicator for health status is perceived health as reported on a scale ranging from `excellent' to `poor'. If QOL and health status represent the same construct, then one would expect to ®nd that the domains and domain eects are also the same for each construct. But if these constructs are indeed dierent in patients' minds, then the pattern of relationships between domains and constructs should also be dierent. Our analysis therefore focuses on the relationships between domains, QOL and health status. Methods Study identi®cation We collected data about the relationships in the structural equation model from a number of dierent studies in order to maximize the generalizability of our results. To be eligible for our analysis, studies had to meet three criteria. Eligible studies were required to: (1) employ a QOL instrument that produced scores for multiple domains, (2) include global ratings of both `QOL' and `perceived health', and (3) provide a correlation matrix showing the associations between all domain scores and between domains and global ratings.
449 We searched all articles appearing since 1990 in four leading journals that frequently publish quantitative evaluations of QOL instruments (Quality of Life Research, Medical Care, Journal of Clinical Epidemiology, Social Science and Medicine). Papers cited in these journals were also reviewed if it was suggested that they met the eligibility criteria. We also included the results of several studies we have conducted. Analyses of individual studies Our ®rst analysis objective was to identify the domains associated with QOL and health status. This was accomplished by performing a series of forward stepwise regression analyses using QOL ratings or perceived health ratings as the dependent variable and individual domain scores as predictors. The stepwise procedure begins by identifying the domain that is most strongly related to the global rating. It then selects the next most strongly related domain after controlling for the ®rst domain, and so on. We retained all domains that had standardized regression coecients, or b, of 0.10 or greater, regardless of statistical signi®cance. Smaller b were considered to be trivial eects. Backward stepwise analyses gave nearly identical results to the forward method. These analyses were repeated for each study. To summarize regression results across studies, we counted the number of times a particular domain had a b value greater than 0.10 and the number of times it was less than 0.10. We also computed the mean b for each domain across studies. Meta-analysis procedures Computing mean standardized eects across studies provides an indication of the strength of the relationship between domains and constructs, but this approach has several serious methodological shortcomings. First, the same set of domains does not appear in every regression. Thus, b values are statistically adjusted for certain domains in some studies, but not in others. Second, simply averaging b across studies ignores dierences in sample size that aect the precision of the results. Third, these means may be biased by covariances among domains within the same study. To address these shortcomings, we performed a meta-analysis of the available studies. While metaanalyses have traditionally focused on a single association, the technique has recently been extended to the synthesis of correlation matrices. The metaanalysis was limited to the two constructs of interest and the three domains ± mental health, physical functioning and social functioning ± that were common to all studies. An ecient method for pooling correlations is to combine the correlations via generalized least squar-
es. We followed the method described by Becker and her colleagues [21, 22]. This approach involves three steps. First, the variance±covariance matrix for each study was computed. Second, the individual matrices were concatenated into a single matrix. We computed a block-diagonal variance±covariance matrix for all the studies in which the main diagonal is the variance associated with each observed correlation and odiagonal elements surrounding the diagonal are covariances among variables within the same study. This matrix also included covariance terms for two studies in which the same patients were administered two dierent QOL instruments. Otherwise, matrix elements representing dierent studies were assumed to be zero. Third, a synthesized correlation matrix was estimated by generalized least squares. This technique merges the data from all studies into a single matrix in which synthesized correlations are weighted to adjust for both within- and betweenstudy variances. The estimates were computed using Mathcad [23]. We also conducted a statistical test of the hypothesis that correlations were homogeneous across studies. The homogeneity hypothesis is tested by the Q statistic, which has an approximate v2 distribution [24]. The synthesized matrix was then used to estimate two path models suggested by Figure 1. Coecients in the path models were estimated using SPSS/PC+ [25]. Results Study characteristics We found ®ve studies in the literature that met all three eligibility criteria [26±30]. We rejected one other study [31] because global assessments were part of the mental health domain. While the selected journals published many articles about QOL instruments, most either did not contain a global measure of QOL or did not provide a correlation matrix for all of the relevant variables. We supplemented the published studies with seven QOL analyses we conducted [32, 33], for a total of 12 studies. Six dierent instruments were used in these studies. These included the RAND-36, which is based on the same items as the SF-36, and two forerunners from the Medical Outcomes Study, the HIV-30 and the SF-20; the European Organization for Research and Treatment of Cancer-Core Quality of Life Questionnaire (EORTC QLQ-30); Hornquist's QL-status and change (QLsc); and the QLI-Mental Health instrument. Also included were two new instruments we developed, the Multidimensional Index of Life Quality (MILQ) and the Multidimensional Quality of Life Questionnaire for people who have AIDS or are HIV+ (MQOL-HIV). The 12 studies encompassed a wide range of chronic diseases including cardiovascular disease, cancer, HIV infection, diabetes and
450 schizophrenia (see Tables 1 and 2). Sample sizes ranged from 37 to 327 patients. In these studies, QOL was measured either by a 100 mm visual analog scale (one study), by a single item referring to `QOL' or `life satisfaction' (4 studies), or by a composite score combining two or more of these ratings (7 studies). Most studies measured perceived health using the standard question: ``In general, would you say your health is excellent, very good, good, fair, or poor?'' The EORTC instrument asks patients to make overall assessments of `physical condition' and `quality of life', which it then combines into a single scale. We separated these two items for the purpose of this analysis. Analyses of individual studies Table 1 shows the results of the stepwise regression analyses of QOL in each study. All domains with b (standardized regression coecients) greater than or equal to 0.10 are shown in order in this table, along with the variation explained by the regression model, and a list of domains that did not help to predict QOL. The results of the analyses for perceived health are shown in the same format in Table 2. The results of these analyses of the individual studies are summarized in Table 3. Nine domains appeared in at least ®ve studies. The domain having the greatest impact on global ratings of QOL was the mental health/emotional well-being domain. All of the studies included this domain and it had a b value greater than 0.10 in every study but one. The mean b for mental health across studies was 0.36. Physical functioning also played a role in predicting QOL ratings. b for this domain averaged 0.16, or less than half size of the b for mental health. The energy/fatigue b averaged 0.12, but this domain appeared in only ®ve studies. None of the other domains tested had mean b exceeding 0.10. The results for perceived health were quite dierent. These ratings were most strongly related to the physical functioning, energy/fatigue and pain domains. Mental health was much less important for perceptions of perceived health than it was for QOL. This pattern has also been found in a large population-based study of self-rated health [34]. Synthesized correlation matrix The synthesized correlation matrix for the ®ve common measures in these studies is reproduced in Table 4. All ®ve variables in the meta-analysis were positively correlated. The strongest correlations were between QOL and perceived health, and between QOL and mental health. The smallest correlation (0.44) was for mental health and physical functioning. This correlation is nearly identical to that found by Hays and Stewart [11] in their analysis of nearly 2000 patients in the Medical Outcomes Study. The
homogeneity statistic for this matrix was highly signi®cant (Q = 307.6 with 110 df), leading us to reject the null hypothesis that the individual matrices were drawn from a common population matrix. However, this statistic is extremely sensitive to sample size and will usually be rejected in studies as large as this, which involved more than 1500 patients. Path models Two path models were estimated from the synthesized correlation matrix. Figure 2 shows the eects of the domains on each of the two major constructs. The synthesized correlation between QOL and perceived health across these studies was 0.72. Thus, one of these constructs explains about half of the variation in the other. The pattern of domain eects was clearly dierent. When rating QOL, patients give much greater emphasis to mental health (0.47) than to physical functioning (0.28). This pattern is reversed for appraisals of perceived health in which physical functioning is more important than mental health. The social functioning domain has only a minor eect on either construct. Another way of testing the distinction between QOL and health status is to ask whether QOL perceptions are purely a function of perceived health or whether the eects of other domains persist after controlling for perceived health. The path model in Figure 3 is constructed to test for such eects. The relationships among the individual domains are likely to be reciprocal, but in this model we assume that both mental and social functioning are explained in part by physical functioning. This ordering is consistent with longitudinal assessments of these variables [35]. The path estimates show that mental health aects QOL ratings even after controlling for perceived health (coecient = 0.34). Moreover, the total eect of mental health on QOL (0.47, direct eect plus the indirect eect through perceived health) exceeds the total eect of perceived health (0.41). This is further evidence that QOL and health status are in fact distinct constructs. The direct eects of physical and social functioning on QOL are quite small; most of their in¯uence is mediated by perceived health. Discussion The results of our meta-analysis indicate that, from the patient's perspective, QOL and health status are two distinct constructs. We found that individual domains do not have the same eects on ratings of QOL that they have on ratings of perceived health. Mental health has a much greater impact than physical functioning on QOL ratings; this pattern is reversed for perceived health. These dierential eects imply that patients are evaluating two dierent constructs.
Table 1. Analyses of quality of life Study
Avis et al., 1996 [32]
Avis et al., 1996 [32]
Smith et al., 1996a
Ringdal and Ringdal, 1993 [29]
Wu et al., 1991 [30]
Instrument
MILQ (Multidimensional Index of Life Quality)
MILQ (Multidimensional Index of Life Quality)
MOS SF-20
EORTC QLQ-30
MOS HIV-30
Global quality of life criterion measure
Combined Ladder of Life (10-pt.) and Campbell Satisfaction (7-pt.) ratings
Combined Ladder of Life (10-pt.) and Campbell Satisfaction (7-pt.) ratings
Combined Ladder of Life (10-pt.) and Campbell Satisfaction (7-pt.) ratings
Overall quality of life; 7-pt. visual analogue scale
Single overall QOL item (5-pt. scale)
Predictors
Domain Mental health Physical function Health professionals
Domain Mental health Physical function Cognitive function Spouse/partner
Domain Mental health Health perceptions Physical function
Domain Emotional function Personal functioning Cognitive function
Domain Mental health Pain
b 0.53 0.35 0.11
b 0.40 0.23 0.14 0.12
b 0.36 0.35 0.24
b 0.40 0.32 0.24
Physical health
b 0.47 0.21 0.15
R2
0.61
0.47
0.57
0.49
0.39
Other domains tested
Physical health Financial situation Spouse/partner Cognitive function Productivity Social function
Physical health Social function Financial situation Health professionals
Social function Role function Pain
Pain Social function Fatigue Nausea
Physical function Role function Social function Cognitive function Energy/fatigue
Sample
N = 327 cardiology clinic patients
N = 196 patients undergoing cardiac procedures
N = 122 cardiology clinic patients
N = 177 cancer patients
N = 117 HIV+ patients
451
452
Table 1. (Continued) Study
Hornquist et al., 1993 [28]
Becker et al., 1993 [26]
Bjordal and Kaasa, 1992 [27] Smith et al., 1997 [33]
Smith et al., 1997b
Instrument
QLsc (Quality of Life: status and change)
QLI-MH (Quality of Life Index for Mental Health)
EORTC QLQ-30
MQOL-HIV
RAND-36
Global quality of life criterion measure
Single global life satisfaction item
Composite of 10 indicators of life quality developed by Andrews and Withey
Overall quality of life, 1 = very poor, 7 = excellent
Factor score combining Ladder of Life (10-pts), Campbell satisfaction item (7-pts.), and 100 mm visual analogue scale
Factor score combining Ladder of Life (10-pts.), Campbell satisfaction item (7-pts.), and 100 mm visual analogue scale
Predictors
Domain Psychological
b 0.38
Domain Economics
b 0.27
Domain Social function
b 0.27
Domain Social function
b 0.34
Social
0.21
0.21
Fatigue
0.22
Mental health
0.27
0.38
Activity level
0.15
Occupational activities Psychological
Domain Emotional well-being Energy/fatigue
0.20
0.21
Financial status
0.16
General health
0.12
Health habits
0.14
Psychological Social relations
0.20 0.19
Physical function Role function
0.14
Physical function Partner intimacy
0.14 0.13
b 0.46
R2
0.58
0.50
0.40
0.54
0.58
Other domains tested
Illness symptoms Bodily health
Physical health Activities of daily living Illness symptoms
Cognitive functioning Emotional functioning Emesis Pain
Physical health Social support Cognitive functioning Sexual functioning Medical care
Physical functioning Role limitations-physical Role limitations-emotional Social functioning Pain
Sample
N = 52±73 insulindependent diabetes mellitus patients
Convenience sample of N = 37 schizophrenics receiving antipsychotic medication
N = 126 head and neck cancer patients
N = 113 HIV+ men and women
N = 108 HIV+ men and women
Table 1. (Continued) Study
Smith et al., 1997 [33]
Smith et al., 1997b
Instrument
MQOL-HIV
RAND-36
Global quality of life criterion measure
Factor score combining Ladder of life (10-pts), Campbell satisfaction item (7-pts.), and 100 mm visual analogue scale
Factor score combining Ladder of Life (10-pts.), Campbell satisfaction item (7-pts.), and 100 mm visual analogue scale
Predictors
Domain Mental health Physical function Social support
Domain Emotional well-being General health Role-physical
R2
0.46
0.55
Other domains tested
Physical health Social functioning Cognitive functioning Financial status Partner intimacy Sexual functioning Medical care
Physical functioning Role limitations-emotional Energy/fatigue Social functioning Pain
Sample
N = 87 HIV+ men
N = 89 HIV+ men
b 0.40 0.32 0.12
b 0.45 0.30 0.14
b are standardized regression coecients. Unpublished analyses of data in Ref. [32]. b Unpublished analyses of data in Ref. [33]. a
453
454
Table 2. Analyses of perceived health Study
Avis et al., 1996 [32]
Avis et al.,1996 [32]
Smith et al., 1996a
Ringdal and Ringdal, 1993 [29] Wu et al., 1991 [30]
Instrument
MILQ (Multidimensional Index of Life Quality
MILQ (Multidimensional Index of Life Quality
MOS SF-20
EORTC QLQ-30
MOS HIV-30
Perceived health measure
5-pt. perceived health scale 5-pt. perceived health scale (1 = poor, 5 = excellent) (1 = poor, 5 = excellent)
5-pt. perceived health scale (1 = poor, 5 = excellent)
7-pt. overall physical condition scale (1 = very poor, 7 = excellent)
5-pt. perceived health scale (1 = poor, 5 = excellent)
Predictors
Domain Physical function
b 0.34
Domain Role functioning
b 0.34
b 0.36
Domain Energy/fatigue
b 0.38
Mental health Health professionals Productivity
0.25 0.13
Pain Mental health
0.30 0.25
Domain Personal functioning Fatigue Emotional function Social function Pain
0.18 0.16
Physical function Pain
0.27 0.13
Domain Physical function
b 0.63
0.13
0.14 0.11
R2
0.42
0.40
0.47
0.49
0.39
Other domains tested
Financial status Spouse/partner Cognitive function Social functioning
Mental health Social functioning Financial status Spouse/partner Cognitive functioning Productivity Health professionals
Physical functioning Social functioning
Nausea Cognitive functioning
Role functioning Social functioning Mental health Cognitive functioning
Sample
N = 326 cardiology clinic patients
N = 160 patients undergoing cardiac procedures
N = 122 cardiology clinic patients
N = 177 cancer patients
N = 117 HIV+ patients
Table 2. (Continued) Smith et al., 1997 [33]
Smith et al., 1997b
QLI-MH (Quality of Life EORTC QLQ-30 Index for Mental Health)
MQOL-HIV
RAND-36
7-pt. scale of bodily health, (1 = very bad, 7 = very good)
12-item index of physical health
7-pt. overall physical condition scale (1 = very poor, 7 = excellent)
5-pt. perceived health scale (1 = poor, 5 = excellent)
5-pt. perceived health scale (1 = poor, 5 = excellent)
Domain Psychological Health habits Social
Domain b Social relations 0.40 Economics 0.14
Domain Physical function Social function Fatigue
Domain Physical function Social function Cognitive function Medical care
b 0.41 0.26 0.16
Domain Physical function Energy/fatigue Pain
b 0.36 0.24 0.19
0.11
Emotional well-being
0.13
Study
Hornquist et al., 1993 [28]
Becker et al., 1993 [26]
Instrument
QLsc (Quality of Life: status and change)
Perceived health measure Predictors
R2
b 0.45 0.21 0.19
0.55
Bjordal and Kaasa, 1992 [27]
b 0.44 0.24 0.17
0.23
0.43
0.49
0.50
Other domains tested Activity level Illness symptoms
Psychological Activities of daily living
Role functioning Cognitive functioning Emotional functioning Emesis Pain
Mental health Social support Financial status Partner intimacy Sexual functioning
Role-physical health Role-emotional Social functioning
Sample
Convenience sample of N = 37 schizophrenics receiving antipsychotic medication
N = 126 head and neck cancer patients
N = 113 HIV+ men and women
N = 108 HIV+ men and women
N = 52±73 insulin-dependent diabetes mellitus patients
455
456
Table 2. (Continued) Study
Smith et al., 1997 [33]
Smith et al., 1997b
Instrument
MQOL-HIV
RAND-36
Perceived health measure
5-pt. perceived health scale (1 = poor, 5 = excellent)
5-pt. perceived health scale (1 = poor, 5 = excellent)
Predictors
Domain Physical function Medical care Mental health Sexual function
Domain Pain Social function Energy/fatigue Physical function
R2
0.36
0.46
Other domains tested
Social functioning Social support Cognitive functioning Financial status Partner intimacy
Role limitations-physical Role limitations-emotional Emotional well-being
Sample
N = 87 HIV+ men
N = 89 HIV+ men
a
b 0.35 0.20 0.14 0.12
Unpublished analyses of data in Ref. [32]. Unpublished analyses of data in Ref. [33]. c Criterion measures were recorded for regression analyses so that higher scores indicate better health. b
b 0.32 0.20 0.18 0.11
457 Table 3. Summary of domain eects on quality of life and perceived health in 12 studies Quality of lifea
Mental health/Emotional well-being Physical functioning/ADLs Energy/Fatigue Physical health Financial situation/Economic status Social functioning/Social relations Role functioning/Role limitations Cognitive functioning Pain a
Perceived healtha
P0:10
