آیا تغییرات سلامت خود ادراکی طول عمر را پیش بینی می کند؟

Researchers can rely either on retrospectively reported or on prospectively measured health changes to identify and quantify recent changes in respondents’ health status. The two methods typically do not provide the same answers. We compare the validity of prospective versus retrospective measures of health changes by investigating their predictive power for subsequent mortality. Data from a cohort study conducted in the Netherlands are used to compare the ability of changes in self-assessed health (SAH) – either reported retrospectively or measured prospectively in three waves (1991, 1993 and 1995) – to predict survival until 2004. We examine the relationship between health changes and mortality with a proportional hazard models controlling for individual unobserved heterogeneity, with and without control for pre-existing chronic conditions and the onset of new chronic diseases. For a high proportion of reports (39.8%), prospectively measured health changes in SAH do not concur with retrospectively reported health changes. Our results show that both measures of health changes are predictive of mortality in the model controlling for levels of SAH and socioeconomic characteristics only. Controlling for SAH, prior presence of chronic conditions, the onset of new conditions and unobserved characteristics, we find that prospectively reported health changes still predict longevity, whereas retrospective changes do not. These results suggest that the collection of longitudinal information on health changes has advantages over the – easier and cheaper – option of retrospective collection of the same information.

t is well known that self-assessed health (SAH) at one point in time has substantial predictive power for behavior including medical care utilization (Van Doorslaer et al., 2000 and Van Doorslaer et al., 2004), labor force participation (Bound, 1991), as well as for subsequent health outcomes, like survival (Dowd and Zajacova, 2007, Huisman et al., 2007, Idler and Benyamini, 1997, Mackenbach et al., 2002 and Van Doorslaer and Gerdtham, 2003), even after controlling for other, more objective, health indicators. Much less is known about the predictive value of health dynamics, i.e., changes in SAH. In many instances, researchers are interested in such changes, especially the negative ones – often referred to as ‘health shocks’ – because these may be equally (or even more) important precursors of later outcomes as (than) health levels. They are also indicators of the degree of persistence of health status. The two questions that we seek to answer are: (a) do changes in health levels have predictive ability over and above the information contained in health level itself?; and, if so, (b) how can such changes best be elicited? A priori, the answer to the first question ought to be affirmative, and this can easily be seen from the graph in Fig. 1, which depicts health trajectories for two hypothetical individuals A and B. Clearly, the information about a difference in the level of health at time t+1 has predictive power for the likelihood of each person’s health falling below a critical level. If all else is equal, including the health level at t, then person A, with the lower health at t+1, is likely to reach the minimal critical health level sooner and exhibit shorter expected survival. In the particular case depicted, knowing that both persons started off at the same health in t will lead to very different predictions of future health paths than knowing that they were already in different health states at t and moved along parallel trajectories between t and t+1 (such as A′ and B). It seems therefore obvious that information on recent health changes does add to the information on health levels. Health trajectories of hypothetical individuals a, a′ and b. Fig. 1. Health trajectories of hypothetical individuals a, a′ and b. Figure options Regarding the second question, there are basically two main approaches to eliciting health changes from self-reports. The first and easiest option is to simply ask retrospective questions about health changes: respondents then rate their health compared to a reference point in the past. This health transition question asks respondents to rate their general health compared with a previous period, with three response categories: “better”, “same”, and “worse”. It represents a simple and straightforward way of obtaining health change information from cross-sectional surveys when there is no opportunity to follow respondents over time. However, it only provides a proper alternative to the prospective, longitudinal collection of health change data, if the information obtained is similar, if not identical. The prospective health changes can simply be obtained by computing the changes in SAH between two consecutive waves in longitudinal data. However, it has been shown that retrospectively reported health changes between point 1 and 2, assessed at point 2, do not always concur with prospectively assessed changes between point 1 and 2 in time. Benitez Silva and Ni (2008) discuss several reasons for the possible incongruence. First of all, the incongruence may occur due to the reporting heterogeneity bias and cut-point shifts in SAH. 1 In particular, cut-point shifts of SAH, for a given individual, over time may be one possible source of bias in the prospectively health change measure. This means that, for a given true but unobserved health state, individuals may report health differently depending on their health expectations at two different points in time. In addition, health changes identified by retrospective health changes may not be large enough to cause a category jump in SAH in the next period and may not show up in prospectively measured changes in SAH. On the other hand, some biases have been reported regarding retrospective health change elicitation. The direct health change question forces individuals to provide a comparison of their current health with a different point in time. This may cause reliability and recall problems, and individuals may use different reference points in time when recalling the previous health state (Norman, Sridhar, Guyatt, & Walter, 2001). There is also some evidence that retrospective self-reports of health are biased towards the respondent’s present health state (Knox and King, 2009 and Norman et al., 1997) indicating that respondents with good health currently are more likely to report that their health has recently improved, and respondents with poor health currently are more likely to report that it has worsened. Given possible biases with both of the health change measures, the empirical question then becomes: which of the two change measures appears to perform better in predicting future hard health outcomes like mortality? In spite of an abundance of studies demonstrating high predictive ability of SAH for mortality, only a few have done this for health changes. Ferraro and Kelley-Moore (2001), for instance, found that SAH predicted mortality risk over 20 years follow-up only when treated as a time-dependent covariate, highlighting the importance of using dynamic models when multiple observations are available. Han et al. (2005) studied the impact of SAH as a time-dependent covariate in Cox regression model among older people and found change in SAH to be a stronger predictor of mortality than SAH at baseline. Strawbridge and Wallhagen (1999) also used SAH as a time-dependent covariate and found that change in SAH was a significant predictor of mortality among women. More recently, Lyyra, Leskinen, Jylhä, and Heikkinen (2009) showed that the use of SAH as a time-dependent covariate in a Cox regression model enables advantage to be taken of all the information in a longitudinal study design. Using data from the German Socio-Economic Panel, Schwarze, Andersen, and Anger (2000) found that mortality was not only affected by the level of SAH but also by changes compared to a previous year. On the other hand, the only study we could find which analyzed simultaneously the effect of SAH and retrospectively reported health declines, found only the latter to be significant (Deeg, Van Zonneveld, Van der Maas, & Habbema, 1989). While all of these studies do examine the relationship between SAH changes and mortality, none of them has compared the predictive ability of different measures of health change. One such direct comparison of the predictive ability of self-reported retrospective versus prospective changes was done by Benitez Silva and Ni (2008) using data from the US Health and Retirement Survey (HRS) but for subjective survival expectations as outcome measure, not actual mortality. Their results have favored the use of retrospectively reported health changes instead of prospectively computed changes in SAH but the measurement of survival expectations has been shown elsewhere to be noisy and subjective itself (Bassett & Lumsdaine, 2001). We also believe their results stem in part from the inappropriate control for initial health (cf measures section below). In this study, we exploit the simultaneous availability of four waves (1991-93-95) of longitudinal health data from the GLOBE study (Mackenbach, van de Mheen, & Stronks, 1994) and a mortality follow-up until 2004 to examine and compare the validity of alternative measures of health levels and changes for predicting mortality. These health measures include the level of SAH, computed changes in SAH, retrospective assessments of health changes, a set of self-reported chronic conditions and changes in self-reported chronic conditions. They enable us to answer the two main questions of this study. First, is there any value added of including retrospective/prospective health changes for mortality prediction, over and above health levels? Second, are prospectively and retrospectively reported health changes equally predictive of subsequent mortality? Methods Data Our data were taken from the longitudinal GLOBE study that was conducted since 1991 in a region in the Southeast of the Netherlands. The study is based on a cohort of non-institutionalized Dutch nationals, aged 15–74 years, living in the city of Eindhoven and surroundings. The GLOBE study is widely used and has contributed to the understanding of the explanation of socioeconomic inequalities in health in the Netherlands (see e.g., Van Lenthe et al., 2004, and Van de Mheen, Stronks, Schrijvers, & Mackenbach, 1999). Study results were a main source of information in the development of policy measures aimed at the reduction of socioeconomic inequalities in health. More information on the design and objectives of the GLOBE study can be found elsewhere (Mackenbach et al., 1994). A baseline random sample of approximately 27.000 individuals (stratified by age and postal code) was drawn from the population registries of the participating municipalities. These individuals received a postal questionnaire in the Spring of 1991, which had a response rate of 70.1% (n = 18973). Two sub-samples drawn from the baseline sample were re-interviewed in the autumn of 1991 and in subsequent years. The first sub-sample, hereafter called the “healthy sample”, was randomly drawn from the baseline sample (n = 2800). These individuals were re-interviewed in 1993, 1995 and 1997. In the second sub-sample, which we call the “frail sample”, individuals who reported to have at least one of four chronic conditions (asthma, severe low back pain complaints, diabetes mellitus and heart disease) were overrepresented (n = 2867). The frail sample was re-interviewed annually in 1992–1995 and 1997. In our analysis, we only use data from years 1993 and 1995 in which both samples were interviewed (so, we exclude 1992 and 1994) and for which the main variables of interest (in particular retrospective changes) were available. This means that we further exclude data from 1997 as none of the samples was asked to report retrospective health changes in this year. For the same reason, we use the 1991 wave only to compute prospective changes in SAH between 1991 and 1993. In the analyses presented in this paper, we pool the healthy and the frail samples in order to increase the sample size. Dropping cases with missing values for at least one of the variables used in the analyses leaves us with a total of 6148 observations (3242, in 1993, and 2906, in 1995), of which 3238 belong to the healthy sample (1705, in 1993; 1533, in 1995) and 2910 to the frail sample (1537, in 1993; 1373, in 1995). The GLOBE dataset was augmented through linkage to the national register of cause-of-death, until the end of 2004. During the follow-up period of 11 years, 8.6% (282/6148) of the respondents in our sample died. Measures We use several measures of health levels and of health changes in our analysis. The main measure of health level is derived from the standard question: “How is your health in general?” with response categories very good, good, fair, sometimes good/sometimes poor and poor. Further measures of health levels are whether individuals suffer from a range of chronic physical and mental conditions, described in detail below. The retrospective health change variable is the directly reported change in respondent’s health compared to a previous year (cf below). Response categories are “health remained the same”, “health has improved” and “health has worsened”. In 1995, respondents in the frail sample were asked to compare their health to that one year before, while for the healthy sample the reference period was two years. In 1993, the reference period was two years for both samples. Since we use the pooled sample in our analysis, we tested whether our results are sensitive to consideration of a different reference period for 1995 for the frail sample. This was done by including in the model interaction terms between an indicator of these observations and the indicators of reported changes. These interaction terms were jointly not significant (p-value = 0.395), supporting our use of the pooled sample. We compute prospective health changes (better, same or worse) as the difference between currently and previously reported SAH. Better or worse computed changes reflect jumps of one or more health categories between waves. The reference periods considered for the computation of prospective health changes are consistent with those described above for reported health changes. A set of interaction terms between an indicator of observations from the frail sample in 1995 and indicators of computed changes was also jointly not significant (p-value = 0.334). Several recent papers have favored the use of prevalence (Adams, Hurd, McFadden, Merrill, & Ribeiro, 2003) or the onset of new chronic conditions (Smith, 1999, Smith, 2007 and Wu, 2003) as arguably more objective indicators of health and health changes than those derived from self-assessed health and self-reported (prospective or retrospective) health changes. The GLOBE data also provide information on the presence of self-reported chronic physical and mental conditions, and whether they were diagnosed during the 12 months prior to the survey. We are therefore able to assess whether the association between self-assessed health – and retrospective/prospective health changes – and mortality can be explained by prevalence and onset of these conditions. We use information on fourteen conditions: high blood pressure, back pain/problems, diabetes, heart diseases, stroke, cancer, anchylosis, rheuma, lung diseases, stomach diseases, nervous diseases, intestine diseases, skin diseases. Additional emotional and mental health information is taken from Nottingham Health Profile questions (we consider an indicator of presence of at least 1 emotional or mental problem). We use both prevalence and onset of new chronic conditions between waves. Since we are interested in the effects of health changes (i.e., self-reported retrospective/prospective and onset of chronic conditions) controlling for health levels, we need to define these in a consistent manner. Note that the effects of a health change can be obtained in two different ways, as the effects of: i) current health, H2, controlling for previous health, H1; or ii) a health change, H2 − H1, controlling for previous health, H1 (it can be easily shown that a.H2 + b.H1 = a.(H2 − H1) + (a + b).H1, i.e., that the effect of (H2 − H1) in ii) equals the effect of H2 in i)). Since retrospective changes are a direct measure of changes, H2 − H1, and for comparative purposes, we opted for option ii), rather than i), to evaluate effects of all health changes. We have therefore considered previous health levels (self-assessed health and prevalence of chronic conditions) observed at the reference point in time for health changes (that is, two years previously for both the sub-samples in 1993 and for the healthy sub-sample in 1995; and the previous year for the frail sample in 1995), rather than current levels. If, on the other hand, the model included (H2 − H1) and H2, then the effect of the former could not be interpreted as that of a health change because a.(H2 − H1) + (a + b).H1 = −b.(H2 − H1) + (a + b).H2, i.e., the latter effect of (H2 − H1) does not equal that of the other two options and will, in fact, be of the opposite sign if effects of current and previous health levels are of the same sign in i), which is a reasonable expectation. We also included two indices based on activities of daily living (ADL). The first is an ADL index and is simply a sum score of problems with the following activities: move at the same floor, get in/out of bed, eat and drink, getting (un)dressed, washing face/hands, washing completely. The second is an index of mobility problems, which adds up indicators of difficulties with each of the following activities: walking down/up stairs, moving outdoors, leaving/entering house, sitting down/getting up from chair. We also control for demographic and socioeconomic characteristics, like gender, marital status (unmarried, married, divorced, widowed), education (at most primary, low vocational, middle education, high education), working status (employed, unemployed, disabled, retired, housework, student, living from investments) and income (low, middle, high). GLOBE respondents reported their total monthly household income in 13 categories ranging from ‘0–1000 guilders’ to ‘above 5800 guilders’ (the guilder – 1 guilder = 0.45€ – was the currency in The Netherlands during the period under analysis). We consider values under 1900 guilders as low income (first four categories), 1900–3500 as middle income (5th to 9th categories) and more than 3500 as high income. All demographic and socioeconomic characteristics are measured both in 1993 and 1995, except for education (measured only at the baseline). Descriptive statistics for all variables used in the analyses are shown in Table 1. Recommended articles What is self-rated health and why does it predict mortality? Towards a unified conceptual model 2009, Social Science & Medicine more Health status and health dynamics in an empirical model of expected longevity 2008, Journal of Health Economics more Income-related health inequalities in the Nordic countries: Examining the role of education, occupational class, and age 2010, Social Science & Medicine more View more articles » Citing articles (8) Related book content