Trajectories of Health for Older Adults over Time: Accounting Fully for Death
P. Diehr,D. Patrick
DOI: https://doi.org/10.7326/0003-4819-139-5_Part_2-200309021-00007
IF: 39.2
2003-09-02
Annals of Internal Medicine
Abstract:The study of healthy aging is essentially the study of how health-related variables, such as cognition, change over time for persons with different characteristics. One of the best ways to demonstrate this change is a graph, such as a plot of average cognitive ability over time for men and women. Unfortunately, if some people die during the study, their longitudinal health variable (which we refer to generically as health) is missing after their death. The usual analytic approach is to perform a complete case analysis or an available data analysis, both of which restrict study to the healthiest subgroup of the study sample and may give an overly optimistic picture of the trajectory. Our goal is to develop a graph of health over time that accounts appropriately for death. In this paper, we demonstrate how omission of the deaths gives incorrect results and explain and illustrate two approaches for including deaths in longitudinal graphs. There are 3 components of a longitudinal analysis: a longitudinal variable (health) measured frequently (every year), a measure of time, and an independent variable (x) whose association with health over time is of interest. Time could be a calendar year, years since recruitment, or age at the time of the measurement. Alternatively, time could be measured in relation to some event, such as years before death or before and after stroke or myocardial infarction. Finally, we would usually be interested in how persons differ by some factor, such as treatment or control status or sex, indicated by the variable x. Figure 1 shows a generic health analysis based only on available data and shows the mean of health over time for two groups. Figure 1. Mean of health over time. x Although the two lines in Figure 1 appear to be parallel, the upper line increased by 7.77 points in 9 years and the lower line increased by 8.98 points in 9 years. One might predict that the two lines would cross at some future date. This is a foolish conclusion, since the health-related variable in this case is age and x denotes age 75 years or older at baseline versus age 65 to 75 years at baseline. The finding implies that younger persons age faster than older persons and will eventually catch up in age. Fortunately, the truth is known in this situation: Both groups age 1 year per year. The discrepancy occurs because the oldest people are most likely to die. The mean age of persons alive at each point does not increase 1 full year per year. In this case, one has mistaken a pattern of dying for a pattern of change in health. This particular mistake would never be made, but similar mistakes could easily be made when the true model is not known, which is usually the case. We suggest some ways of graphing data that account fully for death and avoid such problems. Methods We explored incorporating deaths into 2 components of the longitudinal analysis: health and time. We operationalized the approaches and illustrated them using data from the Cardiovascular Health Study. The Cardiovascular Health Study is a population-based longitudinal study of 5888 adults 65 years of age and older at baseline (1). Participating investigators and institutions from the study are listed in the Appendix. Participants were recruited from a random sample of the Medicare eligibility lists in four U.S. counties. Extensive baseline data were collected for all participants. After baseline, participants had an annual clinic visit and provided additional information by mail and telephone. Two cohorts were followed, one cohort (n = 5201) with 10 years of follow-up and the second cohort (all African Americans, n = 687) with 7 years of follow-up to date. Data collection began in 1990 and follow-up was almost complete for all surviving participants in 2000 (2). At baseline, the mean age was 73 years (range, 65 to 100 years), 58% were women, and 84% were white. The Cardiovascular Health Study has longitudinal data on many older adults, with 9 years of follow-up and very little loss to follow-up. Many participants died (30% in the first 9 years), but there were relatively few missing data. We imputed values for the missing data (except those missing because of death), explained in detail elsewhere (3, 4). In the current paper, we refer to excellent, very good, or good health as being healthy and fair or poor health, or dead, as not healthy. Results A few measures, such as the Quality of Well-Being scale (5) and the Health Utilities Index (6), were constructed so that death has a value of 0, but most health-related variables do not have a value for death. We consider adding a category or value for death, which results in a new variable that has a value for death, is a monotone transformation of the original variable, is interpretable, and is on a ratio scale. Diehr and colleagues (7, 8) have considered the ubiquitous measure How is your health: excellent, very good, good, fair, or poor? When this measure is used longitudinally, it seems logical to add a sixth category, dead. One approach to exploring the trends in health over time is to calculate transition probabilities among the 6 health states (9). By using such probabilities, the health of a hypothetical cohort of men in poor health at age 70 years was projected to improve until about age 76 years, after which it declined (10). Another approach is to graph the frequencies of the 6 categories over time. The top panel of Figure 2 shows the percentage of Cardiovascular Health Study participants in each health state by time from baseline. All persons were alive at baseline, but over time a substantial percentage of the original cohort died. Tracing the top of the good layer shows the regular decrease in the percentage of persons who are healthy (the percentage in excellent, very good, or good health) over time. Figure 2. Persons' health in the first 9 years of the Cardiovascular Health Study ( top ), in the 2 years before or after stroke ( middle ), and before or after stroke with no deaths ( bottom ). Such stacked bar graphs can be used with different measures of time, such as time before and after a stroke, as shown in the middle panel of Figure 2. The percentage of persons who are healthy is flat in the 2 years before the stroke, decreases by about 30 percentage points at the time of the stroke, and does not indicate recovery (the percentage does not increase after the stroke). This decrease is caused, in part, by the deaths. A qualitatively different picture emerges in the bottom panel of Figure 2, which does not include the deaths and indicates a smaller decrease at the time of the stroke, followed by some recovery. This panel may be misleading because it does not represent the trajectory of health for all persons experiencing a stroke. The stacked bar graphs show the distribution of all 6 health states over time but are inconvenient for analysis because it is difficult to compare the trajectories of different groups. Diehr and colleagues (7) considered two ways of assigning values to the 6 levels (Table). The second column shows how the 6 states could be transformed to create an ordinal variable. This variable could be used in nonparametric analyses but is unsatisfying to plot over time because its mean is not interpretable. The third column transforms the 6 categories into a binary variable that takes on the value of 100 if the person is healthy and 0 if the person is not healthy. Dead is coded as 0. The mean of this variable is the percentage who were healthy, or PCTH. This value is completely interpretable, but it may lose information in going from 6 categories to 2. Furthermore, it may seem inappropriate to group fair or poor health with death. Table. Three Ways of Recoding Self-Rated Health To Include Death The fourth column of the Table gives coding for a new variable that overcomes both these shortcomings: the estimated percentage probability of persons being healthy 2 years later, or the expected percentage of healthy persons (EPCTH). Under this transformation, excellent is coded as 95 because 95% of older adults who are in excellent health are healthy 2 years later, poor is coded as 15 because only 15% are healthy 2 years later, and dead is coded as 0. These probabilities were developed from several data sets, including the Cardiovascular Health Study, and were nearly the same if a 1-year time horizon was used (8). The mean of EPCTH is the percentage of persons expected to be healthy 2 years later. Mean EPCTH is consistently lower than PCTH because it refers to health 2 years in the future and health declines on average for older adults. If mean PCTH is plotted versus time, the area under the curve is the mean number of years spent in the healthy state, or the years of healthy life. The interpretation of the area under the EPCTH curve is the expected years of healthy life beginning 2 years in the future (8). Mean PCTH and EPCTH can be examined before and after a stroke, and the trajectories can be compared for different types of events. Figure 3 shows the smoothed trajectories before and after death, stroke, myocardial infarction, and a random event for a comparison group, which was constructed in a manner similar to that of a figure published elsewhere (3). The trajectories are very different: The random event group declines only a little over time, the myocardial infarction shows previous decline and some recovery, the strokes show no previous decline and no recovery, and the deaths have the lowest initial value and the steepest decline in the 2 years before the event. Figure 3. Expected percentage of persons who are healthy before or after an event. The PCTH and EPCTH transformations may be appropriate for disability measures, as well as health status. For example, the Cardiovascular Health Study measured activities of daily living each year and assessed which of 6 activities could be performed without limitations (4). We defined healthy as having no activities of daily living disa