Advance care planning is complex: Patients may not fully understand treatments and health outcomes, or they may have difficulty formulating and expressing preferences about future situations. Communication between patients and their proxies and clinicians may therefore be compromised. Patients may also change their minds when faced with the actual decision to choose or forego life-sustaining treatment. Thus, some physicians are skeptical about the validity of preferences expressed in advance directives [1].

Ideally, treatment preferences expressed before life-threatening illness occurs should be compared with those chosen at the actual moment of decision. This comparison is difficult to make because advance directives are invoked only when the patient cannot participate in decision making. It is therefore important to design alternate strategies to ensure the validity of prospective treatment preferences.

One alternative is to examine the stability of treatment preferences, especially in persons whose health status has changed or who have had a new experience with a treatment or a health state [2, 3]. Another is to test whether persons with a particular condition have different treatment preferences from persons who must evaluate the same health state from a hypothetical point of view [4].

A third alternative is to explore the logical relation between treatment preferences and ratings of the outcomes of those treatments. This process can be viewed as a type of convergent validation. Evidence of convergent validity is seen when measures of concepts that are thought to be related are in fact correlated [5, 6]. Discordance is the degree to which the hypothesized relation between treatment and health state preferences does not correspond (Figure 1). By applying this logic to advance care planning, treatment preferences can be linked theoretically to health states that are rated as worse than death or better than death (Figure 1). We hypothesized that persons would refuse treatments that resulted in health states rated as worse than death and accept treatments that resulted in health states rated as better than death (Figure 1).

View larger version (14K): [in this window] [in a new window]   Figure 1. Model for exploring the validity of treatment preferences. Preferences are concordant when patients refuse treatments in health states considered worse than death and accept treatment in health states considered better than death (A and D). Discordant preferences (B and C) should be discussed to clarify misunderstanding or explore patient values.

We evaluated the convergent validity of prospective treatment preferences by comparing them with patient ratings of health states as worse than death or better than death. We evaluated validity cross-sectionally at the levels of treatment decisions and persons. We also examined the stability of concordance over time and participants' reasons for discordant preferences.

Methods

Top Methods Results Discussion Author & Article Info References

Participants

We assembled a diverse convenience sample of participants from seven groups of volunteers in the greater Seattle area: well adults (21 to 65 years of age), older well adults (>65 years of age), older adults with chronic conditions, persons with terminal cancer, persons with AIDS, stroke survivors, and nursing home residents. Participants had to be 21 years of age or older, have no major vision or hearing impairments, show cognitive ability according to the Telephone Interview for Cognitive Status, and speak English [7]. In addition, well adults could not have any health condition that had lasted longer than 1 year, be receiving regular treatment by a clinician, or be taking medications more than twice monthly. Persons with chronic conditions had to be 65 years of age or older and have at least one chronic condition. Persons with terminal cancer must have had a physician-estimated life expectancy of 6 to 24 months. Persons with AIDS had to have received a diagnosis of class IV HIV infection or AIDS. Stroke survivors must have had a stroke within the past 10 years and had to have residual impairment. Nursing home residents were expected to have been in residence for more than 6 months.

We recruited well adults by sending letters to addresses that were randomly selected from the 1990 Seattle telephone directory. Patients were recruited with the help of community- and university-affiliated physicians and social service intermediaries, who identified eligible persons. Potential participants were sent a cover letter with an information statement and return postcard. All persons who returned postcards were called and screened; those who were eligible were enrolled in the study. Informed consent was obtained at the first interview.

Procedures

As part of a longitudinal study, participants completed a two-stage interview. In the first stage, participants reported social and demographic characteristics, including age, sex, race, marital status, and level of education [8]. They also completed the 136-item Sickness Impact Profile, the 20-item Center for Epidemiologic Studies-Depression scale, and the Perceived Quality of Life scale [9-11]. In the second stage, participants were given a health state classification system and were asked to choose the level from four areas of functioning that most closely resembled their current health (Appendix Table). This system was also used to describe four hypothetical health states: permanent coma, dementia, severe pain, and severe stroke. These hypothetical states represent scenarios in which advance care directives may be invoked; previous research has determined that many persons consider these states worse than death [12-16].

Preferences for six life-sustaining treatments (antibiotics, long-term mechanical ventilation with tracheostomy, hemodialysis, feeding by jejunal tube, short-term mechanical ventilation, and cardiopulmonary resuscitation) were elicited in each of the five health states. Treatment preferences were measured on a five-point scale by using a visual aid (Figure 2). Figure 2 shows how preferences for tube feeding were elicited in the severe stroke health state. To elicit preferences for antibiotics, short-term mechanical ventilation, hemodialysis, and tube feeding, interviewers placed an identical copy of a card with a health state description in the spaces for state of health and outcome. For long-term mechanical ventilation, the card for the outcome health state differed from the card for the baseline health state in that it described the most likely outcome scenario as being confined to bed and being dependent for self-care. To elicit preferences for cardiopulmonary resuscitation, we presented a modified version of the visual aid that showed three outcomes with different probabilities: a 20% chance of returning to the baseline health state, a 5% chance of a permanent coma, or a 75% chance of death. Participants also rated their current health and the four hypothetical health states using a seven-point scale: 1, much better than death; 2, somewhat better than death; 3, a little better than death; 4, neither better nor worse than death; 5, a little worse than death; 6, somewhat worse than death; and 7, much worse than death.

View larger version (23K): [in this window] [in a new window]   Figure 2. Sample of the visual aid used to elicit treatment preferences.

At the 30-month follow-up interview, participants were asked supplementary open-ended questions to elicit reasons for discordance between health state ratings and treatment preferences. To minimize participant burden, a maximum of four discordant decisions were chosen for review. For each participant, we used stratified random sampling to select discordant decisions for discussion. As many as two of these decisions were sampled when participants rated a health state as worse than death and wanted treatment. The other two represented decisions in which participants rated a health state as better than death and did not want treatment. Participants were asked to give all their reasons for making a decision. Reasons were recorded and coded by the interviewer and one of the investigators.

Statistical Analysis

Differences in demographic characteristics and measures of health status across participant groups were tested by using analyses of variance and chi-square tests. We also tested whether health state ratings differed across the five health states and whether treatment preferences differed across the health states and treatments preferences. Because treatment decision was the unit of analysis for these tests, we used generalized estimating equations to account for the fact that multiple decisions made by one person are not independent observations [17]. The hlogit command in Stata (Stata Corp., College Station, Texas) fits a logistic regression model as if all decisions are independent observations and then adjusts the SEs to account for clustering [18].

Each participant rated 5 health states and chose 6 treatment preferences for each health state, for a total of 30 treatment preferences. Concordance between treatment preferences and health states was analyzed at the levels of decisions and persons. For each treatment, a logistic regression with SE adjustment was used to test how well the health state ratings on the seven-point scale predicted treatment refusal. For this analysis, the five-point treatment preference scale was collapsed to "refuse treatment" (definitely and probably no) and "do not refuse treatment" (not sure, definitely yes, and probably yes). A goodness-of-fit test was used to assess the linearity assumption. To facilitate the application of our study to advance care planning, we also collapsed the ratings of treatment preference and health state according to a 2 x 2 decision model (Figure 1).

We report the percentage of decisions that were concordant among health states rated as either worse than death or better than death. At the person level, we calculated three concordance scores: the percentage of concordant decisions among all states worse than death, the percentage of concordant decisions among all states better than death, and a total concordance score that included all states that were worse than death or better than death. Participants had high concordance if their scores were 80% or greater: this criterion was met if at least 5 of 6 decisions were concordant.

The stability of convergent validity was evaluated in two ways. First, a logistic regression with SE adjustment was used to test whether the odds ratio that related health state ratings to the probability of refusing treatment changed over time. Second, we evaluated whether participants who had high concordance at baseline also had high concordance 6 months later. These analyses included participants who had at least one state that was considered worse than death or better than death at baseline and at 6 months. We used chi-square analyses to test the null hypothesis that the probability of high concordance at follow-up was independent of whether the person had high concordance at the previous interview.

Role of Funding Source

Our funding source, the Agency for Health Care Policy and Research, had no role in gathering, analyzing, or interpreting the data nor in recommending how to publish the results.

Results

Top Methods Results Discussion Author & Article Info References

Participants

Table 1 shows the baseline characteristics of the 341 participants included in the analyses. Like the general population of Seattle, our sample was predominantly white (93%) and well educated. The sample was diverse, and all of the variables shown in Table 1 differed across participant groups (P < 0.001).

Health State and Treatment Preferences

Participants' ratings of their current health differed markedly from their ratings of the four hypothetical health states (P < 0.001 for all groups) (Table 2). Only two participants rated their current health as worse than death (one nursing home resident and one person with terminal cancer). Both of these participants reported severe dysfunction and extremely low perceived quality of life. Ratings differed across groups for the severe pain (P = 0.02) and coma (P = 0.002) health states only. Permanent coma was rated as worse than death by 52% of participants and as better than death by 31% of participants.

Several patterns can be seen in the distribution of treatment refusals (Table 3). For every health state, participants were more likely to refuse invasive and long-term treatments (P < 0.001). Treatment refusal also differed by health state: Coma had the highest rate of refusal; current health state had the lowest rate; and dementia, severe stroke, and severe pain had similar rates of refusal in the middle range for every treatment (P < 0.001).

Convergent Validity

Decision-Level Analyses

The relation between treatment preference and health state rating is shown in Figure 3. For all treatments, the percentage of treatment refusals increased as the health state rating went from much better than death to much worse than death. The odds ratio for a one-unit change in health state rating varied from 1.94 for antibiotics to 1.70 for long-term mechanical ventilation (P < 0.001 for all treatments): that is, a change in health state rating from a little worse than death to somewhat worse than death, for example, was associated with a 70% to 94% increase in the odds of treatment refusal.

View larger version (54K): [in this window] [in a new window]   Figure 3. Relation between treatment preference and rating of health state. For each treatment, ratings for all five health states were combined. White indicates participants who accepted treatment; diagonal shading indicates participants who were unsure whether to accept or reject treatment; and black indicates participants who rejected treatment. Each treatment preference was associated with the strength of the health state rating (P < 0.001).

Concordance was high for health states rated as much better than death or much worse than death (Figure 3). Preferences were more discordant when health states were rated more equivocally (a little better than death or a little worse than death). If the health state was rated as worse than death, 13% of participants decided to accept treatment with antibiotics and 5% were not sure. If the baseline state of health was rated as better than death, many participants decided to refuse invasive or long-term treatment.

According to the decision model shown in Figure 1, 85% of decisions were concordant for states considered worse than death (treatment was rejected). For states considered better than death, 62% of decisions were concordant (treatment was accepted). These percentages were computed by using the treatment decision as the unit of analysis and by summing across all treatments and all health states.

Person-Level Analyses

Table 4 shows the distribution of concordance scores at baseline. Concordance scores were higher for states worse than death than for states better than death. The percentage of participants who had high concordance did not differ across the groups for states considered worse than death (data not shown). For states better than death, the percentage of participants who had high concordance differed across groups (nursing home residents, 11%; elderly outpatients, 21%; stroke survivors, 31%; terminally ill persons, 35%; older well adults, 43%; persons with AIDS, 44%; and younger well adults, 65%) (P < 0.001 for all groups).

Stability of Concordance

At the decision level, the strength of the relation between health state ratings and treatment refusals did not differ over time. The logistic regression analyses of all decisions made by the 189 respondents who had complete data at all four interviews resulted in the following odds ratios for all treatments combined: 1.95 at baseline, 2.10 at 6 months, 1.95 at 18 months, and 2.14 at 30 months (P = 0.08 for difference across time).

Stability of concordance within persons was higher for states worse than death than for states better than death. For states worse than death at baseline, 88% of the 129 participants with high concordance still had high concordance 6 months later. For states better than death, 68% of 109 participants still had high concordance. For states worse than death or better than death, 75% of 122 participants remained highly concordant. Persons who dropped out at 6 months or who considered no state worse than death (n = 33), better than death (n = 11), or better or worse than death (n = 16) were not included in these results. These rates of 6-month stability were higher than those expected by chance (P < 0.001). Repeating these analyses for the intervals from 6 to 18 months and 18 to 30 months produced similar results. Of the 71 persons who survived, participated in all four interviews, and rated at least one health state as worse than death at all interviews, 82% had high concordance at all time points. Of the 74 surviving participants who rated at least one health state as better than death at all interviews, 39% were highly concordant at all time points.

Reasons for Discordant Preferences

Of the original 341 participants, 189 were available for the 30-month interview. Of the 152 participants not available for interview, 10 were lost to further follow-up, 41 refused participation, 18 were too ill, and 83 had died. The stratified random sampling method for discordant decisions revealed 71 instances in which participants rated health states as worse than death but wanted treatment and 150 in which participants considered the health state better than death but did not want treatment. When participants were asked to explain their reasons for discordant preferences, their answers differed according to the type of discordance. Discussions about discordance for states worse than death led to a change of preference almost two thirds of the time once the relation between treatment preference and health state rating was made explicit. Participants changed either their treatment preference or health state rating (41% and 23%, respectively) to make the two concordant. Reasons for changing treatment preference or health state rating included the following: Participants would agree to a trial treatment because the treatments were simple or short term (such as antibiotic therapy or mechanical ventilation for 2 days [16%]) and because even though they rated personally the health states as worse than death, they had to consider their family members' desires that they remain alive (10%). One person said that he would accept short-term treatment to stay alive long enough to say goodbye.

For the 115 participants who rated their current health state as better than death but did not want treatment, the most common reasons stated were that receiving treatment led to a worse outcome state of health (such as machine dependency or increased dysfunction [71%]) and that the treatments themselves were unacceptable (46%). When discussing the 35 hypothetical scenarios rated as better than death, 60% of participants said that they would not want treatment that would prolong life under those circumstances because the quality of life was marginal. Being a burden to others was another reason given for not wanting treatment in both current and hypothetical health states (22% and 23% of participants, respectively). Participants changed either their treatment preference or health state rating at a much lower rate than for states worse than death: Only 10% changed treatment preferences, and 1% changed health state ratings.

Discussion

Top Methods Results Discussion Author & Article Info References

Consistent with the results of previous research [12, 15, 19-21], more than half of the participants in our study considered permanent coma a health state worse than death and were more likely to reject all treatments in coma, including antibiotics and short-term mechanical ventilation. Only at the lowest level of function and quality of life (and rarely even then) did participants consider their current situation worse than death.

We tested the hypothesized relation between treatment preferences and ratings of health states. At the level of treatment decisions, we found high rates of concordance between treatment preferences and health state ratings; this supports our hypotheses of convergent validity. Decisional concordance varied by strength of preference: Higher concordance was found when health state ratings were stronger (that is, much worse than death or much better than death) (Figure 3). These results suggest that treatments are easier to accept or reject if participants have strong preferences about an outcome health state.

Concordance was higher for states worse than death than for states better than death, both at baseline and across all interviews. These results suggest that judging health outcomes as worse than death leads to the rejection of treatment, regardless of the treatment. However, patients may also consistently rejected treatment if the outcome is considered better than death, particularly if the treatment is invasive. If invasive treatments are long term, such as permanent dependence on a mechanical ventilator, the outcome health sate and the treatment can be confounded. When evaluating treatment, persons may consider their baseline health state to be better than death but consider continued existence with long-term treatment less desirable than their pretreatment condition.

Concordance was remarkably stable across the observational periods for health states worse than death. This result is consistent with earlier reports that the choice to forego treatment is more stable than the choice to receive treatment [2, 22]. Preferences for health outcomes considered worse than death and the rejection of life-sustaining treatment are grounded in a consistent value system that is stable over time.

Deliberation about discordant preferences helped participants understand and organize their thinking and gave them the opportunity to justify or change their decisions. Reasons for discordant preferences included poor outcomes; invasiveness of treatment; and values that overrode foregoing treatment, such as concern for the wishes of the family. Time did not permit examination of concordant decisions, but such deliberation might yield insight for patients and clinicians by defining additional core values.

The logic of relating the value of outcome health states to preferences for life-sustaining treatment has important clinical implications. In advance care planning, this logic can help structure the discussion of values and preferences [23, 24]. If a patient identifies an outcome of treatment as worse than death, clinicians can validate treatment preferences by asking whether this means that the patient would not want any treatment to keep him or her alive in that situation. If preferences are discordant, clinicians can explore the patient's definition of states worse than death or values that may override preferences for outcomes.

Full concordance for health states considered better than death implies that the patient wishes to remain alive under all circumstances. Clinicians should confirm this assumption. If patients do not want specific treatments in states considered better than death, clinicians should explore what core values influence their patients' decision making. In particular, clinicians may want to ask about treatments that impose significant burdens, such as long-term mechanical ventilation or dialysis.

The design of our study limits its generalizability. The selection bias of a volunteer sample could not be assessed. Nonwhite persons and persons with low educational levels were underrepresented compared with the general population of the United States. The participants, however, represented a broad range with respect to age, functional status, and life experience.

Our descriptions of treatments and health states posed cognitive challenges for some participants. Preferences were undoubtedly influenced by how the questions were asked [25]. Study participants were told to assume that all treatments, except for cardiopulmonary resuscitation, had a 100% probability of success if chosen. They were also told to assume that the life-sustaining treatments would not change the attributes described by the outcome health state. Hence, the judgment task was simpler than it would be in real clinical decisions.

Our results indicate that health state ratings and treatment preferences, elicited in tandem, are useful for identifying situations in which treatment is clearly wanted or not wanted and for identifying concordance and discordance. Given the constraints of a busy practice setting, clinicians will not be able to elicit all health state ratings and treatment preferences. It is possible, however, to inquire about the presence or absence of any health states that the patient would consider worse than death or any treatments that the patient might not want regardless of their health state. If patients cannot coherently explain discordance, further discussion is needed to clarify misunderstandings or to help patients integrate many values into a reasonable decision policy. Such discussion will provide insight into patient values and provide information for future (and possibly different) situations. When the relation between preferences is concordant or discordant with good reason, clinicians and proxy decision makers can be more confident that they will act in accordance with the patient's wishes should they be required to do so.

Dr. Pearlman: Veterans Affairs Puget Sound Health Care Systems, 1660 South Columbian Way (182). Seattle, WA 98108.

Ms. Starks: Department of Medicine. Box 358280, University of Washington, Seattle, WA 98195-8280.

Dr. Cain: Department of Biostatistics. Box 357232, University of Washington, Seattle, WA 98195-7232.

Dr. Cole: Information Design Seattle, 4218 50th Avenue Northeast, Seattle, WA 98105.

Dr. Uhlmann: 425 West Bannock, Boise, ID 83702.