Using Conjoint Analysis to Estimate Healthy-Year Equivalents for Acute Conditions: An Application to Vasomotor Symptoms

Background

Quality-adjusted life years (QALYs) provide a common metric for measuring health impacts of diseases and their treatments. They integrate mortality, morbidity, and preferences into a single index number representing the equivalent years of perfect health. QALYs require estimates of health-state utilities, which are values between zero and 1 that indicate the relative importance or seriousness of health outcomes. Aggregating utilities over time and affected individuals yields the number of QALYs achieved by a particular health-care intervention. This index of health-care benefits is widely used in cost-utility analysis for evaluating health-care investments and in guiding formulary and reimbursement decision-making.

Although health-state utility estimates have proven useful for evaluating a wide range of diseases, obtaining and interpreting health-state utilities for acute, self-limiting, or nonfatal conditions can be problematic. For example, Bala and Zarkin evaluate the utility gain from using anesthesia in a 1-hour root-canal operation [1]. Scaling a health-state utility for chronic pain from a year to an hour yields a QALY gain for anesthesia of about 0.0001. Using a cost-effectiveness threshold of $40,000 per QALY yields a corresponding societal value of about $4. Clearly, dental patients would pay far more than $4 to avoid an hour of intense pain, suggesting the QALY gains of anesthesia based on chronic-pain health-state utility estimates are inappropriate in this application. Methods commonly used to estimate health-state utilities for chronic conditions may be of limited use in obtaining valid estimates for acute, self-limiting, or nonfatal conditions. Researchers therefore employ somewhat ad hoc procedures such as “chaining” to derive QALYs for these conditions [2–4].

Health-state utility elicitation methods include visual analog scale (VAS), standard gamble (SG), and time trade-off (TTO). Most visual analog or “thermometer” scales ask subjects to rate a specified health state on a scale ranging from death to full health. Although the rating scale is easy to administer, it involves no trade-offs among health states or between health states and time. Thus, this method has no basis in utility theory [5].

Standard gamble, which is derived from von Neumann and Morgenstern's theory of utility under uncertainty, is a method for eliciting theoretically consistent utility estimates [6]. Subjects choose between a pair of health states A and B, in which A represents a chronic condition of less-than-perfect health and B is a gamble between perfect health with probability P or instantaneous painless death with probability 1 to P. The value of P that leaves subjects indifferent between the certain and probabilistic alternatives is the utility of the certain condition. Unfortunately, many subjects are reluctant to accept trade-offs between mortality risks and acute, self-limiting conditions [7]. SG methods also suffer from the limitation that subjects must evaluate potentially small probabilities of a clinically implausible outcome, that is, instantaneous painless death. The fact that most people have poorly developed computational skills and have difficulty working with probabilities is well known [8,9].

Because subjects may find the SG elicitation format difficult, many researchers employ the TTO elicitation format [10]. Subjects evaluate living the remainder of their life in an impaired health state versus living a shorter number of years in full health. TTO appears to be an easier task than SG for subjects, but longevity trade-offs also may be unacceptable for acute conditions where reduced life expectancy is not a clinically relevant outcome. All health-state preference elicitation methods require defining the duration of the condition, most commonly as “the rest of your life.” The TTO method explicitly relies on well-defined health-state-duration preferences to quantify the relative seriousness of the condition. To simplify using preference estimates in health-economic evaluations, analysts conventionally apply health-state utility estimates to independent time periods. Although all health-state preference elicitation methods incorporate positive time preference to some degree, the TTO method is especially vulnerable to this phenomenon because it is explicitly based on time preference logic. Although in and or itself, capturing positive time preference in health-state utility elicitation is not a problem, applying positive-preference health-state utility estimates to independent time periods, as is commonly performed in the estimation of QALYs in health economic evaluations, can confound the results of such studies. This problem is exacerbated when analysts apply discounting to health effects and thus double-count time preference in their models [11].

The objective of this study is to demonstrate a practical method of obtaining healthy-year equivalents (HYEs) as an alternative measure of the value of treatments for acute, self-limiting, or nonfatal conditions [12–14]. HYEs, like QALYs, measure perfect-health time equivalents for a given condition; however, unlike QALYs, HYEs do not require that a health state is independent of the time spent in that health state. Rather, the quality-adjusted time equivalence is obtained directly for specified outcome durations. Although HYEs avoid restrictive assumptions required to estimate health-state utilities, the HYE method has not gained acceptance because it requires obtaining a separate utility estimate for each possible duration of a given health condition. This study demonstrates the feasibility of using readily available conjoint-analysis methods to obtain duration-specific TTOs for clinically relevant durations of acute vasomotor symptoms associated with menopause.

Conceptual Framework

Conjoint-analysis methods have been widely used and validated in marketing research, transportation, and environmental economics [15–19]. Conjoint analysis, as applied to health-care decision-making, is a systematic method of eliciting trade-offs to quantify the relative levels of satisfaction people assign to various treatment attributes. It is based on the premise that medical interventions are composed of a set of features or attributes and that the attractiveness of a particular intervention to an individual is a function of these attributes [20]. Conjoint analysis is increasingly used in health economics to calculate patients' and physicians' stated preferences for health-care interventions, treatment alternatives, and health-care services (see Ryan and Gerard [21] for a recent review of health applications).

Conjoint-analysis surveys yield discrete-choice data. The random utility model (RUM) has been used extensively in analyzing individual choice data [22]. RUMs assume the conjoint utility associated with a particular alternative is expressed as a function of the attributes of the alternatives. The conjoint utility of individual i from alternative j is:

(1)

V indicates the deterministic part of the utility function and ε denotes a random error term. X is a vector of treatment or symptom attributes, Z is a vector of individual characteristics, and β is a vector of marginal-utility parameters.

Unlike SG and TTO health-state utilities that are anchored by zero and 1, conjoint utilities are defined on an interval scale. Therefore, only relative differences in conjoint utilities are meaningful. Conjoint marginal utilities, or partworths, can be used to indicate subjects' willingness to trade symptom severity for symptom duration. Define total conjoint utility for a health profile described by a period of less-than-perfect health combined with a period of good or normal health as:

(2)

where H_k is a set of symptoms, T_k < T̄ is the symptom duration, T̄ is a clinically relevant maximum symptom period, H_o is good health and T̄ − T_k is the duration of good health in the profile. We are interested in finding the time-equivalent duration in good health, HYE that provides the same conjoint utility Ū as a given symptom profile:

(3)

Figure 1 illustrates the relationship between time equivalence and conjoint utility for combined durations of both symptoms and normal or perfect health over a clinically relevant time period T̄.

In general, we are interested in the change in HYEs for clinically relevant symptom improvements. Using Eq. 3 and solving for ΔHYE for an improvement from health profile j to health profile k:

(4)

Equation 4 is a straightforward analog to calculating ex-post willingness to pay when the conjoint task involves money trade-offs instead of TTOs (when trade-offs involve risks, a similar welfare analog can be calculated in terms of maximum acceptable risk) [23]. If conjoint utility is linear in parameters, and holding individual characteristics Z in Eq. 1 constant, then

(5)

And Eq. 4 becomes:

(6)

Calculating ΔHYE using Eq. 4 for nonlinear utility specifications may require numerical approximations.

Application to Vasomotor Symptoms

Survey Instrument

We developed a graded-pairs conjoint-analysis survey that incorporates the most salient attributes of vasomotor symptoms. The attributes were determined from a review of the literature and consultations with a medical expert, a health economist, and a psychologist with extensive experience in clinical trials of hormone-replacement therapy (HRT) for the treatment of vasomotor symptoms. Although we used input from these advisors in the development of the attributes and levels, we did not seek formal consensus. Based on this information, we included the following characteristics of vasomotor symptoms: severity and frequency of daytime hot flushes, frequency of night sweats, and duration of symptoms.

The maximum duration of symptoms included in the study was 7 years. This duration was chosen to represent a clinically meaningful duration for an acute, self-limiting condition such as vasomotor symptoms. We consulted the literature on the epidemiology of vasomotor symptoms. The mean duration of vasomotor symptoms is 2 to 3 years. Nevertheless, 60% of women experience hot flushes for up to 5 to 7 years [24].

In addition to symptom attributes, the survey included 10-year risks of serious adverse events, including risk of hip or back fractures, risk of heart infarct, and risk of breast cancer. Adverse event risks were used to address a separate survey objective and are presented elsewhere [25]. Table 1 lists the treatment attributes and levels used in the survey.

Table 1. Attributes and levels for HRT for the relief of vasomotor symptoms

Treatment feature	Levels
Severity of daytime hot flushes	▪ No daytime hot flushes ▪ Mild: a fleeting warm sensation with no sweating that does not disrupt normal daily activity ▪ Moderate: a warm sensation with sweating that does not disrupt normal daily activity ▪ Severe: a hot sensation with sweating that can disrupt normal daily activity
Frequency of daytime hot flushes	▪ None (0 times) during the daytime ▪ 1–2 times during the daytime ▪ 3–6 times during the daytime ▪ More than 6 times during the daytime
Frequency of night sweats	▪ None (0 times) per night ▪ 1–3 times per night ▪ 4 or more times per night
Duration of hot flushes and night sweats	▪ 1 year ▪ 2 years ▪ 4 years ▪ 7 or more years
10-year risk of hip or back fracture	▪ 15/1000 (1.5%) or 50% reduction ▪ 30/1000 (3%) or no change
10-year risk of heart infarct	▪ 38/1000 (3.8%) or 25% reduction ▪ 50/1000 (5%) or no change ▪ 65/1000 (6.5%) or 30% increase
10-year risk of breast cancer	▪ 23/1000 (2.3%) or 25% reduction ▪ 3/1000 (3%) or no change ▪ 39/1000 (3.9%) or 30% increase

• HRT, hormone-replacement therapy.

The survey was finalized after conducting two sets of pretest interviews. Sixteen women completed interviews using a “think-aloud” protocol. Interviews lasted between 45 and 75 minutes, depending on the responsiveness of the subject. Some of the pretest interview subjects had difficulty completing conjoint tasks with three alternatives. Thus, to obtain more realistic and reliable responses, the conjoint tasks were revised to include only two, rather than three, treatment options. The question format was changed from choice to rating (graded-pairs) to improve the power of the two-alternative format. The final survey consisted of a series of paired treatment comparisons created by assigning different levels to corresponding pairs of treatment attributes. The study was reviewed and approved by the Office of Research Protection and Ethics at Research Triangle Institute.

Subjects were asked to assume that they were seeking treatment for severe and frequent vasomotor symptoms and that only two treatments were available. Relative treatment preferences were indicated using a 5-point scale. Risk levels were described using relative and absolute scales in separate versions of the survey instrument (see Table 1). Figure 2 is an example trade-off task stated in absolute terms. We report only the symptom severity and duration trade-off results here.

The experimental design combined specific attribute levels to describe a set of hypothetical treatment profiles. These profiles were paired to provide statistically efficient comparison sets for the trade-off tasks. We employed a variation of a commonly used algorithm to construct a near-optimal experimental design using 27 treatment-pair trade-off tasks [26–28].

Empirical Model

The conjoint utility of a given treatment outcome k is a function of the categorical β coefficients for symptom severity, symptom frequency, and 10-year risks of serious adverse events. Because preferences for the frequency and severity of daytime hot flushes and night sweats may depend on duration, we specified efficacy as interaction terms between vasomotor symptom attributes and duration. Duration also was included independently in the model as a linear, continuous covariate. Thus, the empirical specification for utility V_k is:

(6)

where	T is symptom duration in years.
	Severity is symptom severity level.
	DayFreq is number of hot flushes per day.
	NightFreq is number of night sweats per night.
	HipRisk is relative or absolute risk of fracture.
	HeartRisk is relative or absolute risk of heart attack.
	CancerRisk is relative or absolute risk of breast cancer.

The different treatments of risk levels as relative and absolute were controlled for in our estimates. Nevertheless, we report only the main-effects risk variables in this model because we are only interested in efficacy variables which depend on duration.

Because each subject provided responses to multiple trade-off questions, we estimated a random-effects panel model to account for within-subject rating correlation.

We used multivariate, ordered-probit regression to estimate conjoint utility parameters for each attribute level, while controlling for subject demographics and other characteristics [30]. Ordered-probit estimation is the appropriate procedure for ordered category data because it does not require that intervals between rating categories be equal. Estimating continuous variables as categorical avoids the requirement of assuming any particular functional form. All statistical analyses were conducted using Stata version 8.2 (StataCorp, College Station, TX, USA).

The maximum duration of symptoms included in the survey was 7 years. The shortest duration of symptoms included in the study was 1 year. Normal health is defined as no symptoms for the full 7-year period. The worst health state is severe vasomotor symptoms for the full 7-year period where severe vasomotor symptoms are defined as severe daytime hot flushes more than six times per day and night sweats more than six times per night, on average. Thus, the HYEs estimated in this study represent changes in HYE relative to 7 years of severe vasomotor symptoms.

Results

Partworth Estimates

The rescaled partworth estimates and 5% confidence intervals for efficacy attributes for 1 year are presented graphically in Figure 3. The length of the line segments between the best and worst levels of an attribute indicates the relative importance of that attribute. Among the efficacy attributes, the severity of daytime hot flushes is the most important attribute with a score of 10, and is followed by the frequency of daytime hot flushes and the frequency of night sweats, respectively.

Figure 3 indicates that estimated partworths are consistent with the natural ordering of the categories, so no symptoms or milder symptoms and fewer symptoms of each type generally have significantly higher utility values than less attractive outcomes. These results indicate that subjects understood the natural ordering of attribute levels and could discriminate effectively among them.

Discussion

Our study of women's trade-off preferences for vasomotor symptom control yielded several results of interest. Women in our sample were willing to trade longer durations of milder symptoms for shorter durations of severe symptoms, indicating that duration of menopausal symptoms is not independent of symptom severity and frequency. Women's past experience with vasomotor symptoms and socioeconomic characteristics significantly affected their preferences for treatment alternatives. Older women and/or women with menopausal experience have larger HYE gains than younger women for the same improvements in health.

We found clinically significant preference differences between older, more symptom-experienced women and younger, less symptom-experienced women. Older women had better-defined preferences and higher HYEs across health states relative to younger women. These results might suggest that younger women experience greater benefits from any given vasomotor symptom intervention. Nevertheless, therapeutic benefits depend on changes in HYEs, not their absolute values. The distribution of gains between age groups thus depends on the particular kinds of improvements realized.

Our results are subject to several qualifications and limitations. The potential for hypothetical bias is a common concern in conjoint-analysis studies. Choices involving hypothetical treatments do not have the same clinical, financial, and emotional consequences of actual treatments. The women in our study evaluated hypothetical improvements in vasomotor symptoms relative to a severe reference condition. Although the hypothetical outcomes and durations were clinically relevant, some subjects may have found it difficult to imagine health states that they had not experienced. Nevertheless, possible measurement error did not manifest itself in the form of systematic departure from theoretically valid choice patterns. Parameter estimates also were well-ordered and generally significantly different than one another.

Our sample was drawn from a large consumer panel used for market research and political polling. The panel is not strictly representative of the general population. Minorities are somewhat underrepresented and the sample is somewhat better educated and has higher income than the general population. The statistical analysis presented here employs demographic weighting to account for possible oversampling, or undersampling of particular groups relative to the general population. Data were weighted to reflect the national population of women aged 46 to 60 years for key demographic variables including age, race and ethnicity, geographic region, education, and income. Demographic weights were based on US Census data obtained from the March 2004 Current Population Survey [31]. It is important to note, however, that the prevalence of hormone replacement therapy use is higher among nonhispanic white women, more educated women, and women with higher incomes [32]. Our sample thus may be more representative of women likely to seek treatment for vasomotor symptoms.

The results of this study likely are sensitive to the maximum duration of symptoms defined in the study and are thus not generalizable to other conditions for which a different duration of symptoms may be appropriate. In addition, to be generalizable across conditions, the best level of each attribute should be defined so that the combination of these levels corresponds to an asymptomatic health state comparable to perfect health and the survey includes a mortality risk attribute. Although we include an asymptomatic health state, the absence of death in this approach represents a systematic departure of this method from conventional health-state utility estimation.

The specification of health states and longevity in any health-utility analysis has implications for the interpretation and generalizability of the results. In a conjoint analysis, what are considered to be clinically relevant levels and ranges of levels of each attribute are condition-specific and some may disagree as to whether any given set of attribute levels is appropriate to describe the condition. The effect of different specifications of attributes and attribute levels on utility estimation is impossible to determine without conducting a study that compares directly two or more alternative specifications. It is important to note that, however, the same is true for any method of health-state utility elicitation or evaluation.

Other studies have estimated the impact of vasomotor symptoms on health-state utility. In one of the earliest studies of the cost-effectiveness of hormone replacement to treat vasomotor symptoms, Weinstein et al. used the authors' collective judgment to determine the health-state utility impact of HRT [33]. Noting the lack of empirical evidence in Weinstein's approach, Daly et al. administered a rating scale (VAS) and a TTO instrument to a convenience sample of women at a clinic in the UK [34]. Daly et al. defined mild and severe menopausal symptoms and estimated both VAS and TTO utilities. Zethraeus et al. used the health states defined by Daly et al. with a larger sample of women [35]. The utility impact of VMS in nearly all of these studies is large, perhaps much larger than would be expected when compared with utility estimates for diseases that are generally believed to have a significant impact of quality of life such as cancer. Nevertheless, our results are consistent with earlier findings that VMS has a large impact on women's quality of life.

The most obvious limitation of the approach we suggest here is its relative complexity compared to SG and TTO approaches. Most health-state utility studies elicit trade-offs for predefined health states or apply utility function algorithms to data collected using quality-of-life instruments. The resulting health-state utilities are combined arithmetically with durations and numbers of affected patients. The conjoint approach requires a separate time-consuming, resource-intensive study. The study also requires specialized skills in survey research, experimental design, and statistical analysis.

There are three potential advantages to conjoint-based HYEs rather than SG- or TTO-based QALYs to evaluate improvements in acute self-limiting conditions. First, because SG and TTO health-state utilities estimated to assess chronic conditions may yield unrealistically small QALY gains for acute conditions. Likewise, it may be difficult to elicit trade-offs involving instantaneous painless death or life expectancy for conditions for which death is not a realistic outcome. Conjoint studies can be designed to elicit trade-offs in clinically relevant trade-off ranges and, thus, may avoid understating gains. Second, unlike conventional techniques used to estimate QALYs, conjoint analysis can be used to estimate interactions between symptom or disease severity and duration. Finally, conjoint analysis does not impose linearity assumptions that are inconsistent with neoclassical utility theory. Despite these theoretical advantages of conjoint analysis, further studies directly comparing the conjoint-analysis approach to estimating HYEs to conventional SG and TTO approaches to estimating QALYs are required to provide more definitive evidence of the value of each of these approaches to assessing health interventions.

Source of financial support: This study was funded in part by a contract with Wyeth Pharmaceuticals, Collegeville, Pennsylvania.