With requests for greater involvement and participation of patients and the general societal shift toward greater self-responsibility, self-determination, and autonomy, the role of the patient is changing (Reference Klemperer and Rosenwirth1). To fully and objectively assess clinical needs and treatment decisions, patient perspectives should be included.

With regard to optimal treatment of neuroendocrine tumors (NETs), decisions on approval, appraisal, utility assessment, and ultimately the pricing of health technologies are made by experts, with judgments grounded on evidence of clinical effectiveness obtained from (randomized) clinical trials. So far, decisions important from the perspective of NET patients and their next-of-kin, and how they are weighted against each other, have not yet been considered. As different preference studies have shown, decisions from experts do not always correlate with the subjective preferences of the patient, and patient criteria are not necessarily congruent with the endpoints set by experts (Reference Mühlbacher and Juhnke2). In addition, endpoints can be weighed differently in decision-making. Finally, subgroup differences in terms of preferences between different patient populations can also occur (Reference Mühlbacher and Bethge3).

NETs are slow-growing tumors that arise in the gastrointestinal tract, endocrine pancreas, lung, or thymus and are often malignant (Reference Granberg4). Many secrete hormones and thus provoke endocrine symptoms. Up to 10 percent of all NETs (Reference Kulke and Mayer5), especially midgut NETs, secrete high concentrations of hormones (Reference Kvols6). Symptoms caused by these NETs are collectively known as carcinoid syndrome (Reference Kulke and Mayer5).

NETs are relatively rare, having an annual incidence of approximately 2.5 to 5 per 100,000 people in Germany (approximately 1,600–4,000 per year) (Reference Arnold7) and prevalence of approximately 35 per 100,000 (Reference Öberg and Castellano8). Recent estimates show a steady increase in the incidence of NET and a dramatic increase in its prevalence; however, survival rates have increased, which is probably due to better diagnosis and treatment (9;10). The treatment of NET is multi-modal (Reference Melmed, Casanueva and Cavagnini11) and aims to control symptoms and reduce circulating hormone levels, thus preventing further tumor growth. Reducing tumor size may improve the chances of survival and quality of life of patients (Reference Colao, Faggiano and Pivonello9).

Drug treatment for patients with NET has several aims: treat the hormonal symptoms (anti-symptomatic treatment), improve the quality of life, and delay tumor growth (anti-proliferative therapy). The choice of a drug depends on the particular tumor type, because there are few drugs that are suitable for universal use (7;12). Biotherapies include mainly two groups of substances: somatostatin analogues and interferon-alfa (4;7;12). Several other new drugs are also available with different mechanisms of action, including cytostatics, tyrosine kinase inhibitors, and anti-angiogenic drugs; recently, radiolabeled substances have been introduced. Many of these compounds have shown promising results in patients with NET (Reference Granberg4).

Due to lack of information on patient preferences regarding drug treatment options of NET in Germany, identification and weighting of patient-relevant outcomes for decision-making are necessary. Obtaining preference criteria from patients should improve decisions about usage of health technologies in the short term (as preference-based treatment alternatives lead to higher compliance and adherence), strengthen patient orientation in the medium term (as patient-centered treatment alternatives and communication enhance patient understanding), and improve clinical effects in terms of morbidity and mortality in the long term (as higher compliance and adherence will lead to better treatment outcomes).

Against this background, the following questions are to be answered in the context of an empirical study with German participants: What are the key decision criteria for the selection of an optimal drug therapy from the perspectives of patients as well as their relatives? How do patients and relatives weigh these different decision criteria when selecting the optimal NET drug therapy? Are there differences in the weighting of the relevant decision criteria for different sub-groups? In this empirical study, we documented patient preferences regarding the drug treatment of NET. A discrete choice experiment (DCE) was used to rank patient-relevant treatment characteristics. The goal of the survey was to provide a basis for patient-oriented evaluations of innovative treatment options in NET and to assess patient perspectives on the basis of scientific evidence. Our overall aim was to expand the scientific clinical findings in the field of NET through assessments of treatment from the perspectives of patients and family members by means of preference data.

Methods

Elicitation of Patient Preferences through DCEs

The DCE is a choice-based version of the conjoint analysis, which was made possible by the theoretical work of Lancaster (Reference Lancaster13) and McFadden (Reference McFadden and Zarembka14). Instead of ranking or rating different therapeutic features (as in traditional importance elicitation formats and conjoint analyses), DCEs perform a pairwise comparison of hypothetical alternatives (differently configured therapy options) and ask participants to choose between them (Reference Ben-Akiva and Lerman15). Thus, respondents are forced to make trade-offs between attributes and their levels. This method offers practical advantages, such as closeness to reality, as trade-off decisions are part of everybody's everyday life. The implementation of pairwise comparisons considerably reduces the degree of complexity of the tasks for participants (Reference Hensher, Rose and Greene16–Reference Lancsar and Louviere18). Therefore, DCEs have been increasingly used in health economics and health service research (19;20). The structure of a DCE and its analysis are multi-stage (Reference Bridges, Hauber and Marshall21–Reference Mühlbacher and Johnson23). Several checklists are available and were considered during design of this study (16;18;24).

Decision Model: Attributes and Levels

A literature search was conducted on indications for NET (by means of PubMed and Medline) to document available state-of-the-art treatment options. Effects tables were used to structure the relevant literature. An effects table is a qualitative tool for displaying a concise summary of the key benefits and risks of a drug and summarizes key results from clinical studies in a certain indication (Reference Pignatti25).

Before the main survey was conducted, a preliminary qualitative study in nine NET patients who underwent semi-structured personal interviews was conducted. Within these, a list of treatment characteristics extracted from the literature was discussed and tested for relevance from the patient's perspective. Patient priorities were elicited using an analytic hierarchy process (AHP) (26;27).

From these preliminary results, a DCE survey was developed using the most relevant aspects from the AHP, which was pretested in another nine qualitative personal interviews. These interviews were used to evaluate the clarity of the questionnaire design, the quality of the scales used, and the comprehensibility of the attributes, levels, graphics, and trade-offs included. This adaptive approach allowed the quality of the instrument to be optimized and questionnaire finalized (Reference Helm and Steiner28). A scope test to test for possible recoding of attribute levels on an ordinal scale was not included. However, visualizations used in the DCE to present attribute levels were also tested during the interviews. Figure 1 shows the final framework of the attributes, the characteristics with their explanations, and the associated icons.

Fig. 1.

Overview of the decision model with visualizations.

The description of the attribute levels was created in a style used in the classification scheme by the U.S. Department of Health and Human Services and the Office of Human Research Protection of the Food and Drug Administration (29). Furthermore, a detailed patient-friendly introduction to every attribute and related level was included in the survey. Attributes related to side effects were introduced with respect to severity and risk to explain the different aspects of the attributes and to avoid any free interpretations.

Results

Socio-demographic Data and Health Responses

Participants could choose whether they filled in the online questionnaire or the identical paper-and-pencil questionnaire. Most (74.6 percent) opted for the online version. Of the total sample of 275 participants, 51.6 percent were female. Mean age was 58.4 years (range, 28–78 years; standard deviation [SD] = 10.05). As shown in Supplementary Table 1, 77.1 percent of the participants were married. Of total participants, 26.2 percent held an intermediate high-school diploma or technical college qualification and 42.5 percent had a university degree. Of total patient responders, 82.1 percent rated their current health as “good,” “very good,” or “excellent,” with 17.5 percent indicating that their current health was “not so good” or “poor.”

Mixed Logit Model (Random Parameter Logit Model) Results

Because the linearity assumption cannot be taken for granted, a nonlinear model was estimated. Thus, effects coding was used, and no function was assumed about the attribute levels. With this coding, a coefficient could be calculated for each level. The random parameter logit (also “mixed logit”) takes unobserved heterogeneities between the survey participants into account. Figure 2 and Supplementary Table 2 illustrate the corresponding values. In the mixed logit model, the significance of the individual therapeutic properties must be evaluated by differences between the coefficients of the best and worst levels. If one compares this more differentiated analysis with the previously presented analysis of the conditional logit model, the relevance of the decision criteria shows a similar structure. If the different risks (of occurrence) are analyzed separately from the effect-coded model shown here, it becomes apparent that no clear rankings for side effects can be derived due to overlaps among confidence intervals.

Fig. 2.

Random parameter logit model (95 percent CI and SD).

AIC and BIC were used to select the models. Both information criteria follow the idea that a model should not be unnecessarily complex and balance the quality of fit of the estimated model to the available empirical data (sample) and its complexity measured by the number of parameters. The number of parameters is considered “punitively,” meaning that, for both information criteria, the model with the lowest value of the information criterion is best. Within the analyses, the model quality, measured by AIC and BIC, improved significantly in the combined model. Thus, for further analyses and for ease of illustration, the compound attributes with six levels (side effects) were summarized according to the primary dimension of the attributes to three levels. The summary was based on the severity of occurrence (“mild” to “severe”) for the occurrence of “diarrhea” and “nausea/ vomiting.” In the “occurrence of abdominal pain,” the summary was based on the probability of occurrence (15 percent to 25 percent).

The meaning of each therapy characteristic must be evaluated by differences in coefficients of the levels. In addition, the mixed logit model allows an analysis of existing heterogeneities within the individual attributes and levels. Because all attributes have been adopted as “random parameters” within the calculation model, the extent of heterogeneous preferences within the sample can be inferred from the SD. The corresponding SDs are shown in Supplementary Table 2.

The random parameter logit model resulted in the following order. The preference analysis (considering the 95 percent confidence interval) showed dominance for the attribute of “overall survival” (level difference [LD] = 3.229), which was followed by “response to treatment” (LD = 1.440), “stabilization of tumor growth” (LD: 1.236), “occurrence of nausea/ vomiting” (LD = 1.218), and “occurrence of diarrhea” (LD = 0.852). In the therapy decision of NET patients, “abdominal pain” (LD = 0.150) did not seem to play a decisive role, which is expressed by the lack of significance for two of the three levels.

Subgroup Testing Through Latent Class Analysis

Because the mixed logit model revealed significant SDs for most levels, resulting from heterogeneities within the sample, an exploratory analysis was generated by means of latent class models. A latent class model divides the population into several classes with differences in sensitivities across classes. With the class membership being treated as a latent component, a probabilistic class-allocation model is used. The class-membership probabilities are a function of socio-demographic attributes of the decision maker, meaning that any taste heterogeneity across respondents can be linked directly to variations across individuals. The standard mixed model has a rather low degree of flexibility in terms of the analysis of the extent of heterogeneity, and the possibility of identifying variations due to patient characteristics is quite limited (35;36). The latent class model has significant advantages in interpretation over the mixed model. Using LC, one is able to link taste heterogeneity to socio-demographic indicators instead of just knowing that a certain sensitivity follows a certain (assumed) random distribution in the study sample (35;37).

When reviewing model accuracy using AIC and BIC, a three-class model was identified as suitable to reflect existing heterogeneity in responses. In addition, based on the probability of class membership, structure variables could be analyzed regarding affiliations and possible correlations between socio-demographic variables, and preference weightings were derived. To determine class membership, significant differences in group distribution were first analyzed using cross tables, which were confirmed by mean value comparisons. From the significant structural parameters, a regression model was calculated. In total, five structural and personal characteristics were identified that influenced the respective probability of class membership: “gender,” “marital status (married),” “living in a household with a partner,” “worries about the future,” and “self-rated health.”

The preference pattern of the first group (class 1 [n = 54; 19.5 percent]) was characterized by a very high rating of the three outcome attributes (Figure 3). In class 1, order of preference was “response to treatment” (LD = 1.989), followed by “overall survival” (LD = 1.613) and “stabilization of tumor growth” (LD = 1.347). Members of class 1 were predominantly male. More than in the other two classes, participants in class 1 indicated “less/not at all” when asked about concerns regarding future health. In the other two classes, participants were moderately or very concerned regarding future health.

Fig. 3.

Preference patterns for class 1, 2, and 3 (95 percent- CI).

Class 2 (n = 95; 34.84 percent) showed significantly different preference patterns versus class 1. Side effect attributes were significantly more important, with “nausea/vomiting” (LD = 1.677) being most relevant for this subgroup, which was followed by “occurrence of diarrhea” (LD = 1.166) and “overall survival” (LD = 0.964) (Figure 3). Participants in class 2 were predominantly and significantly female. In class 2, an above average number of participants lived in single households (as opposed to the results of class 3). Accordingly, class 2 included the lowest proportion of married participants (Table 1). This effect was statistically significant.

Table 1.

Latent Class Analysis (Three-Class Model)

−−, low number of respondents; ++, high number of respondents.

In class 3 (n = 126; 46.0 percent), a noticeable preference was shown for “overall survival” (LD = 4.519) (Figure 3). This most important preference was followed by “stabilization of tumor growth” (LD = 1.130) and “response to treatment” (LD = 0.949). Participants in class 3 mainly lived together with a partner and were mainly married (Table 1). We observed a significant association between the structure variable “married” and class membership.