Physician preferences for non-metastatic castration-resistant prostate cancer treatment

The DCE is a common approach used in healthcare research to elicit preferences for treatment characteristics (i.e., attributes) and the willingness to accept trade-offs among these attributes [22]. In a DCE, respondents are presented with a series of recurring questions in which they choose a preferred option from sets of hypothetical medication profiles with systematic variations in the levels of attributes. The attributes included in this DCE (Table 1) were selected based on a literature search and qualitative interviews. The literature search focused on preferences and/or outcomes studies and clinical trials of SGARIs [8‐17, 23]. Using the preliminary list of search-identified attributes, individual phone interviews were conducted with 5 nmCRPC-treating physicians, 5 nmCRPC patients, and 5 caregivers of nmCRPC patients to evaluate the relevance from all perspectives. During the interviews, respondents were asked to look at a list of preliminary attributes and levels, rate their importance, and comment on their relevance, meaningfulness and comprehensibility. Verbatim quotes were noted and analyzed qualitatively. Based on the feedback from the interviews and discussion with the clinical experts, a final list of attributes and levels were derived and comprised 2 efficacy-related attributes (OS and time to pain progression [TPP]) and 5 AE-related attributes (frequency or severity of fatigue, skin rash, cognitive problems, risk of serious fall, and risk of serious fracture). This research focused on OS instead of MFS (the primary endpoint in the recent SGARI clinical trials) because the MFS definition varies slightly between trials and is a fairly recent endpoint, whereas OS is assumed to be universally understood by most physicians. The qualitative interviews also verified that physicians, patients and caregivers viewed OS as the most important efficacy attribute.

To create the alternative choices (i.e., medication profiles), a widely used algorithm (in SAS 9.4) was adapted to create a design that maximized statistical efficiency [24]. The final design was composed of 4 survey versions, each with 14 medication choice questions (i.e., pairs of hypothetical medication profiles; see example in Fig. 1). Each respondent was randomized to 1 of the versions. The survey proceeded as follows: sociodemographic and practice characteristics, treatment attribute descriptions, and 14 medication choice questions. An additional question was added whereby one medication profile was dominant over the other (i.e., higher levels of efficacy and lower levels of AEs) to assess whether respondents were paying attention. Attribute descriptions (appendix Tables 1 and 2, [additional files 1 and 2]) were included before the choice questions to ensure uniformity of interpretation by physicians. Descriptions of AEs and their respective severity levels were adapted from the common terminology criteria for adverse events (CTCAE) [25]. Severe AEs were based on CTCAE grade 3 and above, and mild-to-moderate AEs were based on CTCAE grades 1 and 2. In general, a severe AE affects self-care and a mild-to-moderate AE may affect ability to maintain daily activities.

The initial survey draft was pre-tested via phone interviews with 5 physicians (2 urologists, 3 oncologists) to assess clarity of survey instructions, comprehensibility of attribute descriptions, and whether physicians were willing to make trade-offs amongst attributes. Based on the pre-testing feedback, minor wording changes were made for clarity, and the ranges of levels for 4 attributes (OS, TPP, serious fall, serious fracture) were slightly expanded.

Published preference-related studies in prostate cancer have included sample sizes of 65 to 285 respondents [11, 13‐16], and the target sample size of 150 completed responses in this study followed recommendations in DCEs (additional file 3) [26]. The survey was administered online in January 2019 to a convenience sample targeting 75 urologists and 75 oncologists recruited from existing online panels which update their member profiles annually. Eligible respondents were required to have experience treating patients with nmCRPC, be a urologist or oncologist, be ≥18 years of age, and practice within the US. Physicians were excluded if they had participated in an online survey on nmCRPC within the prior 6 weeks, or if they were unwilling to provide consent. All study procedures and materials were reviewed and approved by a centralized US Institutional Review Board.

Descriptive statistics were used to analyze sociodemographic and practice characteristics. Quality checks for DCE responses were evaluated by checking for variability in responses (i.e., respondents who always chose the same medication, such as ‘medication A’, across all 14 choice questions) and assessing the dominance test. Respondents who failed both checks were excluded.

Effects-coded random parameter logit (RPL) models were used to analyze the DCE responses [27]. The RPL model yields a preference weight (i.e., coefficient) for each attribute level, which can be interpreted as the relative preference strength for each attribute level; a greater preference weight indicates a stronger preference [27]. The difference between highest and lowest preference weights (i.e., vertical distance between an attribute’s best and worst levels) is also a measure of the attribute’s overall relative importance over the ranges represented in the DCE [27]. Relative attribute importance scores were also calculated by expressing this vertical distance as a percentage of the total variance observed (i.e., summation of vertical distances across all attributes). The estimated preference weights were used to calculate the rate at which physicians trade between efficacy and treatment risks (i.e., the marginal rate of substitution [MRS]). Specifically, we looked at the amount of OS that physicians would be willing to forego to achieve a given reduction in AE risk or severity. The MRS was calculated as the change in preference weights from a given improvement in AE that could be offset by the change in preference weights from a reduction in OS. In the MRS calculation, we specified OS as a continuous variable, as this was supported by model fit statistics. Separate models were estimated for each physician specialty to examine if there are differences in willingness to trade OS for AE risk/severity reductions. All DCE analyses used STATA/IC 14.2 and R Studio 3.5.0.

Of the 150 physicians who completed the study survey, 1 oncologist was excluded from the analysis due to failing both quality checks (Fig. 2). Respondents’ sociodemographic and practice characteristics are summarized in Table 2. Compared to oncologists, urologists were slightly older and had a higher proportion of males (P < 0.05). Among all physicians, the average time in practice was about 17.9 years. The most frequently reported medications prescribed for nmCRPC were enzalutamide (70.5%), leuprolide (67.1%), and bicalutamide (47.0%).

The estimated preference weights and relative importance scores are presented in Fig. 3. Physicians’ preferences were naturally and logically ordered (i.e., higher efficacy was preferred over lower efficacy, and lower risks were preferred over higher risks). The difference in preference weights of adjacent levels indicates the relative impact of shifting from one level to the next, where a larger difference signifies a greater impact on treatment choices. The bar graphs in Fig. 3 represent the each attribute’s relative importance to physicians’ treatment choices. Of the efficacy attributes, physicians placed more importance on a 9-month improvement in OS compared to a 9-month improvement in TPP. Physicians viewed the risk attributes in the following order of importance: a reduction in cognitive problems from severe to none; a reduction in risk of a serious fracture from 8% to none; a reduction in fatigue from severe to none; a reduction in risk of a serious fall from 8% to none; and a reduction in rash from severe to none. Of note, physicians valued a reduction in cognitive problems (from severe to none) and serious fractures (from 8% to none) more than a 9-month improvement in OS. For the same amount of risk reduction, physicians placed more importance on avoiding serious fractures compared to serious falls.

Table 3 presents the amount of OS that physicians were willing to forego for various improvements in the other 6 attributes. Specifically, physicians were willing to trade more OS months for AE reductions that they viewed as more important (i.e., cognitive problems, serious fracture, and fatigue). For cognitive problems and fatigue, physicians were on average willing to trade 9.1 and 6.6 months, respectively, to reduce the AEs from severe to mild-to-moderate. To reduce these AEs from mild-to-moderate to none, physicians were willing to trade 2.5 and 0.8 months of OS for cognitive problems and fatigue, respectively, between hypothetical medication profiles. Physicians were willing to trade 3.9 months and 5.3 months of OS, respectively, to reduce the risk of a serious fracture from 8 to 5% and 5% to none. They were also willing to trade 2.9 months of OS to reduce risk of a serious fall from 8 to 5%, and 2.3 months of OS to reduce the risk from 5% to none.

The amount of OS that urologists and oncologists were willing to forego, for their patients, in return for reduction in AEs were similar (95% confidence interval [CI] overlap) except with respect to rash (Table 3). Specifically, urologists placed relatively more importance on avoiding rash compared to oncologists, and were willing to forego about 7.0 months of OS (95% CI: 4.3–11.0) to reduce rash from severe to none, whereas oncologists were not significantly willing to trade OS (CIs cross 0) to reduce rash (Table 3).