MCDA was used to predict which S1PR modulator clinicians would prefer on the basis of clinical management convenience and why. To inform the MCDA, clinicians’ preferences for different attributes of S1PR modulator clinical management were quantified in a DCE administered within an online survey. In the MCDA, partial MCDA scores, which captured how well an S1PR modulator would perform on a given attribute, were calculated by weighting real-world S1PR modulator profiles by their corresponding preference weight as estimated in the DCE. An overall clinical management MCDA score was calculated for each S1PR modulator by summing its partial MCDA scores.

Attributes to be included in the DCE were directly informed by US Food and Drug Administration (FDA) product labels for fingolimod, siponimod, ozanimod, and ponesimod [2,3,4,5] (Table 1, Supplement 1). These product label attributes were discussed in detail at two MCDA workshops (Supplement 2), and their final definitions were confirmed by our clinician co-authors with experience in the field. The profiles of S1P receptor modulators are generally similar in terms of their benefit–risk and safety profiles but require differential clinical management events, the convenience and feasibility of which pose a preference-sensitive situation. The eight attributes included were: first-dose observations, eye exams, genotyping prior to treatment, liver function tests, drug–drug interactions, interacts with antidepressants, interacts with foods high in tyramine, and immune system recovery time. This study focused exclusively on lifestyle management, changes, or tests that could impact a patient’s daily life. Therefore, efficacy, side effects, and treatment administration were not examined for this study. The inclusion of drug–drug and antidepressants interactions refer to the necessity of having to avoid or cease taking other drugs, impacting patient life, rather than the possibility of side effects. Interacts with antidepressants was included as a distinct attribute from drug–drug interactions because mental health conditions are a common co-morbidity for patients with multiple sclerosis (MS) [12,13,14].

Example choice task. Example choice task presented to participants. On each choice task, participants viewed profiles of two hypothetical treatments and were asked to choose which they would prefer. Attribute levels were varied on each task.

Attribute levels used in the DCE were identical to those of real-world S1PR modulators, with two exceptions. First, the DCE added an additional level for the drug–drug interactions attribute to allow for a more even increase between levels. Fingolimod, siponimod, ozanimod, and ponesimod have have 1, 2, 2, and 6 types of drugs causing DDIs, respectively. The levels appearing in the DCE choice tasks were 1, 2, 4, and 6 to use linear interpolation to generate the weights. Linear interpolation was applied to DCE data to generate weights for the drug–drug interaction attribute. Second, the eye exams attribute included levels for one eye exam and two eye exams, even though ozanimod only requires an eye exam for patients with a history of diabetes mellitus or uveitis [4]. This conditional eye exam was collapsed into the one eye exam level to limit experimental complexity and to focus on preferences for the number of eye exams. Immune system recovery time levels were based on FDA labeling for the median time for peripheral blood lymphocytes to return to the normal range after the drug is stopped. Where a range of time was given, the upper limit was always selected, and all times were converted to weeks.

Cognitive pilot interviews of ten clinicians were conducted to confirm the relevance of the selected attributes and ensure that the survey was comprehensible and clear (Supplement 2). Participants completed the survey in its entirety with a moderator assessing comprehensibility. All participants understood the attributes and the impact these CMEs would have. All attributes were determined to be important and relevant to clinicians and were therefore included in the final survey.

A quantitative pilot study with 80 clinicians was then conducted to examine data quality and to assess whether participants would be able to distinguish the different levels of the attributes and make trade-offs between attributes (Supplement 2). No additional questions were asked of participants in this phase and no changes to the choice design, attributes, or levels were made after the quantitative pilot.

The study was approved by Ethical & Independent Review Services, a fully accredited institutional review board (study numbers 22022-01 and 22022-01A) and conducted in accordance with US regulatory requirements.

This study included neurologists with experience treating ≥ 5 patients with MS/month in the last year. Participants had to currently reside in the USA and be able to read and understand English. Potential participants were identified via online databases, panels, and social media, invited to participate by email, and provided online consent before taking part in the study. The main DCE study targeted 200 clinicians, a sample size comparable to that used in other DCEs [15, 16].

Participants completed a screening form to confirm their eligibility and a questionnaire collecting information about their demographic characteristics and professional experience. Participants were then provided text introducing and defining the clinical management attributes used in the DCE. Following this, participants completed the DCE. Within each DCE choice task (Fig. 1), participants were asked to choose between two hypothetical S1PR modulators described by the eight selected clinical management attributes.

The levels of these attributes were varied according to a D-efficient experimental design generated using Ngene software (ChoiceMetrics, Sydney, Australia) [17] to identify the subset of choice tasks that ensured all effects of interest can be estimated independently [18]. In the main study, the design used estimated parameters on the basis of the probability distribution of the quantitative pilot data (i.e., Bayesian priors). The design was split into three equal blocks to minimize the burden for each individual participant: each participant was randomly allocated a block of 12 experimental choice tasks, in line with the number of tasks presented in similar DCEs [10]. The order of the experimental choice tasks and the order of the attributes in the task were randomized between participants.

A practice task was given prior to the experimental tasks to familiarize participants with the DCE format, and two tasks were given following the experimental tasks to assess the internal validity of participants’ responses [19]. The first internal validity test repeated the participant’s third experimental choice task to explore whether the participant made consistent choices. The second validity test presented a choice alternative that outperformed the other on all attributes to assess whether participants would select the dominant (i.e., superior) option. Choices made in these tasks were not included in the main analysis.

Statistical analyses were conducted using R (version 4.2.1, R Foundation, Vienna, Austria). Statistical tests were two-sided and used a significance level of 0.05.

Data from the experimental DCE tasks were analyzed following the random utility maximization framework, which estimated the effect of changes in attributes on preferences as part-worth utilities (Supplement 2) [20, 21]. Analyses used a mixed-logit model to account for unobservable heterogeneity (i.e., preference differences) that could not be explained by observed participant characteristics [22]. Preferences were assumed to vary among participants according to a normal distribution [22].

The relative attribute importance (RAI) scores were calculated to measure the maximum percentage contribution of an attribute to a preference relative to all other attributes and their levels [20]. RAI scores sum to 100% and capture the relative importance that participants placed on each attribute and the proportion of trade-off decisions that were influenced by a particular attribute [20].

The clinician preference data generated from the main DCE study were used in an MCDA to estimate (1) which among the approved S1PR modulators clinicians would choose on the basis of clinical management convenience and (2) the impact of each clinical management component in driving this decision. The MDCA weighted the real-world clinical management profiles of the four FDA-approved S1PR modulators by the marginal utility (i.e., incremental preferences) for each clinical management attribute (Supplement 2). Partial MCDA scores quantified how much an S1PR modulator gained for its performance on a given attribute. Partial MCDA scores were calculated on the basis of the marginal utilities derived from the DCE model that corresponded to the treatment performance measures; formulas for calculating the partial MCDA scores can be found in Supplement 2. MCDA overall MCDA scores were derived by summing the partial MCDA scores across all attributes. S1PR modulators with higher overall clinical management MCDA scores represented more strongly preferred treatments, and attributes with higher partial MCDA scores represented aspects of clinical management that more strongly influenced participants’ preferences.