A discrete choice experiment is a research technique where participants are presented with hypothetical scenarios and asked to choose their preferred option from a set of alternatives. Each scenario consists of various attributes with different levels, providing a detailed picture of what factors are most important to respondents. In some DCEs, the options are labelled with specific names (like brand names), whilst in non-labelled ones, only the attributes are shown without specific identifiers. Participants typically go through multiple repeated choice task, comparing and selecting their preferred option several times. This method helps researchers understand preferences and trade-offs people are willing to make.

An online DCE survey was administered to a sample of female breast cancer survivors who had completed treatment within the last five years in Australia. Participants were predominantly recruited through the Breast Cancer Network Australia Review and Survey Group (a national group of 1374 Australian women living with breast cancer who are interested in receiving invitations to participate in research and members of one of the largest consumer advocacy networks for people with breast cancer in Australia). Additionally, the survey was advertised through existing networks of the research team. Emails were sent out to potential participants in August 2022, with reminders sent in November 2022 and January 2023 and the survey closed in 2023. Ethical approval was obtained for this study from the Queensland University of Technology Human Research Ethics Committee on 28/10/2021 (ID: 4567- HE31).

The DCE was a non-labelled survey with respondents presented with two hypothetical scenarios at a time, known as choice sets. The selection of the final set of attributes and levels for this DCE was based on a review of the literature, focus groups with consumers and health service providers, a quantitative structured prioritisation exercise, and an expert panel discussion [6, 13]. The breast cancer follow-up care models were described by five attributes and their corresponding levels (Table 1).

In the study, the concept of “follow-up care” was conceptualised as a comprehensive care model for individuals who have completed cancer treatment. This model aims to enhance the overall well-being of cancer survivors by including various aspects, such as surveillance for cancer recurrence and screening for secondary cancers, monitoring of physical late effects resulting from cancer and/or treatment, management of psychosocial concerns, promotion of health and addressing comorbidities. The second attribute, “allied health” professionals, were defined according to the Medicare scheme. This includes Aboriginal health workers or Aboriginal and Torres Strait Islander health practitioners, audiologists, chiropractors, diabetes educators, dietitians, exercise physiologists, mental health workers, occupational therapists, osteopaths, physiotherapists, podiatrists, psychologists and speech pathologists. The number indicated in the levels corresponds to the number of sessions or consultations provided by these professionals. In the scope of this study, the term “Survivorship Care Plan” denotes a comprehensive document shared amongst the survivor, oncologist and the broader care team. This document encompasses various elements, namely: a summary of the cancer treatment received, a clearly outlined schedule for follow-up appointments and screening tests, including the contact information of the healthcare professionals involved in the treatment and ongoing care, a compilation of potential symptoms to be vigilant of and the potential long-term side effects to anticipate, identification of medical, emotional, psychological or social needs post-treatment, along with strategies for their management, clarification of the roles and responsibilities of different members within the healthcare team and the appropriate points of contact in case of concerns, and recommendations for adopting a healthy lifestyle post-treatment.

The online DCE survey consisted of three sections. First, the respondents received information on completing a DCE task and were shown a sample task (Fig. 1). Demographic information, such as gender, age and education level were also collected to summarise the characteristics of participants. The second section consisted of DCE tasks including one repeated and one dominant choice task to assess response validity. The third-choice task, which was one of the ten choice tasks, was repeated at the end of the ten main choice tasks. The proportion of participants who provided identical responses to both tasks was used as an estimate for assessing the internal reliability and consistency of responses. Additionally, a choice task featuring an apparent dominant alternative was administered following the main DCE tasks. The proportion of participants who correctly identified this dominant option served as a secondary estimate for evaluating the internal reliability and consistency of responses[14, 15]. In the third section, respondents completed the 5-level EuroQol 5-dimension (EQ 5D-5L) multi-attribute utility instrument. EQ 5D-5L is a generic quality-of-life instrument that has been used amongst people with cancer [16], and recently published Australian-specific tariffs were used to estimate the EQ 5D-5L utility scores [17].

Example of a choice task seen by respondents

The final attribute list and corresponding levels would result in 26,244 (38 × 22) possible choice tasks. Since it is not feasible to present all possible combinations to all respondents, 20 choice tasks were selected using a fractional factorial design, which makes up the choice sets. The main aim of using a fractional factorial design was to have a manageable number of choice tasks whilst maximising the design’s statistical efficiency [18]. Prior research [12, 19] indicates that individuals are capable of effectively responding to ten-choice sets simultaneously. Therefore, the fractional factorial design in this study was divided into two blocks, with each participant being presented with only ten of the possible 20 choice tasks.

The final DCE design used a Bayesian D-efficient design with normally distributed priors generated using Ngene software [20]. The priors were derived through pilot testing with 28 respondents. Since there was no existing data available regarding the coefficients for the various attributes, non-informative priors—small positive or negative priors, or zero priors—were utilised to develop the D-efficient design for the pilot (Supplementary Table 1 and Supplementary Fig. 1). The design was constructed based on a priori hypotheses. The final Bayesian D-efficient design was optimised using the Modified Federov algorithm employing 1000 Halton draws (Supplementary Table 1 and Supplementary Fig. 2). The Bayesian D-error for the design is 0.1355. In addition to the D-error, the final design was assessed based on attribute level overlap—where attribute levels are present in both choice tasks, and attribute level balance—the distribution of attribute levels across the two choice tasks. A lower degree of attribute level overlap and equal distribution of attribute levels are indicative of an optimal DCE design (Supplementary Table 2).

To accommodate individual preferences, a latent class modelling (LCM) approach under a random utility framework was used for the analysis [21]. The random utility framework assumes that the participants chose the alternative that maximised their utility. LCA identifies unobservable, or “latent,” subgroups within participants’ preferences. LCA postulates that within each latent class, preferences are homogeneous, yet these preferences differ distinctively across classes [22]. The assignment of respondents to these classes is probabilistic, estimating the likelihood of each respondent belonging to a specific class based on their responses. Incorporating respondent characteristics enhances the precision of these probabilistic class assignments. The selection of an optimal number of latent classes was informed by statistical fit measures, such as the Akaike Information Criterion (AIC) and log-likelihood values, ensuring the identification of genuine preference patterns, rather than mere stochastic variations (Supplementary Table 3). Socio-demographics and other data collected as part of the online survey were analysed using descriptive statistics. Mixed logit and multinomial logit regression models were also fitted, but they were inferior to the latent class model (Supplementary Table 3).

Furthermore, we estimated the willingness to pay (WTP) values for each attribute level. The WTP metric, commonly utilised in DCE, quantifies the monetary value respondents assign to specific attributes or their changes. The WTP is estimated by dividing the utility coefficient of a given attribute by the utility coefficient of the cost attribute. This ratio, termed the marginal rate of substitution, indicates the monetary value respondents associate with a unit change in the attribute in question. These WTP estimates indicates the relative importance respondents assign to distinct healthcare attributes and reflect the monetary trade-offs they are predisposed to make in their healthcare decisions.

All analyses were conducted using the NLOGIT 5 software [23].