2.1

Conjoint analysis

Conjoint analysis is a type of tradeoff analysis that can provide insights on how participants value different attributes in complex decisions and has been validated in men with PC [ , , ]. We created a conjoint analysis tool for clinically localized PC using an adaptive choice-based conjoint analysis (ACBC) platform(Sawtooth, North Orem, Utah). This computer-based exercise asked participants to work through 12 scenarios and choose either “definitive local treatment” (corresponding to choices with lower 5-, 10-, and 15-year cancer mortality risk with higher risk of side effects) or “nonaggressive” treatment (corresponding to choices with no reduction in cancer mortality risk and lower risk of side effects) in each scenario ( Fig. 1 , Appendix Fig. 1 ). Participant demographic data was collected as well: age, race/ethnicity, education level, and income. Health literacy was also evaluated using a single-item literacy screening tool [ ]. The exercise was piloted by 10 nonphysicians whose feedback informed survey design and whose results were excluded from final analysis.

Example of a single conjoint analysis exercise.

2.2

Conjoint attributes and levels

Because best practices for conjoint analysis recommends 4–5 attributes for each scenario, we included risk of cancer death and the most common side effects of radical prostatectomy (ED, UI) and radiation therapy (IUS) as decision attributes. Participants were also asked to consider a defined LE based on their age and self-defined comorbidities in the Charlson Comorbidity Index. Unlike other attributes, LE remained static for all scenarios and was considered as a fixed effect in models, consistent with how LE remains constant in real-world decisions about treatment.

The conjoint exercise was based on the key decision attributes as described in the AUA guidelines for prostate cancer shared decision making.[ ] We used outcomes data from prospective case series and randomized controlled trials comparing definitive treatment to watchful waiting to define a plausible range of risk of key attributes in the setting of definitive local treatment and CM [ , ]: (1) cancer mortality (5/10/20-year 0%/0%/5% for low-risk disease, 2%/5%/12% for favorable-intermediate-risk, 4%/8%/16% for unfavorable-intermediate-risk, and 5%/11%/18% for high-risk); (2) UI (5%/10%/15%/20%/25%/30%/35%), (3) ED(5%/20%/40%/60%/80%/100%), and (4) IUS (5%/10%/15%/20%/25%/30%/35%/40%).

2.3

Covariates

Covariates included in data analysis were age, race/ethnicity, household income, education level, health literacy, region, and self-reported comorbidities.

2.4

Participants and online crowdsourcing

We conducted an online crowdsourced conjoint analysis survey of men who demographically match the SEER dataset of US PC patients in terms of age, race/ethnicity, and regional demographics [ ]. We elected this approach to avoid bias associated with sampling men with pre-existing prostate cancer, who would be biased based on their previous counseling and experiences. All participants provided written consent. The conjoint analysis exercise was performed in collaboration with Cint®, a global survey research firm with a marketplace of diverse individual users who are compensated to share their opinions.[ ] From the over 1 million available survey respondents in the US, we used the Cint platform to identify a target population of 2,000 US men matching the key demographic characteristics noted above. The number of users queried and survey response rate was not provided by Cint. For completing the exercise, participants were rewarded with a prespecified number of marketplace points to be redeemed through Cint’s online payment partners.

2.5

Statistical analysis

A discrete choice modeling approach was used to assess individual treatment preferences across 4 key tradeoffs: risk of cancer mortality, ED, UI, and IUS. We then utilized a multivariable multinomial conditional logistic regression model to examine treatment choice across 12 binary choice tasks across attributes, while adjusting for demographic variables (age, comorbidity, race, ethnicity, education, and income). Additionally, an interaction term was introduced to assess the impact of LE on treatment choice. We then estimated marginal effects of LE on treatment choice overall and across tumor risk. We intentionally departed from the traditional application of conjoint analysis, which primarily focuses on quantifying the part-worth utilities of decision attributes, given that we wished to assess the impact of LE, a static attribute at the patient level, on treatment choice across varying risk of cancer risk and side effects. All statistical analyses were performed using R software version 4.2.3, with statistical significance determined by 2-sided P-values of less than 0.05.