The recruitment and interview of participants for the qualitative component took place between December 2021 and March 2022. Convenience sampling was undertaken to include the Singapore population aged ≥40 years with varied engagement with CRC screening services. Recruitment was conducted via the NUS social media platforms and its email blast services, as well as other participant recruitment channels and word of mouth. Interested potential participants contacted the researchers, who verified their eligibility before taking informed consent.

The interviews took place either on the web via a videoconferencing application or in a quiet room within the NUS that was convenient for recording. The interviews adopted a semistructured format using a topic guide. Each interview lasted approximately 30 to 45 minutes and was conducted in English.

The interviews were transcribed verbatim, and the data were analyzed using thematic analysis. A preliminary codebook with emerging themes of relevance from the first 5 transcripts was developed upon full familiarization of the transcripts. A deductive and semantic approach was used in the clustering of codes into metacodes and categories of interest. The coding framework was subsequently applied to the remaining transcripts. The identified themes were also further reviewed to ensure their usefulness and accuracy in representing the data.

DCEs have increasingly become a popular method for investigating and eliciting patient and population preferences for health care [28]. The method is based on consumer choice theory [29], which posits that respondents make choices between hypothetical products or scenarios comprising of ≥2 alternatives based on the importance they place on the characteristics of these attributes. In a DCE, a product or scenario is described with a fixed number of attributes with varying combinations of levels. Per this paradigm, in choosing the ideal product or scenario, the respondent evaluates the overall desirability of the alternatives and makes trade-offs among the attributes. From the respondents’ choice, their preferences are indirectly revealed, determining the attributes that drive the respondents’ preferences as well as the way variations of the attributes and levels may affect the respective preferences [30].

The selection of attributes and levels must be relevant to the policy process and the study population, while being consistent with the random utility theoretical foundation of DCEs [31]. An initial set of attributes and levels for the DCE was based on a scoping review of the existing literature, which yielded 13 attributes. During the aforementioned qualitative interviews, participants were asked to rank 3 attributes that they valued the most, and the weighted preferences of all participants helped shortlist the final attribute list for the DCE. Following that, a quantitative survey with eligible health care professionals (n=11), who had at least 1-year working experience with patients with CRC or RC screening, was conducted to ensure the validity of the selected attributes and their corresponding levels. After these iterative processes, six attributes were identified and ultimately used in the DCE: (1) procedure, (2) pain level, (3) sensitivity, (4) recommendation, (5) out-of-pocket cost, and (6) risk of test. Each attribute was assigned various levels based on the best information available. The blood-based test was one level of the procedure attribute.

To optimize the choice sets, a pilot study was conducted with 12 participants. Adjustments were then made to the text for the attributes and levels to improve clarity for the participants. The final set of 6 attributes and levels is presented in Textbox 1. The total number of appearances and selections for each attribute and level may be found in Multimedia Appendix 1.

The DCE questionnaire was designed using Sawtooth Lighthouse Studio (version 9.13.2), and a 2-stage design was used. For each task, participants first selected the preferred choice from 2 test profiles and were then asked to choose between taking the test or opting out of it in real life. Correspondingly, the parameter labeled “Opt-Out” represents the utility associated with declining the preferred test in the first stage. A negative value thus indicates a participant’s preference to undergo the screening test. The questionnaire was designed using the random task generation method provided by Sawtooth. The DCE included a total of 20 blocks with 10 choice sets each. Each study participant saw 1 block of choice sets, consisting of 10 choice sets, from the 20 blocks. To test for internal validity, 1 fixed choice set offering 2 alternatives is common across all blocks, of which one is intended to be strictly dominant over the other.

The levels of cost were selected to reflect the costs of different procedures in reality. Participants in qualitative interviews also demonstrated a similar perception regarding the cost of the tests. Certain within-concept prohibitions were also specified to provide combinations of attributes that were realistic. This included prohibiting high out-of-pocket payment costs and the presence of mild pain for stool-based tests. However, we allowed the blood-based test to appear together with a higher cost given that commercial companies may set higher prices [32]. The coverage matrix of the DCE design was examined using Sawtooth Lighthouse Studio to ensure all the parameters can be estimated. In addition, considering the large range of costs, we treated cost as discrete variables rather than a continuous variable in our analysis. An example of a DCE choice task is presented in Figure 1.

Participants reported any familial history of CRC and if they had attended any type of CRC screening in the past. Various sociodemographic information was also collected. All variables were coded as categorical variables, and some variables were subsequently recoded as binary or ternary variables to form meaningful subgroups for analyses. Psychosocial inventories—the Zimbardo Time Perspective Inventory [33], the Intolerance of Uncertainty Scale [34], and the Duke UNC Functional Social Support Questionnaire [35]—were also included to measure participants’ degree of present orientation, intolerance of uncertainty, and social support, respectively.

This study was conducted as a web-based survey hosted on REDCap from May 19, 2022, to May 28, 2022. The target population for this study was Singapore citizens and permanent residents aged ≥40 years. The survey was sent to the participants from Singapore Population Health Studies web-based panel. This web-based panel consists of a cohort that is broadly representative of the general Singapore population. Participants gave their implicit consent by participating in the survey and were reimbursed with Singapore $10 (approximately US $7.50) for every successful completion of the survey.

In calculating the minimum sample size, we made use of the proposed formula of n>500*c/(t*a) by Johnson and Orme [36], where 500 is a fixed variable, c demotes the largest number of levels for a certain attribute, a indicates the number of DCE choice sets per block of questionnaire, and t refers to the number of alternatives per DCE choice set (excluding “Opt-out”). Accordingly, the sample size required for this study should be >125 participants (500*5/(2*10)=125). In addition, Lancsar and Louviere [30] suggested 20 responses per block or questionnaire, which led to a minimum sample size of 400. Considering that the DCE may be relatively difficult to answer, we expected a relatively high nonresponse rate of 20% to 30%. Hence, we set the target sample size to be 600. We subsequently tested the sample size using simulation functions with Sawtooth Lighthouse Studio, which deemed the sample size of 600 as sufficient. Owing to the overwhelming response, our final sample size of 1189 participants not only meets the minimum required sample size but also allows for the possibility of conducting flexible subgroup analyses.

A total of 2 models were tested: a mixed logit (MXL) model and a mixed-mixed multinomial logit (MMML) model [37]. The MXL model was selected to account for the correlation introduced by the repeated observations from each participant and to relax the assumptions of independence from irrelevant alternatives [38]. This model assumes that the choices made by the same participant are correlated, and preference heterogeneity exists across the population sample. The interpretation of the mean preference weights is made in relation to the chosen base level, and the SD of each level indicates the variability in the mean preference weights. The MMML model was also selected as it incorporates both MXL and the latent class logit model. Unlike latent class logit where a homogeneous fixed preference is assumed within each latent class, a distribution of random coefficients is specified in the MMML model. Within each class, preference weights and the average probability of each demographic within each class can be derived. The number of classes was chosen by examining the Bayesian information criterion.

In addition, the conditional relative importance (CRI) for a given attribute, defined as the difference between the highest preference weight of the attribute level and the lowest preference weight of the attribute level, was reported. A higher CRI indicates the attribute is more important in designing the CRC screening program. Profile-based normalization was then applied to normalize the sum of CRI of all the attributes to 1.

A left-specific constant was included in each regression, with a statistically significant coefficient indicating a left-right bias in the study [39]. Participants in DCE may take shortcuts and use simplifying heuristics when answering DCE questions, which can introduce an unintended source of variability in the data. When using reading order heuristic, study participants may tend to choose the choice on one side [40]. Incorporating a variable indicating the position of the choice in the regression can disentangle the associated bias [41,42].

All levels were coded as dummy variables. Reference levels were selected based on the current recommendation for CRC screening for the average-risk adult population. However, we intentionally designated “No Recommendation” as the reference level for the “Recommendation” attribute to investigate the nuanced preferences arising from diverse recommendations made by different individuals. Continuous psychological variables were demedian by subtracting the median value from each data point, resulting in a centered distribution. As for the classification of the continuous sociodemographic variables, age bands were set according to the definition of senior citizen locally, which is at the age of ≥60 years. Household income was grouped to ensure a sufficient sample size in both groups.

Statistical significance was set at P<.05. All quantitative data analyses were carried out using the statistical software R (version 4.2.1; The R Foundation) [43].

A total of 30 participants completed the interview. A summary of their sociodemographic characteristics is presented in Multimedia Appendix 2. Of these participants, 14 had undergone a colonoscopy, 20 had taken a stool test, and 6 had taken a blood-based test. A total of 5 participants had not undergone any screening for CRC. Five main themes that elucidated the participants’ motivations and important attributes of CRC screening were identified: (1) accuracy, (2) cost, (3) perceived risk and pain, (4) convenience of the test, and (5) the method of sample collection.

When it came to blood-based tests, many participants were uncomfortable with its inferior accuracy relative to other tests. They saw it as “pointless” due to its possibility of giving rise to many false positives and negatives:

For participants who felt that they had no symptoms, the level of accuracy of the stool-based tests provided sufficient assurance that they felt a colonoscopy was unnecessary. Many cited an approach of escalation—to undergo a colonoscopy only in the event of a positive stool-based test:

Participants who had undergone invasive tests such as colonoscopy either (1) were diagnosed with hemorrhoids or (2) had discovered blood in their stools. They had opted for the colonoscopy under their physician’s or friend’s advice, seeing that a colonoscopy was a more “comprehensive” or “complete” test. Many felt that the relatively higher accuracy of the colonoscopy provided them greater ease of mind:

The stool-based and blood-based tests were more favorable to most participants as the process was “simple” and “pain-free.” Furthermore, many who had undergone a blood-based test did so as part of their annual comprehensive health check-up and saw little extra risk or pain in doing it:

Many of these participants who had no prior experience undergoing colonoscopy were more likely to express fear of the risk of colon perforation from the procedure. Due to its invasive nature, many were also “scared” and “uncomfortable” with the pain the procedure might induce:

Interestingly, most participants who had undergone colonoscopy had little qualms about the risk and pain of the procedure. Instead, many expressed difficulties adhering to the bowel preparation instead:

While being relatively easy to conduct, some participants shared reservations about the collection of stool samples. They saw the process as “dirty,” “disgusting,” or “troublesome,” especially when 2 separate stool samples were required. While some complained about the uncomfortable experience of stool collection, some expressed personal concerns about improper collection:

Nevertheless, participants who have done it across many years expressed little of such concerns, seeing it as a routine exercise that was necessary:

As colonoscopies and blood-based tests require medical professionals to perform them, some participants felt that the need to arrange a doctor’s appointment was time-consuming. This was especially true for colonoscopies, where a referral from a primary care physician is required to receive a subsidized rate for the colonoscopy:

However, participants shared that stool-based tests were relatively more accessible, with kits being easily obtained at pharmacies or mailed to them on request. Even when returning stool samples, some participants shared that mailing services to the laboratory were available, which saved them the shame and hassle of dropping them off at a clinic:

Many participants were aware of the stark difference in cost between a colonoscopy and a stool-based or blood-based test. While many participants, especially those younger and working, had employer and private insurances to subsidize a colonoscopy procedure, they highlighted that out-of-pocket cost was still substantially higher. Many cited that a higher willingness to pay must come in tandem with either higher accuracy and lower frequency of testing:

A total of 1189 participants completed the study. Of these, 127 (10.7%) participants did not have complete sociodemographic information, while 44 (3.7%) participants failed the fixed choice task. This led to 168 (14.1%) participants dropping out, leaving a sample of 1021 participants for analysis. The demographic characteristics of all participants are presented in the Multimedia Appendix 2.

The main results of the MXL model are presented in Table 1.

All mean coefficients were significant at P<.05. On average, participants exhibited a higher preference for blood-based tests relative to a 2-day stool-based test (coefficient=0.40, 95% CI 0.24-0.55). Participants also exhibited a higher preference for a 1-day stool-based test relative to a 2-day stool-based test (coefficient=0.27, 95% CI 0.10-0.45). The preference for these 2 procedures were however not statistically different from each other (coefficient=0.12, 95% CI –0.04 to 0.29).

The profile-based normalized CRI of the 6 attributes based on the MXL model is presented in Figure 2. Ranking the attributes, participants were most concerned with the cost and sensitivity of the screening test. This is followed by the procedure type, the risk level, the pain level, and ultimately the recommendation received for the screening test.

The results from the MXL model suggested preference heterogeneity across the participants, with the SDs of several preferences being statistically significant. Thus, we ran an MMML model while also estimating the posterior class membership probabilities. A 2-class MMML model was the most appropriate based on the BIC. The classes were labeled post hoc as (1) strong supporters and (2) weak supporters, based on the coefficient value for “None.” A more negative coefficient value means people are more willing to take the screening test in real life. The class shares for strong supporters and weak supporters are 38.09% (n=339) and 61.91% (n=632), respectively. The main results of the MMML model are presented in Table 2. The full table with SD is presented in Multimedia Appendix 3.

Strong supporters did not exhibit a higher preference for blood-based tests relative to a 2-day stool-based test. However, weak supporters had a higher preference for blood-based tests compared with a 2-day stool-based test (coefficient=0.66, 95% CI 0.44-0.88). Similar to the results in the MXL model, strong and weak supporters were most concerned with the cost and sensitivity of the screening test. Weak supporters were more likely than strong supporters to be concerned with the procedure, pain level, and risk of test. Compared with an existing national screening program that is 2-day stool-based, has no pain, 80% sensitivity, recommended by the government’s Health Promotion Board, costs Singapore $5, and has no risk, the relative uptake rate of a blood-based screening test with all else constant will increase by 0.2% (95% CI −1.2% to 1.6%) for strong supporters and increase by 5.9% (95% CI 3.6% to 8.2%) for weak supporters.

Weak supporters were more likely to be women, ethnic Chinese, and people aged 40 to 60 years. However, strong supporters were more likely to be working, have higher income levels, and live in private housing. Strong supporters were also more likely to have a family history of CRC and to have opted for CRC screening in the past. Compared with weak supporters, strong supporters were also found to have higher scores for the perceived safety of CRC screening tests, social support, present orientation, and intolerance of uncertainty.