Public preferences for pharmacogenetic testing in the NHS: Embedding a discrete choice experiment within service design to better meet user needs

Genetic testing can be used to improve the safety and effectiveness of commonly prescribed medicines—a concept known as pharmacogenetics. This study aimed to quantify members of the UK public's preferences for a pharmacogenetic service to be delivered in primary care in the National Health Service.


| INTRODUCTION
There is significant inter-and intraindividual variation in response to medicines, a phenomenon related to several factors, including age, comorbidities and polypharmacy.2][3] Several healthcare services around the world have made significant progress in deploying pharmacogenetics in practice, although models for implementation are notably heterogenous. 4,5 the National Health Service (NHS), the limited pharmacogenetic testing currently offered is delivered reactively, where variants in single genes are tested and results are then returned to inform prescription later.One example is testing for variants in DPYD to guide prescription of fluoropyrimidine chemotherapy agents, an approach mandated by the Medicines and Healthcare products Regulatory Agency (MHRA) to avoid toxicity. 6Typically, patients provide blood samples for testing and the results, returned approximately 1 week later, can be used to guide prescribing.
Recognizing the limitations of single-gene testing, healthcare systems are increasingly adopting multigene panel-based testing. 4This involves testing individuals for many common pharmacogenetic variants, before embedding these data within their medical records to inform life-long prescription.This represents a complex healthcare intervention and, although much can be learned from existing panelbased pharmacogenetic programmes, it is highly unlikely that existing service models could be simply replicated, without adaptation, in nationalized healthcare systems such as the NHS in England. 7In part, this is due to the particular need to share data across institutional boundaries and the heterogeneous costing models and pressures that operate across different healthcare settings. 8,9e Pharmacogenetics Roll Out: Gauging Response to Service (PROGRESS) Programme (ISRCTN15390784) aims to develop, validate and pilot a panel-based pharmacogenetic service for the NHS in England, with the initial phase focussed on deployment in primary care.The programme aims to capture the views stakeholders and discrete choice experiments (DCEs), a survey-based method used to quantify stated preferences for health, states, goods or services, represent an approach by which to do this. 10,112][13][14][15] This study aimed to identify public preferences for a panel-based pharmacogenetic testing service by using a DCE, embedded within the PROGRESS programme.

| DCE design
An online DCE was designed to quantify the preferences of members of the public for a pharmacogenetic testing service in primary care for the NHS.The design process followed published guidelines. 16The DCE used 2 case studies where patients were asked to consider 1 of 2 clinical scenarios in which they presented to their primary care physician requiring medicinal treatment and pharmacogenetic testing could be used to inform prescribing.The first scenario was a patient presenting with joint pain, the second was a patient presenting with low mood.
Two distinct clinical scenarios were chosen in order to explore whether preferences for pharmacogenetic testing depended on the potential urgency of drug prescription.For each scenario, the choice question was framed as "If you had to choose one of these pharmacogenetic tests (Test A or Test B) to help guide your treatment, which would you choose?" (Figure 1).
The first scenario (pain) represented a situation where treatment may be required rapidly, whereas the latter (low mood) represented a scenario where there is less time pressure to commence therapy.For both these scenarios there are established clinical guidelines to support pharmacogenetic informed prescribing.For example, genotyping of CYP2D6 can be used to inform opioid prescription whilst CYP2C19 variants can be used to inform selection of selective serotonin reuptake inhibitors antidepressants. 17,18

| Establishing attributes and levels
The DCE used an unlabelled design with generic attributes common to both alternatives and included an opt-out option, representing no pharmacogenetic test so that the potential uptake of a pharmacogenetic test in primary care could be investigated.Attributes and levels for the DCE (Table 1) were developed with support of public stakeholders and the process is described in detail in Appendix S1.

What is already known about this subject
• Panel-based pharmacogenetic testing could be used to improve the safety and effectiveness of prescribing.
• Many healthcare services are beginning to integrate panel-based pharmacogenetic testing into routine clinical practice.
• There is no consensus how panel-based pharmacogenetic testing might be implemented in the National Health Service, and existing implementation models cannot be easily replicated.

What this study adds
• Predicted uptake is highly sensitive to design adaptations, with respondents consistently preferring noninvasive samples, and data to be reused and accessible.
• Preference heterogeneity was observed, with only a small proportion of patients consistently declining pharmacogenetic testing.
• These findings can be used to support development of future pharmacogenetic programmes.

| Experimental design
Based on the number of attributes and levels, a complete factorial design would result in 6561 (n = (1 3 Â 3 4 ) 2 ) possible choice questions.
These were rationalized to a manageable number informed using a D-efficient design, developed using Ngene software (ChoiceMetrics, Sydney, Australia).This approach selected a subset of the total possible choice sets (questions; n = 16) by generating a mathematical design whilst minimizing the D-error.The D-error represents a summary statistic that reflects the average predicted variance in the estimation of the coefficients produced by a given experimental design.
The surveys were created by allocating the 16 choice sets to 1 of 2 blocks (Block 1 and Block 2).Each respondent was asked to complete 8 choice sets, resulting in 4 surveys.There were 2 choice sets (16 questions) for each clinical scenario, with respondents only answering 1 choice set each (8 questions).

| Survey design
The survey was presented in English and programmed using Hyper-Text Markup Language (HTML) for online administration using SSI  Participants were made aware during the training material for the DCE that turnaround time represented the time from being seen in clinic by their GP to the respondent receiving their prescription.In these hypothetical scenarios, a prescription would not be issued prior to the pharmacogenetic results being available.
F I G U R E 1 Example of a question in the discrete choice experiments.
Appendix S2.After viewing the study background and the animated training material (Appendix S3), respondents were randomized on a 1:1 ratio to either the joint pain or low mood clinical scenario. 19thin each clinical scenario, respondents underwent a further 1:1 randomization to either Block 1 or Block 2 of the choice questions, meaning each respondent answered 8 questions in total.This design ensured that there were an equal number of respondents across each of the 16 choice sets, and between the 2 choice scenarios.

| Sample frame
The relevant study population were individuals from the general UK population over the age of 18 years.An online panel company (Pureprofile) was used and instructed to recruit a 50:50 male:female split which was otherwise a representative sample of the general UK population, distributed across age ranges and with representation from different ethnic groups.Eligible panel members received a link to the online survey and were reimbursed through the panel incentives at rates set by Pureprofile.

| Sample size
The number of completed surveys required for analysis depends on the number of attributes and levels, the number of choice questions and choice of analysis.There is no consensus on the best way to successfully estimate the sample size requirements for a DCE.One approach is to use the rule of thumb of Orme (500*C [Max levels]/ [Tasks*Alternatives]) which estimated that a sample size of at least 125 per DCE (250 across both scenarios) would be sufficient to estimate parameters.
Preference heterogeneity (also known as taste heterogeneity) is defined as the degree to which preferences vary across respondents.
Individual preferences for healthcare are not identical across the population, and models that present without considering that heterogeneity potentially provide an incomplete assessment of the dataset. 20ale heterogeneity is defined as the differences in error variance of respondent's preferences, a phenomenon that can occur for several reasons including different survey designs or conflicting interpretation of attributes and levels by respondents. 21,221][22] Therefore, this study aimed to recruit at least 750 respondents per DCE (1500 across both scenarios).

| Piloting
Once a final version of the survey with 2 case studies was developed, a quantitative pilot was undertaken with 50 members of the public,

| Analysis
Individual choice responses were used as the dependent variable in the model.A CLM was created for the whole dataset in the first instance and the coefficients for each attribute indicated the direction of preference for each attribute and level.Qualitative variables were effects coded and the continuous variable was treated as linear in the baseline analysis.Two distinct CLMs were first developed for each of the 2 scenarios (pain and low mood).The presence of scale heterogeneity, as the result of using 2 clinical scenario versions of the survey, was tested for using the methods suggested by Swait and Louviere (Appendix S4). 23If scale existed, then the datasets should not be merged, and each clinical scenario analysed separately.If there was no evidence of scale heterogeneity, preferences for the test would be deemed to be the same across clinical applications and the data combined.The model was then rerun to identify the best functional form (Appendix S5).A parsimonious approach taken to select the preferred model, by comparing Akaike information criterion (AIC) and the

Bayesian information criterion (BIC).
Latent class analysis was performed to identify the presence of preference heterogeneity and underlying subgroups of respondents with similar preferences using Stata (Version 14.0).The number of classes was chosen in an iterative process using the AIC and BIC estimates for each number of classes until inclusion of additional classes became noninformative (Table S2).

| Predicted uptake and comparator services
The coefficients from the final selected model and the latent class analysis were used to estimate uptake in the context of different hypothetical pharmacogenetic services.Limited pharmacogenetic testing is currently available in the NHS, but it was possible to reach consensus on what a panel-based pharmacogenetic service might look like if testing was to be delivered using current infrastructure and approaches, representing a base-case service for comparison (Table 2).
The base-case service was developed based on the DPYD testing programme, currently used in the NHS, which requires blood testing, does not make the data available for future reuse, and does not routinely make the data available to patients.Although the laboratory turnaround time for DPYD testing in the UK is approximately 5 days, it was agreed that the turnaround time for a panel-based service using current infrastructure and capacity would be best estimated at 20 days (Table 2).This turnaround time reflects the time from the patient being seen by their healthcare professional (HCP) to the point when they are issued their prescription, informed by their pharmacogenetic results.
T A B L E 2 Participant characteristics and attitudes.Predicted uptake of the base-case service was then compared against other hypothetical service templates.The level of each attribute with the largest positive coefficient was added, in turn, to the base-case service, providing an estimation for uptake if a single attribute were improved (Table 2).Finally, an optimized service was created, representing a service template where each attribute within the service had been augmented so that the level of each attribute with the largest positive coefficient added simultaneously.For this optimized service, turnaround time was set at 10 days.There is no standard approach to understand the extent of uncertainty around predicted uptake.However, we approached this by generating confidence intervals around the predicted uptake for different tests using a simulation-based approach (Appendix S5).An online web-application was developed allowing stakeholders to interact with the data and assess uptake for every possible combination of attributes (https:// shorturl.at/aehDZ).3 | RESULTS

| Participant characteristics
In total, the survey was completed by 2224 respondents.The responses of 231 respondents were excluded as their response times were under 5 min, which were considered a-priori too rapid to have properly considered and completed the survey.As such, the analysis dataset included 1993 completed surveys (Table 2).The median response time in the analysis dataset was 10.8 min.

| Model selection
The baseline CLM for each case study is provided in Appendix S6.A CLM with a single constant and linear specification of the turnaround time attribute was chosen.To assess for the presence of scale heterogeneity between the 2 scenarios, a Swait and Louviere plot was created. 23This analysis did not identify the presence of scale heterogeneity (Appendix S4), suggesting that preference was not impacted by the exemplar clinical scenario (i.e.pain vs. low mood).As such, the subsequent analysis was undertaken by merging data for the 2 sets of respondents.The best model fit for the merged datasets (pain and low mood) was found using a random parameter logit where the random coefficients were prespecified as correlated (Table 3).The selected model showed that, in general, respondents preferred quicker turnaround times to shorter turnaround times, but caution should be maintained when contextualizing this due to the likely nonlinearity of this effect.Respondents preferred a turnaround time of 5 days, compared to 10 days, and a turnaround of 10 days compared to 15, but beyond this point the relative impact of turnaround time is less prominent.The coefficient for a turnaround time of 20 days was lower than at 10 days, but higher than at 15 days, which was an unexpected finding.

| Preferences for pharmacogenetic testing
Respondents completing the surveys consistently chose a pharmacogenetic test to guide therapy over opting for standard of care, as shown by the positive coefficient for the constant term in the estimated model (Table 3).Assuming that a pharmacogenetic service was developed using current infrastructure (base-case service; Table 4), predicted uptake across all respondents was 51.0% (Figure 2).Assuming the use of blood sampling, whatever model was estimated, reduced uptake of pharmacogenetic testing, as demonstrated by the negative coefficient of this level.Respondents generally preferred saliva sampling to blood tests but had the strongest preference for cheek swabs.Introducing a cheek swab into the base-case service, in lieu of blood sampling, could potentially improve uptake by 8 percentage points to 59% (Figure 2).
The coefficient representing no future reuse of data from the pharmacogenetic test was estimated to have a negative value.This suggested that respondents were averse to their data not being reused for future prescribing decisions.Though the coefficient representing reuse by their general practitioner was positive, respondents preferred to see their data available regionally or nationally.Regional reuse of data was marginally preferred to national reuse and adding this level into the base-case service saw predicted uptake increase from 51.0 to 90.1%.
Respondents preferred being able to access their own results, with positive coefficients for receiving their results via post, email and via an online portal, the latter being the level that had the largest coefficient of any level in the model.Not providing patients access to their results had a large negative coefficient.Integrating an online portal for patients to access their results into the base-case model increased uptake from 51.0 to 90.1%.Uptake for every possible combination of attributes can be viewed using the web-application (https://shorturl.at/aehDZ).

| Willingness to wait
Respondents were willing to wait an additional 0.

| Preference heterogeneity
The latent class analysis identified that there was heterogeneity in preferences and the optimal selected model inferred that there were 4 sub-groups of preferences within the total sample (Tables S1 and   S2).

Turnaround time 20 days 10 days
Note: The base-case service was developed based on the DPYD testing programme, currently used in the NHS, which requires blood testing, does not make the data available for future reuse and does not routinely make the data available to patients.The optimized service represents a service template where each attribute within the service had been augmented so that the level of each attribute with the largest positive coefficient (Table 3) added simultaneously.
their preference for having access to their own data, and they preferred these data to be made available via a web application (Figure 4).Individuals in Class 3, 16% (n = 319) of the sample, preferred pharmacogenetic testing over no testing but it was not clear that a specific attribute or level drove their observed choices and estimated preferences.Class 4 represented a smaller proportion of respondents (3.9%; n = 78).Respondents who were members of this class were predicted to have a low intrinsic value for pharmacogenetic testing and preferred not to have the test.When respondents in this class did choose to have a test, they were particularly sensitive to the turnaround time, wanting a result as rapidly as possible.

| DISCUSSION
This study aimed to quantify the preferences for a sample of members of the public for a pharmacogenetic testing service, implemented in primary care, for the NHS.It was observed that across both clinical scenarios that respondents preferred pharmacogenetic testing over no pharmacogenetic testing.This DCE was designed, in part, to assess whether there might be differences in respondents' preferences based on the clinical context in which pharmacogenetic testing was offered.
Pain was chosen to reflect a clinical scenario where there may be a time pressure to issue a prescription, whereas low mood was chosen to represent a contrasting scenario where that same time pressure was not present.Identifying differences in preferences between these scenarios could indicate that different models of service delivery might be required dependent on the clinical scenario.However, no differences in responses were identified; therefore, further analysis was undertaken across the whole dataset.
In the final selected model, respondents wanted their pharmacogenetic data to be reused to guide future prescribing, preferring these data to be shared regionally rather than nationally, and the majority of respondents wanted to have access to their data.Respondents preferred noninvasive sampling techniques, specifically cheek swabs, over blood testing.Preference heterogeneity was observed via latent class analysis.The largest subgroup, representing half of respondents, could be characterized by their preference for having access to their own data, whilst the second largest group representing nearly 1/3 of respondents could be categorized by their preference for the data to be reused to support future prescribing.Very few respondents preferred not to have pharmacogenetic testing.
The results from this DCE highlight areas of priority to improve the uptake of a pharmacogenetic service in primary care.Within an initiative such as the NHS England PROGRESS programme, these findings can be used to develop a service designed based on the requirements of clinical stakeholders, resulting in a more impactful intervention.Relatively simple adaptations to the theoretical basecase service, such as moving from blood testing to a noninvasive sampling approach, were predicted to have a meaningful impact on predicted uptake.If pharmacogenetic testing is to be introduced at scale within any health system, reliance on blood testing via traditional phlebotomy services could represent a major barrier.
Access to phlebotomy is seen as a major challenge in primary care, and the additional burden of many thousands of pharmacogenetic tests is likely to overwhelm capacity without significant investment in infrastructure and workforce. 24Our findings suggest that the use of self-sampling approaches could improve uptake of the overall service whilst reducing reliance on a finite phlebotomy service.This could also improve turnaround time as the patient could complete the sample immediately, rather than needing to wait for a phlebotomy appointment.[27][28] Respondents in this study had strong preferences for having access to their own data, and for those data to be reused to guide prescribing in the future.In the 2022 Data Saves Lives data strategy, the NHS made a commitment to provide patients better access to F I G U R E 2 Predicted uptake by attribute.Predicted changes in uptake based on the adaption of the base-case service (blood test, no reuse of data, no patient access and a turnaround time of 20 days) by adding the level of each attribute with the highest coefficient.An optimized service reflects a model with check swabbing, regional reuse of data, an online portal for patient access to their information and a turnaround time of 10 days (Table 2).Error bars represent 95% confidence intervals.
their own data through shared care records and connected platforms, such as the NHS app. 29Giving patients access to their own medical records has been shown to deliver a number of benefits, ranging from individual reassurance and reduced anxiety, or improving selfreported levels of engagement and adherence to treatment. 30wever, providing full access to results without any explanation or summary is inappropriate and careful management and contextualisation of pharmacogenetic results will be required. 31A need for interpretation of pharmacogenetic data provides an explanation for why, in the latent class analysis, respondents who were categorized by having access to their own data (Class 2; 50.9% of respondents) preferred to access their results via an online portal where they could interact with their results.
Pharmacogenetic data are complex and, without effective contextualisation, considering the specific clinical scenario and contemporaneous guidelines, the usefulness of those data drops markedly.As such, there is a requirement to ensure that pharmacogenetic data are readily available in a clinically meaningful format at the point of prescribing, necessitating clinical decision support integration within electronic healthcare records. 32Deployed correctly, this integration would promote the longitudinal reuse of pharmacogenetic data across multiple healthcare settings, an approach that appears to be positively viewed by members of the public.Making pharmacogenetic data available for future reuse by HCPs was associated with notably increased uptake, and respondents preferred for their data to be shared beyond their own general practice.There have been several high-profile national health data programmes in the UK which have failed because of privacy and data sharing concerns. 33Pharmacogenetics has the potential to improve prescribing across a patient's lifetime, but this will require sharing of genomic and clinical data across institutional boundaries, something that will require large scale informatic and technical investment.This DCE clearly demonstrates that there is a public appetite for the sharing of this information across the NHS, but efforts must be made to learn from past failures.Previous research has shown that support for data sharing for direct care, even without explicit consent, is broad, but is not universal. 34The framing of any initiative to share pharmacogenetic data across institutional boundaries will be critical.
One approach to maintain public support would be to emphasize the role of regional structures in delivering a pharmacogenetic data sharing programme.Respondents consistently preferred regional over national data sharing strategies, a feature that was particularly evident within the latent class analysis in the group of respondents who prioritized data reuse.One driver for this observation is that national programmes could be susceptible to being identified by the public as government-driven initiatives, whereas regionally branded programmes might be seen as being led by the local NHS and local academics.According to the 2022 Public Attitudes to Data and AI Survey, this distinction matters. 35In that report, The NHS and Academic Researchers were seen as the most trustworthy custodians of data, with 73 and 61% of the public saying they trusted these parties to manage their data, respectively.Meanwhile, The Government was seen as the second least trustworthy, ahead of social media companies, with only 40% of respondents saying they trusted the Government to manage their data.

| LIMITATIONS
Although previous estimates suggest that DCEs produce accurate predictions of health-related behaviours with relatively high sensitivity, they remain susceptible to hypothetical bias.This can result in disparities between revealed and stated preferences, potentially leading to an overestimate of uptake or willingness to wait. 36For example, this study did not identify a difference in participant response based on the acuity of the clinical situation, with no preference heterogeneity identified between the low mood and joint pain scenarios.This could well reflect a hypothetical bias as one would expect that there would at least be a reduced willingness to wait for a pharmacogenetic test in the context of pain, compared with low mood.
Because of these limitations, the impact of any service design decisions or policy changes made based on DCE data should be explicitly validated by assessing revealed preferences within the context of pilot programmes.Additionally, although public preferences are a key factor in informing the design of a pharmacogenetic service, this represents just a single perspective.The opinions of HCPs who will be offering the test are also important and should be captured.In addition, this study reports data from a large sample, although it should be noted that the respondents may not be entirely representative of the whole UK population.
The outputs from this study demonstrate how a pharmacogenetic service might be designed based on stakeholder preferences to increase uptake.However, these findings do not address whether such a service should be introduced into routine practice in the NHS, and there remains uncertainty around the clinical usefulness or cost effectiveness of panel-based pharmacogenetic testing.There is increasing evidence from cost-effectiveness analysis that  2]) by adding the level of each attribute with the highest coefficient.An optimized service (Table 2) reflects a model with check swabbing, regional reuse of data and an online portal for patient access to their information.Error bars represent standard error of the mean.
8][39] However, no cost-effectiveness analysis has been undertaken to assess the impact on costs to the NHS, and outcomes for patients, of implementing pharmacogenetic panel testing in primary care.

| CONCLUSION
This study finds that the predicted uptake of a pharmacogenetic service varies considerably depending on how the service is designed.
The theoretical base-case service has a predicted uptake of 51%, increasing to over 99% if the optimized service was developed with noninvasive testing approaches, providing patients access to their data, and sharing of results for future reuse.These data can also be used to inform a decision-analytic model, allowing for an estimation of imperfect uptake to be factored into any future modelling exercise.
In a resource scarce environment, recognizing that the optimized service may be challenging to develop and maintain, these results provide a framework to support the prioritization of attributes within a service.Within the PROGRESS programme, based on these findings, noninvasive sampling has now been integrated into the recruitment pathway and a strategy for persisting and messaging pharmacogenetic data across the NHS ecosystem has been developed, with the responsibility of regional initiatives being strongly emphasized.Undertaking this DCE within the PROGRESS programme also provided a framework structured engagement of stakeholders at an early stage of the project. 40Through these findings, the opinions of patients and service users will contribute towards the development of this initiative, ensuring that it is designed to meet their needs and deliver the greatest benefit for individuals and the healthcare system, maximizing the chance of successful translation into routine clinical practice.

Web 8 . 3 . 8
Sawtooth software.The survey design is detailed in T A B L E 1 Attributes and levels developed for the discrete choice experiment (DCE).
with 25 respondents completing each scenario.These individuals were recruited from Pureprofile.A conditional logit model (CLM) was run on this dataset to assess whether the DCE had been designed and formatted appropriately.These responses were not included in the final analysis.A prespecified minimum completion time of 5 min was established based on observing colleagues, with no prior awareness of pharmacogenetics or DCEs, complete the survey.Any completion time under 5 min was considered too rapid to have meaningfully digested the introductory material and completed the survey.Any responses with a completion time under 5 min were excluded from the analysis.
This study was approved by the University of Manchester Research Ethics Committee following proportionate review on 21 June 2023 (REF: 2023-16921-29848).
The final selected model (correlated random parameter logit) assumed that turnaround time is linear and continuous.Effects coding turnaround time within the initial CLM with visual inspection of the data (Appendix S7 and FigureS2) suggested that the effect of this attribute was nonlinear.This observation was further confirmed by the superiority of the quadratic model (AIC = 24222.9,BIC = 24319.43)over the CLM (AIC = 24278.67,BIC = 24366.42).
Within the total sample of 1993 respondents, 29.2% (n = 582) of the sample were predicted to belong in Class 1.The size of the estimated constant (5.75) suggests that individuals in the sample predicted to be a member of this class placed a relatively high intrinsic value on having pharmacogenetic testing and were particularly sensitive to how their data was reused.Around half (50.9%; n = 1014) of the sample were predicted to belong to Class 2. Individuals in this class could be characterized by T A B L E 3 Correlated random parameter logit model.

F I G U R E 3
Predicted willingness to wait.This figure demonstrates the estimated additional number of days participants would be willing to wait for the inclusion of a given level in a pharmacogenetic service, with all other attributes modelled assuming the mean coefficient for that attribute.TAT, turnaround time.Error bars show 95% confidence intervals.The willingness to wait for the overall test (62.98 days [95% confidence interval 54.52 to 71.45]) is not shown on this figure.

F I G U R E 4
Latent class analysispredicted uptake.Labels 1-4 represent the 4 individual classes and T represents the total sample.(A) Description of the 4 classes estimated by the latent class analysis.(B) The predicted uptake for each class across different theoretical service models.Changes in uptake based on the adaption of the base-case service (blood test, no reuse of data, no patient access and a turnaround time of 20 days [Table The correlated random parameter logit model was estimated using categorical variables that were effects coded.Each estimated coefficient represents the impact of that level on predicted uptake in the context of the mean effect of each attribute.Because of this the coefficient of levels within a single attribute can be directly compared (i.e.blood vs. cheek swab), but comparison of levels from different attributes (i.e.blood vs. online portal) cannot, though inferences around directionality and relative importance can still be made.P values assess whether including each level results in statistically significant (P < .05)differencefrom the average service, based on the mean effect of all other attributes.Abbreviation: GP, general practitioner.aThe constant represents respondent's preferences for some form of pharmacogenetic service vs. receiving no pharmacogenetic test (i.e.opt out).T A B L E 4 Comparator services.