Patient preferences for stratified medicine in psoriasis: a discrete choice experiment

Background: New technologies have enabled the potential for stratified medicine in psoriasis. It is important to understand patients’ preferences to enable the informed introduction of stratified medicine which is likely to involve a number of individual tests that could be collated into a prescribing-algorithm for biologic selection to be used in clinical practice. Objective : To quantify patient preferences for an algorithm-based approach to prescribing biologics (‘biologic-calculator’) in psoriasis. Methods: An online survey comprising a discrete choice experiment (DCE) was conducted to elicit the preferences of two purposive samples of adults living with psoriasis in the UK, identified from a psoriasis patient organisation (Psoriasis Association) and an online-panel provider (Dynata). Respondents chose between two biologic-calculators and conventional prescribing described using five attributes: treatment delay; positive and negative predictive values; risk of infection; cost-saving to the NHS. Each participant selected their preferred alternative from six hypothetical choice-sets. Additional data including socio-demographic characteristics were collected. Choice data were analysed using conditional logit and fully correlated random-parameters logit models. Results: Data from 212 respondents (Psoriasis Association = 67; Dynata = 145) were analysed. The signs of all estimated coefficients were consistent with a priori expectations. Respondents had a strong preference for high predictive accuracy and avoiding serious infection but there was evidence of systematic differences in preferences between the samples. Conclusion: This study indicates that individuals with psoriasis would value a biologic-calculator and suggested that such a biologic-calculator should have sufficient accuracy to predict future response and risk of serious infection from the biologic. prediction algorithm), informed by known patient and/or genetic characteristics, into the prescribing pathway of biologics for people with psoriasis. a model of service delivery to enable clinicians to collect information to feed into the biologic-calculator and the patient of the subsequent treatment choice. Further research, using methods from implementation science, 44,45 should be undertaken to understand how the biologic-calculator could be used in clinical practice. This study aimed to quantify the preferences of patients for algorithm-based prescribing (biologic-calculator) compared with conventional prescribing of biologics for people with psoriasis. The results suggested that patients assigned the greatest value to the ability of the biologic-calculator to predict response (PPV) and non-response (NPV), followed by the risk of serious infection from the biologic. These findings have important implications for the implementation of stratified medicine in psoriasis and suggest that tools should be designed with the goal of reaching a sufficient level of predictive accuracy given the cost of implementing these into clinical practice.


Introduction
Targeted biological therapies ('biologics') are a highly effective addition to systemic treatments available for moderate-to-severe psoriasis. 1 The use of biologics, however, may be linked to adverse events (AEs) such as injection site reactions and infections (tuberculosis, lower respiratory tract, skin and soft tissue). [2][3][4] Not all patients will respond to the selected biologic, and secondary failure complicates treatment in an important subset. Given that biologics are expensive and delays in achieving effective treatment are undesirable, there is a sizeable interest in the development of tools to help inform clinicians about targeted treatment selection (stratified medicine).
Ongoing programmes of work seek to develop 'stratified medicine' approaches to the prescribing of biologics with the objective of enabling cost-and time-savings through improved response rates and decreased probability of AEs. 5,6 There have been significant advances in recent years, suggesting that targeted biologic selection may be feasible in psoriasis through therapeutic drug monitoring (TDM) and potentially by genomic testing. 7,8 The information from the results of these individual assessments and patient characteristics could be collated into a prescribing-algorithm (hereafter termed 'biologic-calculator') to aid clinicians' and patients' decision-making when choosing an appropriate biologic. Using such a biologic-calculator would, in theory, result in a more efficient use of healthcare resources and enhanced quality-of-life for people with psoriasis.
Prescribing algorithms, in general, and a biologic-calculator specifically, may be characterised by their ability to accurately predict who will [positive predictive value (PPV)], or who will not [negative predictive value (NPV)] safely respond. It is possible to improve the predictive value of a prescribing-algorithm by including specific variables (such as body mass index, smoking status, gender, location of psoriasis 7 as well as relevant biomarkers (e.g. HLA-C*06:02 genotype status)). 5 The introduction of such variables may delay treatment initiation and increase financial burden due to additional tests, such as those to determine genotype status. Researchers developing a biologic-calculator must weigh the incremental benefit gained from additional information against the incremental cost of collecting it when determining the required predictive values of a prescribing algorithm.
Discrete choice experiments (DCEs) are a potentially useful method to use to understand the benefits, harms and risks associated with new interventions such as a prescribing algorithm. 9 Published studies have used DCEs to quantify patient preferences for biologics in psoriasis but to

Accepted Article
This article is protected by copyright. All rights reserved our knowledge, preferences for an algorithm-based approach to the prescribing of these biologics have not been quantified. 10,11 Including predictive (positive and negative) values as an attribute in a DCE can provide information on the required level of predictive (NPV and/or PPV) accuracy for a biologic-calculator to be deemed sufficiently acceptable to inform prescribing. Such evidence could help those involved in the development of stratified medicine approaches to guide the informed introduction into clinical practice. This study aimed to quantify the preferences of people with psoriasis for a 'biologic-calculator' to aid selection of a first-line biologic.

Materials and methods
A DCE to elicit the preferences of a sample of people with psoriasis for a biologic-calculator compared with the conventional prescribing approach to select a biologic was embedded in an online survey. Survey respondents were asked to choose between two algorithm-based approaches (biologic-calculators A and B) and an opt-out alternative of 'conventional prescribing.' The optout was phrased to represent current prescribing without an algorithm. The algorithm-based approach was framed as representing predictive information in addition to current clinicianinformed prescribing. Ethical approval was obtained from The University of Manchester's Research Ethics Committee (reference: 2016-0172-470).

Survey Design
The DCE was designed and analysed in line with published recommendations. 12,13 The survey was programmed for online administration using SSI Web 8.3.8 Sawtooth software. 14 This survey was developed parallel to, and shared many design features with, a version for people with rheumatoid arthritis (RA). 15 The final survey version for people with psoriasis (Supplementary Material S1; see Supporting Information) comprised three sections: training materials to help the respondents understand the rationale behind the survey; the choice questions; and questions asking the respondents about themselves.

Designing the DCE
Five attributes and relevant levels (see Table 1) were selected to address the choice question: 'If these were the only approaches to prescribing biologics, which, if any, would you choose?' An iterative process, conducted alongside developing a similar survey for people with RA, identified the relevant attributes. 15 The results from interviews conducted as part of a qualitative study in

Accepted Article
This article is protected by copyright. All rights reserved RA 16 and five focus groups (attended by a total of 51 individuals with RA) were supplemented with a psoriasis support group meeting (n = 7 individuals), literature review of psoriasis and DCEs, and two clinical expert interviews to inform the selection of attributes and to ensure that participants understood the survey. The psoriasis support group meeting involved collating views of the online survey by presenting and discussing the training materials and the framing of the attributes and levels. The findings from the psoriasis group meeting were consistent with those from the RA group meetings. <insert table 1 here> Four levels were assigned to each of these five attributes (Table 1)

Experimental design
It was not possible to present all potential scenarios for a DCE using five attributes, each with four levels ( × ( -1) / 2 = 523,776) and a main-effects fractional factorial design was used. This to suggest an 'obvious' best option, to check that respondents were answering in line with economic theory. Each respondent was, therefore, asked to complete six choice-sets but data from five of them were used in the analysis.

Piloting
The DCE survey went through an extensive piloting process (pilot survey with 82 patients; consultation with two academic dermatologists) that was run in parallel with a similar survey designed for people with RA. 15 Changes were made to the levels and their associated images for 'cost saving to the NHS' based on the results from the quantitative pilot.

Accepted Article
This article is protected by copyright. All rights reserved Training materials 15 were used at the start of the survey to provide respondents with sufficient information required to make choices in the DCE. Bespoke training materials (see https://mindbytes.be/our-work/patient-preference-survey-psoriasis/) were created using a narrative storyline in collaboration with MindBytes© 23 because this study required respondents to become familiar with complex attributes for a biologic-calculator described in terms of predictive values (NPV and PPV), infection risk as well as potential cost saving to the NHS. Respondents were asked to indicate if anything was unclear after being shown the narrative storyline by answering a specific question about whether they understood the information provided.

Background questions
To be able to describe the sample, respondents were asked to complete key socio-demographic questions including age, gender, employment status, psoriasis history (time since diagnosis, experience of biologics), a self-reported generic measure of health status (EQ-5D-5L) 24 and a disease-specific measure (Dermatology Life Quality Index (DLQI)). 25 Their responses to the EQ-5D questionnaire were valued using a published UK-specific set of preference weights where the resulting score is anchored on zero (representing being dead) and 1 (representing full health) with the possibility of scores below zero (equivalent to worse than being dead) for serious health conditions. 26

Study population and sample
Individuals with psoriasis, aged 18 years or older, were recruited from two sampling frames: a UK patient organisation for people with psoriasis (the Psoriasis Association) 27 and an online-panel provider (Dynata previously known as ResearchNow). 28 Respondents were sent a link to the online survey (no reminders were used). The first question was a screening question used to exclude those who did not have a diagnosis of psoriasis. No restrictions were placed on the date of diagnosis, disease severity or treatment experiences for current patients to be eligible.

Data analysis
A pre-specified analysis plan was created at the design stage of the DCE which stated that respondents who did not complete the survey, failed the dominance check question or always chose either biologic calculator A or B in all choice sets would be excluded. The dominance check question is a 'test' question that is used to verify whether the respondents are engaging with the

Accepted Article
This article is protected by copyright. All rights reserved questions and/or understand the questions. 29 The 'correct' answer to the dominance check question should be obvious to the respondent. The decision to exclude those who failed the dominance check question was taken because this question had quantitative attributes with levels that showed a logical direction of impact. Therefore, if a respondent failed the dominance check question with an obvious direction of preferences then they were clearly not engaging with the survey.
Descriptive statistics were produced for respondents that were included in the final sample.
In the base-case analysis all attributes were specified as linear, continuous variables and the choice data were analysed using conditional logit models 30

Balancing benefits and harms
The observed balance between the specified benefits (improved predictive value) and harms (delay to treatment and risk of serious infection) was quantified by generating estimates of marginal rates of substitution (MRS) and their associated 95% confidence intervals (CIs) using the delta method. 35 The MRS corresponds to the amount of an attribute respondents were willing to accept in exchange for higher levels of another attribute (see Supplementary Material S4 for additional information).

Results
In total, a purposive sample (comprising a mix of gender and age groups) of 250 people with psoriasis completed the survey. The final sample size of 212 respondents was available for analysis after excluding those who failed the dominance check question (n = 33; three of whom originated from the patient organisation sample) and those who always chose either biologic

Accepted Article
This article is protected by copyright. All rights reserved calculator A or biologic calculator B in every choice set (n = 7). Out of those who failed the dominance check question, only one respondent did not have any formal qualifications which implied that failure of dominance check was not related to lower educational attainment in this sample. The study results were based on a final sample size of 145 respondents from the onlinepanel provider and 67 respondents from the patient organisation.
Descriptive statistics for sample characteristics are reported in Table 2 for all respondents and the two subsamples. On average, people from the patient organisation were more likely to be male, Sample-reported experience of psoriasis and biologics (see Table 3) indicated that those in the online-panel provider group were more likely to have received their diagnosis in the past 5 years and reported more recent flare-ups compared with those from the patient organisation. The vast majority of respondents in either group had never been prescribed biologics.

Patients' preferences
The results from the conditional logit models for each sample and the 'Swait and Louviere' plot 31 confirmed the presence of potential scale and preference heterogeneity (Supplementary Material

Accepted Article
This article is protected by copyright. All rights reserved S3; see Supporting Information). Therefore, a fully correlated RPL model was used to estimate in this sample preferred the biologic-calculator to conventional prescribing when attribute levels were set to be the same for all alternatives. The negative ASC term for the patient organisation sample failed to reach statistical significance meaning that these respondents did not have a strong preference for either of the alternatives when attribute levels were set to be the same.

Balancing benefits and harms
The MRS were calculated using 'delay to treatment' (see Table 5) as the denominator because this attribute appeared to be the closest to a linear functional form (see Supplementary Material S4 for further details). Respondents collated from the patient organisation were willing to delay the start of treatment by 3.25 days (statistically significant) and those from the online-panel provider by 3.89 days (statistically significant) for a £100 cost saving. The most valued attribute in both samples was the ability of the biologic-calculator to determine who will not respond to treatment

Accepted Article
This article is protected by copyright. All rights reserved (NPV), as both groups were willing to wait 23-29 days for a 10% increase. Respondents collated from the online-panel provider were willing to delay treatment by 22.95 days compared to 28.84 days in the patient organisation sample for an increase of 10% in NPV but this was not statistically significant in either group. Another important attribute in both samples was the ability of the biologic-calculator to determine who will respond to treatment (PPV) as respondents from the patient organisation were willing to delay treatment by 19.22 days and those from the online-panel provider by 14.09 days (statistically significant in both groups). The patient organisation group of respondents displayed stronger preferences for predictive accuracy of the algorithm. The MRS values for the ability to predict response (PPV) and non-response (NPV) were not statistically different from one another in either of the samples.

Discussion
This study was designed to quantify the preferences of individuals with psoriasis for an algorithmbased approach to prescribing biologics. All five attributes (NPV, PPV, risk of serious infection, delay to treatment and cost saving to the NHS) were consistent with a priori expectations in terms of the direction and magnitude of the estimated coefficients.
The ability of the algorithm to determine response (PPV) and non-response (NPV) were the two most important attributes driving preferences in both samples relative to the other attributes in the DCE. However, NPV was not statistically significant in the patient organisation sample. The next most influential attribute was the risk of infection. These data on the trade-offs that patients were willing to make are informative to researchers involved in the development of prescribingalgorithms to introduce stratified medicine into practice. Importantly, this study suggests that NPV was as important as PPV to patients although it was not statistically significant in the patient organisation sample. This suggests that patients showed a clear preference to avoid being prescribed a biologic treatment that will not work for them. This finding is important since most research aims to identify markers of response (rather than non-response). 37,38 The observation that probability of non-response was a key factor driving preferences has been shown in other DCEs. For example, in a DCE comparing algorithm-based prescribing to conventional prescribing in RA, the authors reported that NPV was a predictor of preferences. 15 Another DCE that elicited preferences of neurologists for pharmacogenetic testing in epilepsy also

Accepted Article
This article is protected by copyright. All rights reserved suggested NPV to be a strong predictor of preferences. 39 This suggests that NPV is important not only for people with psoriasis, but also for physicians and for people with RA and other autoimmune conditions.
The presence of scale and preference heterogeneity indicated that there were variations in the preferences of the samples. In such cases, it would be incorrect to form conclusions from merging the data from both samples and using a pooled conditional logit model. 40  The potential contribution of eliciting patient preferences is to use these results to inform the subsequent design of a biologic-calculator that takes account of the need to achieve adequate levels of, for example, PPV and NPV. Currently, the types and number of tests to include in a prescribing-algorithm are unknown. Future development would involve developing a prediction algorithm and embedding the biologic-calculator (using the results of tests as an input into a

Accepted Article
This article is protected by copyright. All rights reserved prediction algorithm), informed by known patient and/or genetic characteristics, into the prescribing pathway of biologics for people with psoriasis. Therefore a model of service delivery will be required to enable clinicians to collect information to feed into the biologic-calculator and inform the patient of the subsequent treatment choice. Further research, using methods from implementation science, 44,45 should be undertaken to understand how the biologic-calculator could be used in clinical practice.
This study aimed to quantify the preferences of patients for algorithm-based prescribing (biologic- 148.

Accepted Article
This article is protected by copyright. All rights reserved

Accepted Article
This article is protected by copyright. All rights reserved

Accepted Article
This article is protected by copyright. All rights reserved

Accepted Article
This article is protected by copyright. All rights reserved