2.1 Recruitment
Treatment preferences of patients with RA were assessed using a discrete choice experiment (DCE). An invitation to participate in the study was advertised to members of the Swedish Rheumatism Association via email, newspaper, newsletter, social media, mobile application, and the association’s website. The invitation to participate was also distributed to patients attending ten rheumatology clinics in Sweden and via an online research panel of patients with RA. A printed copy of the survey was distributed by the Rheumatology clinic at Uppsala University hospital. All participants received information about the study and provided their informed consent before completing the survey. The following inclusion criteria were used: established RA diagnosis, 18-80 years of age, and the ability to understand and answer the questions. Data were collected from November 2018 to August 2019. The survey was approved by the regional ethics review board in Uppsala, Sweden (Reg no. 2017/521, 2018/156). Data generation, storage and sharing were governed by the General Data Protection Regulation (GDPR) Act, Uppsala University data protection and security policies and ethical consent provided.
2.2 Methodology of discrete choice experiment
DCEs, a cross-sectional survey method used to assess preferences, allows for quantitative assessment of patient preferences for health care policies, services, and interventions [15]. DCEs, which uses random utility theory (RUT), aims to quantify the relative importance of one treatment characteristic over another treatment characteristic. RUT assumes that the value (utility) of a product can be determined by the value (utility) of the characteristics of that product (i.e., attributes) and their levels. Respondents in a DCE are presented with hypothetical scenarios (choice questions) with varying attributes and levels. Respondents are asked to choose their preferred option for each question [16]. The utility can be estimated by modelling the choices that respondents make between alternatives of treatments that are described by different choice questions [17]. DCEs can also be used to measure and explain heterogeneity within the preferences of patients [18].
2.3 Attributes and levels
Using a step-wise approach, we identified attributes and levels for inclusion in the DCE. First, the analysis of a literature review of previous studies of patient preferences for DMARDs resulted in 12 potential treatment attributes [2, 3, 5, 14, 19-24]. Second, the attributes and levels identified in the literature review were discussed with a rheumatologist to make sure that they reflected current clinical practice. Third, three focus groups using the nominal group technique (NGT) were conducted with patients with RA (n=7); these patients were asked to identify new attributes and rank all potential attributes from most to least important [25]. The focus groups were audio recorded, lasted for about 90 minutes, and conducted using an interview guide. Fourth, results from the focus groups were discussed during several validation meetings with one rheumatologist, the research team, and two patient research partners. These meetings revealed seven attributes: route of administration, frequency of use, probability of mild short-term side effects, probability of side effects changing appearance, probability of psychological side effects, probability of severe side effects, and effectiveness of treatment. Each attribute was revealed to have three levels based on current clinical knowledge of existing biologics and JAK inhibitors. Detailed information regarding the selection and description of the attributes and levels is available in the Supplementary material. All attributes and levels included in the DCE are displayed in Table 1.
Table 1. Attributes and levels
|
Attribute
|
Level 1
|
Level 2
|
Level 3
|
Route of administration
|
Tablet
|
Injection
|
Drip
|
Frequency of use
|
Daily
|
Weekly
|
Monthly
|
Probability of mild short-term side effects (nausea, vomiting or headache)
|
Common 1 in 10
|
Uncommon 1 in 100
|
Rare 1 in 1000
|
Probability of side effects changing appearance (hair loss, weight changes or skin rash)
|
Common 1 in 10
|
Uncommon 1 in 100
|
Rare 1 in 1000
|
Probability of psychological side effects (anxiety, mood changes, depression or sleep disturbance)
|
Common 1 in 10
|
Uncommon 1 in 100
|
Rare 1 in 1000
|
Probability of severe side effects that requires hospitalisation such as severe infections or allergic reactions
|
Common 1 in 10
|
Uncommon 1 in 100
|
Rare 1 in 1000
|
Effectiveness (the ability to decrease inflammation and swelling of the joints, also pain and other symptoms)
|
30 % improvement
So out of 100 persons taking the treatment, 30 will get enough improvement, the rest will get a small or no improvement
|
50 % improvement
So out of 100 persons taking the treatment, 50 will get enough improvement, the rest will get a small or no improvement
|
70 % improvement
So out of 100 persons taking the treatment, 70 will get enough improvement, the rest will get a small or no improvement
|
2.4 Experimental design and survey
The survey started with information about RA and available treatment options before entering the DCE. The last section of the survey consisted of demographic and disease-related questions, health literacy [26], and numeracy [27]. The DCE had an attribute-based experimental design. Respondents were asked to choose their preferred treatment from two alternatives (see figure 1. Example of a choice question). The choice questions also included a hover function with further explanations of the attributes and the levels (see Supplementary file for full text explanations of attributes and levels).
The survey was pilot tested with a subgroup (n=22) of patients with RA and patient research partners. Six of the pilot tests were ‘think aloud’ interviews. The respondents were encouraged to articulate their thoughts while completing the survey. The language and the layout of the survey were slightly changed after the pilot test. Using the pilot test data, we fitted a Multinomial logit (MNL) model and used the beta estimates as priors for the final experimental DCE design generated by NGene 1.0 (ChoiceMetrics, 2011), which is a d-efficient (Bayesian) design [28]. A constraint was posed on the design: route of administration and frequency of use (e.g., if the route of administration was a tablet, the frequency of use could not be ‘monthly’). A total of 60 unique choice questions were divided into four blocks. Each respondent had to answer 15 unique choice questions. All attributes were displayed in each of the choice questions; three attributes were identical across the two offered alternatives to reduce the cognitive burden to respondents. We applied the decision-making scenario ‘think of yourself in a situation where your treatment is not working, your joints are swollen, you have pain or unbearable side effects and need to change to a second-line treatment’.
2.5 Statistical analysis
SPSS® Statistics 20 and Nlogit® were used for analyses. Demographic data were analysed using descriptive statistics. Results were considered statistically significant if P<0.05. The patients’ preferences were determined by attribute level estimates using a MNL model [29]. Latent class analysis (LCA) models were used for further analysis of the DCE data. Such models account for the multilevel structure of the data (i.e., every respondent answered multiple choice questions) and account for the investigation of preference heterogeneity. LCA models assume that there are two or more latent classes of data with different preferences. The classes are characterised by unobserved latent variables that can be related to a set of choice patterns. Once choice patterns have been stratified into classes, it is possible for the model to determine the probability that a respondent with certain characteristics will be assigned to each class [30]. The attributes were dummy coded (i.e. the mean effect for each attribute was normalized at zero) except for effectiveness that was effects-coded. The ‘likelihood ratio test’, the Akaike information criterion (AIC) were used to determine the most appropriate model A three-class model based on the utility is displayed below:
Vrta|c = β0|c + β1|c Route of administration Tablet rta|c + β2|c Route of administration Injection rta|c + β3|c Frequency of use Daily rta|c + β4|c Frequency of use Weekly rta|c + β5|c Mild short-term side effects 1 in 10 rta|c + β6|c Mild short-term side effects 1 in 100 rta|c + β7|c Appearance side effects 1 in 10 rta|c + β8|c Appearance side effects 1 in 100 rta|c + β9|c Psychological side effects 1 in 10 rta|c + β10|c Psychological side effects 1 in 100 rta|c + β11|c Severe side effects 1 in 10 rta|c + β12|c Severe side effects 1 in 100 rta|c + β13|c Effectiveness rta|c
The utility component (V) describes the utility that respondent ‘r’ belonging to class ‘c’ reported for alternative ‘a’ in choice question ‘t’. β0 represents the constant of the model. The attribute level estimates of each attribute level are represented by β1 – β13. A class assignment model was fitted after the specified utility function. Several demographic and disease-related variables were tested for their potential impact on class membership in the LCA: age, gender, numeracy, health literacy, education level, disease duration, occupational status, and experience with DMARD treatment and side effects. The final class assignment utility function was:
Vrc = β0|c + β1|c disease durationr + β2|c and mild side effectsr
A significant attribute estimate within a certain class indicates that this attribute contributes to the decision-making process of respondents who belong to that class. The sign of the beta indicates whether the attribute level has a positive or negative effect on the utility.
To calculate the relative importance of the attributes, the difference between the highest and lowest estimates of the attribute level was calculated for each attribute. The largest difference value was given a 1, representing the attribute that was deemed most important by respondents. The other difference values were divided by the largest difference value, resulting in a relative distance between all other attributes and the most important attribute.
A minimum acceptable benefit (MAB) for changes in attribute levels was calculated. The MAB is interpreted as the minimum change in effectiveness that respondents would require (on average) to accept changes to a less desirable level in another attribute (probability of getting a certain side effect by 10%, 1%, and 0.1%). MAB was estimated as the difference between the preference weights (parameters) for two levels ‘l’ of an attribute divided by the preference weight, βk =effectiveness, which is the unit change in the level of benefit: