1B
ORAL ABSTRACTS: STATED PREFERENCES AND DISCRETE-CHOICE EXPERIMENTS
Method: This study estimated the value of WGS information using double-bounded contingent-valuation methods. An online survey (n=406 adults from US general population) was used to evaluate willingness to pay (WTP) for a basic WGS report consistent with current guidelines versus a report containing non-actionable findings. Respondents first indicated whether they would purchase an augmented report for a randomly assigned dollar amount. A follow-up question increased or decreased the starting value conditional on the initial response. The resulting no/no, no/yes, yes/no, and yes/yes responses define double-bounded segments for interval-regression analysis.
Result: Interest in non-actionable, fatal diseases: 34% (n=139) would want to know this information, 38% (n=154) would not want to know, and 28% (n=117) were not sure. 55% of respondents (n=224) were not willing to pay anything for such non-actionable genetic information. The Table summarizes mean incremental WTP by respondent characteristics.
Mean WTP ($US) for information on non-actionable mutations (95% CI) |
|
Would want to know if they had gene variants that lead to a fatal disease |
$226 (157, 295)*** |
Underwent prior genetic testing |
$198 (74, 321) *** |
Would want to know if they had gene variants that lead to memory loss |
$155 (90, 220) *** |
Prefers to decide themselves which results to include in report |
$94 (20, 168) ** |
No health insurance |
$86 (-38, 211) |
Education (bachelor’s degree or higher) |
$20 (-43, 84) |
Race/ethnicity (white, non-Hispanic) |
-$65 (-125, -6) ** |
**p<0.05; ***p<0.001 |
|
Conclusion: A majority of respondents were unwilling to pay for non-actionable genetic information. Among the largest mean WTP values were those for mutations linked to a fatal condition or memory loss. Heterogeneity in perceived values of non-actionable genetic information suggests that considerations should be given to both actionable and non-actionable genetic information depending on patient preferences. Our findings suggest that quantifying patient preferences could provide useful information to inform WGS recommendations and policies.
Method: Following piloting (n=50), an on-line hybrid (linked conjoint analysis (CA) and discrete choice experiment (DCE)) stated choice experiment, was completed by a public sample (recruited via an internet panel provider). Two survey versions (A: NBSP for 9 conditions; B: NBSP for 20 conditions) comprised four tasks each: a validated measure of attitudes towards involvement in decision-making; six CA questions (11 information attributes); ten DCE questions (4 attributes: 3 process and the ability to make an informed decision); demographic questions. Literature reviews and 13 semi-structured interviews informed attribute/level selection. The design criteria were orthogonal foldover arrays (CA) and Bayesian D-efficiency using Ngene (DCE). The CA and DCE data were analysed separately and then linked using ordered logit and logit models. Marginal willingness-to-pay (WTP) values with 95% confidence intervals (CI) were calculated.
Result: The sample comprised 700 respondents (58% female; mode age band 25 to 34 years; 48% with university degree: 48% parents). A high proportion (37%) of respondents indicated wanting to make decisions about screening after the midwife provided information and a recommendation. Responses differed between versions A and B. In version A, respondents positively valued: all information attributes except the possibility of receiving false-positive results or how parents can prepare their baby; information early in pregnancy (WTP £14.32; CI: £8.25 to £21.43); receiving information in an individual discussion (WTP £10.57; CI: £5.84 to £16.44); the ability to make a decision about screening (WTP £13.51; CI: £10.72 to £16.52).
In version B, respondents positively valued: all information attributes (CA); information early in pregnancy (WTP £15.20; CI: £7.72 to £23.68); the ability to make a decision about screening (WTP £16.05; CI: £12.28 to £20.82). Respondents completing version B had no significant preferences for how information is given.
Conclusion: This hybrid CA-DCE was able to elicit preferences for information provision in the context of a NBSP. Respondents stated a need for different types of information to allow them to make an informed decision and had clear preferences about which information was more important and what format of information provision was preferred, which was in some instances affected by the number of conditions included in the NBSP.
To measure individuals’ preferences for potential baskets of health outcomes and medical expenditures, we designed a discrete choice experiment (DCE) questionnaire, pre-tested it for clarity of presentation, ease and feasibility of administration, and assessed its validity.
Method:
We recruited 31 participants by snowball sampling. The instrument included: (1) Information section in which we carefully described the context of the decision-making; (2) 10 choice tasks each including 2 generic alternatives (Situation A vs B) described by 2 attributes (total costs and health outcomes (in Healthy Years Equivalent (HYEs)); (3) Feedback type questions. In addition we added 2 tasks to test both dominance and stability properties. Participants’ preferences were estimated using different specifications of conditional logit model, and answers to the feedback questions were descriptively analysed. Predictive performance of the best choice model was investigated using a bootstrapping procedure with 1,000 replicates.
Result:
Overall difficulty of the questionnaire: 19.4% found it “very difficult”. The topic was considered interesting (“moderately to extremely”) by 80.7%. 73.3% considered the amount of information conveyed acceptable (“moderately to extremely”), only 3.2% considered it difficult (“very to extremely”) to understand. 87.1% and 74.2% declared taking into account HYEs and costs respectively, in all the vignettes when they made their choices. 3.2% of respondents failed stability test, 12.9% failed dominance and 3.2% failed both stability and dominance. The best specification of the choice model included an interaction effect between preferences for health outcomes and medical expenditures. The estimated preferences were in line with a prioriassumption regarding both the sign and magnitude of the estimates. Respondents positively valued increase in health outcomes (β=0.306) and negatively valued increase in level of medical expenditures (β=-0.113). The interaction effect was significant and negative. The mean predictive performance of this model was high: 80.84% [95% CI: 80.72-80.96].
Conclusion:
The results of the pre-test for clarity of presentation, ease and feasibility of administration were positive; they also indicate that responses were valid. This questionnaire, once administered to a representative sample of the population, can generate a population based net loss or net benefit functions to be used, for example, in a framework recently published in Health Economics that describes how to assess and manage the risk of potential undesirable outcome in the context of resource allocation.
Method: An experiment with 4 distinctly different framings was conducted among a Dutch nationally representative sample of 1200 respondents. These framings comprise both DCE approaches (i.e. DCEduration and DCEdeath) with and without the addition of lead time (LT) to the EQ5D5Lhealth profiles. A Bayesian efficient design consisting of 8 sets of 30 (matched pairwise) choice tasks was used. The design was jointly optimized for all framings, thereby keeping all aspects of the DCE design except for the framing constant. Respondents were randomly assigned to one of the 4 study arms. Mixed logit models were used to analyze the DCE data, and the resulting estimates of the utility decrements associated with the severity levels within each dimension were compared between the 4 arms to establish the impact of the framing effects.
Results: The estimation results revealed substantial framing effects. While the DCE death approach classified just 8% of the health states as worse than death, much higher percentages were found in the other arms: 28% (duration), 57% (LT-death) and 81% (LT-duration). Relative distances between health states on the latent scale were not affected by adding LT, but anchoring on death altered the values. We observed less dispersion for mild to moderate states, and a more stretched distribution for severe states.
Conclusion: Estimation results were substantially altered by the framing of death as explicit or implicit, and immediate or postponed. These framing effects may help to explain the commonly observed discrepancies between values derived using Time Trade Off and the popular DCE duration approach. While one may argue against the use of a death alternative in DCE tasks for health state valuation on basis of theoretical and statistical considerations, it would seem to be an essential component for those who aim to reconcile DCE and TTO results.
Method: An online DCE was designed and piloted (n=124) to elicit the preferences of female members of the public (recruited via an internet panel provider) for a NBSP described by two risk attributes (probability of detecting a cancer and risk of unnecessary follow-up per 100 women screened) and a cost attribute (out-of-pocket expense). Two survey versions presented the risk attributes as: (A) percentage only or (B) percentage and an icon array. The unlabelled DCE was blocked into two surveys, each containing 11 sets of choices between two screening programmes and an opt-out. The design, generated using Ngene, included an internal validity test through the inclusion of a dominant choice set. The DCE data were analysed using heteroskedastic conditional logit (HCLM) and scale-adjusted latent class (SALC) models.
Result: 1007 women (version A =501; B=506) completed the DCE. The results of the HCLM suggested that all attribute coefficients, but no two-way interactions, were significant and had the expected signs. Interactions of attributes with risk framing version were not significant and the risk framing version had no significant impact on the scale parameter. SALC analysis revealed heterogeneity in preferences, with five latent classes and three scale classes providing the best fit. The class probabilities indicate 84% of respondents were members of three large classes where all scale-adjusted attribute coefficients were significant: 31% in class 1 (probability of detecting a cancer most important), 27% in class 2 (cost of screening most important), 25% in class 3 (risk of unnecessary follow-up most important). The remaining 17% were split between classes 4 (9%) and 5 (8%). Class 1 members tended to be aware and concerned about their risk of breast cancer. Class 2 members were less likely to be older (over 50). Class 3 individuals tended to be younger (25-34) and have experience of cancer in their family. Risk framing version was not a predictor of class membership.
Conclusion: This study found the framing of risk attributes did not impact respondents’ choices in a DCE. However, other sources of heterogeneity were found in women’s preferences for the balance between the risks and benefits of a NBSP.
Method: Eligible caregivers had a child aged 4-14 years being treated for ADHD, and were recruited from community clinics and support groups. Treatment preferences were elicited using Case 2 Best Worst Scaling (BWS), implemented as part of a larger survey that captured demographic and clinical information, including time since ADHD diagnosis. The 18 BWS profiles were identified from a main-effects orthogonal array spanning seven attributes (i.e., medication, therapy, school, caregiver training, provider, communication, and costs). The dependent variable was caregivers’ choice of a best and a worst attribute for each profile. Preference heterogeneity was examined by: a) stratification of time since ADHD diagnosis (<4 years/4+ years), since illness experience can influence preferences; b) latent class analysis (LCA). In the stratified analysis, the best-worst score method was estimated separately for each strata. Latent Gold®Choice was used to conduct the LCA. The latent segment solution was determined using model fit statistics and theoretical interpretability. Bivariate statistics tested for statistically significant differences in demographic and treatment by diagnosis duration strata and across latent segments.
Result: Caregivers (n=184; 84% mothers; 43% college-educated) reported for children who were on average 9 years old and taking stimulants (75%). Regardless of time since the child’s ADHD diagnosis, medication use seven days a week, therapy in a clinic, and an individualized education program were most preferred (p<0.001). Irrespective of ADHD duration, out-of-pocket costs and caregiver behavior training were most important (p<0.001), but the conditional importance of medication and school accommodations differed. Demographic and treatment variables were similar between the two groups. The LCA generated a three-segment solution: 1) cost-sensitive (53%); 2) multi-modal treatment (23%); 3) medication-oriented (24%). Medication administration, therapy location, school accommodation, caregiver behavior training, and out-of-pocket costs discriminated each class. ADHD duration did not differ across segments (p>0.05), and few other factors were significantly different.
Conclusion: Latent class analysis provides a nuanced understanding of preferences for medical treatment, and can be applied to other medical conditions. Future research will build upon this work to investigate the correlation with adherence.