Line of therapy and patient preferences regarding lung cancer treatment: a discrete-choice experiment

Abstract

Objective

A growing literature on patient preferences informs decisions in research, regulatory science, and value assessment, but few studies have explored how preferences vary across patients with differing treatment experience. We sought to quantify patient preferences for the benefits and risks of lung cancer treatment and test how preferences differed by line of therapy (LOT).

Methods

Preferences were elicited using a discrete choice experiment (DCE) following rigorous patient and stakeholder engagement. The DCE spanned five attributes (each with three levels): progression-free survival (PFS), short-term side effects, long-term side effects, risk of developing late-onset side effects, and mode of administration (MOA) – each defined across 3 relevant levels. A D-efficient design was used to generate 3 survey blocks of 9 paired-profile choice tasks each and respondents were asked which profile they preferred and then if they preferred to have no treatment (opt-out). A mixed logit model, controlling for opt-out, was used to estimate preferences. Preferences and trade-offs between PFS and other attributes were compared across two groups: those receiving ≤1 LOT and those receiving ≥2 LOT.

Results

Of the 466 participants, 42% received ≤1 LOT and 58% received ≥2 LOT. Stated preferences differed between the groups overall (p<.001) and specifically for 18 months of PFS (p<.001), moderate short-term side effects (p<.001), no long-term side effects (p=.03), and 30% chance of late-onset side effects (p=.02). Those receiving differing amounts of LOT were willing to trade different amounts of PFS to change from moderate to mild short-term side effects (p<.001), moderate to no (p<.001) and mild to no (p<.001) long-term side effects. There were also differing amounts of tradeoff acceptable between the groups for a 10% decrease in risk of late-onset side effects (p=.016), a decrease in MOA from infusion every 3 weeks to pills taken daily at any time (p=.005) and from pills taken daily without food to pills taken daily at any time (p<.001).

Conclusion

We demonstrate differences in preferences based on experience with LOT, suggesting that patient treatment experience may have an impact on their preferences. As patient preference data become an important component of treatment decision making, preference differences should be considered when recommending therapies at different stages in the treatment journey. Understanding patient preferences regarding treatment decisions is essential to informing shared decision-making and ensuring treatment plans are consistent with patients’ goals.

Keywords:

• Lung cancer
• patient preferences
• discrete choice experiment
• line of therapy

Introduction

Lung cancer is the leading cause of cancer deaths in the United States¹ and the most common cancer diagnosis¹. Globally, lung cancer is the most common type of cancer for both diagnoses and deaths². The burden of lung cancer is largely due to its high mortality rate – it has a 19% five-year survival rate, which is among the lowest of all cancers³. Almost 80% of lung cancer cases are diagnosed at a late stage (stage III or IV) due to the asymptomatic nature of the disease in early stages and often the malignant tumor has already spread to other parts of the body⁴. Line of therapy (LOT) is defined as the number of therapies a patient receives to treat their lung cancer (excluding maintenance therapy) in their lifetime. Patients often receive multiple LOT in order to slow disease progression and prolong survival⁵. LOT is an important parameter both from a drug development and a clinical practice point of view. LOT is often an eligibility criterion for lung cancer clinical trials. Some clinical trials, especially phase 1 clinical trials, may be exclusively available to patients who have been on multiple LOTs. More LOTs are associated with more side effects (higher accumulation) the more treatment a patient receives and therefore will impact how clinical decisions can be made.

New treatment options have become available for non-small cell lung cancer (NSCLC) patients in recent years with the development of targeted therapies and immunotherapy. Since 2015, the US Food and Drug Administration (FDA) approved several new drugs for the treatment of NSCLC, the most frequent histological subtype of lung cancer. However, the side effects and toxicities of these treatments can be significant. With the appearance of new treatment options for lung cancer, the complexity of treatment decisions for people living with lung cancer has increased. Preference research finds that patients with lung cancer are willing to endure prolonged treatment in exchange for longer survival^6,7 and that experiences concerning physical, psychological and organizational areas can affect preferences⁸. Lung cancer is at the forefront of precision medicine with biomarker-driven care becoming the mainstay for treatment of patients with advanced-stage disease. Patients with lung cancer with targetable driver mutations such as Epidermal Growth Factor Receptor (EGFR)⁹ or Anaplastic lymphoma kinase (ALK)¹⁰ have multiple treatment options available and are living longer with targeted oral treatments.

While research on patient preferences in lung cancer has increased over time^11–14, patient preference heterogeneity is still not well understood. Prior to the advent of biomarker-driven care, willingness to pay for a new treatment for advanced NSCLC has been shown to be higher among those with a history of chemotherapy compared with those who had not received chemotherapy¹⁵. A more recent study showed a statistically stronger preference for “clearly longer survival” in a latent class analysis of patients with all cancer types who had received surgery, but patients with either colon or lung cancer were included in that analysis⁶. There is a paucity of evidence on how clinical factors such as stage, histology at diagnosis, and presence of targetable driver mutations interact with LOT to affect patient preference.

Diverse stakeholders, including regulatory organizations such as the FDA, understand the importance of incorporating the voice of patients into drug development and regulatory benefit-risk analysis. Patient testimony has been a mainstay in patient advocacy and is the primary emphasis of patient-focused drug development (PFDD). In light of the recent twenty first Century Cures Act, the FDA now supports the collection of patient experience data including quantitative patient preference information¹⁶. Guidelines for lung cancer treatment written by the National Comprehensive Cancer Network (NCCN) recognize patient heterogeneity in recommendations, with some treatments only recommended in certain patient populations¹⁷. Preferences vary across patient groups, and decision makers have stressed the importance of studying this heterogeneity¹⁸.

Cancer treatment decisions are complex and patient preferences for treatments can vary widely. It is critical that treatment decisions for NSCLC take into account the risks patients are willing to accept in order to attain or avoid benefits and side effects. We sought to quantify the preferences of people with NSCLC and explore variation of preferences by LOT. Understanding preference heterogeneity is a key requirement of preference studies as outlined by the FDA guidance and current good research practices. It can allow decision makers to better understand decision populations and to customize decision rule to meet the needs of specific populations. While it was not our primary intention, evidence on variation in treatment preference by LOT can demonstrate to practicing clinicians how treatment experience can affect patient preferences.

Materials and methods

Survey development

The project was guided through rigorous engagement with patients with lung cancer, modeled after previous patient-engagement exercises¹⁹. A 27-member patient action committee (PAC) was formed through the survivors’ board of a patient advocacy organization, LUNGevity Foundation, which guided the development of the survey (LUNGevity website). PAC members represented a range of demographic and clinical backgrounds. The engagement period spanned multiple years and consisted of PAC meetings, individual interviews, and email communication. Engagement activities were first geared towards acquainting the study team and PAC members²⁰ and defining project goals²¹, and then shifted towards developing a pilot survey²², and the final DCE survey.

The survey was developed through an iterative mixed-methods process^23,24 and followed good-research practices for attribute identification and survey development^25,26. Attributes were identified from the literature¹² and interviews with PAC members^20,21. Levels were selected based on efficacy and risks associated with existing lung cancer medications. The survey was emailed to all PAC members for comment, with pre-test interviews conducted with four PAC members to identify usability issues. The survey was then pilot tested at three in-person lung cancer survivorship summits. Appropriate and iterative adaptations were made after the pretest interviews and each pilot test round.

The DCE consisted of five attributes that could be presented at three levels each: progression free survival (PFS; 12, 18, 24 months), severity of short-term side effects (mild, moderate, severe), severity of long-term side effects (none, mild, moderate), risk of late-onset side effects (10, 20, 30%), and mode of administration (pill taken daily at any time, pills taken daily that you need to take 1 h before eating or 2 h after eating, infusion in the hospital or clinic every 3 weeks). PFS was chosen over overall survival (OS) as previous studies demonstrate that patients are often not willing to trade between OS and other attributes²⁷. In addition, pilot testing of the survey also revealed that lung cancer patients valued PFS over PS as an attribute for the DCE.

From the five attributes at three levels, 243 separate profiles or more than 50,000 profile pairs could be generated. To narrow the number of choice options and balance response burden and statistical efficiency²⁸, the DCE employed a blocked D-efficient main-effects experimental design using Ngene software (ChoiceMetrics). Participants were randomly assigned to one of the three blocks of 9 paired comparisons each (27 in total). Priors were not incorporated in the experimental design, as correlation introduced by including priors reduced our ability to estimate independent preference coefficients in previous studies²⁹. A fixed choice task, which allowed for testing of version consistency, was added to each survey block³⁰. Therefore, participants completed 10 paired-comparison choice tasks each.

Participants were shown a series of paired hypothetical treatment profiles and asked to select the profile they preferred. As participants were at varying stages of disease, profiles were presented as fixed health profiles and asked participants to select a profile while considering that they are newly diagnosed with lung cancer. The use of a fixed profile has been successfully used to reduce measurement bias^11,31. An opt-out option was added after each task to evaluate the potential selection of no treatment^32,33. This dual response design was chosen to include the possible choice of no treatment, while also ensuring efficiency³⁴. Full scenarios were shown to maximize information on acceptable tradeoffs in each choice³⁵. An example choice task is presented in Figure 1.

Figure 1. Example of choice task.

We examined rates of nonresponse and task nonattendance of the choice tasks. Acceptability of the DCE was assessed to ensure patient-centeredness, relevance, and response burden in accordance with stated-preference method guidance^18,36. Participants were asked to indicate their level of agreement on a 5-point Likert scale ranging from strongly disagree to strongly agree on four statements that assessed response burden (ease of understanding and ease of answering), response consistency, and relevance of the DCE^37,38. The survey also included self-report questions about demographic information and disease/treatment history, personality traits, patients’ experiences with side effects, and their concerns about lung cancer treatment.

Survey distribution and study sample

The survey was programmed in Qualtrics (Provo, Utah) and distributed online. Multiple recruitment channels were chosen to ensure a widespread distribution of participants. This included recruitment through the patient advocacy and research organizations LUNGevity Foundation, Cancer Support Community, Patient Advocate Foundation, and Edge Research. Participants were required to be at least 18 years of age, live in the US, and have been diagnosed with NSCLC. Underrepresented minorities were oversampled based on race and ethnicity.

Statistical analysis

Responses from patients with NSCLC who completed at least once choice task were included in the analysis. Analyses were conducted separately for participants that had received ≤1 LOT and ≥2 LOT. Demographic and clinical characteristics, concerns regarding lung cancer treatment, and side effect experience were analyzed descriptively. Differences between groups with varying number of LOT were assessed using Fisher’s exact test or a t-test, as appropriate. All tests of significance were two-sided and statistical significance was established at p<.05. Statistical analyses were performed using Stata version 15 (StatCorp, College Station, TX, USA) and R, version 3.6.1 (R Core Team, 2019).

Subjective evaluation

We explored the acceptability of the DCE instrument by evaluating the Likert responses for the DCE choice task evaluation questions and comparing to an agreement threshold of 75%. This threshold was based on previous research reporting agreement with these items^37,39. The questions were summarized as the percentage of participants that chose either “Agree” or “Strongly Agree” to each of the statements. Each of these percentages were compared to a 75% agreement benchmark using a one-proportion, two-sided z-test wherein a p-value of >.05 indicated that item met the benchmark, and a p-value <.05 indicated that the item either was statistically significantly above or below the benchmark³⁷. Differences in the endorsement of debriefing items across the two LOT groups were compared using two-sample t-tests.

Preference model

The DCE results were analyzed using a mixed logit model, which allows for the examination of preference heterogeneity⁴⁰. This model assumes that preferences vary across individuals according to a continuous distribution. The model provides mean coefficients as well as a measure of their distributions in the form of a standard deviation (SD). A statistically significant SD indicates that there is significant preference heterogeneity for that attribute. Aggregate models were used to evaluate overall preferences and to explore heterogeneity identified from the mixed logit model. Stratified analyses were conducted to determine if preferences differed between participants that had received different numbers of LOT (≤1 LOT and ≥2 LOT).

The dependent variable was the participants’ preferred treatment in each choice task and the independent variables were the levels of each of the attributes in each profile. Effects coding was utilized for each attribute to provide parameter estimates for each of the three levels. An alternative specific constant (ASC) was used to control for the opt-out option and provided a “no choice” option. After comparing the model fits (based on Akaike Information Criteria and Chi-square) of different mixed logit models including different (sets of) random parameters, all parameters were set at random with a normal distribution for the final analysis, including the ASC. Overall models and individual parameter estimates between those receiving ≤1 LOT and those receiving ≥2 LOT were assessed via Wald tests⁴¹. We also looked at the additional stratifications of age, race, gender, and education. Raw parameter estimates were shifted so that a chosen level of each attribute was set to 0.

Healthy month equivalence model

The random parameters from the mixed logit model were estimated in the “preference space”, which means that each estimate represents the marginal utility of that attribute level^42–44. Parameters in the “willingness to pay” (WTP) space are easier to interpret and they can be estimated directly or can be obtained from the marginal utility by dividing it by the coefficient of a covariate expressed in monetary value (a price for example), taken as a non-random parameter. The ratio can then be interpreted as a monetary value (or willingness to pay). Since we have a time attribute as opposed to a price attribute, we were able to calculate time equivalents in this same fashion. That is, the utility (U) participant n acquires from choosing treatment j for task t can be expressed by: (1) $$U_{\mathit{njt}}=\mathrm{ }{c_n}^{\mathrm{\prime }}{X}_{\mathit{njt}}+{\lambda }_n{p}_{\mathit{njt}\mathrm{ }}+\mathrm{ }{\epsilon }_{\mathit{njt}}\mathrm{ }$$(1)

Where ɛ is participant-specific random error and ${\mathit{c}}_n$ and ${\lambda }_n$ incorporate both preference estimates and variance-scale for X, a vector of treatment characteristics (short-term side effects, long-term side effects, and mode of administration) and p, the time benefit of PFS, respectively. Trade-offs between changes in attributes and time^45,46 were estimated consistent with healthy-months equivalents (HME) approach^22,47,48 so that if we let ${{\mathit{HME}}_n=w}_n=\mathrm{ }\frac{{\mathit{c}}_n}{{\lambda }_n},$ Equation (1)(1) $$U_{\mathit{njt}}=\mathrm{ }{c_n}^{\mathrm{\prime }}{X}_{\mathit{njt}}+{\lambda }_n{p}_{\mathit{njt}\mathrm{ }}+\mathrm{ }{\epsilon }_{\mathit{njt}}\mathrm{ }$$(1) can be rewritten as: (2) $$U_{\mathit{njt}}=\mathrm{ }{{({\lambda }_{n}w}_n)}^{\mathrm{\prime }}{X}_{\mathit{njt}}+{\lambda }_n{p}_{\mathit{njt}\mathrm{ }}+\mathrm{ }{\epsilon }_{\mathit{njt}}\mathrm{ }$$(2)

These time equivalents represent the number of additional months of PFS a patient would be willing to give up for a treatment that provides a decrease in side effects or a more desirable mode of administration. Positive HME values indicate that the attribute is desirable and negative HME values indicate that an attribute is undesirable.

PFS and chance of late-onset side effects were recoded as continuous variables to estimate HME values, while short-term side effects, long-term side effects, and mode of administration was recoded using dummy coding. The reference level for each of the dummy coded variable was the level that was hypothesized to be least desirable; severe for short-term side effects, moderate for long-term side effects, and infusion every 3 weeks for mode of administration. This specification allows us to examine how much PFS participants are willing to trade to move from a worse attribute level to a better attribute, i.e. for short-term side effects we evaluated how much PFS participants were willing to trade to go from experiencing severe short-term side effects to experiencing either moderate or mild short-term side effects. A mixed logit was used as before, but PFS was set as fixed⁴⁹ while short-term side effects, long-term side effects, risk of late onset-side effects, and mode of administration were modeled as random parameters with normal distributions. Difference in time equivalents between those receiving one or less LOT and those receiving two or more LOT were assessed via t-tests.

Results

Descriptive statistics

A total of 466 of the 520 eligible patient participants (90%) completed as least one DCE task. 1% had not yet received a therapy, 41% had received only one LOT, and 58% had two or more LOT. A total of 82 choice tasks were not completed among 17 participants (2.0%). Twelve participants (2.6%) always selected “Drug A” or “Drug B” across all choice tasks indicating task nonattendance. The majority of patients were female (74.4%) and the average age of participants was 58.1 years old (standard deviations [SD] = 13.0). Participants were primarily white (83.8%), with 68.5% being married and 81.1% having children. Clinical factors significantly associated with the number of LOTs received were years since diagnosis (p=.04), stage at diagnosis (p<.001), current stage (p<.001), biomarker presence (p<.001), clinical trial participation (p<.001), and doctor-led decision-making style (p=.004). Patient characteristics can be seen in Table 1.

Table 1. Patient characteristics by number of lines of therapy.

Variable	Lines of therapy		p-Value
	1 or fewer (N = 196)	2 or more (N = 270)
Demographic characteristics
Female	75.1%	73.9%	.78
Age	59.2 (12.3)	57.3 (13.4)	.13
Hispanic	7.4%	9.9%	.35
White	83.7%	84.0%	.94
Primary language is English	98.4%	99.2%	.42
Armed Forces	12.1%	10.3%	.56
Married/domestic partnership	66.3%	70.1%	.39
Has children	80.5%	81.5%	.79
Highest level of education is high school/GED or less	10.0%	5.4%	.06
Unemployed	4.8%	3.5%	.49
Annual household income more than $75,000	39.5%	41.8%	.63
City/town population less than 2,500	10.7%	4.7%	.02
Clinical characteristics
Years since diagnosis	4.4 (4.0)	5.1 (3.4)	.04
Chronic conditions at diagnosis	27.6%	27.4%	.97
Adenocarcinoma	51.0%	61.1%	.03
Stage at diagnosis			<.001
Stage I/II	45.7%	25.8%
Stage III/IV	54.3%	74.2%
Current stage			<.001
No evidence of disease	66.1%	28.1%
Stage I/II	7.9%	9.2%
Stage III/IV	26.1%	62.7%
Actionable biomarker present*	29.6%	47.8%	<.001
Private Health Insurance	65.6%	66.8%	.79
Medicare/Medicaid	36.3%	44.8%	.07
Other health Insurance	9.1%	9.6%	.85
No health Insurance	2.7%	0.8%	.11
Clinical trial participation	8.2%	28.9%	<.001
Doctor-led decision making style	18.4%	9.3%	.004
Personality characteristics
Optimistic	80.4%	80.1%	.93
Self-control	79.4%	77.4%	.62
Improving health	86.8%	80.1%	.06
Risk taker	37.0%	51.3%	.003
Good with numbers	70.9%	71.3%	.93

*Actionable biomarkers include ALK, BRAF, EFGR, NTRK, ROS1.

Treatment preferences

Aggregate mixed-logit results are presented in Figure 2. Participants valued all treatment attributes (p < 0.05). Results met face validity as preference estimates moved in the hypothesized and intuitive direction^50,51; participants favored PFS and wanted to avoid short-term side effects, long-term side effects, risk of late-onset side effects and more restrictive modes of administration. The coefficient for the ASC was −5.7 (SD = 4.1), participants preferred some treatment to no treatment at all. The large SD associated with the ASC indicates that there was a lot of heterogeneity across participants in the decision between no treatment and the hypothetical treatments presented.

Figure 2. Aggregate preferences: mean preference weights with 95% confidence interval and standard deviation of preference distribution.

The direction of the preference estimates remained similar to that of the aggregated mixed logit model when stratified by LOT (≤1 or ≥2), but the two models were statistically different as assessed by a Wald test (p<.001) [41]. Stated preferences were statistically different on the following treatment attributes: 18 months (p=.017) and 24 months (p=.004) of PFS, moderate short-term side effects (p<.001), moderate long-term side effects (p=.012), and 30% chance of late-onset side effects (p= <.001). Preferences did not significantly differ across 12 months of PFS, mild or severe short-term side effects, long-term side effects, or mode of administration. The two models also did not differ in the ASC. Re-scaled coefficients are displayed in Table 2 and graphed in Figure 3. Additional stratifications of age, gender, race, and education were not found to be statistically significant (p>.05) when assessed by a Wald test.

Figure 3. Stratified mixed logit results. *p<.05; **p<.01; ***p<.001

Table 2. Mixed logit results stratified by LOT.

Attribute	Level		LOT ≤ 1 (N = 196) coefficient	LOT ≥ 2 (N = 270) coefficient	p-Value
Progression free survival	12 Months	Mean	0 (0.17)	0 (0.13)
		SD	1.53 (0.19)***	1.57 (0.14)***
	18 Months	Mean	1.78 (0.09)***	1.5 (0.07)***	.017
		SD	−0.12 (0.17)	0.12 (0.13)
	24 Months	Mean	2.55 (0.14)***	3.11 (0.13)***	.004
		SD	−1.4 (0.25)***	1.69 (0.2)***
Short-term side effects	Mild	Mean	0 (0.14)	0 (0.1)
		SD	0.95 (0.15)***	0.85 (0.11)***
	Moderate	Mean	−0.91 (0.09)***	−0.34 (0.08)***	<.001
		SD	0.32 (0.14)*	0.89 (0.13)***
	Severe	Mean	−2.41 (0.15)***	−2.53 (0.13)***	.532
		SD	−1.27 (0.21)***	1.73 (0.19)***
Long-term side effects	None	Mean	0 (0.11)	0 (0.08)
		SD	0.61 (0.17)***	0.47 (0.11)***
	Mild	Mean	−0.42 (0.09)***	−0.22 (0.08)**	.092
		SD	0.23 (0.21)	0.52 (0.12)***
	Moderate	Mean	−1.68 (0.13)***	−1.28 (0.09)***	.012
		SD	−0.84 (0.26)**	0.99 (0.16)***
Risk of late-onset side effects	10%	Mean	0 (0.1)	0 (0.08)
		SD	0.97 (0.14)***	0.7 (0.11)***
	20%	Mean	−0.33 (0.09)***	−0.22 (0.07)**	.333
		SD	0.63 (0.17)***	0.06 (0.18)
	30%	Mean	−1.33 (0.12)***	−0.78 (0.08)***	<.001
		SD	−1.6 (0.25)***	0.76 (0.2)***
Mode of administration	Pills daily anytime	Mean	0.5 (0.09)***	0.62 (0.07)***	.305
		SD	−0.49 (0.17)**	0.14 (0.14)
	Pills daily without food	Mean	0.31 (0.09)***	0.37 (0.07)***	.572
		SD	0.31 (0.16)*	0.31 (0.1)**
	Infusion every 3 weeks	Mean	0 (0.09)	0 (0.08)
		SD	0.18 (0.22)	0.16 (0.18)
ASC		Mean	−6.35 (0.69)***	−5.56 (0.48)***	.346
		SD	4.44 (0.52)***	3.72 (0.33)***

Standard errors (SEs) are denoted in parentheses; *p<.05; **p<.01; ***p<.001.

Healthy-month equivalents

Figure 4 presents the healthy-month equivalents results stratified by LOT. Participants were willing to trade between 1.3 and 11.4 months of PFS in order to decrease side effects or change between modes of administration. Participants that had received ≤1 LOT and ≥2 LOT were willing to trade a different amount of PFS to go from moderate to mild short-term side effects (p<.001), moderate to no (p<.001) and mild to no (p<.001) long-term side effects. There were also differencing amounts of tradeoff acceptable between the groups for a 10% decrease in risk of late-onset side effects (p=.016), a decrease in mode of administration from an infusion every 3 weeks to pills taken daily at any time (p=.005) and from pills taken daily without food to pills taken daily at any time (p<.001). The two groups did not differ on amount of PFS willing to trade to go from severe to mild short-term side effects, but biggest trade-off occurred in both groups here with ≤1 LOT willing to trade 11.4 months of PFS and ≥2 LOT willing to trade 10.8 months. Those with ≤1 LOT were willing to trade more for decreases in long-term side effects and chance of late-onset side effects, while those with ≥2 LOT were willing to trade more for a decrease in moderate to mild short-term side effects, and less restrictive modes of administration.

Figure 4. Healthy month equivalents. *p<.05; **p<.01; ***p<.001; PFS- progression free survival

Debriefing items

The majority of patients found the choice tasks easy to understand (84.8%), easy to answer (66.1%), consistent with their preferences (86.6%), and relevant to them (66.5%). Figure 5 shows the agreement rate with each of these statements by LOT grouping. Acceptability measures met or exceeded the 75% threshold for easy to understand (p<.001) and consistent with preferences (p<.001) for both groups. However, they failed to meet the a priori benchmark for easy to answer and relevant to participants. The groups differed on agreement for finding the choice tasks relevant to them, with the ≤1 LOT group finding the DCE less relevant to them than the ≥2 LOT group (53.7 vs 75.7%, p<.001). The groups did not differ on easy to understand (81 vs 87.5%, p=.06), easy to answer (64.7 vs 67.0%, p=.61), and consistent with their preferences (85.3 vs 87.5%, p=.49). We also explored factors related to finding the choice tasks relevant within the ≤1 LOT group, as the degree of acceptance was so low. We found that current stage was related to relevance with 79.7% of those finding the tasks irrelevant to them to have no evidence of disease.

Figure 5. Debriefing items stratified by LOT. *p<.05; **p<.01; ***p<.001

Discussion

A DCE was used to show that patients with NSCLC have varying preferences for lung cancer treatment. The study sample was composed of a clinically heterogeneous group of patients with lung cancer. Participants had different histologies of NSCLC, were at different disease stages, and had undergone different treatments. We showed that patients with different numbers of LOT displayed different preferences for lung cancer treatment. Significant preference heterogeneity was present. Patients with different number of LOT were willing to trade different amounts of PFS for varying levels of treatment attributes. Thus, efforts to deliver more desirable lung cancer treatments may be more effective if tailored to individuals’ preferences and treatment history.

We found that years since diagnosis, stage at diagnosis, current stage, biomarker presence, lung cancer type (adenocarcinoma vs not), clinical trial participation, and doctor led decision-making style were all significantly associated with LOT. Biomarker-driven lung cancer, specifically adenocarcinoma, tends to have more treatment options^9,10 and patients having ≤1 LOT are more typically early-stage patients who have received surgery and neoadjuvant/adjuvant chemotherapy with curative intent, while those having ≥2 LOT include more advanced-stage patients with more aggressive tumors. Those patients having ≤1 LOT may also be more willing to try more toxic treatments or treatments with less benefit, since this is a scenario they have not yet experienced. There also may be less patient choice in first line therapy, as indicated by a higher percentage of those in the ≤1 LOT group adopting a doctor-led decision making style about treatment. There are more clinical trials for advanced-stage patients⁵² and therefore such patients might be more attuned to a longer PFS for a hypothetical drug. It is therefore not surprising that patients with ≤1 LOT in our cohort were more likely be to early-stage patients with no evidence of disease (NED) and whose tumors lack an actionable biomarker. Conversely, almost 50% of the patients in our cohort who had received ≥2 LOT were more likely to have biomarker-driven advanced-stage lung cancer which tends to be aggressive, have more treatment options, and more clinical trial options available. These factors and their role in preference heterogeneity are not well understood. Further research is warranted to understand the relationships between these factors and number of LOT and how that relates to treatment preferences.

We were able to explore whether participants prefer a treatment with both risks and benefits to no treatment at all by including a treatment opt-out choice. In our models, there was significant variability around the ASC representing the opt-out, but overall treatment was preferred to no treatment. The large negative coefficient of the ASC displays that there is significant disutility association with the no treatment option. This was the case in both LOT sub-groups.

Modeling the short-term side effects and long-term treatment effects attributes as continuous would have resulted in a more parsimonious model. By allowing these variables to be modeled as categorical, we were able to explore whether patients assigned different time equivalence estimates to a change from a severe to a moderate treatment effect or a change from a moderate to a mild treatment effect. Using this specification, the model did not have to assume that patients assigned the same importance to both changes.

Healthy-month equivalents varied over LOT, indicating that those receiving different amounts of therapies place value on different treatment attributes. We found that those with more LOT were willing to trade more PFS for less severe short-term side effects, indicating that those patients may place a greater emphasis on the present. Patients with a greater number of LOT were also willing to trade more PFS for a less intensive mode of treatment administration than those with fewer LOT. These differences may be due to the build-up of treatment-related side effects over time in those patients receiving multiple LOT (i.e. neuropathy). Those with fewer LOT were willing to trade more PFS for less severe long-term side effects and less risk of late-onset side effects, demonstrating that those patients may place more emphasis on the future and may be less worried about the impact of short-term side effects. As LOT is related to disease progression, this may offer some understanding as to where the groups were willing to make differing amounts of trade-off.

Participants were recruited in the US through several methods, including patient advocacy and research organizations. This method of recruitment may result in larger preference heterogeneity, but by using mixed logit models, we were able to capture this variation. While other sample strategies were examined, this strategy was chosen in order to provide the largest sample size, geographic variation, and a better representation of patients. We also excluded family members and caregivers from our analyses. Previous research shows that patient and caregiver preferences may differ^37,53 and thus we chose to focus on the patient perspective. This allowed us to gain insight about patients’ preferences and trade-offs they were willing to make. This recruitment method also resulted in a sample that was majority female which was consistent with other advocacy based studies^53,54. Greater involvement by women is explained by many factors, including volunteerism and humanistic tendencies. We do not believe that this is presents a bias in our work, but reflects reality in advocacy groups. We made many efforts to be inclusionary in our study, including having diverse patient engagement through the development of our study.

Difficulty in answering DCEs is consistent with previous literature⁵⁵. Other methods of preference elicitation, including simple choice or ranking techniques, have been shown to be easier to answer, but a DCE was used here in order to evaluate explicit trade-offs patients were willing to make between PFS and other treatment attributes. DCEs have also been shown to be more consistent with preferences, as was illustrated in our study as well. Recent guidelines for the application of a DCE analysis in health were applied to help direct the analyses and minimize bias³⁶.

This study does not explore reasons why preferences differ over participants with differing amounts of therapies and as number of LOT is associated with other clinical factors, there may be other factors influencing the preference differences we have seen here. This limits implications for the tailoring of lung cancer treatment to patients based on lines of therapy received. Further research is needed including methods such as prospective studies and qualitative interviews that explore this observed difference in preferences between patients with different quantities of therapies. This necessary research will ultimately help clinicians individualize lung treatment.

Conclusion

The study results suggest that lung cancer may be preference-sensitive and that preferences might differ by clinical and demographic characteristics. We showed that LOT is related to differences in preferences and that patients with NSCLC consider both benefits (PFS) and risks in their treatment decision making. Short-term, long-term, and late-onset side effects each factored into patient’s preferences and treatment decisions should consider each of these different types of side effects. Therefore, it is important that patients have access to relevant information and tools to help inform the shared decision-making process regarding their treatment options to ensure patient-centric care.

Transparency

Declaration of funding

This work was supported by LUNGevity (with funding from Celgene, Merck, and Boehringer Ingelheim). Celgene, Merck, and Boehringer Ingelheim had no influence on the design of the project, data collection, analysis, interpretation, or writing of the report.

Declaration of financial/other relationships

The authors report no potential conflicts of interest. Peer reviewers on this manuscript have no relevant financial or other relationships to disclose.

Compliance with ethical standards

This research was conducted in accordance with the Declaration of Helsinki, and the study protocol was reviewed by the Institutional Review Board. Dr. Presley is a Paul Calabresi Scholars supported by the OSU K12 Training Grant for Clinical Faculty Investigators (5K12 CA133250-09)

Acknowledgements

The authors sincerely thank LUNGevity and the members of the Patient Action Committee (PAC) for their participation, engagement, and valuable contributions to Project Transform.