What Matters Most to Patients with NSCLC When Choosing Anticancer Treatments in China? A Best–Worst Scaling Study

Introduction

The incidence and mortality of malignant tumors in China have been steadily increasing. According to data from the National Cancer Center of China, in 2022, lung cancer remained the leading cause of both cancer incidence and death in the country.^1,2 Non-small cell lung cancer (NSCLC), the most common histological subtype, accounts for approximately 80%–85% of all lung cancer cases.^3,4 With the advent of precision medicine, survival among NSCLC patients has been significantly prolonged, and therapeutic strategies have grown increasingly complex. In addition to conventional surgery and chemotherapy, molecularly targeted therapies and immunotherapies have emerged as novel options for patients with advanced disease.^5–7 Current pharmacological regimens encompass cytotoxic agents, targeted therapies, and immune checkpoint inhibitors.⁸ The therapeutic landscape for lung cancer has expanded rapidly in recent years, leading to an increasingly diverse array of treatment regimens. These options differ markedly with respect to clinical efficacy, safety profiles, modes of administration, and their broader impact on patients’ quality of life.^9,10 Such heterogeneity underscores the importance of tailoring treatment decisions to individual patient characteristics. Accordingly, current clinical guidelines advocate for personalized strategies informed by tumor stage, molecular testing, and evidence-based standards.¹¹ Beyond these clinical determinants, however, the integration of patient values and preferences is indispensable, as they represent a cornerstone of truly individualized and patient-centered care.¹²

In recent years, a patient-centered model has become a core objective of healthcare reform in China, shifting decision-making from a physician-directed approach to a shared, collaborative process. Studies have shown that patient involvement in clinical decisions is crucial for improving treatment outcomes and satisfaction,¹³ and that patient preferences are closely linked to adherence and quality of life.¹⁴ Chinese Expert Consensus on the Multidisciplinary Team Diagnosis and Treatment of Lung Cancer emphasizes that the optimal treatment plan should integrate individual patient data with the latest evidence and fully respect patient values and preferences.¹⁵ Against this backdrop, quantitatively characterizing patient treatment preferences is essential for accurately identifying individual therapeutic goals and health needs, thereby advancing personalized care. Nevertheless, research on medication preferences among NSCLC patients in China remains scarce.

Stated-preference methods, including best-worst scaling (BWS) and discrete choice experiments (DCE), are widely used to quantitatively elicit patient preferences through choice modelling techniques. Although DCEs have been increasingly applied in oncology, the use of BWS in this field remains relatively limited.^16–19 However, DCE typically require respondents to evaluate multiple attribute levels simultaneously and may impose a higher cognitive burden when the number of attributes is large, whereas rating/ranking tasks are susceptible to scale-use heterogeneity and limited discrimination.²⁰ The BWS grounded in random utility theory, asks respondents to repeatedly identify the most and least important items within a set, thereby enabling efficient estimation of relative attribute importance while reducing cognitive burden and mitigating common scale biases.^21,22 Moreover, accumulating evidence suggests that BWS may be more reliable for eliciting patient medication preferences.²³ In particular, object-case BWS (BWS-1) is well-suited for prioritizing a relatively large set of treatment attributes when the objective is to obtain a stable importance ranking and preference weights.²⁴

Therefore, this study applied BWS-1 to estimate preference weights and rank the relative importance of key anticancer drug attributes among Chinese patients with NSCLC. By quantifying the relative importance of attributes and examining preference heterogeneity, the results aim to support shared decision-making in clinical practice and provide patient-centered input to broader policy and health-technology assessment deliberations.

Materials and Methods

Study Participants and Elicitation Method

A convenience sample of patients was recruited from Hainan Cancer Hospital between July and October 2022. The questionnaire comprised three sections: disease characteristics, patient demographic information, and thirteen BWS-1 choice tasks assessing drug-selection preferences. Eligible participants were aged ≥18 years, had histologically confirmed NSCLC, had received or were candidates for chemotherapy or immunotherapy, and had provided written informed consent. Patients with psychiatric disorders or those unable to complete the survey were excluded. Although no consensus on the optimal sample size for BWS studies exists, prior studies have indicated that at least 100 respondents provide a robust basis for preference modeling.^24,25 Consequently, a minimum target of 100 patients was set, and data were collected via face-to-face administration to ensure quality. Given the sample size, analyses of preference heterogeneity were prespecified as exploratory.

Prior the formal survey, a pilot test and a field-based cognitive interviews were conducted to evaluate the questionnaire’s logic flow, layout, and respondent comprehension. The field survey was carried out by a trained research team composed of faculty members and graduate students from Fudan University and Hainan Medical University, together with medical staff from the hospital. All interviewers received systematic training covering study background, methodological principles of BWS, questionnaire design, and interview techniques. Attribute explanation cards and sample tasks were provided to respondents to facilitate comprehension.

To ensure the quality of data collection, the survey was conducted face-to-face. The questionnaire was administered by trained investigators, with support from clinical staff when appropriate. Before participants completed the tasks, one or two investigators explained the rationale of the survey and provided standardized instructions, and clarified technical terms as needed. Completed questionnaires were entered into an EpiData 3.1 database using a double-entry procedure, and any discrepancies were resolved before analysis. The study received ethical approval from the School of Public Health, Fudan University (IRB# 2021–07-0911), and written informed consent was obtained from all participants in accordance with the Declaration of Helsinki.

Experimental Design

The final list of thirteen attributes was determined through a structured multi-step process to ensure comprehensiveness and clinical relevance. First, we conducted a literature review of NSCLC preference studies, treatment guidelines, and pivotal clinical trials/drug labels to compile an initial attribute pool covering efficacy, adverse events, treatment convenience, and economic burden. Second, an in-person expert consultation involving five clinical physicians in March 2022 was conducted to refine the attribute scope, improve clinical interpretability, and ensure coverage of both overall quality-of-life–related domains and specific adverse events (eg., nausea/vomiting). Third, adverse-event attributes were prioritized using multiple criteria: (i) mapping candidate symptoms to commonly used health-related quality-of-life symptom domains; (ii) prioritizing adverse events with greater documented impact on quality of life in the literature; (iii) considering adverse-event frequency profiles extracted from RCT studies trials and drug labels; and (iv) cross-checking adverse-event attributes commonly included in prior NSCLC stated-preference studies. Based on these steps, thirteen attributes were finalized (Table 1).

Table 1 Attributes of the BWS-1 Study in NSCLC Patients

For reporting and interpretation, attributes were grouped into four domains—efficacy (OS, disease control rate, PFS), safety/symptoms (eg., dyspnea, pain, fatigue, nausea/vomiting), convenience (mode of administration), and economic burden (monthly out-of-pocket expenditure)—to align with how treatment trade-offs are typically framed in clinical decision-making and stated-preference research.²⁶ Domain grouping was used for descriptive organization only; estimation was conducted at the attribute level.

Following the ISPOR Task Force’s ten design criteria, a Balanced Incomplete Block Design (BIBD) was implemented in R 4.1.3 to generate thirteen BWS-1 choice sets, each containing four attributes. This design ensured that every attribute appeared an equal number of times and co-occurred with every other attribute in a balanced manner, allowing reliable estimation of relative importance while minimizing respondent burden. To reduce cognitive fatigue, the order of tasks was fixed across all questionnaires, consistent with previous BWS studies in China.²⁴ Attribute definitions are provided in Table 1, and a sample task is illustrated in Table 2. Detailed experimental design and analytical procedures have been reported elsewhere.^22,24,27 Finally, thirteen choice sets per respondent were chosen to balance statistical precision with respondent burden. This design allowed each attribute to appear an equal number of times across tasks, ensuring reliable estimation of relative importance while maintaining participant engagement.

Table 2 Example of BWS-1 Question

Statistical Analysis

Two analytical approaches were applied: counting analysis and a conditional logit model (CLM). For counting analysis, the number of times each attribute i (among the thirteen core attributes, eg., OS, disease control rate) was chosen as “best” or “worst” was tallied at both individual and aggregate levels. Standardized Best–Worst (BW) scores, defined as the difference between best- and worst-selection counts, quantified relative attribute importance. A McFadden CLM was fitted via maximum likelihood estimation to derive preference weights, with monthly out-of-pocket cost designated as the reference attribute; model coefficients represent marginal utility differences relative to this baseline. To explore heterogeneity in preferences, K-means clustering was employed to partition respondents into two subgroups (Cluster 1 and Cluster 2), optimizing the ratio of between-cluster to within-cluster variance.

Results

Participant Characteristics

A total of 102 respondents provided informed consent and completed the survey; all questionnaires passed the embedded quality-control task and were included in the final analysis The mean age was 60.24 ± 10.64 years; 72.5% were male, 71.6% were urban residents, and 98% were married. More than half (54.9%) had a smoking history of over 10 years. Annual household income was predominantly below 150,000 CNY. Insurance coverage was primarily provided by Urban Employee Basic Medical Insurance (53.9%) and Urban–Rural Resident Basic Medical Insurance (37.3%). Regarding disease characteristics, 60.8% were diagnosed with lung cancer in 2022; adenocarcinoma accounted for 71.6% and squamous cell carcinoma for 22.5%; most patients presented at stage IIIA or later. Driver mutations were detected in 57.8% of patients (Table 3).

Table 3 Demographic and Disease Information for BWS-1 Respondents (N=102)

Treatment Preferences in NSCLC Patients

Counting‐Analysis Results

Table 4 summarizes the counting‐analysis outcomes. At the aggregate level, OS received the highest “best” count (358) and the lowest “worst” count,² whereas fatigue demonstrated the opposite pattern (“best” = 2; “worst” = 257). The top three attributes by standardized BW score were OS (BW = 356; standardized = 0.873), disease control rate (BW = 219; standardized = 0.537), and PFS (BW = 170; standardized = 0.417), indicating that survival benefit and disease control were most highly valued. Conversely, fatigue (BW = –255; standardized = –0.625), mode of administration (BW = –186; standardized = –0.456), and cough (BW = –108; standardized = –0.265) were most frequently rated as least important. At the individual level, the smallest standard deviation was observed for OS (SD = 0.265), reflecting consensus regarding its importance, whereas monthly out‐of‐pocket cost exhibited the greatest variability (SD = 0.641), indicating heterogeneous economic preferences. Overall, the attributes most often selected as “most important” were OS, disease control rate, PFS, dyspnea, and pain; those most often selected as “least important” were objective response rate, hemoptysis, fever, nausea/vomiting, monthly out‐of‐pocket cost, cough, mode of administration, and fatigue.

Table 4 BWS-1 Results

Conditional Logit‐Model Results

CLM (Table 4) corroborated the primary role of survival-related attributes. The model showed good overall fit (likelihood ratio χ²=1932, df=12, P<0.001) and strong discrimination based on predicted choice probabilities (AUC=0.92, 95% CI: 0.90–0.94), which is generally considered outstanding (AUC>0.90).²⁸ OS (β= 3.537; preference weight = 0.574), disease control rate (β= 2.025; preference weight = 0.126), and PFS (β= 1.574; preference weight = 0.081) held the highest coefficients and weights, consistent with counting analysis. Although monthly out‐of‐pocket cost served as the reference attribute, its relatively low preference weight (0.017; rank 9) suggests a modest negative impact on treatment choice, potentially owing to insurance coverage or sample heterogeneity. Adverse events such as nausea/vomiting (β= –0.005), fever (β= 0.048), and cough (β= –0.185) did not reach statistical significance (P ≥ 0.05), indicating limited influence on overall preferences.

Preference Heterogeneity

K-means clustering was applied to the counting‐analysis results to examine subgroup differences. As illustrated in Figure 1, the greatest heterogeneity was observed for dyspnea, pain, mode of administration, and objective response rate—particularly for mode of administration—while other attributes exhibited relatively homogeneous preferences across clusters.

Figure 1 K-means Clustering Results Based on Counting Analysis.

Discussion

This study applied a BWS-1 approach to systematically characterize anticancer drug-attribute priorities among NSCLC patients in China. Comparative analyses using counting and CLM yielded largely consistent patterns, supporting the stability of the main findings across analytical approaches. Minor discrepancies were observed for a small number of attributes (eg., monthly out-of-pocket cost and nausea/vomiting), which likely reflect differences in model parameterization and the fact that some attributes had weaker or non-significant effects in the CLM.

Patients’ preference intensities were found to be markedly greater for survival‐related attributes—namely, OS, disease control rate, and PFS—than for other considerations. Given the importance-ranking nature of BWS-1 tasks, the strong dominance of OS may partly reflect a ceiling effect whereby universally salient outcomes attract repeated “best” selections. In contrast, attributes related to treatment convenience and symptom management, such as cough, mode of administration, and fatigue, were deemed relatively less important. These findings are consistent with international studies that have examined lung cancer patients’ preferences using DCE. Similar to our results, those studies also emphasized the importance of survival benefits and disease control.^29–32 Although targeted therapies and immunotherapies have substantially prolonged survival, long‐term outcomes remain suboptimal: many advanced NSCLC patients, even when treated with combinations of targeted agents, chemotherapy, and immunotherapy, exhibit median OS of less than two years.^33–35 Moreover, the challenge of balancing treatment efficacy against side effects remains, as tumor heterogeneity often drives resistance to molecularly targeted drugs, thereby reducing therapeutic benefit.³⁶ Current modalities therefore continue to fall short of fully meeting patients’ expectations for durable disease control.

Interestingly, objective response rate—a commonly reported endpoint in clinical trials—received relatively low scores. This may reflect patients’ limited understanding of this technical term, making it less relevant to their perceived treatment priorities. This suggests that, in future preference studies, attribute selection should strive to maintain scientific rigor while avoiding overly specialized clinical or biochemical metrics, thereby ensuring that patients can readily comprehend the significance of each attribute.

In addition, adverse reactions such as dyspnea, fever, and nausea/vomiting had varying impacts on NSCLC patients’ preferences, with dyspnea being the most strongly avoided symptom. For NSCLC patients who already suffer from impaired pulmonary function, the onset of dyspnea can significantly compromise their quality of life and overall physical and mental well-being. Therefore, in the selection and optimization of treatment regimens, patients’ desire to avoid adverse effects like dyspnea should be carefully considered. A balance must be struck between the incidence of side effects and the symptom burden, and proactive measures should be taken to prevent and alleviate such adverse events.

With the widespread adoption of innovative therapies such as targeted treatments and immunotherapy, NSCLC patients now face an increasingly diverse range of drug options. However, clinical decision-making and reimbursement evaluations still lack systematic evidence from the patient perspective. Our findings indicate that NSCLC patients in China place the highest value on survival-related outcomes (overall survival, disease control, progression-free survival), while showing relatively less concern for endpoints such as objective response rate, treatment mode of administration, or specific side effects. This pattern was observed in the aggregate results, but some attributes (eg., fever, nausea/vomiting, cough) did not reach statistical significance in the CLM and should be interpreted cautiously. This contrasts with the emphasis often placed by oncologists on biomarker-driven efficacy and minimizing adverse events. Previous studies have reported similar discrepancies: physicians tend to focus on reducing toxicity, whereas patients are more concerned with treatment benefits and the treatment process itself.³⁷ In China, survey evidence also revealed that while clinicians and family caregivers prioritize survival, patients emphasize financial burden and quality of life.³⁸ These divergences highlight the need to strengthen shared decision-making in NSCLC care, ensuring that patients’ voices and priorities are integrated into clinical consultations. Our heterogeneity findings further suggest that preferences for symptoms and treatment convenience may vary across individuals, underscoring the value of individualized preference elicitation in practice.

From a broader perspective, patient preference data can also inform health policy and reimbursement decisions. Internationally, frameworks such as value-based insurance design demonstrate how patient values can be systematically incorporated into coverage evaluations by incentivizing high-value care. In China, health preference evidence can provide patient-centered input to policy and health technology assessment deliberations by clarifying the relative importance of outcomes and treatment features, and by supporting more transparent communication of value.

This study has several limitations. First, participants were recruited from one hospital and the sample size was relatively small, which may limit generalizability. Given the sample size, we did not conduct subgroup analyses across different patient groups; future studies should use multi-center sampling and larger cohorts to enable more robust preference heterogeneity assessment. Second, despite pilot testing and interviewer support, some attributes—especially those describing clinical endpoints and disease-related symptoms—may have been interpreted heterogeneously across respondents. This highlights the importance of clear and concise attribute formulation in preference surveys. Future studies should further optimize attribute descriptions and implement additional comprehension checks to minimize misinterpretation. Third, compared with DCE, BWS-1 allows fewer straightforward ways to incorporate explicit quality-control items, such as clearly dominant choice tasks. As a result, identifying inattentive responses may be more challenging, and additional design or analytic measures may be needed to support data-quality assessment.

Conclusion

Our study found that OS, disease control rate, and PFS are the most important attributes for NSCLC patients in China when selecting anticancer drug treatments. Heterogeneity in cost preferences was observed, but these subgroup patterns should be interpreted as exploratory given the sample size. Overall, the findings provide patient-centered evidence on treatment priorities that can support shared decision-making and the development of patient-centered communication and care pathways.