Methodology of health-related quality of life analysis in phase III advanced non-small-cell lung cancer clinical trials: a critical review

The Food and Drug Administration (FDA) considers overall survival (OS) benefit as the foundation for the approval of new anticancer drugs in the United States [1]. Nevertheless, the increasing number of effective salvage treatments available in many types of cancer (i.e. subsequent lines of treatments) has resulted in the need for a larger number of patients to be included and/or the need of a more prolonged observation period to attain sufficient events that can achieve planned statistical power; this increases the cost of clinical trials and requires a longer duration to obtain results [2]. Consequently, intermediate endpoints such as progression-free survival are often used as primary endpoints because they are assessed earlier. However, there is a lack of consistency in their definitions [3] and they are not systematically validated as surrogate endpoints for OS.

In this context, HRQoL could constitute an alternative endpoint. HRQoL is recognized as a endpoint for cancer therapy approvals by the American Society of Clinical Oncology and the FDA [1, 4] HRQoL reflects the patient-perceived evaluation of one’s health, including physical, emotional, and social dimensions as well as symptoms due to disease or treatment. Several publications have underlined some key issues regarding the heterogeneity of HRQOL reporting in randomized clinical trials (RCTs) in oncology [5, 6]. Moreover, the statistical longitudinal analysis of HRQoL remains unstandardized which compromises the comparison of results between trials [7]. To improve the reporting of HRQoL in oncology randomized clinical trials, an extension to the CONSORT statement regarding HRQoL reporting was published. Nevertheless, these guidelines do not detail the methodology of statistical analysis of HRQoL. Claassens et al. reported that some aspects of HRQoL reporting (missing HRQOL data, a priori hypotheses of HRQOL, rationale for instruments used) remain underrepresented in non–small-cell lung cancer (NSCLC) RCTs [6].

In this review, we showed that HRQoL was declared as an endpoint in only 49 % of the phase III clinical trials in advanced NSCLC published between 2008 and 2014. Moreover, we clearly demonstrated the heterogeneity and the weakness of the methodology of HRQoL measurement, statistical analysis and reporting. First, two questionnaires are the most widely used: the QLQ-C30 plus LC13 module (48 %) and the FACT-L (44 %). As an example, the OPTIMAL [8, 9] and the Lux-Lung 3 [10] trials compared the efficacy and tolerability of two tyrosine kinase inhibitor (erlotinib and afatinib, respectively) versus chemotherapy in first-line line treatment of patients with advanced EGFR mutation-positive NSCLC. HRQoL was assessed by the FACT-L questionnaire in OPTIMAL trial [9] while EORTC QLQ-C30 and LC13 module were used in LUX-Lung 3 trial [7]. These two clinical trials could hardly be compared since EORTC and FACT questionnaires for lung cancer do not contain the same dimensions in regard to impact of lung cancer on HRQoL. Moreover, two studies used the EuroQoL EQ-5D generic questionnaire which comprises only five short questions and is suitable since it limits patient burden and in that way also encourages response rate. Nevertheless this questionnaire has not been tailored to the special requirements of patients with cancer. At this time, the foremost challenge would be to promote, through cancer site and treatment modalities, guidelines for selecting the best questionnaires allowing for direct comparison of results across trials. Moreover, it is also still necessary to develop some new tools to evaluate HRQoL. Already validated questionnaires may not be adapted to new targeted biotherapy agent which can induce some long-term moderate toxicities.

Then, the number of HRQoL measures and intervals between two consecutive measures vary from one study to another. HRQoL is often captured until tumor progression, nevertheless, in advanced NSCLC, we could wonder if it would be more appropriate to measure HRQoL until death. Recommendations on the schedule of HRQoL assessment should be provided. At least three HRQoL assessment are recommended: at baseline, during treatment and at the end of the study (http://​groups.​eortc.​be/​qol/​eortc-qlq-c30). However, a more intensive HRQoL assessment is preferable to better capture the longitudinal trajectory of HRQoL level and to capture any relevant changes. In this review, most of studies evaluate HRQoL at baseline and then every treatment cycle which allow a good appreciation of the impact of treatment of HRQoL over time, but it depends on the number of cycles received by the patient. Moreover attention should be paid to the timing of diagnostic procedures which could influenced HRQoL results, especially with lifethreatening cancers. Patients are likely to be experiencing stress in anticipation of the yet unknown results. After the procedure, the patients will either be experiencing great relief or anxiety depending on the results. This point should be taken into account in HRQoL assessment design and be carefully documented in the protocol and emphasized during training.

In confirmatory clinical trials with multiple endpoints, the use of multiple test procedures is mandatory and CONSORT Statement recommends a multiplicity adjustment in case of multiple testing [11]. However, only one study clearly stated the procedure to control the type I error [12].

Prior to longitudinal HRQoL data analysis, the MCID should be a priori determined [13]. In our review, only nine studies (33.3 %) clearly specified it. The MCID represents the smallest changes/differences in HRQoL score, which is perceived as clinically important. For the EORTC questionnaires, a 5-point to a 10-point difference in scores could be considered as the MCID [14]. In patients with NSCLC, Maringwa et al. [15] tried to determine the smallest changes in HRQOL scores in a subset of the EORTC QLQ-C30 scales, which could be considered as clinically meaningful. They concluded that the estimates of 5 to 10 units of the QLQ-C30 scales may be used as guidance for clinicians and researchers to classify patients as improved or deteriorated. In our review, the 5 studies which used the QLQ-C30 and stated the MCID, all used a 10-point decrease in the HRQoL scores as the MCID. A sensitivity analysis, with a MCID of 5 points could have been proposed to assess the robustness of the results.

Missing data, considered as missing not at random, can bias the longitudinal analysis if it is not adequately taken into account [16]. Patients may drop out before the planned end of the study, resulting in the absence of any available HRQoL data after the patient’s drop out (i.e. attrition). Moreover, drop out occurs generally due to a deterioration of patient health status or death. Patients may also be too tired to fill the questionnaire entirely at a specific measurement time. This induces the potential risk to select subpopulation of patients with better HRQoL levels and with available HRQoL data. Not adjusting for missing data can limit the robustness of the results and the confidence in the HRQoL conclusions. Therefore, the profile of missing data at baseline and the number of HRQoL at subsequent time points for each group must be specified. In our review, these two information were reported in only 1 (3.7 %) and 7 (25.9 %) trials, respectively. Furthermore, only 5 studies (18.5 %) described the statistical approaches for dealing with missing data and the 5 methods were different. Thus, the reporting of missing data and the statistical approaches of analysis of missing data need to be standardized.

There are a number of possible ways of analyzing longitudinal HRQoL data. Currently, the two main robust methods are: the LMM and the TTD [6, 17]. In our review, the most widely used is the mean change from baseline (33.3 %). This method summarizes the longitudinal data into a summary statistic before performing a between-arms comparison. This method is rather simple but may overlook important changes in HRQoL along time and should not be applied. This method is not a model of longitudinal analysis. The LMM method was used in 6 studies (22.2 %). In all these studies the fixed effects, the random effects and the correlation matrix between HRQoL measures introduced in the model were not specified. Thus, results are very difficult to interpret. The LMM can estimate a time effect, an arm effect and an interaction between treatment arm and time (reflecting a different evolution of the two treatment arms over time). The LMM contains both fixed effects (reflecting average trends) such as treatment and random effects (individual trends). This model accounts for the association (i.e. correlation) of measures made on the same patient at different times.

Finally, the LMM model generally require a normally distribution of the score studied. All studies using this method did not mention this hypothesis and a priori did not check it. For EORTC questionnaires, scores generally do not respect a normal distribution due to the low number of items per dimension. The TTD approach was used in 5 studies (18.5 %). Currently the definition of the TTD is not standardized; therefore the studies must clearly define it. The definition of the TTD must include: reference score, the event of interest, the censoring process, including death or not [6]. In our review, only one study defined the TTD: “Time to deterioration in patient-reported outcomes was measured in months from random assignment to the first instance of symptom worsening (10 points from baseline). Patients without worsening, including those with disease progression, were censored at the last available patient-reported outcome assessment; those lacking post-baseline assessments were censored at random assignment. Patients who died without documented worsening were considered to have deteriorated at the time of death” [7]. Other statistical methods were used: “distribution of patients whose symptom had improved, remained stable or worsened”, “Fisher test for equal proportion of patients achieving at least two points increase”, “group comparisons of scale at each time”, “rates of symptom palliation”, “percentage of patients with at least two points improvement at the beginning of the cycle two”. These methods can guide on the choice of the most appropriate analytical method to use to analyze longitudinal HRQoL data (e.g., the LMM or the TTD) approach. However, these analyses alone cannot be sufficient to capture all information present in the longitudinal HRQoL assessment and must be completed by a longitudinal statistical approach taking into account the correlation between HRQoL measures.

Finally, it should be acknowledged that journal space is often limited, and authors may not have been able to report all the methodology. Therefore, HRQoL may systematically be reported in separate HRQoL dedicated manuscripts.

Our review demonstrated the poor quality and the heterogeneity of the measurement, analysis, and reporting of HRQoL in phase III advanced NSCLC trials. The heterogeneity between trials limits their cross comparison and the feasibility of meta-analysis.

The HRQoL CONSORT statement regarding HRQoL reporting was published in 2013 [18] and it is true that we reviewed studies published before 2013. The use of the HRQoL CONSORT extensions should be encouraged. Nevertheless, these guidelines do not detail the methodology of statistical analysis of HRQoL. Incomplete or inaccurate statistical analysis of HRQoL can affect the reliability of these outcomes. Therefore, development of guidelines for longitudinal HRQoL analysis of clinical trials is important to facilitate interpretation of HRQoL findings.