Factors contributing to the ceiling effect of the EQ-5D-5L: an analysis of patients with prostate cancer judged “no-problems”

The present study was performed in order to determine factors associated with an EQ-5D-5L CE among Japanese patients with PC [26]. We excluded patients with a maximum score on the EQ-VAS and/or FACT-P in order to remove possible “full-health” patients. This exclusion ensured that the patient sample was experiencing health-related problems that may not be detected by the EQ-5D-5L but could be revealed via alternative instruments.

Among the sociodemographic and medical characteristics examined in Models 1, 2, and 4 (Table 2), age, living alone, ECOG PS = 1, and received treatments (surgery, external-beam radiation therapy, and palliative treatment) were negatively related to a lower CE. The aging process is associated with deteriorated health status and diminished HRQoL. For instance, even among a random sample of Japanese adults, QoL scores are lower for individuals over the age of 60 [36]. In addition, Konnopka and Koeing [12] observed significant associations between age and the EQ-5D-5L dimensions of “morbidity” and “self-care.” In terms of living alone, one study analyzed the associations between living arrangement and HRQoL, revealing that HRQoL among older adults living alone or with adult children was worse than those living with a spouse in an urban area [37]. In addition to living alone, quality of care and support is also associated with HRQoL [26]. Regarding ECOG PS scores [33], a score of 0 or 4 is given when a patient is “fully active, able to carry on all pre-disease activities without restriction” or “completely disabled; cannot carry out any self-care; totally confined to a bed or chair,” respectively. These items are very similar to the EQ-5D domains of morbidity and usual activities. Thus, a high score on the ECOG PS correlates with the EQ-5D. Among the received treatments, three were negatively associated with a CE. Although further analyses are needed, namely at the EQ-5D domain level, these treatments are likely to have negative consequences for patients when compared to other treatments (e.g., active surveillance and/or brachytherapy).

The days from last treatment variable was positively significant in Models 2 and 4. A recent longitudinal study on HRQoL among locally advanced or advanced PC patients was reported by Zajdlewicz et al. [38]. Though their study did not use the EQ-5D and FACT-P, HRQoL fluctuated from diagnosis to a 5-year follow-up assessment. The significance observed in Models 2 and 4 in the present study disappeared when FACT-P subscale scores were added. Thus, it seems that this association may be related to specific FACT-P factors.

Model 3 included FACT-P subscale scores, specifically physical well-being and PC subscale scores. Two prior studies analyzed FACT-P to EQ-5D-3L mapping in metastatic PC. One was a study using data from AFFIRM (A Study Evaluating the Efficacy and Safety of the Investigational Drug MDV3100) [39]. Physical well-being and the PC subscale had the highest predictive value among a healthy participant sample (FACT-P score ≥ 76). The other study was conducted in six European countries. The physical and functional well-being subscales had the highest predictive value for the EQ-5D based on generalized linear models [40]. The discrepancies between these two studies could be due to the analyses and settings employed. Although our results showed significance differences of physical well-being and the PC subscale, just as the AFFIRM study showed, further work is needed to examine similarities and differences between FACT-P and EQ-5D-5L associations.

When we added FACT-P subscales into the sociodemographic and medical factors models, results changed considerably (Model 4 to Model 5 in Table 2). The variables of living statuses, PSA concentration, ECOG PS, maximal CECAE grades, the days from last treatment, surgery, external-beam radiation therapy and dummy variables for registered hospitals were not selected, and age was no longer significant. However, the physical well-being and PC subscale score significance was maintained from Model 3 to Model 5. It is possible that the continuity in significance values could be due to the highly sensitive nature of the FACT-P subscales.

Statistical significance for palliative treatment was observed in Models 4 and 5. Palliative treatment includes care for pain stemming from bone metastasis, spinal paralysis, hematuria, lower urinary tract obstruction, and renal failure [30]. These symptoms work to negatively impact the ceiling effect of EQ-5D-5L. Further studies are needed to clarify which circumstances, situations, and/or statuses caused by the palliative treatment can be explained by factors other than the two FACT-P subscale scores. This viewpoint may also be helpful for facilitating “bolt-on” studies concerning PC patients.

Although Konnopka and Koenig [12] focused on the association between CE on the EQ 5D-5L and morbidity in a general population, to our knowledge, ours is the first study to analyze the association between CE on the EQ-5D-5L and predictor variables among PC patients. Nevertheless, we would like to raise four limitations of this study. First, due to the small sample size, we could not perform more detailed analyses, such as those performed by Konnopka and Koenig (i.e., analyzing the EQ-5D-5L at the dimension level). To better clarify and discuss our afore-mentioned gaps (i.e., which factors contributed to the loss of significance for the living alone, surgery, external-beam radiation therapy variables, etc.), analyses at the EQ-5D-5L domain level and with better statistical power are needed. Second, as we have described elsewhere [26], the present study relied on self-administered questionnaires. We cannot completely rule-out the possibility of selection bias. If patients were unable to complete the questionnaires, they were excluded for practical and ethical purposes. Third, we implemented five LASSO selection and followed logistic regression models. While cautions must be taken for multicollinearity and confounding in multivariate analysis, it has been reported that LASSO selection performs relatively better concerning multicollinearity and robustness [41‐44]. Model 5 showed the lowest AIC among our models, and almost all inputted variables were excluded, which indicates that the LASSO selections in our analysis work well. Lastly, statistically significant ORs of dummy variables for registered hospitals in a few of the models may have occurred due to the different tendency for disease progression of patients registered in each hospital [26]. We also cannot exclude regional differences between the hospitals and/or judgments by physicians in charge. Therefore, there is a possibility that our results do not sufficiently generalize to the entire population of PC patients.