How is the minimal clinically important difference established in health-related quality of life instruments? Review of anchors and methods

Health-Related Quality of Life (HRQoL), a multidimensional construct that assesses several domains (e.g., physical, emotional, social), is an important Patient-Reported Outcome (PRO) in clinical trials as well as in routine clinical practice, and such information is also used by health policy makers in health care resource allocation and reimbursement decisions [1‐3]. A PRO is defined as “any report coming directly from patients about how they function or feel in relation to a health condition and its therapy” [4]. Interpretation of changes in HRQoL scores of Patient-Reported Outcomes (PROs) is a challenge to the meaningful application of PRO measures in patient-centered care and policy [5].

Numerous clinical trials have established the importance of HRQoL in various diseases, and it is increasingly popular to evaluate generic and disease-specific HRQoL in clinical trials as a measure of patients’ subjective state of health [5, 6].

To be clinically useful, HRQoL instruments must demonstrate psychometric properties such as validity, reliability and responsiveness to change [7, 8]. Responsiveness to change is important for instruments designed to measure change over time. However, the statistical significance of a change in HRQoL scores does not necessarily imply that it is also clinically relevant [9‐12]. Indeed, health policy makers need to present clinically meaningful results, to determine if the treatment is beneficial or harmful to their patients and also to know how to interpret and implement those results in their evidence-based method for clinical decision making [13]. Interpretation of clinical outcomes therefore should not be based solely on the presence or absence of statistically significant differences [14]. This highlights the need to define the minimal change in score considered relevant by patients and physicians, called ‘the Minimal Clinically Important Difference (MCID)’.

The MCID was first defined by Jaeschke [15] as ‘the smallest difference in score in the domain of interest which patients perceive as beneficial and which would mandate, in the absence of troublesome side effects and excessive cost, a change in the patient’s management’. MCID values are therefore important in interpreting the clinical relevance of observed changes, at both the individual and group levels. From the patient’s viewpoint, a meaningful change in HRQoL may be one that reflects a reduction in symptoms or improvement in function, however, a meaningful change for the physician may be one that indicates a change in the treatment or in the prognosis of the disease [16, 17].

Several methods have been developed, but no clear consensus exists regarding which methods are most suitable. An extensive review of available methods was published by Wells and colleagues and classified them into nine different methods [18].

Another review proposed three categories of methods for defining the MCID: distribution-based, opinion-based (relying upon experts) and anchor-based methods [19].

On one hand, anchor-based methods examine the relationship between a HRQoL measure with another measure of clinical change: the anchor [20]. Anchors can be derived from clinical outcomes (laboratory values, psychological measures, and clinical rating performance measures) or Patient-Reported Outcomes (PRO) (global health transition scale, patient’s self-reported evaluation of change) [20].

On the other hand, distribution-based methods use statistical properties of the distribution of outcome scores, particularly how the scores differ between patients. The distribution methods may use methods based on Standard Error of Measurement (SEM), Standard Deviation (SD), Effect Size (ES), Standardized Response Mean (SRM), Minimal Detectable Change (MDC), or Reliable Change Index (RCI) [20, 21].

The Delphi method has also been put forth in the literature. It involves the presentation of a questionnaire or interview to a panel of experts in a specific field for the purpose of obtaining a consensus [22]. The expert panel is provided with information on the results of a trial and are requested to provide their best estimate of the MCID. Their responses are averaged, and this summary is sent back with an invitation to revise their estimates. This process is continued until consensus is achieved [23].

To date, methods to determine MCID can be divided into two well-defined categories: distribution-based and anchor-based methods [20, 24‐26]. These two methods are conceptually different. Distribution-based methods are the most used with a meaningful external anchor [20, 24‐26]. Revicki et al. [21] recommended the usage of the anchor-based method to produce primary evidence for the MCID of any instrument and the distribution-based method to provide secondary or supportive evidence for that MCID.

The interest in estimation of MCID for HRQoL instruments has been increasing in recent years, and several reviews focused on estimates of MCID [20, 27‐29]. MCID values have been shown to differ by population and study context as well as choice of anchors. This variability highlights the need to understand how the MCID was statistically established and what kind of anchors have been used, in order to facilitate its application in the Quality of Life field.

A systematic review was conducted to describe, from a structural literature search, the different types of anchors and statistical methods used in estimating the MCID for HRQoL instruments, either generic or disease-specific ones.

Our systematic review identified 47 studies reporting anchors and statistical methods to estimate MCID for generic and disease-specific HRQoL instruments. This review pointed out that the interest of MCID in HRQoL instruments has been increasing in the recent years and the largest work has been done in the field of oncology disorders.

Most studies used anchor-based methods in our review (n = 41), either alone or in combination with distribution-based methods. As discussed by Gatchel and Mayer [77], anchor-based methods are good depending on the choice of the external criteria as well as the methodology used.

We observed multiple anchors chosen by authors, and the most common anchors were non-clinical and from the viewpoint of patients in order to assess how a patient feels about his or her current health status over time. These anchors are well-studied and applicable to a wide range of patients [78]. However, patients may be aware that the phase of their disease is deteriorating, thus they will conclude that their HRQoL is similarly deteriorating. Furthermore, the patients’ subjective experiences are related to the way in which people construct their memories. It is hard for people to accurately recall a previous health state; they will rather create an impression of how much they have changed by considering their present state and then retrospectively applying some idea of their change over time. Hermann [79] described the problem of “recall bias” where events intervening between the anchor points influence the recall of the original status, while Schwartz and Sprangers [80] described “response-shift” where a patient’s response is influenced by a changing perception of their context.

Clinical anchors were not widely applied in the included studies. Changes in Performance Status (PS), in particular the KPS and the WHO PS, were chosen by authors because of accessibility and interpretability [25]. As they do not provide MCID values per se [81], clinical anchors were applied in our review with distribution criterion. However, we did not find any combination with another subjective anchor.

Authors recommended the usage of multiple independent anchors [20]. Anchors must be easily interpretable, widely used and at least moderately correlated with the instrument being explored [8, 32, 33]. According to Cohen’s [82], 0.371 was recommended as a correlation threshold to define an important association. However, anchor-based methods may be vulnerable to recall bias, and as was evident in our review, different anchors may produce widely different estimates for the same HRQoL instrument.

Cut-offs for different anchors were differently assigned by authors. Even for the same anchor, many cutoffs were used. There is no agreement on the exact cutoffs for anchors, they are generally assigned for the purposes of research and depending on study context and anchor used [29].

Once the anchor has been chosen, different statistical methods were applied to estimate the MCID. The most established method in our review was the mean change score, also called the Change Difference (CD) method. This latter is defined as the mean change of patients who improved and, therefore, authors can set its cutoffs on the basis of the change score of patients who were shown to have had a small, moderate, or large change. MCID corresponds to the difference between two adjacent levels on the anchor. MCID would depend on the number of levels on the anchor: the larger the number of levels, the smaller the difference between two adjacent levels, and the smaller the MCID [83].

Each of the statistical methods has its specific concepts and produces a MCID value different from the other methods. Some authors pointed out that the largest threshold value is most often generated from the average change method, whereas the smallest threshold from the CD and MDC methods [48‐50].

Therefore, the usage of patient point of view’ anchors by most studies may be explained by the lack of satisfying objective scales, which incite the usage of subjective anchors in first place. In addition, perhaps the CD method is simple to apply by many authors, but we cannot affirm that this is the most relevant method. We conclude that there are many faces to the MCID, it is not a simple concept, nor simple to estimate.

In addition, distribution-based methods, derived from statistical analysis, were also applied in few studies. In accordance with literature [24‐29], most often fractionations in our review include 0.2 SD, 0.3 SD, 0.5 SD and 1 standard error of measurement (SEM).

Some studies determined that MCID corresponded closest to the 0.5 SD estimate. The 0.5 SD was the value in which most meaningful changes fall, as previously proven in a study by Norman et al. [84].

Distribution-based methods also produced different values of MCID depending on the distributional criterion. Nevertheless, distribution-based methods do not address the question of clinical importance and ignore the aim of MCID, which is to define the clinical importance distinctly from statistical significance. Authors recommended the usage of these methods when anchor-based calculations are unavailable [20].

In this review, MCID values were defined for Patient-Reported Outcomes (PRO) measuring HRQoL using the two methods: anchor and distribution-based methods. Some studies had been developed to determine MCID for the Patient-Reported Outcomes Measurement Information System (PROMIS) instruments, using the same reported statistical methods. In recent years, the PROMIS Network (www.​nihpromis.​org), a National Institutes of Health Roadmap Initiative, has advanced PRO measurement by developing item banks for measuring major self-reported health domains affected by chronic illness. Therefore, further studies should be developed to determine the meaningful change in HRQoL for PROMIS.

Summing up, we did not observe a single MCID value for any HRQoL instrument in our review. Several factors may influence this variability. On one hand, we found many available methods that produced many MCID values for the same HRQoL instrument. Authors applied, for the same instrument and in the same cohort, four different methods and reported four different MCID values [48‐50, 52, 59], which suggest that variation could be explained further than differences in disease severity or disease group since the same cohort of patients was analyzed. On the other hand, even with the same methodology for the same instrument, MCID values vary since the variability may be related to study population, in particular, patient demographics and patient baseline status. Wang et al. [85] stated that MCID scores are context-specific, depending on patient baseline and demographic characteristics. Therefore, factors affecting MCID values are specific to the population being studied and are non-transferable across patient groups, also related to the multiple reported conceptual and methodological differences.

Our review exhibits some limitations that deserve mention. First, the search strategy only focused on Pubmed and Google scholar, which might have caused the loss of some papers. However, the inclusion of grey literature is a useful source of relevant information, which ensures a certain standard of quality of the selected papers. In addition, our review was limited to the nine last years, which can also lead to the loss of some papers. To our knowledge, there is no review published recently to define MCID in the QoL field, our objective was therefore to provide researchers the new statistical methods to be applied for further researches.

The following question remains to be answered “which is the best method for MCID”? Sloan J (2005) [86] stated that “there are many methods available to ascertaining an MCID, none are perfect, but all are useful”. The MCID can be best estimated using a combination of anchor and distribution measures to triangulate toward a single value. Using several methods enables to assess the robustness of the results. This corresponds to a sensitivity analysis not on the data but on the methods. Anchor-based methods should be used as primary measures with distribution methods as supportive measure.

We conclude that many methods have become available, which lead to different estimations of MCID. MCID should be based on the context of each clinical study. Therefore, in order to stay cautious while interpreting MCID in the field of Quality of Life, close collaboration between statisticians and clinicians may be critical and necessary in order to integrate an agreement regarding the appropriate method to determine MCID. Moreover, as performed for the data, a sensitivity analysis on method, ie performing the analysis with several methods is highly recommended to assess the robustness of the results.