Our first concern is regarding the items Mulhern et al. have selected for their health-state classification system (DEMQOL-U). We question the content validity of the items, because these do not cover the full spectrum of dementia health-related quality of life (HRQOL). In our opinion, the authors place too much emphasis on mood-related items. It is generally accepted that health is composed of three domains: physical, mental, and social [3,4]. One would therefore expect any HRQOL measure to cover all three to at least some degree. Disease-specific HRQOL measures will most likely put more emphasis on one or two of the three domains depending on the disease. In the case of dementia, one would expect an instrument to emphasize mental and social well-being. This is exactly what the original DEMQOL does. In contrast, not all these domains are covered by the DEMQOL-U.

The authors of the original article describing the development of the DEMQOL measure used a conceptual framework of five domains: 1) daily activities and looking after yourself, 2) health and well-being, 3) cognitive functioning, 4) social relationships, and 5) self-concept [5]. By comparison, three of the five items of the DEMQOL-U fall under “health and well-being” and two under “cognitive functioning.” This suggests that domains 1, 4, and 5 that Mulhern et al. intended to cover were omitted from their classification system. In the article by Smith et al. [5], the authors present a preliminary factor analysis solution for the field test data. It covers four factors: 1) daily activities, 2) memory, 3) negative emotion, and 4) positive emotion. The final field test revealed a different four-factor structure, which the authors found more difficult to interpret.

The study by Mulhern et al. advances a five-factor solution replicating the original factors 2, 3, and 4. In addition, it identifies two new factors, Social relationships and Loneliness. The omission of a domain that describes physical functioning (and thus impacts daily activities), however, is a major concern. Dementia denotes a class of illnesses that occur mostly in frail elderly people. Accordingly, many patients with dementia suffer from physical comorbidity. Thus, the omission of a physical domain could lead to the overestimation of patient-reported utilities.

The absence of several relevant conceptual domains is not our only concern. The items covering cognition and relationships that are part of the DEMQOL-U classification system might indicate some aspects of “worrying” in addition to or instead of the intended item content because of the way they are framed. The DEMQOL measure was framed in aspects of worrying because this stem was most easily understood during pretesting. The authors of the DEMQOL, however, allocated “being worried or anxious” to the domain of health and well-being. In that light, framing separate items in terms of “being worried about …” raises the possibility of confounding for these items.

Our second concern is regarding the analytical strategies applied by Mulhern et al., namely, factor analysis and Rasch analysis. Factor analysis seems unnecessary, because the same technique was used to develop the original DEMQOL instrument, albeit the number of observations was substantially lower in the earlier study. This may explain the differences between the solutions found in each study. It should be noted that factor analysis is fully directed by relationships (i.e., correlations) between variables (i.e., items). This means that two items with more or less equal distributions of responses (i.e., frequencies) will load on the same factor. However, factor analysis results will not tell us anything about the weight (i.e., importance) of these items.

Subsequently, the authors apply Rasch analysis on the items for each factor derived by the factor analysis. In the Rasch analysis, they perform several tests, one of them testing for the unidimensionality of each dimension. This does not yield much information, because the basic feature of factor analysis with varimax rotation is that it produces orthogonal (i.e., unidimensional) factors. Apart from this, the standard Rasch analysis may not be the correct response model for the type of data that are obtained in the setting of reported health levels. Mulhern et al. apply the Rasch model to Likert items, although these items do not have the correct response structure. We have noticed this incongruence in many other studies directed at transforming descriptive HRQOL questionnaires or instruments into preference-based HRQOL instruments [6,7,8,9,10]. Rasch analysis requires a “cumulative” data structure (if a respondent agrees with a statement of a certain level, this means that this person also agrees with the statements that precede this level). In standard descriptive HRQOL questionnaires, we are dealing with an “ideal point” or “single-peaked” data structure. When persons whose attitudes are to be measured agree or disagree with a statement, the implied response function is single-peaked. In other words, it is expected that a person will agree with the statements that are close to the person's own attitude and disagree with those statements (e.g., categories of the item) that are far from the person's location on the scale in either direction.

Coombs [11] developed this implied response process within a deterministic framework and coined the term “unfolding” for the simultaneous processes of locating persons and items on a scale from the agree/disagree responses. This term continues to be used in the literature. Unfolding, however, became extremely cumbersome for more than four statements. Therefore, it did not pose a challenge to the Likert approach as the favored procedure in practice. Nonetheless, modern extensions of this unfolding model that can deal with a large number of statements now exist. Given the response options for the items in the DEMQOL, we believe that such a probabilistic unfolding model for polytomous responses may have been a more valid method for item selection [12].

In summary, we feel that the introduction of the DEMQOL-U represents an important step in the right direction. The instrument, however, still has two important weaknesses: insufficient comprehensiveness and limited validity of data analysis. These weaknesses require further consideration before its use in research and clinical practice is warranted.