Development of DEMQOL-U and DEMQOL-PROXY-U: generation of preference-based indices from DEMQOL and DEMQOL-PROXY for use in economic evaluation

Article history

This issue of the Health Technology Assessment journal series contains a project commissioned/managed by the Methodology research programme (MRP). The Medical Research Council (MRC) is working with NIHR to deliver the single joint health strategy and the MRP was launched in 2008 as part of the delivery model. MRC is lead funding partner for MRP and part of this programme is the joint MRC–NIHR funding panel ‘The Methodology Research Programme Panel’.

To strengthen the evidence base for health research, the MRP oversees and implements the evolving strategy for high quality methodological research. In addition to the MRC and NIHR funding partners, the MRP takes into account the needs of other stakeholders including the devolved administrations, industry R&D, and regulatory/advisory agencies and other public bodies. The MRP funds investigator-led and needs- led research proposals from across the UK. In addition to the standard MRC and RCUK terms and conditions, projects commissioned/managed by the MRP are expected to provide a detailed report on the research findings and may publish the findings in the HTA journal, if supported by NIHR funds.

The authors have been wholly responsible for all data collection, analysis and interpretation, and for writing up their work. The HTA editors and publisher have tried to ensure the accuracy of the authors’ report and would like to thank the reviewers for their constructive comments on the draft document. However, they do not accept liability for damages or losses arising from material published in this report.

Declared competing interests of authors

none

Copyright statement

© Queen's Printer and Controller of HMSO 2013. This work was produced by Mulhern et al. under the terms of a commissioning contract issued by the Secretary of State for Health. This issue may be freely reproduced for the purposes of private research and study and extracts (or indeed, the full report) may be included in professional journals provided that suitable acknowledgement is made and the reproduction is not associated with any form of advertising. Applications for commercial reproduction should be addressed to: NIHR Journals Library, National Institute for Health Research, Evaluation, Trials and Studies Coordinating Centre, Alpha House, University of Southampton Science Park, Southampton SO16 7NS, UK.

Background

Overview

This is necessarily a complex multiphase study. For ease of understanding this chapter will set out the overall plan of the investigation and development of the preference-based measures from the DEMQOL system. Each of the major elements will then be reported and discussed in its own chapter before bringing the data together to draw conclusions from the programme as a whole.

The project had four linked phases:

• phase 1 – derivation of the health-state classification systems

• phase 2 – general population valuation survey and modelling to produce values for every health state defined by each classification system

• phase 3 – patient and carer valuation survey

• phase 4 – application of measures to trial data.

The methods of deriving a preference-based measure from an existing condition-specific measure are the subject of a recent HTA review. 37 This identified six stages for this process (Figure 1). These stages are used as a guide to the key components in the development of a preference-based measure rather than a prescriptive methodology as it is not always practical or possible to follow each stage separately or sequentially. Furthermore, the precise technique used at each of the development stages used may differ depending on the size and structure of the original instrument. Derivatives of this approach have been applied successfully in deriving a preference-based index from the generic Short Form questionnaire-36 items (SF-36)25 and Short Form questionnaire-12 items (SF-12)38 and condition-specific measures including the King’s Health Questionnaire,39 the Asthma Quality of Life Questionnaire,40,41 and the Overactive Bladder Questionnaire. 42

FIGURE 1.

The six stages for deriving a preference-based HRQL measure.

The problem with deriving preference-based measures from existing measures of HRQL such as DEMQOL and DEMQOL-Proxy is that they are simply too large. The instruments include multiple dimensions and numerous items measuring a range of HRQL constructs and therefore would define many millions of potential health states, which would be too many and complex for preference valuation by respondents. The first challenge in this type of work is therefore to fashion a health-state classification system that is amenable to valuation by sampling items from the original instruments (like the EQ-5D in structure with a small number of dimensions and minimum number of items per dimension). The process of developing a health-state classification system is stages I–IV in Figure 1. To carry out stages I–IV, we identified and used existing data sets that included DEMQOL and DEMQOL-Proxy data (however, we were unable to identify data to carry out the validation required for stage IV).

The aim of phase 1 is to develop two brief health-state classification systems, one from DEMQOL and one from DEMQOL-Proxy, that could be used to generate health states for valuation. These classification systems consist of items taken from DEMQOL and DEMQOL-Proxy, respectively, and so can be derived from any study that has used the existing DEMQOL and/or DEMQOL-Proxy instruments. The classification systems and corresponding preference weights have been named DEMQOL-U and DEMQOL-Proxy-U, with the ‘U’ referring to the utility scores generated for every health state.

Organisations such as NICE and other similar agencies around the world charged with providing judgements on the cost-effectiveness of health-care interventions mostly require HRQL to be valued using a choice-based technique. 43 There are a range of choice-based techniques available including time trade-off (TTO), standard gamble (SG), ranking, and discrete choice experiments (DCEs). In TTO, respondents are asked how many years they would be willing to sacrifice in order to be in full health in comparison with a set number of years (usually 10) in an impaired health state described by the classification system. The time spent in full health is varied until the respondent is indifferent between the options. SG asks respondents how big a risk of death they are willing to take in order to have a chance of ending up in full health. Ranking asks respondents to rank a number of health states from best to worst and DCE asks respondents to choose which is better out of two health states produced by the classification system. For the valuation studies described in this report we used TTO and ranking. TTO was selected for this study over SG to avoid asking respondents how much they would be willing to risk their life. We used the TTO protocol developed in York,44 which was used to derive the EQ-5D value set currently recommended for use in studies of cost-effectiveness by NICE. 45 This was done to allow for some level of comparability between the utility values produced by the DEMQOL measures and the EQ-5D value set and to hence best meet the reference case of NICE.

The question of whose values should be used to value health is an interesting one, but ultimately political,39 and so beyond the scope of this project. The main valuation survey in this study used a representative sample of the general public to conform to the requirements of NICE and the majority of other reimbursement authorities around the world. General population values are typically used because they reflect the views of society as a whole, which funds the services, and to enhance comparability between programmes. Data from the general population survey were modelled using regression techniques to estimate preference-based scoring algorithms that can be applied to existing and future DEMQOL and DEMQOL-Proxy data to generate a utility score for each instrument.

However, there are important questions about whether or not the general population is the right valuation group in dementia given the complex nature of the disorder and the error often inherent in public attitudes and understandings of dementia. 11 In preparing the design of this study we engaged in conversation with the Alzheimer’s Society through its Quality Research in Dementia forum. It was clear that it wanted the values of people with dementia and their family carers to be investigated directly as part of this research. We therefore took this opportunity to complete a supplementary valuation survey of people with dementia and their carers to explore the size and direction of any deviation from the general public valuations.

In the final phase of the study we applied the general population algorithms to a trial data set that was completed during the project [HTA Study of Antidepressants for Depression in Dementia (HTA-SADD), a placebo-controlled RCT of the treatment of depression in dementia, Chief Investigator SB]46 to investigate the psychometric performance of the DEMQOL utility measures. We had hoped to do the same for the MRC-DOMINO trial data set [DOnepezil and Memantine IN mOderate to severe Alzheimer's Disease, a placebo-controlled RCT of donepezil and memantine (Ebixa^®, Lundbeck) alone and in combination for the treatment of patients whose treatment response is questioned, Chief Investigator RH]. However, the data set was not available for analysis by the end of this project and so we focused our analyses on the HTA-SADD data alone. This provided an opportunity to examine the psychometric properties of the indices in comparison with EQ-5D, the original DEMQOL measures and also other dementia-specific indicators.

Source of data

The data for the analyses shown in this chapter were obtained from the HTA-SADD study of the use of antidepressants for depression in dementia. Clinical measures available included the Cornell Scale for Depression in Dementia (CSDD),49 the MMSE,50 the Bristol Activity of Daily Living Scale (BADLS),51 the Neuropsychiatric Inventory (NPI),52 DEMQOL and DEMQOL-Proxy14,17 and the EQ-5D. 53

Statistical analyses

Summary statistics were used to describe the distribution of responses on the self-report EQ-5D, carer report EQ-5D (CEQ-5D), DEMQOL-U and DEMQOL-Proxy-U. Agreement between measures (EQ-5D/DEMQOL-U and CEQ-5D/DEMQOL-Proxy-U) was assessed. The utility values generated from self-report HRQL data were then compared with those generated from carer reports to examine the extent of agreement between the two sets of values. Construct validity was next examined in light of the absence of a gold standard for utility measurement in populations with cognitive impairment. Specifically, the construct validity of the EQ-5D, DEMQOL-U and their proxy equivalents was examined in terms of convergent validity to quantify the association between the utility values and measures of cognitive impairment (MMSE), depression in dementia (CSDD), neurobehavioural problems (NPI) and daily functioning (BADLS). Construct validity was also further examined using known group differences based on recommended thresholds for the MMSE54 and CSDD. 55 Responsiveness to change was examined using the minimal clinically important difference (MCID) thresholds recommended by the DOMINO trial group. 56

Outline of chapters in this report

The remainder of this report presents the methods used in more detail, the rationale for the methods used, the findings and the implications of the findings for the conduct of economic evaluation in this area. Chapter 3 presents a quantitative evaluation of the DEMQOL and DEMQOL-Proxy dimensional structure. This evidence was used to establish the dimensions used for the health-state classification systems. Chapter 4 presents the Rasch analyses completed to inform the selection of items used to construct the health-state classification systems. The general population valuation survey is presented in Chapter 5 along with the econometric analyses generating utility values for all states described by the classification systems. The comparison of patient and carer health-state values with those of the general population is described in Chapter 6. The DEMQOL-U and DEMQOL-Proxy-U indices are then evaluated in Chapter 7. Chapter 8 provides a brief discussion of the work and how the resultant preference-based measures can be used in economic evaluation.

Background to the DEMQOL system

The DEMQOL system14,17 is a measure of HRQL in dementia. It consists of two interviewer-administered instruments: DEMQOL (self-reported by the patient) and DEMQOL-Proxy (proxy reported by a carer). DEMQOL and DEMQOL-Proxy can be used for all types of dementia and were developed from a five-domain conceptual framework. This was based on in-depth qualitative interviews with people with dementia and their carers47 and includes health and well-being, cognitive functioning, social relationships, daily activities and self-concept. Items were drafted to represent each of the conceptual framework domains. The items developed were piloted with patients and carers and subsequently evaluated in two-stage field testing (item reduction followed by psychometric evaluation).

DEMQOL contains 28 items reported on a four-point Likert scale (a lot/quite a bit/a little/not at all); all items refer to the last week. A global quality-of-life item is also included but does not contribute to the overall score. Items are scored from 1 to 4, with higher scores indicating better HRQL. In the original psychometric evaluation there was some evidence of content validity (four of the original conceptual domains were represented in the item-reduced version). DEMQOL was also found to have high reliability (internal consistency and test–retest) and moderate validity (convergent and discriminant) in mild and moderate dementia. 14,17 In terms of acceptability, there was some evidence of missing data but floor and ceiling effects were not apparent. Factor analyses established a four-factor solution (defined as daily activities, memory, positive emotion and negative emotion), but the factor model was inconclusive and did not fully support the original conceptual framework.

DEMQOL-Proxy has 31 items reported on a four-point Likert scale (a lot/quite a bit/a little/not at all) and also includes an additional global quality-of-life item. The original psychometric evaluation found that DEMQOL-Proxy has good reliability (internal consistency and test–retest) and acceptable content validity (all five conceptual domains are represented in the item-reduced final version) and that there is some evidence for convergent and discriminant validity across the full range of severity. There is also evidence for acceptability in mild/moderate and severe dementia. Factor analysis suggested a two-factor solution (functioning and emotion), but this did not support the original conceptual framework.

DEMQOL-Proxy has 31 items reported on a four-point Likert scale (a lot/quite a bit/a little/not at all) and also includes an additional global quality-of-life item. The original psychometric evaluation found that DEMQOL-Proxy has good reliability (internal consistency and test–retest) and acceptable content validity (all five conceptual domains are represented in the item-reduced final version) and that there is some evidence for convergent and discriminant validity across the full range of severity. There is also evidence for acceptability in mild/moderate and severe dementia. Factor analysis suggested a two-factor solution (functioning and emotion), but this did not support the original conceptual framework.

Method

The current analysis used DEMQOL (n = 1189) and DEMQOL-Proxy (n = 1223) data drawn from two sources: routine data collected from a memory service48 and data collected from an unpublished study assessing HRQL in dementia. To be consistent with the original DEMQOL development, we excluded those without a definite diagnosis of dementia (n = 451) and also those with a MMSE score < 10, indicating severe memory problems (n = 80). Although DEMQOL-Proxy can be used for those with severe dementia, data from carers of patients with severe dementia were also excluded from the analysis so that the patient and proxy samples used for the development of the preference-based measures were consistent in terms of diagnosis and severity. Analyses reported below on non-imputed data are based on this sample of 658 for DEMQOL and 692 for DEMQOL-Proxy.

For the factor analyses we used imputed data. We therefore excluded those for whom imputation (using the standard rule outlined in the DEMQOL scoring) could not be undertaken (DEMQOL: n = 14; DEMQOL-Proxy: n = 10). The final sample for factor analysis was therefore 644 for DEMQOL and 682 for DEMQOL-Proxy. The demographic characteristics of the sample used for the factor analyses are shown in Table 1.

First, we conducted item analysis on non-imputed data. To do this we evaluated rates of missing data, maximum endorsement frequencies (MEFs), adjacent endorsement frequencies (AEFs) and redundancy using the criteria outlined below. Next, we imputed responses for missing data (using the imputation rule specified in the standard scoring instructions for DEMQOL) to evaluate the factor structure of DEMQOL and DEMQOL-Proxy and to determine the extent to which the conceptual domains are supported. The criteria for item analyses are based on well-established techniques and also our own previous psychometric work. 14,57–61

Results

Discussion

This chapter provides an evaluation of each of the items and also the dimension structure of DEMQOL and DEMQOL-Proxy to determine the extent to which the original conceptual framework is supported by quantitative data. The analysis replicates part of the original development work14 for DEMQOL and DEMQOL-Proxy using a large, clinically representative sample. This factor analysis provides the basis for the dimensions represented in the dementia-specific preference-based measures developed.

We have identified two separate five-factor models for DEMQOL and DEMQOL-Proxy. Both models reflect aspects of the original conceptual framework but also highlight important differences between self- and proxy reports. The potential for one dimension (self-concept) to be represented was limited by the particular items that were retained in the final versions of DEMQOL and DEMQOL-Proxy. The four original domains (daily activities, health and well-being, cognitive functioning, social relationships) that were operationalised in the questionnaire were represented in the model for either DEMQOL or DEMQOL-Proxy, but neither model supported all four of these original domains. This supports using DEMQOL and DEMQOL-Proxy as complementary measures in trials and research settings. For example, the domain of health and well-being split into two factors (representing positive and negative emotion) for both DEMQOL and DEMQOL-Proxy. The domain of daily activities was more strongly supported in DEMQOL-Proxy than in DEMQOL, probably reflecting the seven items retained for this domain in the final DEMQOL-Proxy compared with two items in DEMQOL. Similarly, social relationships is evident as a factor in DEMQOL but not in DEMQOL-Proxy, reflecting the five items retained for this domain in DEMQOL compared with two items in DEMQOL-Proxy. However, the emphasis on daily activities and looking after yourself in DEMQOL-Proxy and social relationships in DEMQOL is also consistent with the differences found between self- and proxy reports in previous qualitative work. 14 In addition, the factor analyses reported here suggest that in self-reported data (DEMQOL) social relationships consists of two separate parts (which we have labelled ‘social relationships’ and ‘loneliness’) and in proxy-reported data (DEMQOL-Proxy) daily activities and looking after yourself consists of two separate parts (which we have labelled ‘daily activities’ and ‘appearances’).

The factor structure reported here better supports the original conceptual domains than the earlier factor analysis conducted during the development of the DEMQOL system. This more positive result is possibly due to the much larger sample used for this study (approximately six times larger than that used for the original validation).

In the development of the classification systems we used EFA rather than confirmatory factor analysis (CFA) to investigate the dimensionality of DEMQOL and DEMQOL-Proxy. We recognise that CFA would also be a valid way to investigate the dimensionality of the DEMQOL system. However, we used EFA because the factor analysis carried out during the development of the DEMQOL system on a much smaller sample was inconclusive and did not clearly match the original conceptual framework, and therefore we did not have an established a priori factor structure to confirm. There are some features of the EFA method used that may impact on the dimension structure established for DEMQOL and DEMQOL-Proxy. First, factors with more items are more likely to be clearly defined using EFA, and this has implications for the item loadings of the smaller factors both within and across factors. Second, factors are based around inter-item correlations and so a minimum of two items is required to generate a factor. However, factors with only two items are not strong, and five or more strongly loading items are recommended. 63 In this study we used EFA as a guide to the dimensionality and after running a range of models selected the five-factor models that fitted the data conceptually and did not contain a high number of non- or cross-loading items. The DEMQOL and DEMQOL-Proxy models both contain a factor with two items but we believe that the items in each of these factors can be clearly defined as a dementia-specific HRQL concept. Also, as the aim is to develop a multiattribute instrument that retains many of the key HRQL concepts from the original instrument, we believe that using the factor model including only strong factors with five or more items would have resulted in a classification system with reduced sensitivity, which has implications for the preference weights derived from the valuation stage.

The factor analysis used orthogonal rotation, which assumes independence between the factors extracted. It could be argued that, as the instruments are measuring constructs relating to dementia, the factors are related, and therefore we should use oblique rotation, which takes into account the relationship between factors. For the DEMQOL system, the factor structures produced using both rotation methods were similar and so orthogonal was used. For both DEMQOL and DEMQOL-Proxy there were items that loaded on more than one factor. However, as we use factor analysis to also exclude items from consideration for the reduced classification system, we followed guidelines used in the development of DEMQOL to exclude items (i.e. items loading within 0.2 of each other across factors or < 0.4 on any factor).

There are a number of other methods that we could use to investigate the factor structure of the DEMQOL system, including the use of a polychoric matrix or IRT techniques such as Rasch analysis. Using a multiple-factor IRT approach within the six-step guide to developing a condition-specific preference-based measure would mean that both the investigation of dimensionality (step I) and item elimination and selection (steps II–IV) could be based around the same underlying latent model. Further research may assess differences in the factor structures produced using IRT and traditional factor analysis to inform the use of both processes in the development of condition-specific health-state classification systems.

The item analyses revealed that all of the DEMQOL items were robust in terms of missing data and redundancy, although 10 items failed the AEF criterion. There are 17 items in DEMQOL-Proxy that failed at least one of the item analysis criteria. This does not necessarily mean that they are uninformative items, and indeed all items passed all of the criteria in the original development study. These classical item analyses are considered alongside the Rasch analyses in Chapter 3 to inform the final selection of items for the system.

Conclusion

The factor structures established here provide support for dimensions that are similar but not identical to the domains of the original conceptual framework. The work reported here has not evaluated whether or not the dimensions can be used as scores alongside the overall scores for DEMQOL and DEMQOL-Proxy. This would require a prospective study including carefully chosen validating measures. The factor structures are robust enough to provide the basis for the development of dementia-specific preference-based measures for patient self-report and proxy report by carers.

Introduction

This chapter presents stages II and III of the development process described previously (see Figure 1). The aim of the analysis was to identify one item to represent each dimension in the health-state classification system. This involved the application of Rasch methods to evaluate the performance of each of the items within each dimension.

Rasch analysis is one of a number of IRT techniques that could be used to assess the performance of items included in a range of questionnaires and tests, including measures of HRQL. Rasch converts ordinal or categorical responses to items into a continuous unidimensional latent scale. This is done using logit modelling. When items fitted to the scale assess HRQL, the latent scale represents a continuous measure of quality of life covering the full severity range for the particular HRQL construct being measured. Respondents are also modelled on the logit scale, and an individual's position on the scale represents their level of severity in terms of the construct being measured. Responses to items are assumed to be a function of both the position of the person and the item on the overall latent scale. This means that respondents at the more severe end of the logit scale should be more likely to indicate more severe problems on the items included in the dimension than those at the lower end of the scale. The Rasch model for each dimension allows for an assessment of item performance at the overall dimension and individual response category levels. Therefore, Rasch can be used to inform the selection of items from existing condition-specific measures of HRQL to generate a condition-specific health-state classification system. 40,64 A range of Rasch statistics are used to first eliminate poorly performing items and subsequently select items for each dimension. The tests used are outlined below.

Method

To identify the most robust items to use in the health-state classification system, Rasch models were fitted separately to each of the dimensions established by the factor analysis at stage I (Tables 4 and 5). The Rasch model assumes unidimensionality and, as DEMQOL and DEMQOL-Proxy are not unidimensional (see Chapter 3), it would not be appropriate to fit a single Rasch model encompassing all DEMQOL/DEMQOL-Proxy items. Five separate Rasch models were therefore generated for both DEMQOL and DEMQOL-Proxy. Stage II of the development process described in this chapter includes two parts. First, we eliminated items across each dimension by assessing the performance of the items included in each model. We assessed item response level ordering, DIF and goodness of fit of items to the Rasch model. Next, we used Rasch criteria to select one item for each dimension. This included the range of item responses on the logit scale and the spread of item responses at logit 0 (the average item difficulty on the severity scale). Each of these processes is described in detail below. Rasch analysis was carried out using Rasch Unidimensional Measurement Models (RUMM2020^©, 1997–2004 RUMM Laboratory Pty Ltd, www.rummlab.com.au/).

Stage II: selecting items

Item elimination

Item-level ordering

For each model we evaluated the ordering of the responses to each individual item. Item responses are disordered if respondents cannot differentiate between the response choices and therefore the observed response is not in line with the expected response (Figure 2). This can be assessed as item responses are mapped across the logit or severity scale and should be endorsed by respondents who display similar amounts of severity of the underlying construct. If respondents can distinguish between the response choices included on an instrument, responses to the items will be ordered as the observed response will be in line with the expected response at any point of the severity scale. If disordering occurs (i.e. the observed response is different to the expected response at that point of the logit scale), adjacent response levels are collapsed to artificially impose ordering, and the Rasch model is reapplied (Figure 3). Items for which disordering occurs are not considered for inclusion in the health-state classification system as it is important that respondents can distinguish between the different levels of the health dimensions used for the preference-based measure. Although items are excluded from further consideration for use as part of the health-state classification system they remain in the Rasch model to allow for the whole dimension to be modelled.

FIGURE 2.

Item with disordered response options.

FIGURE 3.

Item (Figure 2) with reordered response options.

Examination of differential item functioning

We investigated each item for DIF. DIF occurs when item responses systematically differ according to a range of patient characteristics when equal amounts of the underlying characteristic are present (at different points of the latent scale). In this study the characteristics analysed were gender and age (split into two groups:  < 65 and ≥ 65 years). We considered two categories of DIF – uniform and non-uniform. Uniform DIF occurs when one of the subgroups belonging to each characteristic consistently exhibits a difference in response across the full severity range of the construct being assessed by the item. For example, women may consistently display higher levels of depression as assessed by an individual item than men, and this is the case across the full severity scale of the underlying construct, from mild to severe. Non-uniform DIF occurs when responses between characteristic subgroups systematically diverge depending on the level of the attribute present. For example, responses between genders to a depression item may systematically differ depending on the underlying severity. Items for which either form of DIF is present are split into component factors (e.g. male and female subgroups) and the Rasch model is reapplied with each subgroup included separately. Items displaying DIF are not considered for the health-state classification as systematic differences in responses to the items across person characteristics compromises their ability to compare across subgroups.

Examination of goodness of fit

We evaluated the goodness of fit to the Rasch model of individual items included in each factor. Goodness of fit is assessed by analysing fit residuals and item–trait interactions. Fit residuals assess the discrepancy between the expected and observed responses at both the respondent and the item level. Divergence residuals > l2.5l were considered high. Respondents outside these boundaries were also removed from the analysis and the model refitted. When all of the misfitting respondents had been removed, the fit of items was assessed. Items with residuals above the minimum level were excluded from the descriptive system. The overall mean fit residual for each dimension should be approximately 0 and the SD should be approximately 1.

Item–trait interactions measure overall differences between observed and expected responses for subgroups of responders (which are grouped dependent on where responders lie on the logit scale). The chi-squared test statistic (which is > 0.01 for a well-fitting model, i.e. non-significant) was used to assess item–trait interactions. Items with the highest level of significance (i.e. the largest divergence between the observed and expected responses) were removed one by one, with the Rasch model reapplied after every item was removed. This process continued until only items that fitted the model remained and the overall goodness of fit statistic was non-significant.

Item selection

To select one item per dimension two main criteria were used. We assessed the range of the item on the logit scale and the spread of the item at logit 0. A large range indicates that an item covers the full severity range of the underlying construct. Items were selected that incorporate both positive and negative values. This indicates that the item is sensitive to responses for both more severe and less severe respondents. In Rasch analysis, the latent scale is centred at zero. This point is the average difficulty on the latent scale relating to a particular item included in the final dimension model. Item levels should be distributed across the latent space depending on the severity of the response level. Therefore, a large spread of levels at logit 0 indicates items for which respondents endorse the full range of possible responses at the average severity or difficulty level. Item goodness of fit statistics and classical psychometric analyses, including MEFs, AEFs and missing data rates (alongside input from clinicians, dementia experts and the original instrument developers), were also used to guide the item selection process.

Results

Discussion

This chapter describes the development of condition-specific descriptive systems for dementia for patient self-report (DEMQOL-U) and carer proxy report (DEMQOL-Proxy-U). We selected items to represent all of the five DEMQOL dimensions and four of the five DEMQOL-Proxy dimensions established during stage I of the development process. The methodology used builds on previous work that has applied Rasch analysis to non-preference-based condition-specific instruments to develop condition-specific classification systems that are amenable to valuation. 64,67 This is the first part of the process in developing a condition-specific preference-based measure. The second part is to obtain preference weights so that the measure can be used in the economic evaluation of interventions for dementia. This stage is described in Chapter 5.

This chapter also describes the first attempt to derive a condition-specific health-state classification system specifically for proxy report by carers. The descriptive systems reflect both similarities and differences in the factors that are the key focus for patients and carers in terms of evaluating HRQL in dementia. For example, both measures included cognition and emotion dimensions and this reflects the importance of cognitive functioning and mood both for the person with dementia and for those involved in their care. There are discrepancies between patient and proxy report of HRQL in dementia68 and some evidence suggests that agreement is lower for more subjective domains such as emotional well-being. 69 The development of a proxy-specific measure in which all of the dimensions are meaningful to carers may address some of the concerns about using generic measures for proxy report by carers in dementia.

The results of the Rasch analysis used to select the items for the descriptive system have not been validated on an external sample and this is a limitation of the development process described in this chapter. It was not possible to carry out validation analyses as the sample size was not sufficient to randomly allocate responses to two subgroups. Furthermore, the data used in Chapter 7 of this report to assess the psychometric performance of DEMQOL-U and DEMQOL-Proxy-U were not available to validate the descriptive system. The DEMQOL system is being used in a number of studies and so further work validating the classification systems may be possible in the future.

It is also possible to carry out factor analysis using IRT-based techniques and this would enable us to combine stage I of the development process with stages II and III and base the full classification system development on one underlying model. This approach and the differences in factor structure that might lead from it need further investigation. Indeed, other IRT techniques could be used to select items for the classification system and further research could investigate differences in the items selected using a range of techniques and determine which technique produces the classification system that retains the most information from the original HRQL measure.

There are also concerns around the use of condition-specific preference-based measures, including the extent to which they capture comorbidities. 39 This issue may be addressed by investigating the performance of condition-specific preference-based measures in relation to generic measures such as the EQ-5D in trials and settings where both instruments have been used together. Both DEMQOL-U and DEMQOL-Proxy-U cover a broad range of HRQL issues in dementia. The absence of an activity limitation dimension may be a concern; this was because the Rasch analysis found that the original daily activities items did not meet the minimum threshold for inclusion. Another important limitation is that it has not been possible to consider item responsiveness during the development of the descriptive system. The ability of items to detect change over time is an important psychometric characteristic and so responsiveness needs to be addressed in future research that includes DEMQOL and DEMQOL-Proxy as outcome measures at multiple time points.

It could also be argued that the resultant classification systems reflect a mild description of dementia. This is a result of both the development process (in which the items that performed well reflected mild aspects of HRQL in dementia) and also the sample used, which was patients with a mild or moderate diagnosis. Future work could investigate qualitatively the acceptability and characteristics of the classification system with independent clinical experts and patient groups.

There are a number of strengths of the process described in this chapter. The descriptive systems were based on representative samples from memory and community services where many of the clients have mild to moderate dementia. Although it is widely acknowledged that people with dementia are often able to self-report it is often also necessary to use a proxy report. 31 Administering measures in an interview setting with response cards (as is done with the DEMQOL system) may help to maximise the reliability and validity of such reports. The data also included demographic information that enabled us to investigate DIF characteristics and this helped to strengthen the item selection process and provide a set of descriptive systems that cover as broad a range of HRQL issues in dementia based on the non-preference-based instruments as possible.

Conclusion

In conclusion, item selection procedures have identified robust items to represent all five of the DEMQOL dimensions and four of the five DEMQOL-Proxy dimensions. Using validated measures (DEMQOL and DEMQOL-Proxy) along with a representative sample of patients and carers we have created a classification system that is an appropriate representation of HRQL in dementia. The descriptive systems are amenable to valuation and the next part of the process in developing a condition-specific preference-based measure is to obtain preference weights so that the measure can be used in the economic evaluation of interventions in dementia.

Introduction

Cost–utility analysis measures the benefits of treatment using the QALY, which captures changes in both quantity of life and HRQL. The ‘Q’ quality adjustment weight is typically derived using a preference-based measure. A preference-based measure has two components: first, a health-state classification system that can be used to categorise all patients with the condition of interest; second, a means of obtaining a utility score for all states defined by the system. This chapter reports on work to estimate a preference-based single index for each classification system. First, this chapter reports the valuation study in which a representative sample of the general population valued a sample of health states derived from each classification system using ranking and the TTO elicitation technique. Second, the chapter reports on modelling using the valuation results to produce utility values for all health states described by each classification system. This analysis uses a range of multivariate regression models to produce a single-index measure from each classification system anchored on a full health–dead 1–0 scale, in which a value of 1 is equal to full health and 0 is equal to being dead. Values were obtained from the general population in accordance with recommendations of agencies such as NICE43,54 and the Washington Panel70 for use in economic evaluation.

Method

Health-state descriptions for the valuation study were generated using the DEMQOL-U and DEMQOL-Proxy-U classification systems described in Chapter 4. Table 17 presents an overview of the two classification systems. Each health state is generated using one level of each dimension. Box 1 includes an example health state for each classification system.

The DEMQOL-U health-state classification describes 1024 health states and the DEMQOL-Proxy-U classification describes 256 health states. The large number of health states means that it is impractical to value every state and therefore only a sample of states were selected for valuation. Respondents cannot value a large number of health states because of the nature of the preference elicitation task. In accordance with many previous valuation studies,37,40,42,67,70–72 each respondent valued eight health states. The study design needs to determine both the total sample of all health states to be included in the valuation study and the combinations of eight health states to be valued by respondents. To get more precise estimates of the worst state and to have a common state valued by all respondents, each combination of health states for valuation included the worst state plus seven other states. Previous valuation surveys have used a variety of methods to select health states for valuation, such as an orthogonal array,37,71 balanced design40,42 and the Rasch vignette approach. 66,73 None of these approaches selects the combinations of health states to be valued by respondents, and only the Rasch vignette approach avoids implausible health states. Here, a new approach was used to select health states, which offered the following advantages over existing approaches: (1) it selected the combinations of health states to be valued by respondents; (2) it avoided implausible health states; and (3) it selected the health states that produce the most accurate modelled utility estimates using simulation. The last advantage is gained because the process involves simulation of alternative selections of health states that could be chosen and hypothetical values for those health states (given some assumptions about how people value health states). The hypothetical values were modelled using similar modelling to the analysis that was applied to the actual utility values in the general population valuation survey. The accuracy of the model predictions for each of the alternative selections of health states was compared using predictions for health states that would not be valued. The best performing selection of health states was chosen. For each selection of health states these are separated into combinations of health states for valuation. Each combination is called a block and each block consisted of seven health states of different severity plus the worst state. The steps used in this process are explained further below; the process was undertaken for each classification system.

The selection of health states can involve a different number of blocks, for example five blocks of seven health states plus the worst state, or six blocks of seven health states plus the worst state, or seven blocks, etc. Furthermore, within each block design, alternative health states could be selected, for example there are many different combinations of health states that could be selected within the five-block design, within the six-block design, etc. Therefore, the first step in the process derived 100 approximate optimal Federov designs74 for each block design of 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 blocks (i.e. 100 designs for five blocks, 100 designs for six blocks, etc.). For each block size the optimal design was selected using the D-criterion and the worst state was added to each block. There were 11 designs remaining for inclusion in step 2, one for each block design.

Assuming that the sample size of the valuation study will remain the same, this means that each health state in the 15-block design will be valued fewer times than health states in the 14-block design. It is expected that each block design will involve different health states as well as different numbers of health states in total (e.g. the five-block design will have 35 health states and the six-block design will have 42 health states, assuming that there is no repetition in health states). All of these factors may impact on the reliability and robustness of the modelled utility values. Therefore, the second step in the process involved the simulation of hypothetical utility values for each of the block designs selected in step 1. Utility values were simulated for each block design for 50 sets of 300 individuals using utility functions. The utility function assumed a logarithmic transformation between ‘severity’ and utility in which severity of a health state was generated using a linear function of the dimensions with equal weights for each dimension and equidistance between levels with random measurement errors. As respondents do not always trade life-years to avoid mild health states, non-trading behaviour was simulated using a logistic model of severity [such that approximately 10% of the scores would be 1 to represent non-traders (proportion selected using previous valuation studies)].

The third step in the process involved the estimation of a linear regression model for each block design using the appropriate simulated utility data set. The regression model was used to produce predictions of health-state utility values for all health states not selected in the block design. These predictions were compared with the ‘true values’, the hypothetical values that were generated using the same utility functions explained above, using the sum of squared residuals and average mean sum of squares.

Fourth, the block design was selected that achieved the best predictions of health-state utility values for health states outside the set of health states to be valued (i.e. all health states that were not selected in that block design).

The selected designs involved a sample of 87 unique states across 13 blocks for DEMQOL-U and a sample of 70 unique states across 12 blocks for DEMQOL-Proxy-U. For each measure one block was selected to be valued by a minimum of 71 individuals to enable mean values for these states to be statistically compared with mean values elicited from people with dementia and carers of patients with dementia as reported in Chapter 6.

Valuation study

Sample

Each classification system was valued by a representative sample of the general population. This is in accordance with recommendations by agencies such as NICE43,54 and the Washington Panel,70 who recommend the use of general population values to produce QALY estimates for economic evaluation. The sample size for each measure was selected on the basis of previous valuation studies for similar measures [e.g. for the cancer-specific EORTC-8D (European Organisation for Research and Treatment of Cancer Core Quality of Life Questionnaire) a sample of 344 respondents valued 85 states; for the asthma-specific AQL-5D (Asthma Quality of Life Utility Index) a sample of 307 respondents valued 99 states; and for the overactive bladder-specific OAB-5D (Overactive Bladder Questionnaire-5 Dimensions) a sample of 312 respondents valued 99 states]. The sample size was selected as 310 for DEMQOL-U and 290 for DEMQOL-Proxy-U. The sample size for DEMQOL-U was larger as the classification system describes more states and had a larger selected study design.

The sample was obtained by sampling 600 households in urban and rural areas in northern England using the AFD Names and Numbers version 3.1.25 database (AFD Software Limited, Ramsey, UK). The sample was balanced to the UK population according to geodemographic profiles. Trained and experienced interviewers conducted home interviews in which respondents valued states from one of the classification systems determined using a card block system. Interviewers worked systematically through blocks; odd and even blocks contained DEMQOL-U and DEMQOL-Proxy-U states respectively. This approach was used to try and ensure that there were no systematic differences across the geodemographic profiles of the samples for each classification system.

Procedure

At the start of the interview respondents first self-completed the EQ-5D to report their own health and then whichever classification system would be used subsequently in the valuation task (DEMQOL-U or DEMQOL-Proxy-U). Respondents were asked to complete the questions for their own health and this was done to help familiarise them with the classification system. Respondents were not informed that the health state was associated with dementia as there are concerns that naming the condition can affect elicited utility values. A recent study found that introducing condition labels into health-state descriptions impacted on health-state utility values elicited from the general population. 37,71 The impact differed by condition and health-state severity but as the study was quantitative the reasoning behind why these differences occurred was not determined. Further qualitative research is recommended to examine why these differences occurred and until these findings are available it is recommended that condition labels are not included in health-state descriptions to ensure that elicited utility values are not affected by factors that should not affect the quality of life of health states.

Second, respondents ranked eight health states, as well as full health and dead, in order from the best to the worst. This rank task further familiarised respondents with the classification system and the health states to be valued later using TTO.

Third, respondents undertook a practice TTO task using a hypothetical ‘practice’ state and subsequently valued the (ranked) eight health states using TTO. The intuition underlying the TTO task is that it is designed to determine whether or not respondents are willing to sacrifice time to avoid living in the dementia health state. Essentially this is undertaken by asking respondents whether they would rather live in the dementia health state for 10 years or full health for a shorter time duration than 10 years. If respondents believe that the health state is severe, they should be willing to sacrifice a larger time duration than if they believe that the health state is mild. For some respondents they may not be willing to sacrifice any time at all to avoid the dementia health state. Some respondents may believe that some health states are so severe that they would rather die than live in that health state.

The TTO task for each health state asks first whether respondents prefer to live in a given health state for 10 years after which they will die, or to die immediately. This determines whether the respondent values the health state as better, worse or equal to being dead. For health states considered better than dead, respondents choose between (a) the health state for 10 years, after which they will die, or (b) full health for x years (x ≤ 10), after which they will die. Years in full health, x, is varied until respondents are indifferent between the two options. For health states considered worse than being dead, respondents choose between (a) the health state for w years followed by full health for x years, after which they will die, or (b) immediate death. Both years in full health, x, and years in the health state, w = 10 – x, are varied until respondents are indifferent between the two options. Utility values are generated using the formula x/10 for states better than being dead and –x/10 for states worse than being dead. 45 A sample TTO scripting for one TTO task is included in Appendix 1. The Measurement and Valuation of Health (MVH) study version of TTO was used, including a visual prop designed by the MVH group (University of York),45 which has been reproduced in Figure 4 (although note that in the MVH study respondents valued more than eight health states). All TTO questions follow an iterative procedure to determine the point of indifference, with the task completed only when the respondent switches his or her choice between (a) and (b) or reaches the point at which he or she is indifferent between (a) and (b). The TTO elicitation technique was selected in accordance with the UK valuation of the EQ-5D,45 which also meets the reference case recommended by NICE54 for use in economic evaluation.

FIGURE 4.

Time trade-off visual prop [MVH group (4)]. Progeny of the University of Sheffield – 58 © Copyright MVH Group University of York, 1992.

Finally, respondents rated how difficult they found the rank and TTO tasks and answered questions about sociodemographic characteristics and health service use.

Before conducting the full survey a small number of interviews were conducted and the interviewer discussed the findings with one of the research team. As no problems were identified with the survey, the main survey was started without amendment and these interviews were included in the main sample. The survey was approved by the School of Health and Related Research (ScHARR) Research Ethics Committee at the University of Sheffield.

Results

Time trade-off descriptive statistics

Tables 19 and 20 report descriptive statistics of observed TTO values for DEMQOL-U and DEMQOL-Proxy-U respectively. The number of observations per intermediate health state varied from 18 to 78, with 306 and 287, respectively, for the worst DEMQOL-U and DEMQOL-Proxy-U states. The range of mean values was larger for DEMQOL-U (from 0.954 to 0.184) than for DEMQOL-Proxy-U (from 0.961 to 0.331), although only two DEMQOL-U states had mean values outside the DEMQOL-Proxy-U range. Each classification system had one or more states with a mean value lower than that of the worst state defined by the classification system. In addition, DEMQOL-Proxy-U had two states with a mean value higher than that of the best state. These apparent contradictions are most likely observed because of the much smaller number of observations for some states in comparison with worst state and best state. Figures 5 and 6 present the distribution of observed TTO values for DEMQOL-U and DEMQOL-Proxy-U respectively. There were a large proportion of TTO values at 1 for both measures (26.9% for DEMQOL-U and 28.8% for DEMQOL-Proxy-U) and the distribution of the data was negatively skewed for both measures.

TABLE 19 - Descriptive statistics of observed TTO values for DEMQOL-U

Health state	Mean	SD	Median	Count
11111	0.954	0.153	1.000	78
14231	0.934	0.093	0.975	19
11231	0.930	0.139	1.000	20
11312	0.918	0.098	0.950	19
13221	0.913	0.105	0.950	18
14311	0.909	0.134	0.950	19
12241	0.903	0.153	1.000	19
12231	0.894	0.160	1.000	78
14431	0.888	0.140	0.950	19
24231	0.860	0.182	0.963	18
13432	0.858	0.187	0.938	18
23411	0.845	0.225	1.000	20
12323	0.845	0.159	0.875	19
21241	0.845	0.253	0.998	19
21143	0.836	0.170	0.875	18
21213	0.835	0.230	0.913	18
41111	0.822	0.228	0.925	19
21233	0.819	0.191	0.875	19
22123	0.818	0.204	0.925	19
12432	0.818	0.278	0.925	18
31331	0.817	0.235	1.000	19
34221	0.812	0.181	0.888	20
12123	0.804	0.118	0.813	20
14233	0.801	0.226	0.850	18
13142	0.800	0.199	0.863	18
32131	0.799	0.251	0.900	19
23413	0.796	0.219	0.813	20
12342	0.790	0.228	0.813	20
32112	0.789	0.217	0.875	19
33431	0.789	0.265	0.925	19
21322	0.786	0.261	0.925	19
23214	0.784	0.271	0.925	37
34123	0.782	0.289	0.925	19
21132	0.773	0.306	0.850	20
14223	0.770	0.221	0.850	19
12243	0.768	0.255	0.800	18
32441	0.767	0.212	0.813	18
23441	0.754	0.240	0.775	19
21342	0.753	0.250	0.800	19
34142	0.751	0.345	0.925	18
14133	0.742	0.433	0.925	19
13242	0.741	0.264	0.788	20
22414	0.732	0.222	0.688	18
32142	0.730	0.280	0.775	19
44322	0.728	0.251	0.700	19
34132	0.728	0.276	0.813	20
34313	0.721	0.315	0.813	18
23314	0.685	0.353	0.725	39
14323	0.685	0.296	0.650	20
24313	0.680	0.301	0.750	19
31433	0.675	0.340	0.775	18
32143	0.673	0.302	0.663	78
13413	0.670	0.474	0.800	19
32314	0.661	0.352	0.800	19
33421	0.661	0.313	0.625	19
31413	0.655	0.216	0.650	20
24144	0.654	0.335	0.675	19
41233	0.640	0.329	0.575	18
43122	0.629	0.299	0.675	19
43231	0.621	0.257	0.500	19
34423	0.617	0.225	0.600	19
41212	0.611	0.364	0.625	78
23344	0.609	0.362	0.675	19
22334	0.605	0.340	0.625	20
42243	0.592	0.357	0.625	19
43321	0.591	0.387	0.663	56
41241	0.589	0.480	0.725	18
31444	0.587	0.399	0.525	19
42321	0.576	0.300	0.513	20
41231	0.573	0.397	0.675	39
24434	0.568	0.432	0.675	19
23424	0.566	0.340	0.600	78
43123	0.548	0.400	0.588	20
44221	0.544	0.367	0.525	78
41423	0.534	0.482	0.500	19
44341	0.528	0.356	0.550	20
34234	0.525	0.486	0.600	18
43432	0.462	0.525	0.600	19
43434	0.449	0.429	0.463	18
44242	0.418	0.549	0.500	19
41314	0.417	0.477	0.475	19
43442	0.403	0.412	0.475	78
43344	0.397	0.422	0.500	19
43343	0.364	0.561	0.500	20
41224	0.339	0.542	0.425	20
44444	0.234	0.501	0.200	306
43244	0.184	0.475	0.200	20

TABLE 20 - Descriptive statistics of observed TTO values for DEMQOL-Proxy-U

Health state	Mean	SD	Median	Count
1132	0.961	0.062	1.000	19
2221	0.944	0.108	1.000	20
1111	0.918	0.185	1.000	77
2141	0.918	0.172	1.000	20
2421	0.896	0.151	1.000	39
1143	0.893	0.149	0.963	20
1141	0.892	0.153	1.000	19
2312	0.884	0.154	1.000	19
1131	0.869	0.218	1.000	18
1222	0.869	0.174	0.925	77
3121	0.855	0.213	1.000	19
2212	0.846	0.248	0.950	20
3311	0.846	0.181	0.925	19
1214	0.839	0.227	0.925	38
1333	0.833	0.191	0.863	20
1231	0.830	0.267	0.950	19
2131	0.821	0.241	0.888	18
1233	0.821	0.194	0.925	19
1341	0.806	0.236	0.900	77
3112	0.804	0.255	0.925	77
1223	0.799	0.256	0.888	20
3221	0.796	0.224	0.900	19
1422	0.791	0.308	0.900	38
1322	0.789	0.190	0.875	19
2431	0.763	0.396	0.900	19
3331	0.763	0.221	0.775	19
1432	0.751	0.269	0.825	20
2422	0.750	0.349	0.825	19
1313	0.721	0.277	0.800	38
3432	0.719	0.263	0.775	18
2342	0.716	0.304	0.850	18
3341	0.714	0.309	0.788	20
3431	0.705	0.294	0.800	19
1334	0.700	0.256	0.700	19
3342	0.699	0.287	0.750	38
4111	0.697	0.326	0.775	19
2234	0.692	0.279	0.675	19
1324	0.688	0.398	0.775	19
2243	0.684	0.350	0.788	38
3224	0.684	0.305	0.775	58
1423	0.682	0.341	0.800	19
2224	0.676	0.259	0.725	19
3312	0.674	0.256	0.700	19
3133	0.664	0.239	0.575	19
2424	0.662	0.342	0.725	97
4212	0.655	0.412	0.800	40
3234	0.654	0.287	0.675	95
4322	0.645	0.331	0.638	20
3344	0.634	0.370	0.675	20
4123	0.629	0.271	0.625	19
3144	0.625	0.339	0.663	20
4141	0.614	0.310	0.725	19
4342	0.603	0.468	0.788	20
3424	0.596	0.357	0.538	20
2344	0.593	0.410	0.675	19
3443	0.589	0.288	0.588	38
2214	0.588	0.336	0.675	18
4411	0.580	0.395	0.625	77
4242	0.578	0.289	0.500	19
2423	0.552	0.375	0.488	18
2334	0.533	0.352	0.638	18
4133	0.532	0.330	0.400	18
4214	0.505	0.309	0.425	19
3314	0.484	0.356	0.350	19
4311	0.478	0.421	0.500	36
4333	0.476	0.378	0.488	58
4434	0.434	0.477	0.500	19
4443	0.383	0.458	0.425	18
4444	0.357	0.439	0.325	287
4343	0.331	0.492	0.450	18

FIGURE 5.

Histogram of observed TTO values for DEMQOL-U.

FIGURE 6.

Histogram of observed TTO values for DEMQOL-Proxy-U.

Modelled time trade-off health-state utility values

The coefficients of the main effects variables indicate the utility decrement for every severity level of every dimension of the classification system using level 1 as the baseline. For coefficients to be consistent they must be positive and increasing in size for each subsequent severity level within each dimension. For the DEMQOL-U models all coefficients were positive as expected. The standard OLS models estimated using pooled observations and mean level data [models (1) and (3)] had inconsistent coefficients for negative emotion at levels 2 and 3, but the random-effects GLS model [model (2)] had consistent coefficients. Model (2) was preferred to standard OLS estimated using pooled observations [model (1)] as it takes into account the structure of the data with repeated observations per respondent. The mean model [model (3)] does not suffer from the theoretical bias occurring when estimating models on data containing a large proportion of values at 1, yet performed similarly to model (2), with a large number of significant coefficients and good predictive ability.

Model (2), the random-effects GLS model, was selected as the preferred model because of its strong relative performance. Model (2) had a high number of significant variables, no inconsistent coefficient estimates and good predictive ability with low RMSE and the lowest percentage of AEs at the state level > 0.05 and > 0.10. Figure 7 plots observed and predicted TTO values at the health-state level for model (2). The figure shows that there was no systematic bias in the predictions by severity, but some health states had large prediction errors. The range of predicted values for this model was from 0.986 (best state 11111) to 0.243 (worst state 44444), which was similar to the observed range for these states (from 0.954 to 0.234).

FIGURE 7.

Observed and predicted TTO for DEMQOL-U using model (2).

For DEMQOL-Proxy-U, all models have at least one inconsistent coefficient and at least one negative coefficient for the appearance dimension. These inconsistencies mean that health worsens but predicted utility increases, for example deterioration in appearance from level 1 to level 2 leads to a higher predicted utility value. Other valuation studies in which this has occurred have resolved this issue by merging adjacent inconsistent levels and re-estimating the model to produce a consistent model [see model (10) for example]. This approach was used to estimate models (8) and (10), consistent versions of models (7) and (9), respectively, in which adjacent levels of the appearance dimension have been merged.

For DEMQOL-Proxy-U, the standard OLS model estimated using pooled observations [model (6)], the consistent random-effects GLS model [model (8)] and the consistent mean model [model (10)] performed similarly in terms of predictive ability. As explained above, the consistent version of the random-effects GLS model [model (8)] and the mean-level model [model (10)] are preferred on a theoretical basis to model (6). Both models suffer from autocorrelation between errors to some degree using the LB test, but this is significant only for model (8) (at the 10% level). Despite this finding, Figure 8 demonstrates little difference in the pattern of the predictions by severity for these models. Therefore, model (8) was selected as the preferred model because of its higher number of significant coefficients. The range of predicted values for this model was from 0.937 (best state 1111) to 0.363 (worst state 4444), which was similar to the observed range for these states (from 0.918 to 0.357).

FIGURE 8.

Observed and predicted TTO for DEMQOL-PROXY-U using models (8) and (10).

Discussion

In this chapter we have reported on the estimation of preference weights for the two dementia-specific classification systems derived in Chapter 4: DEMQOL-U derived from DEMQOL and DEMQOL-Proxy-U derived from DEMQOL-Proxy. These preference weights enable a health-state utility value to be estimated for every health state defined by each classification system. These preference weights can be used to generate a utility score for a patient with dementia each time the patient completes the DEMQOL questionnaire or their carer completes the DEMQOL-Proxy questionnaire. These utility scores can then be used as the ‘Q’ quality adjustment weight of the QALY to inform economic evaluation. The use of both the DEMQOL-U and DEMQOL-Proxy-U measures enables utilities to be generated appropriately across the full severity range of dementia. These dementia-specific measures and corresponding utility scores have been designed specifically for use in a cognitively impaired population, offering an advantage over generic measures such as the EQ-5D and SF-36.

The valuation study included rank and TTO tasks. Both rank and TTO data can be used to estimate preference weights that can be used to generate utility values for all health states defined by the classification system. Here, the regression models estimated using the rank data had more significant coefficients than the models estimated using the TTO data, yet had inconsistent coefficients for two DEMQOL-U dimensions. We recommend that the preferred models estimated using TTO data should be used to generate the preference weights. This is in accordance with recommendations from NICE stating that the valuation protocol used to estimate the preference weights should be the protocol used to produce the UK values of the EQ-5D. This promotes comparability of utility values generated using different classification systems from different preference-based measures.

There is currently no accepted standard on the minimum number of observations required for each health state to ensure reliability and robustness of the modelling output. It is possible that more observations per health state may have improved the reliability and robustness of the modelling output and further research into this issue is encouraged. However, the regression models estimated using the TTO data had good predictive ability (e.g. MAE 0.042–0.045) and performed comparably with regression models estimated for other preference-based measures including the generic EQ-5D (MAE 0.039),45 generic SF-6D (MAE 0.073–0.079),38 asthma-specific AQL-5D (0.046–0.057),40 overactive bladder-specific OAB-5D (0.044–0.076)42 and cancer-specific EORTC-8D (0.046–0.054). 72

Each measure taps three common dimensions (although not using all of the same items) in its classification system: positive emotion, negative emotion and memory. DEMQOL-U has two additional dimensions of relationships and loneliness and DEMQOL-Proxy-U has an additional appearance dimension. The appearance dimension had small and inconsistent coefficients across all DEMQOL-Proxy-U regression models, suggesting that deterioration in this dimension did not have a large or stable impact on utility. The coefficients for the dimensions common to both measures were consistently larger for DEMQOL-Proxy-U than for DEMQOL-U (with the exception of negative emotion level 2). This finding is consistent with research which found that adding an extra dimension to a classification system reduced the preference weights for existing dimensions. 78 This suggests that the coefficients of the DEMQOL-U measure with five dimensions may be smaller in the common dimensions than the coefficients of the DEMQOL-Proxy-U measure because of the larger number of dimensions. However, this finding may also be due to differences in the wording of the items selected for each classification system.

Observed and modelled utility scores for DEMQOL-U have a larger range and lower bound than those for DEMQOL-Proxy-U, meaning that DEMQOL-U has the larger severity range. This is likely due to differences in the classification systems: DEMQOL-Proxy-U had fewer dimensions and the appearance dimension had small preference weights. Yet this produces an apparent contradiction as DEMQOL and DEMQOL-U were designed for use in people with mild to moderate dementia whereas DEMQOL-Proxy and DEMQOL-Proxy-U were designed to be appropriate for all levels of dementia. However, the possible range of utility scores for each measure is likely to be different from utility scores of patients because this depends on the distribution of responses to the classification system. This will be evaluated further in Chapter 7.

Conclusion

The data presented here suggest that DEMQOL-U and DEMQOL-Proxy-U, dementia-specific preference-based single-index measures derived from DEMQOL and DEMQOL-Proxy, respectively, are suitable to be used to generate utility scores for use in health technology assessment across the full severity range of dementia when both measures are used. Algorithms are presented on the DEMQOL website that generate DEMQOL-U and DEMQOL-Proxy-U utility scores when applied to DEMQOL and DEMQOL-Proxy data (www.kcl.ac.uk/iop/depts/hspr/research/ciemh/mha/demqol/index.aspx). These algorithms use the preference weights generated using modelled TTO data.

Introduction

In the preceding chapters we have detailed the development of DEMQOL-U and DEMQOL-Proxy-U, dementia-specific preference-based single-index measures from DEMQOL and DEMQOL-Proxy. In Chapter 5 we reported the estimation of a preference-based single index for each classification system using values obtained from the general population. However, such values can also be obtained from other sources including patients, carers and health-care professionals. Values are typically obtained from the general population in accordance with recommendations from agencies such as NICE and the Washington Panel. General population values are often advocated because public preferences are considered appropriate when health care is publicly funded (partly or fully, depending on a country's health system). In addition, the general population is considered to have no vested interest as they do not have prior knowledge about the future health states they may experience. However, it can also be argued that the general population do not fully understand hypothetical health states because of their lack of experience of similar health states, meaning that patients are more able to provide accurate valuations. This is important because patients often provide different values from those provided by the general population. This means that the source of the values may affect the values obtained and thus have an impact on any economic evaluation based on such data. 79

A recent review and meta-analysis found that patient values of health states were significantly lower than general population values using the TTO elicitation technique. 80 However, other papers have found that patients provide higher values for health states using a variety of valuation techniques. 79,81 Reasons for these differences can be summarised as (1) valuing different states because of differences in understanding or interpretation of the description; (2) different scales of measurement because of a response shift in different populations; and (3) adaptation to the state by patients. Patients may better understand what it is like to be in the health state than a member of the general population who has to imagine the health state. Furthermore, patients may value the health state differently because they are bringing to the valuation additional information that is not included in the health-state description. However, patients may also not recall experiences of full health or mild health states, and may have lowered their expectations as a result. Knowledge and consideration of adaptation is one important difference between patients and members of the general population. Patients may also have different sociodemographic characteristics from the general population and this may also affect how they value health states.

Much of the literature examining differences between patient and general population preferences has focused on physical health states, and the relationship may differ for conditions affecting mental health and cognition. In two studies people experiencing a range of different health states gave mental health greater weight than physical health, which was the opposite for members of the general public trying to imagine the health states. 82 Public attitudes and understandings of dementia in particular may mean that patient and general population values differ for dementia health states. There may also be differences in values provided by carers of patients with dementia, as carers have experience of how the condition impacts on the patient without the health problems experienced by the patient that may affect their understanding of the valuation techniques. The contrast in weightings of physical and mental health by the general population and patients is important for policy. If general population values are used to inform policy and these are higher than patient values for mental health and cognition states, then mental health and cognition may be given lower priority than people with these conditions feel is warranted. The higher relative weighting of physical health to mental health by the general population potentially prioritises physical health at the expense of mental health.

The literature is also limited as typically studies ask respondents to value only a small number of states, meaning that the data cannot be examined to determine whether or not there are systematic differences across all health states or whether or not the difference varies by health-state severity. Even those studies that value a large sample of health states have not examined whether or not differences vary by severity. For example, one study estimated different value sets for patients and the general population for EQ-5D but when combining data from both populations to estimate a combined value set included only a dummy variable to capture patient effects rather than including interaction effects for the severity levels of each dimension. 83 However, one study did find that differences in visual analogue scale (VAS) values vary by the severity of the health state between members of the general population with no health problems, those with mild health problems and those with moderate health problems. 84 It is also important to control for sociodemographic characteristics of the samples84 as it has been found that elicited TTO values vary by sociodemographic characteristics,85 and it is possible that this explains some of the variation in values across populations.

To address these general limitations and to investigate the difference in values that might be obtained between the general population and people with dementia and their carers we undertook a comparative study comparing health-state utility values from samples of the general population, people with dementia and carers of people with dementia for a range of dementia health states of differing severity. The analysis here explores whether population, health-state severity and respondent sociodemographic characteristics impact on elicited utility values.

Methods

Results

Descriptive statistics of health-state utility values

Table 24 presents descriptive statistics of observed TTO values for DEMQOL-U elicited from the general population (as also reported in Chapter 5) and from people with dementia. For every health state, the mean and median TTO values are higher for the general population sample than for the patient sample. The range of mean utility values is larger for the patient sample (from 0.816 to –0.023) than for the general population sample (from 0.955 to 0.190). The ordering of states differs by population. The ordering is logically consistent for each population as for many states it is not possible to determine a logical ordering as one state is not consistently better across all dimensions.

TABLE 24 - Descriptive statistics of observed TTO values for DEMQOL-U

Health state	General population (n = 78)			Patient population (n = 71)
	Mean	SD	Median	Mean	SD	Median
11111	0.955	0.153	1.000	0.816	0.241	0.900
12231	0.894	0.160	1.000	0.633	0.297	0.700
32143	0.673	0.302	0.663	0.399	0.398	0.500
41212	0.611	0.364	0.625	0.436	0.347	0.500
23424	0.566	0.340	0.600	0.435	0.327	0.500
44221	0.544	0.367	0.525	0.327	0.439	0.475
43442	0.403	0.412	0.475	0.161	0.455	0.225
44444	0.190	0.484	0.250	−0.023	0.456	0.000

Table 25 reports descriptive statistics of observed TTO values for DEMQOL-Proxy-U elicited from the general population (as also reported in Chapter 5) and from carers of people with dementia. For every health-state mean and median TTO values are higher for the general population sample than for the carer sample. The range of mean utility values is larger for the carer sample (from 0.857 to 0.049) than for the general population sample (from 0.918 to 0.370). The ordering of states differs by population, but the ordering is logically consistent for each population. Figure 9 shows differences in the distribution of TTO values by population for each measure. General population TTO values are negatively skewed with a large proportion of responses at 1, whereas patient and carer TTO values are also negatively skewed but peak at 0.5.

TABLE 25 - Descriptive statistics of observed TTO values for DEMQOL-Proxy-U

Health state	General population (n = 77)			Carer population (n = 71)
	Mean	SD	Median	Mean	SD	Median
1111	0.918	0.185	1.000	0.857	0.211	1.000
1222	0.869	0.174	0.925	0.731	0.251	0.700
3112	0.804	0.255	0.925	0.666	0.313	0.725
1341	0.807	0.236	0.900	0.640	0.250	0.700
3234	0.672	0.272	0.700	0.458	0.361	0.500
2424	0.638	0.341	0.700	0.464	0.364	0.500
4411	0.580	0.395	0.625	0.380	0.406	0.500
4444	0.370	0.482	0.400	0.049	0.465	0.000

FIGURE 9.

Histogram of observed TTO values: (a) DEMQOL-U valued by the general population sample; (b) DEMQOL-U valued by the patient sample; (c) DEMQOL-Proxy-U valued by the general population sample; (d) DEMQOL-Proxy-U valued by the carer sample.

Simple t-tests revealed that health-state utility values for the general population and for people with dementia are significantly different for DEMQOL-U [general population mean (SD) 0.604 (0.410), person with dementia mean (SD) 0.398 (0.447), t = 8.279, degrees of freedom (df) = 1153, p < 0.001] and health-state utility values for the general population and for carers are significantly different for DEMQOL-Proxy-U [general population mean (SD) 0.707 (0.350), carer mean (SD) 0.531 (0.410), t = 7.946, df = 1120, p < 0.001]. Wilcoxon rank-sum tests reached the same conclusions (p < 0.001 and p  < 0.001 respectively).

Discussion

The results demonstrate that the population used to value dementia health states affects health-state utility values elicited using TTO. This is in accordance with the literature. 80 Health-state utility values elicited from people with dementia and carers of people with dementia were lower than health-state utility values elicited from the general population. This is different from many studies which have found that patients provide higher values than the general population for other health states using a variety of valuation techniques, but is consistent with a recent review and meta-analysis which found that patient values were significantly lower using the TTO elicitation technique in particular. 80 Furthermore, the majority of the literature focuses on physical health states whereas this study focused on mental health and cognition, with previous research suggesting that patients and the general population weight problems in physical and mental health differently. 82

Modelling of the health-state utility values indicated that the differences in values were due to the population per se and not to differences in the sociodemographic composition of the populations. This was demonstrated because controlling for sociodemographic characteristics had a minimal impact on the coefficients of the other variables and resulted in only a minor improvement in model performance and few of the sociodemographic coefficients were significant. Regression analysis and descriptive statistics of observed values indicated that the difference in values by population differed for each health state, sometimes indicating a different ordering of the health states. This different ordering of health states is not an inconsistency across populations; it represents differences in the valuation of and trade-off between different dimensions. This suggests that the choice of whose values are used to produce utility values for use in cost-effectiveness analysis will have an impact on the results and could potentially affect resource allocation decisions. This suggests the need to produce a full value set for all DEMQOL-U and DEMQOL-Proxy-U health states using values elicited from people with dementia and from carers of people with dementia respectively.

The finding that general population values are higher than patient values for dementia states suggests that members of the general population may be systematically undervaluing the impact on quality of life arising from problems with cognition and mental health. This may reflect public perception in general of health problems such as dementia. However, it is important to note that it may be due to the wording and labelling of the classification system. General population respondents were not informed at any point that the study was about dementia, whereas people with dementia and carers knew that the study was about dementia. This contextualised health states for the people with dementia and carers, meaning that they had a greater understanding of the health states, but it also meant that they knew the underlying cause of the health state. A recent study examining labelling effects found that introducing condition labels into health-state descriptions impacted on health-state utility values elicited from the general population, and that the impact differed by condition and health-state severity. 37,71 It can be argued that the inclusion of a condition label can better inform the general population about the health state because respondents may value differently, for example, feeling frustration as a result of epilepsy and feeling frustration as a result of dementia. However, it can also be argued that the inclusion of a condition label may mean that respondents take into account preconceptions of the condition or mortality of the condition, and these are factors that should not be taken into account in the elicitation of health-state utility values. As the aim of the general population valuation survey was to elicit values to inform resource allocation decisions across all conditions and patient groups, there must be comparability between the health-state descriptions used here and in other valuation surveys, regardless of the underlying condition. Otherwise, this could imply that, for example, a given generic EQ-5D state has a lower utility value for a patient with epilepsy than for a patient with diabetes. The exclusion of a dementia health-state label in the general population study may therefore have impacted on elicited utility values, and there is a possibility that if general population values were obtained with the inclusion of a dementia label these values may be closer to the values elicited from people with dementia and carers. However, an important difference is likely to remain: the difference in the valuation of and trade-off between the different dimensions in the classification system. Qualitative research or a full-scale valuation study similar to the study outlined in Chapter 5 for people with dementia and carers would indicate differences in the trade-off between and relative importance of the dimensions in comparison with the general population. Further research in this area is encouraged but is likely to be constrained by the difficulties of conducting this type of research in these populations.

One potential limitation in the comparisons made between the general population values and the values from people with dementia and carers is that the general population study involved ranking as a warm-up task prior to the TTO task whereas this was excluded from the survey for people with dementia and carers. Starting the TTO task without the rank task, which helps respondents think about how they would value different health states, may have affected the understanding of the TTO task and may also have affected the values provided by people with dementia and carers. However, it could be argued that people with dementia and carers knew that the states were dementia states and may already have an idea before the interview about how they value different dementia health states.

Among people with dementia and carers there were significantly higher proportions of respondents who were reported by the interviewers as doubtful that they understood the TTO task, but at 12.7% and 7.0% of respondents, respectively, these proportions are not especially high. The preferred regression models were also estimated on samples excluding respondents whose understanding of the TTO task was doubted by the interviewers. The exclusion of these respondents did not have a large impact on the coefficients for health-state severity and interaction effects but some sociodemographic characteristic variables became significant.

Another potential limitation of this study is the involvement of people with dementia with neuropsychological and cognitive problems to value health states using a cognitively demanding elicitation technique. The study was designed at all stages to take into consideration the health and competencies of the patient population: the team was engaged in conversation with the Alzheimer's Society through its Quality Research in Dementia panel prior to designing the study; patients were referred to the study by clinicians specialising in dementia; and the interviews were immediately terminated if the person with dementia suffered from fatigue, misunderstanding or distress. For these reasons the samples are made up of patients with mild dementia and carers of people with mild dementia. This means that the samples may not be representative of people with more severe dementia and their carers but this is a constraint of conducting a valuation study for this patient group. The MVH TTO protocol has been widely used in valuation studies of the general population but may be more challenging for respondents with cognitive problems. The TTO is no more challenging than many other elicitation techniques, such as SG or person trade-off, and arguably is less cognitively demanding than ranking, which requires the simultaneous consideration of multiple health states. Other options that may be less cognitively demanding include ordinal techniques such as discrete choice techniques and best–worst scaling. However, these require many more data and either increase the burden on each respondent and/or require a much larger sample. There are also challenges regarding the anchoring of these values onto the 1–0 full health–dead scale (see reference 19 for an overview). A VAS may be cognitively easier for respondents, but to obtain values anchored onto the 1–0 full health–dead scale would require the valuation of dead alongside each state.

The study used the same valuation protocol as the EQ-5D as recommended by NICE to ensure that elicited utilities are comparable with the UK tariff of the EQ-5D. However, the choice of TTO and the valuation protocol may have impacted on results. A recent meta-analysis of patient and general population values found that results differed by valuation technique80 and other research has found that valuation results may in general differ by protocol. 87 Further research in this area is needed.

Conclusion

Dementia health-state utility values elicited using TTO differ by the population used to elicit the utility values, with people with dementia and carers of people with dementia giving systematically lower utility values than members of the general population. These differences in values were due to the population per se and not to differences in the sociodemographic characteristics of the populations. The general population underestimated the impact of dementia compared with people with dementia and their carers. The ordering of health states also differed by population for some health states, indicating a difference in the valuation of and trade-off between different dimensions. These results suggest that the population used to produce dementia health-state utility values could impact on the results of cost-effectiveness analysis and potentially affect resource allocation decisions.

Introduction

If the DEMQOL-U and DEMQOL-Proxy-U are to be used alongside or instead of generic preference-based measures it is important to assess their psychometric validity, responsiveness and level of agreement between patient and carer report. This can be assessed by applying psychometric methods to data sets containing responses to the DEMQOL system alongside generic preference-based and non-preference-based measures. The issue of how condition-specific preference-based measures compare with generic preference-based measures is particularly important as this indicates the likely impact of using DEMQOL-U and DEMQOL-Proxy-U instead of generic measures to generate QALY values for use in economic evaluation. The psychometric testing of DEMQOL-U and DEMQOL-Proxy-U is reported in this chapter using data from the HTA-SADD trial, a multicentre randomised double-blind placebo-controlled trial of the clinical effectiveness of sertraline and mirtazapine. 46 This study included an assessment of HRQL using the generic preference-based measure EQ-5D and the DEMQOL system.

In this chapter we compared the validity, patient/proxy agreement and responsiveness of the EQ-5D and the DEMQOL-U and DEMQOL-Proxy-U utility measures. Validity was examined in terms of the ability to discriminate between different levels of severity (i.e. known group validity), the convergence between the DEMQOL utility measures and the other measures of dementia-related constructs, and the level of agreement between DEMQOL-U, DEMQOL-Proxy-U and patient- and carer-reported EQ-5D. Agreement between patient- and carer-reported utilities was assessed. Responsiveness was assessed by investigating sensitivity to change in quality of life over time.

Method

Measures

Analysis

Results

Agreement

Figure 10 presents Bland–Altman plots depicting the extent of the agreement between the patient measures (EQ-5D and DEMQOL-U) and between the carer measures (CEQ-5D and DEMQOL-Proxy-U) across the range of possible utility values. For the patient measures at baseline, 4.78% of the responses fall outside the 95% limits of agreement, and there is a small increase at follow-up. The same pattern is apparent for the CEQ-5D and DEMQOL-Proxy-U, for which 3.81% of responses fall outside the 95% limits of agreement at baseline. The charts indicate that the majority of the disagreement occurs at the more severe end of the scale, where a small number of respondents are providing a higher utility score for DEMQOL-U than for the EQ-5D. This may be related to the difference in the ranges of the utility scales, which means that respondents in poor health may generate a substantially lower utility score on EQ-5D than on DEMQOL-U. Agreement between the measures is increased at the higher end of the utility scale, and the mean difference decreases at follow-up. This indicates that measurement convergence between the generic and the condition-specific indicators is better among those in better health at baseline, and also as health status improves over the course of a trial. The mean difference for the CEQ-5D and DEMQOL-Proxy-U is larger, indicating more disagreement between the proxy-reported measures.

FIGURE 10.

Agreement between DEMQOL-U and EQ-5D, and CEQ-5D and DEMQOL-Proxy-U utility values at baseline, week 13 and week 39. (a) DEMQOL-U vs EQ-5D; (b) DEMQOL-Proxy-U vs CEQ-5D.

Table 31 presents respondent scores for DEMQOL-U and DEMQOL-Proxy-U when the patient is in full health according to the EQ-5D (state 11111 is equivalent to a utility score of 1). There is wide variation in the DEMQOL-U and DEMQOL-Proxy-U utility values when the same respondents report optimal health on the EQ-5D (or CEQ-5D). At baseline the DEMQOL-U scores that are found when the EQ-5D utility is 1 range from 0.63 to 0.96. The range increases at the 39-week follow-up (0.48–0.96). The same pattern is apparent for DEMQOL-Proxy-U, where the range at baseline is 0.60–0.81, increasing to 0.65–0.94 at 39 weeks' follow-up.

TABLE 31 - Distribution of DEMQOL-U AND DEMQOL-Proxy-U utility values when EQ-5D is in full health

Measure	n	Mean	SD	Median	Min.	Max.
EQ-5D = 11111
DEMQOL-U week 0	43	0.85	0.09	0.87	0.63	0.96
DEMQOL-U week 13	57	0.82	0.10	0.84	0.61	0.99
DEMQOL-U week 39	45	0.85	0.10	0.89	0.48	0.96
CEQ-5D = 11111
DEMQOL-Proxy-U week 0	8	0.74	0.07	0.76	0.60	0.81
DEMQOL-Proxy-U week 13	15	0.79	0.09	0.77	0.67	0.94
DEMQOL-Proxy-U week 39	18	0.83	0.10	0.86	0.65	0.94

Patient/proxy agreement

In light of the common reliance on carer reports for people with cognitive impairment, the agreement between carer- and patient-reported utility values was examined (Table 32). In terms of mean differences, DEMQOL-Proxy-U utilities differed very little from DEMQOL-U utilities. The difference between the EQ-5D and CEQ-5D scores was relatively larger (0.18–0.20). However, the ICC values indicated that there was better agreement between the EQ-5D and the CEQ-5D than between DEMQOL-U and DEMQOL-Proxy-U. This means that the EQ-5D may have better agreement between patients and carers. However, the patient-and carer-reported EQ-5D utilities did have greater variability than was observed for DEMQOL-U and DEMQOL-Proxy-U, and this may be related to the substantially larger utility range of EQ-5D.

TABLE 32 - Agreement between utility values derived from self- and carer-rated HRQL data

CI, confidence interval; LoA, limits of agreement.

Mean difference = self-rating – proxy rating.

Bland–Altman plots.

CC model: two-way random absolute agreement.

Measure	n	Mean differencea	95% LoAb	ICCc	95% CI
DEMQOL-U and DEMQOL-Proxy-U
Baseline	273	0.02	−0.24 to 0.28	0.15	0.04 to 0.26
Week 13	202	0.01	−0.01 to 0.03	0.25	0.12 to 0.38
Week 39	154	0	−0.26 to 0.26	0.17	0.01 to 0.32
EQ-5D and CEQ-5D
Baseline	290	0.20	−0.46 to 0.85	0.32	0.12 to 0.47
Week 13	216	0.18	−0.42 to 0.79	0.36	0.14 to 0.53
Week 39	170	0.19	−0.43 to 0.81	0.23	0.04 to 0.39

Agreement between patients and carers on DEMQOL-U and DEMQOL-Proxy-U and on the EQ-5D and CEQ-5D was also examined using Bland–Altman plots (Figure 11). Although the mean difference between DEMQOL-U and DEMQOL-Proxy-U is less than that between the EQ-5D and the CEQ-5D, a higher percentage of responses are outside the 95% agreement limits at baseline (6.23% vs 5.86%). However, the magnitude of the difference is not large and this indicates that agreement is similar across the DEMQOL and EQ-5D measure pairs. The difference in the magnitude of agreement remains low across all time points, with the percentage outside the 95% agreement limits lower for the DEMQOL-U and DEMQOL-Proxy-U response pairs. The plots indicate that most of the difference occurs at the more severe end of the scale, where the mean utility score is lower. This means that patients are rating their utility as substantially better than their carer perceives it, or vice versa. The mean difference between ratings is higher for the EQ-5D, but this is explained by the larger range of utility values generated by this instrument.

FIGURE 11.

Agreement between self- and carer-based utility values at baseline, week 13 and week 39. (a) DEMQOL-U vs DEMQOL-Proxy-U; (b) EQ-5D vs CEQ-5D.

Discussion

In this chapter we have documented initial evidence regarding the acceptability, validity, patient/carer agreement and responsiveness of the DEMQOL-U and DEMQOL-Proxy-U dementia-specific preference-based HRQL measurement system (DEMQOL-U and DEMQOL-Proxy-U) and the EQ-5D generic preference-based measure in comparison with external indicators of dementia-related health status and the original DEMQOL and DEMQOL-Proxy. We also assessed agreement between the preference-based instruments and found that overall there is a good level of agreement and that the majority of the disagreement between the measures occurs at the more severe end of the utility scale. There is some evidence for the acceptability of the DEMQOL system, in particular the DEMQOL-Proxy-U, which displays low missing data rates. However, missing data rates are lower for the EQ-5D. There is less evidence regarding the convergent validity of DEMQOL-U and DEMQOL-Proxy-U in comparison with other measures of cognition, depression and activity level in dementia. DEMQOL-U displays a higher level of known-group validity between cognition and depression severity groups than DEMQOL-Proxy-U, but the effect sizes are in the low range and there are differences from the discriminative ability of the patient- and carer-reported EQ-5D. There is no clear pattern regarding agreement between patients and carers. In terms of responsiveness, there is evidence that the DEMQOL utility measures and EQ-5D are less sensitive to change than the original DEMQOL and DEMQOL-Proxy.

The patient- and carer-reported EQ-5D have lower missing data rates than DEMQOL-U and DEMQOL-Proxy-U respectively. This may indicate that the EQ-5D is more acceptable to respondents. However, the differences in missing data rates are not large and may be expected because of the size of the instruments (EQ-5D includes five items whereas DEMQOL and DEMQOL-Proxy, which are used to generate the utility values, include 29 and 32 items respectively). Therefore, it is more likely that data will be missing from the DEMQOL utility measures.

DEMQOL-U displays a lower level of agreement with the patient-reported EQ-5D than DEMQOL-Proxy-U does with the CEQ-5D, and this may be because carers are more likely to report stable responses than people with dementia in terms of the general level of severity reported. However, for both pairs the majority of the scores fall within the 95% agreement range, indicating a good level of agreement. The majority of the disagreement occurs when the mean utility value is at the more severe end of the scale (where DEMQOL utility measure scores are substantially higher than EQ-5D/CEQ-5D scores). This may be due to differences in the classification systems, which may be sensitive to different HRQL issues, and also to large differences in the possible utility scale range.

Exploring the distribution of the utility values of each measure revealed that DEMQOL-U and DEMQOL-Proxy-U utility values tend to be higher than EQ-5D values for poorer health states. This may be expected as the lowest possible EQ-5D utility score is –0.594 whereas the lowest possible utility scores for DEMQOL-U and DEMQOL-Proxy-U are 0.243 and 0.363 respectively. Similarly, EQ-5D utility values tend to be higher than DEMQOL-U and DEMQOL-Proxy-U values for better health states. The EQ-5D had high ceiling effects and therefore may not be sensitive to mild HRQL impairment in patients with better health states. This is also consistent with the considerable variation observed in DEMQOL-U and DEMQOL-Proxy-U utility values, which suggests that patients with full health on EQ-5D may be experiencing varying levels of impaired HRQL. A likely reason for the ceiling effects observed in the EQ-5D is that there are three levels of responses for each dimension, which may result in insensitivity for detecting small differences in health states. 94

As carer report may be a necessary substitute for self-report among people with dementia, a utility measure that allows for good agreement between the carer and the patient perspective would be preferred. Results regarding the agreement between patients and carers across the DEMQOL utility measures and patient- and carer-reported EQ-5D are mixed. There is a higher level of patient/carer agreement between the EQ-5D and the CEQ-5D at each time point, and more agreement between the ratings at baseline; however, DEMQOL-U and DEMQOL-Proxy-U ratings display more agreement at follow-up. However, overall, the agreement as measured using ICC values is low for both measures. As the dimensions included in the EQ-5D are matched across the patient and carer versions, it might be expected that the EQ-5D displays a higher level of agreement between ratings than the DEMQOL measures, for which the dimensions and questions used vary. However, it is well established that dementia patients and proxies give different reports,14,68 and therefore agreement level cannot be seen as a clear indicator of the performance of the measures. Further work in this area could investigate the pattern of agreement further, and this could be done by investigating agreement at different dementia severity levels, both for the DEMQOL utility measures and the EQ-5D.

We psychometrically assessed the convergent and known-group validity of DEMQOL-U and DEMQOL-Proxy-U. DEMQOL-U displayed a moderate correlation with the CSDD (at a higher level than the EQ-5D and DEMQOL), suggesting some convergence in terms of measurement of depression related to dementia. However, correlations between DEMQOL-U and the EQ-5D and the indicators of cognition, daily activities and behavioural disturbances are low. The correlations between the CEQ-5D and the external condition-specific indicators (which are low to moderate) are higher than the DEMQOL-Proxy-U correlations (which are low).

There is no clear pattern to the convergent validity results, which makes interpretation difficult. It could be argued that the lack of convergence for the patient measures is due to differences in the focus of the classification systems in terms of both the dimensions included and the dimensions omitted. However, the EQ-5D has dimensions directly assessing depression and daily activities and DEMQOL-U includes a cognition dimension and so better correlations may be expected. Further analysis should investigate relationships between the measures at the dimension level and also investigate convergence using a range of other indicators. The CEQ-5D correlations are higher, particularly with the other measures completed by carers, and this may mean that patient report in dementia is unreliable and unstable across assessments.

There is also no clear pattern to the known-group validity of the DEMQOL utility measures or the EQ-5D across cognitive impairment and depression severity groups. Overall, DEMQOL-U performs better than DEMQOL-Proxy-U, but the CEQ-5D performs better than the EQ-5D. This may be because the DEMQOL-U descriptive system generates more health states than DEMQOL-Proxy-U and so therefore has more discriminative ability. The better performance of the CEQ-5D may be due to the more stable nature of proxy report in dementia.

For the patient utility measures across the cognitive impairment and depression severity groups (excluding the definite depression group for which the sample size is low), the effect sizes are in the small to moderate range. The magnitude of the change is similar, but the DEMQOL-U effect sizes are larger because the SD of the EQ-5D scores is larger because of differences in utility score range. This means that, in trials, the DEMQOL utility values can be estimated with more precision. This has also been found for other condition-specific preference-based measures in comparison with the EQ-5D. 37

In terms of carer report, the CEQ-5D is more sensitive to differences across cognition severity groups than DEMQOL-Proxy-U. However, both measures display effect sizes in a similar range for depression severity. Again, there is no clear pattern and further research could investigate the discriminative ability of the DEMQOL utility measures using a range of further indicators. Furthermore, more work needs to be carried out on the discriminative ability of the measures among those with severe problems, as the sample size used in this study is too small to draw any conclusions.

Although the DEMQOL utility instruments and the EQ-5D respond to improvements in health status over time, this is not at the same level as with the original DEMQOL and DEMQOL-Proxy. DEMQOL-U was sensitive to deterioration at a higher level than DEMQOL-Proxy-U. The reasons for these differences in results are not clear and more research is required to assess responsiveness in more detail using a range of indicators and different data sets.

For the psychometric analysis reported above we used the utility values derived from the general population, and this allows for a level of comparability with the EQ-5D value set, which was also derived from the general population. However, the difference between the general population and patient/carer valuations of DEMQOL-U and DEMQOL-Proxy-U health states reported in Chapter 6 suggests that a full valuation study with patients and carers would produce a substantially different utility scale. If a full valuation study was to be carried out it would be important to investigate the psychometric performance of the patient- and carer-derived utility scales in comparison with both the general population values and external dementia-related indicators. It was not possible to compare the valuations produced in Chapter 5 as only mean TTO values were produced.

It is important to note that the psychometric performance of the DEMQOL utility measures may be impacted by the HTA-SADD sample used (in which the focus was on depression). DEMQOL-U and DEMQOL-Proxy-U are designed specifically for dementia and therefore further testing on other dementia samples is recommended. Furthermore, this analysis has focused on the psychometric properties of each measure and has not combined the utility values with duration to produce QALY estimates. Further comparative research using QALY estimates is also recommended.

Conclusions

Health outcomes in economic evaluation are often measured using a composite measure of both quality and length of life – the QALY. This composite measure can be used in health policy decision-making to compare the efficiency of different treatment strategies. For utilities (such as those derived here) to be of any value for decision-makers, they must be included in a QALY measure that also considers length of life. The analysis described above presents some early evidence of the level of validity, patient/carer agreement level and responsiveness of the DEMQOL utility measures in comparison with the widely used generic preference-based measure EQ-5D. The results suggest that both the EQ-5D and the DEMQOL utility instruments have advantages and disadvantages over each over, but the pattern is unclear. Further research investigating the complex psychometric performance of the measures is required using different data sets incorporating a range of clinical indicators and dementia severity levels. We originally planned to carry out the analysis described above on a second data set but these data did not become available during the study period. Until further results are available we would recommend using both the EQ-5D and the DEMQOL utility measures alongside each other in dementia studies to allow for both generic and condition-specific utilities to be generated.

To derive health-state classification systems from DEMQOL and DEMQOL-Proxy that can be used to categorise all patients with responses to the measures

This project has used an established multistage process to guide the derivation of preference-based measures from the DEMQOL instruments. The first stage was to identify the dimensions to be used in the health-state classification using factor analysis as a guide to the dimensionality. We chose to use exploratory analysis because of the inconclusive nature of the earlier factor analysis of the DEMQOL system (in comparison with the original conceptual framework), which means that we did not have an a priori structure to confirm. However, the factor structures established here provide some support for the original conceptual framework. The factor structures are robust enough to provide the basis for the development of dementia-specific health-state classification systems: a five-dimensional structure was found for both DEMQOL and DEMQOL-Proxy.

We then used Rasch analyses to select an item to represent each factor. Using stringent criteria to investigate item performance we were able to select one item for each of the five DEMQOL domains and one item for four of the DEMQOL-Proxy domains (one domain was dropped as none of the items performed at an acceptable level in the Rasch analysis). Each dimension has four-level responses corresponding to the response choices for the items in the original DEMQOL instruments. These new instruments were named DEMQOL-U (which produces a possible 1024 health states) and DEMQOL-Proxy-U (which produces 256 possible states). These health-state classifications generate health states that are amenable to valuation and the subsequent generation of dementia-specific QALYs.

To generate utility values for every health state defined by the health-state classification systems derived from DEMQOL and DEMQOL-Proxy

Interviews were successfully conducted with a representative sample of 593 members of the general population who were able to complete the rank and TTO tasks and provide valuations of a sample of health states generated by DEMQOL-U and DEMQOL-Proxy-U. The TTO valuations were modelled to estimate two preference-based algorithms, one each for DEMQOL-U and DEMQOL-Proxy-U. This means that utility values can be estimated for all 1024 DEMQOL-U health states and all 256 DEMQOL-Proxy-U health states. We applied a range of models to the TTO and rank data and selected the TTO model that fitted best overall, with the highest number of significant variables and best-fit statistics. The models fitted to the data performed favourably compared with similar models estimated for other preference-based measures in terms of fit, prediction and the consistency of the coefficients with the descriptive system. The DEMQOL-U model did not have any inconsistent variables (in which a decrease in health status leads to an increase in utility) and DEMQOL-Proxy-U had just one. The modelling process means that it is now possible to generate health-state utility values from any data set in which one or both of DEMQOL and DEMQOL-Proxy have been completed.

To examine whether or not utility values elicited from the general population differ from utility values elicited from patients and carers for dementia health states generated by the classification systems

We interviewed a sample of 71 people with mild dementia and 71 family carers. They were able to complete the TTO task and assign values to a limited number of the health states generated from the new instruments. This is the first time that the values of dementia health states from the general population have been compared with patient and carer values in this way. The main finding was that the values obtained from people with dementia and carers were systematically different from those obtained from the general public. The general public tended to undervalue the impact of dementia compared with people with dementia and carers. This finding is in contrast to the general finding in the literature – that patients tend to give higher values than the general public, at least for physical health conditions. 79 Patient values are thought to be higher because of processes related to adaptating to the condition. In mental health the opposite has been observed, and this may reflect an inability of the general population to understand the consequences for quality of life of problems associated with mental disorders compared with physical health. 82 This finding may reflect a similar inability to understand the consequences of the syndrome of dementia, even though they are described in everyday terms.

It is not possible to model the pattern of the difference between patients and carers and the general population. Too few states were valued to estimate specific algorithms for patients and carers, as the aim of this part of the project was only to test the hypothesis that there are differences. However, it was also observed that there were differences in the ordering of states, which suggests that there may well be differences in the relative weight being given to different dimensions. This requires further investigation.

These results suggest that the population used to produce dementia health-state utility values could impact on the results of cost-effectiveness analysis and potentially affect resource allocation decisions.

To examine the psychometric performance of the dementia-specific preference-based measures using trial data

Finally, we tested the new measures using a data set from a recently completed trial. In comparing the psychometric performance of DEMQOL-U and DEMQOL-U-Proxy with that of their non-preference-based as well as generic preference-based counterparts, this report documents the first early evidence of validity and responsiveness of a condition-specific preference-based measure for people with dementia. We also assessed the utility measures in comparison with clinically based measures. However, these measures may be indirectly related to many of the dimensions covered by the DEMQOL utility instruments and the EQ-5D and further research should test the measures using a range of clinical indicators.

The DEMQOL utility instruments and the EQ-5D have advantages and disadvantages over each other in terms of psychometric performance; however, the pattern is unclear. There is evidence for the acceptability of the DEMQOL system but missing data rates are higher than for EQ-5D. EQ-5D displays high ceiling effects and therefore may not be as sensitive as the DEMQOL utility measures to mild HRQL impairment in patients with better health states.

There is a good level of agreement between the preference-based instruments. The majority of the disagreement between the measures occurs at the more severe end of the utility scale. This may be linked to differences in the dimensions described by the classification system, which may be sensitive to different HRQL issues, and also to large differences in the possible utility scale range. In terms of agreement between patients and carers we found that carer report was more stable. In dementia, carer report is often essential and therefore a utility measure that allows for good agreement between the carer and the patient perspective is important.

There is no clear pattern to the convergent validity results. Correlations between DEMQOL-U and the EQ-5D and the external indicators are low. However, DEMQOL-U displays some convergence in terms of measurement of depression related to dementia. The correlations between the CEQ-5D and the external indicators are higher than the DEMQOL-Proxy-U correlations. It could be argued that the lack of convergence for the patient measures is due to differences in the focus of the classification systems in comparison with the external indicators. Assessment against a wider range of indicators is required.

DEMQOL-U discriminates better than DEMQOL-Proxy-U but the CEQ-5D discriminates better than the EQ-5D. This may be linked to classification system differences in that DEMQOL-U generates more health states than DEMQOL-Proxy-U and so has increased discriminative ability. The better performance of the CEQ-5D may be linked to the more stable nature of proxy report in dementia across matched classification systems.

In terms of responsiveness, there is evidence that the DEMQOL utility measures and the EQ-5D are less sensitive to change than the original DEMQOL and DEMQOL-Proxy. The reasons for these differences are not clear and responsiveness needs to be assessed in more detail using a range of indicators and different dementia-specific data sets.

The psychometric performance of the DEMQOL utility measures may be impacted by the sample used, which focused on depression. The inconclusive nature of the results means that further testing on a range of samples is required.

Limitations of the study

There are a number of limitations in the development process and testing of the DEMQOL utility measures that currently limit the use of the instruments.

The first limitation relates to the lack of an external data set to validate the DEMQOL-U and DEMQOL-Proxy-U classification systems. This is an important step in the process of deriving a condition-specific preference based measure from an existing measure but is reliant on the data available to developers. In this study, no external DEMQOL or DEMQOL-Proxy data were available to validate the system, and the HTA-SADD data were available only towards the end of the study when the psychometric validation stage of the project had been completed and the valuation phase had commenced. It is also possible to use a split-half approach to validation in which a subsample of the development data are used. However, the optimum sample size for Rasch analysis is 500 and splitting the sample used in this study into two would result in data sets that are too small for the task. This problem is made worse for those factors with a small number of items, as the Rasch programme excludes extreme scores within the dimension (i.e. respondents at the floor or ceiling), resulting in a smaller sample size.

Throughout the development process we have made decisions about which methodology to use. There are other valid techniques that could have been used and it is possible to criticise the choices that we have made. First, it is possible to criticise the use of EFA to investigate the dimensional structure of the instruments, and this study may be limited by not investigating a wider range of possible methods for deriving a dimensional structure. We used EFA as the earlier factor analysis of the DEMQOL system was inconclusive and so we were not confirming an existing structure. However, EFA may not be as robust as CFA and it may have benefited the development process to also study some CFA models.

The health-state classification systems that we have derived are a result of the original measure and the sample and analysis techniques used. It is possible that the use of a mild to moderate dementia sample has resulted in classification systems describing a comparatively mild conceptualisation of dementia. This is supported by the results of the valuation study. However, we used the mild to moderate sample as DEMQOL has been validated for use in this group, and it was decided to use a matched sample in terms of severity for DEMQOL-Proxy. It is also possible that the resultant classification systems are missing important dimensions, in particular daily activity limitations (this did not appear as a dimension in DEMQOL-U and was excluded from DEMQOL-Proxy-U because of poorly performing items). However, it should also be noted that the lack of association between dementia-related HRQL and activity limitations has been found elsewhere. 20

Another concern might be the absence of the direct involvement of patients and their carers in the development of the health-state classifications. The process benefited from the results of in-depth interviews with patients and their carers, a detailed psychometric analysis of patient- and carer-reported outcomes and the expertise of clinicians working with patients on a daily basis. However, the process of selection might have been further improved with more direct patient involvement in the decisions and this is something to consider in any further qualitative work that might be undertaken with the measure (see Recommendations for future research).

In the original proposal it was suggested that health-state utility values could also be derived using ranking and a further DCE study. The field has developed since the time that this study was designed and the proposed methods of estimating values on the 0–1 scale required for calculating QALYs have been superseded. 95 Although we have conducted the analysis of the rank data collected in the valuation study, it was decided not to send respondents an additional postal survey given that the results would not be used to inform policy as the TTO results provide greater comparability with the EQ-5D and other measures.

Conclusions for evaluation in dementia

We would recommend that those designing trials and studies on dementia consider the use of DEMQOL-U and DEMQOL-Proxy-U. However, to clarify the nature and extent of the advantages and disadvantages of the new preference-based measures developed here we need further research into the psychometric performance and acceptability of the instruments. Ongoing and future trials and studies using the DEMQOL system will allow for further analyses to be carried out comparing the output of the generic measures with the general population utility values derived from DEMQOL-U and DEMQOL-Proxy-U.

More generally there are important issues surrounding whether or not QALYs estimated from condition-specific preference-based measures can be used to inform resource allocation across programmes. This issue has been examined in detail in a recent HTA report reviewing the development of preference-based measures from existing condition-specific measures. 37 Policy-makers must weigh the potential gains in terms of greater relevance and sensitivity to the condition from using a particular condition-specific preference-based measure with the reduction in comparability. Condition-specific preference-based measures have a number of potential disadvantages including missing side effects of treatment and the impact of comorbidities (where their effect is not additive). There are also concerns that the values for the condition-specific states are subject to focusing effects from respondents failing to take into account their overall health.

Currently there is not enough evidence to be able to determine the likely benefits of using the DEMQOL-U and DEMQOL-Proxy-U measures compared with using the EQ-5D or other generic measures. At this stage it is not possible to recommend that they are used in place of the EQ-5D. We recommend that they are used together with the EQ-5D (or other generic measures depending on jurisdiction) in future studies to ensure that the needs of policy-makers can be met. This will also provide further evidence for testing these new instruments and examining their impact on the cost-effectiveness of interventions.

Recommendations for future research

1. The selection of dimensions and items in DEMQOL-U and DEMQOL-PROXY-U from the original instruments needs to be validated in another data set. The lack of validation is a limitation of the DEMQOL utility measures and we would aim to carry out this process when relevant data become available. This would complete stage IV of the six-step process and provide more evidence regarding the stability and robustness of the classification systems.

2. Further evaluation of DEMQOL-U and DEMQOL-Proxy-U compared with generic preference-based measures such as the EQ-5D, other relevant dementia-specific indicators and the original DEMQOL system should be carried out. The initial psychometric validation reported here is inconclusive, with evidence both for and against the validity and responsiveness of the DEMQOL utility measures, and agreement between preference-based measures and self- and carer report. Therefore, there is a need to expand the analyses completed for the HTA-SADD trial to other trials and evaluations that have used the DEMQOL system using a range of other relevant indicators. This will provide further evidence regarding the psychometric performance of the utility measures, and the samples in which the measures perform best. Further psychometric analysis will allow policy-makers to understand the implications of using these new dementia-specific preference-based measures to estimate QALYs compared with EQ-5D in cost-effectiveness analyses of new interventions.

3. We have shown that it is feasible to obtain health-state values from people with dementia and carers and that the valuations differ from those provided by the general population. A future study is needed to value a sufficient number of health states to estimate patient and carer preference weights for all states defined by the classification system so that economic evaluations can be conducted using both patient and carer and general population utility weights. This will inform an important policy debate about the consequences of using patient or carer values rather than those of the general population.

4. When there are data sets that share the use of the EQ-5D and DEMQOL and the proxy versions, these should be pooled to allow for further analysis of the system. Studies are ongoing that include the DEMQOL system and the EQ-5D and these data may become available to the authors for further analysis.

5. One potential area for future research is to investigate the acceptability and validity of the classification systems qualitatively with dementia patients and their carers.

Contributions of authors

Sube Banerjee (Professor of Mental Health and Ageing, King's College London, Institute of Psychiatry) was the principal investigator for this project overall. He was the principal investigator for the development of DEMQOL and initiated the proposal that led to this work. He led the team at the Institute of Psychiatry, participating in the design and conduct of all stages of the research and report preparation.

John Brazier (Professor of Health Economics, Health Economics and Decision Science, ScHARR, University of Sheffield) was the co-principal investigator of the study, leading the economics team at the University of Sheffield. He participated in the design and conduct of all stages of the research and report preparation.

Brendan Mulhern (Research Associate, Health Economics and Decision Science, ScHARR, University of Sheffield) acted as study co-ordinator, participated in the design and conduct of all stages of the research and report preparation other than the grant preparation and co-ordinated the input of the Sheffield elements of the research. He completed the psychometric and economic work informing the factor structure and item selection.

Donna Rowen (Research Fellow, Health Economics and Decision Science, ScHARR, University of Sheffield) participated in the design and conduct of all stages of the research and report preparation other than the grant preparation. She completed the valuation elements of the study.

Sarah Smith (Research Fellow, Department of Health Services Research and Policy, London School of Hygiene and Tropical Medicine) completed the psychometric work informing the factor structure and item selection. She participated in the design and conduct of all stages of the research and report preparation.

Renee Romeo (Lecturer, Centre for the Economics of Mental Health, Institute of Psychiatry, King's College London) completed the psychometric and economic evaluation of the system using the HTA-SADD data set. She participated in the design and conduct of all stages of the research and report preparation.

Rhian Tait (Research Worker, Institute of Psychiatry, King's College London) recruited and interviewed the south London valuation sample. She participated in all stages of report preparation.

Caroline Watchurst (Research Worker, Institute of Psychiatry, King's College London) recruited and interviewed the south London valuation sample. She participated in all stages of report preparation.

Kia-Chong Chua (PhD student, Institute of Psychiatry, King's College London) assisted RR in the analyses of the HTA-SADD data set. He participated in all stages of report preparation.

Vanessa Loftus (Specialist Registrar, South London and Maudsley NHS Foundation Trust) helped co-ordinate the King's College London team and facilitated recruitment from clinical settings. She participated in all stages of the study conduct and report preparation.

Tracey Young (Senior Research Fellow, Health Economics and Decision Science, ScHARR, University of Sheffield) participated in the design and conduct of all stages of the research and report preparation.

Donna Lamping (Professor of Psychology, Department of Health Services Research and Policy, London School of Hygiene and Tropical Medicine) was the presiding genius behind the development of the DEMQOL system. She participated in the design and conduct of all stages of the research up to her very sad and untimely death. This report is dedicated to her.

Martin Knapp (Professor of Health Economics, Centre for the Economics of Mental Health, Institute of Psychiatry, King's College London) participated in the design and conduct of all stages of the research and report preparation.

Robert Howard (Professor of Old Age Psychiatry and Psychopathology, Department of Old Age Psychiatry, Institute of Psychiatry, King's College London) participated in the design and conduct of all stages of the research and report preparation.

SB and JB are the guarantors for the work reported here.

Disclaimer

This report presents independent research funded by the National Institute for Health Research (NIHR). The views and opinions expressed by authors in this publication are those of the authors and do not necessarily reflect those of the NHS, the NIHR, the MRC, NETSCC, the HTA programme or the Department of Health.

Publications

1. Mulhern B, Smith S, Rowen D, Brazier J, Knapp M, Lamping D, et al. Improving the measurement of QALYs in dementia: developing patient- and carer-health-state classification systems using Rasch analysis. Value Health 2012;15:323–33.

2. Rowen D, Mulhern B, Banerjee S, van Hout B, Young T, Knapp M, et al. Estimating preference based single index measures for dementia using DEMQOL and DEMQOL-Proxy. Value Health 2012;15:346–56.