Future research explored and summarized

Although suggestions for future research are continuously being noted, suggestions are summarized and some new ones are described here.

Exploring multi-morbidity of utility estimates

Though chronic conditions or long-term disorders are a major challenge for health care, “health systems are largely configured for individual diseases rather than multimorbidity… Better understanding of the epidemiology of multi-morbidity is necessary to develop interventions to prevent it, reduce its burden, and align health-care services more closely with patients’ needs” [205, 206]. Multi-morbidity is also important as it puts a large burden on patients; but also since it reflects the reality of actual disease better than a single-view focus on disease whenever it is within research, hospitals or the health-care system in general.

Little is known about estimating or combining utilities for comorbid (or ‘joint’) health states when the only data the analyst has access to the utilities of each condition from separate studies. Several joint health state prediction models have been suggested (for example, additive, multiplicative, best-of-pair, worst-of-pair, minimum and adjusted decrement estimator methods as mentioned earlier etc.), but no general consensus has been reached [181, 186]. From a statistical point of view, the multiplicative model recommended by Ara and Brazier [186] may be the best practical approximation, but not necessarily in line with actual data and reality.

However, paper 3 of present study is not linear, nor multiplicative, but actually incorporates multi-comorbidity in the utility estimates in better alignment with the data due to the unique properties of the ALDVMM regression model; in addition, the enhanced possibilities in regard to multi-morbidity is facilitated by the fact that as much as 199 conditions and subgroups are uniquely included within a single study. These improvements are significant as the results may provide a solution to issues of multi-morbidity.

Future studies should explore the properties of comorbidities and their utility estimates of paper 3, and, if possible, identify systemic patterns or different clusters of conditions and their relationships to a get better understanding hereof. If possible, they could suggest new and differentiated joint health state prediction models based on the findings (may for example be done by simulation based on the earlier mentioned Stata .ster file). The inclusion of all the conditions within a single study gives future studies a unique opportunity as many other “joint health states”

studies are based on a limited number of conditions often from different studies.

Although it may be difficult to identify general patterns given the large number of combinations of the different sets of comorbidities, future studies might identify some patterns within large clusters of for example chosen common diseases.

The register-based definitions of 199 chronic conditions

In general, the register-based framework of 199 conditions can be applied onto an unlimited number of outcomes, including costs, HRQoL and others also mentioned below, as long as the conditions of interest are linked at the micro level to the outcome of interest. Moreover, the framework of definitions can be used by everybody, including health-care administrators within hospitals and governments, and insurance registers, as long as the data contain the ICD-10 codes, medication ATC codes and other variables defined within the framework. Therefore, there is an almost unlimited number of future studies derived from the framework.

Future studies should further validate the register-based definitions based on, for example, comparisons with medical records. This is crucial since there are known diagnostic issues and controversies affected diagnostics such as, for example, shown within depression or similar conditions such as ME/CFS as mentioned [204, 207].

Future studies could also explore clusters of conditions using the registers, in order to support for example prevalence and incidence studies in regards to the importance of multi-comorbidity as mentioned.

Finally, future studies could explore the possibilities and implications of combining SR and RR methods for identifying chronic conditions. For which conditions could this be beneficial or otherwise? Some register conditions are known for not providing valid prevalence estimates, such as diseases like cataracts and tinnitus, for example, which is why there may be some precision to gain for some conditions. The current study has already shown that the two often competing methods may complement each other in certain areas, and that RR-based conditions also have limitations like SR, although the use of SR conditions depends on the purpose of the study and the need for precise diagnosis etc.

Monitoring trends in disease prevalence and incidence

Future studies could explore trends in prevalence – and incidence – rather than the single-point estimate of prevalence provided here, using the register-based framework of 199 chronic conditions. However, these studies should also take the growing number of diagnoses (“diagnostic inflation”) into account and explore whether this is a result of more disease or changes in diagnostic practices or the increasing productivity of country-specific diagnostics, as well, in contrast, possible under diagnosis of some diseases as debated [208–210]. Although the debate has been especially strong within the psychiatric field, an initial test of nationwide somatic ICD-10 codes also revealed an increasing number of somatic conditions by year.

Identifying and monitoring trends in costs and COI

Moreover, future studies could use the register-based framework for identifying the costs of health-care treatments and services by chronic condition.

Future studies could use the prevalence rates of the current study for generating COI [21]. Another approach for generating COI estimates is to use the register framework to identify individual costs and link them to diseases at the micro level within national registers or insurance data that contain the same data and variables defined within the framework. This will most likely enhance precision, but may require some work.

Monitoring trends within costs and COI for chronic conditions could inform decision-makers, but also be of interest to researchers making recommendations and forecasting health-care expenditures or disease burdens.

Generating ratios of costs and EQ-5D/QALY – cost pr. perfect health or QALY

One simple use of the framework and estimates is dividing the mean cost of a condition over (preferable adjusted) EQ-5D essentially generating mean cost pr.

perfect health of each condition for comparison. This would improve the information of traditional COI - expectably easier to generate than CEA as many health care professionals have information on costs – although not a CEA. Future studies could generate national ratios, as well is this might be done by health-care professionals in local municipalities where cost, but not HRQoL, are usually known. However, it is recommended that future studies also explore implications of use for decision makers as it is not a traditional CEA comparing alternatives and interventions; and that were applicable, QALYs are used instead of preference scores. This approach might provide a simply, quick overview with more information for policy and decision makers than common practice.

Improvement of EQ-5D HRQoL estimates and QALY

Future studies could explore the effects of regression interactions between risk factors etc. and conditions, as well as interaction with clusters of multi-morbidities that may influence results. Previous reference studies have recommended that complex epidemiological methods are used for linking conditions and health risk etc. [54] This could also be done in cohesion with reviewing the framework of definitions. There may be different pathways and methods both overlapping and accounting for this.

Future studies could also generate a catalogue of QALYs for the 199 conditions, thereby including mortality in the equation of the burden estimate.

Finally, future studies might incorporate multiple imputations and chained equations along with the ALDVMM as this cannot be done at the present time.

However, the gains thereof may be limited, although it cannot be ruled out that it may have a smaller impact on regression model 4, that is the model with the most missing.