With limited accurate and detailed healthcare data that are not easily available, analyses often rely on more accessible information such as arithmetic average values for the group under study. That approach assumes group homogeneity in all characteristics for which average values are applied. If this assumption is invalid, the summary results of cost and/or Quality Adjusted Life Year (QALY) gain could then be inaccurate with the reality. Other ways are needed to look for estimate of the uncertainty in the outcome results instead of using averages. It is however often unknown whether non-homogeneity in the data causes a substantial difference in the outcome results compared with the assumed homogeneity. To explore that potential size of the difference, this analysis here shows an approach using the extended sensitivity analysis tool, called the ESAP-plan, that may better help understanding what is at stake. The variables selected and studied are known factors causing potential non-homogeneity in the group under study, such as demographic age distribution and infectious disease spread. These points have been mentioned in the literature, but few publications have evaluated the consequences of not assessing non-homogeneity in the data analysis. Non-homogeneity could be of little concern if the variable spread is well balanced in the group, or when the numbers to evaluate are fixed – as illustrated in Step 1 and in Step 6 (linear regression). However, if some or all the variables may have unequal or unbalanced distributions, the overall summary cost estimate may be heavily skewed, as a non-homogeneous factor may become especially critical if it is linked to other unbalanced variables in the group, such as age linked to infection rate, frailty level, disease severity, and cost.
The summary Table 3 illustrates some interesting features of the complete analysis of the ESAP about issues that should be further investigated in a real-life setting. First, it is important to be clear about the reference condition of homogeneity selected for the comparison. One should define upfront whether a full (X1a) or partial (X1b) homogeneity condition is selected for that comparison. This selection may heavily influence the over- or under-estimation of the overall outcome measured with a non-homogeneous spread as indicated in Step 4 versus Step 5 in this exercise. Step 4 has selected a full homogeneous condition of comparison in X1a, leading to a marginal negative net cost-difference (4%) for the non-homogeneous condition, whereas in Step 5 a partial homogeneous condition is selected to compare with, and suddenly the net cost-difference is largely positive (29%) for the same non-homogeneous condition. Second is to check the age-specific demographic change in the study population under study linked to the rate of infectious disease increase by age. That combination results in a bell-shaped frequency of the overall cost as shown in Figs. 2 & 3. The bell shape will be more pronounced when the age-demographic and the disease data are more non-homogeneously spread by age. The third item to consider is the link between the distribution of disease severity levels by age and the cost per severity level. If there is no much of a difference in cost by disease severity level to be expected, then limited effort should be spent to obtain more precise overall costs than the homogeneous dataset.
When the full homogeneity situation has been selected, the results of this study indicate that the range of relative value changes in the overall cost estimates for the disease management can reach a maximum of 15% in the context of extreme situations of demographic age distribution, disease rate increase with increasing age, disease severity distribution by age group, and a high multiplication factor for the high cost in severity level (data not shown in the figures). When less pronounced distributions are considered, the relative cost difference between the fully homogeneous assumption of X1 and the non-homogeneous condition of X2 is likely to be between 2.5–6% overall. It indicates that the level of deviation in the cost summary results is not as large as often suspected. The difference change is limited because the constraints, defined up-front for this analysis, impose strict boundaries on the evaluation. For instance, non-homogeneity of the numbers by age groups is restricted and auto-correlated in the setting defined by the values in the prior age-class and the post-age group using a smooth curve design. This seems reasonable unless catastrophic events may heavily disturb temporarily the age distribution, such as war or natural catastrophe where suddenly many people of a specific age group are lost from the population. Other interesting features, identified through this ESAP, demonstrate the complexity of the problem. Demographic and disease spread alone don’t create a cost difference unless linked to disease severity levels and cost changes by disease severity levels.
An element of concern is the effect of non-homogeneous factors that are unbalanced in the opposite direction across the group, with strange consequences for the summary assessment. For example, with an ageing population the decline in numbers of individuals in progressively older age-groups, due to the increase in mortality with age, is not programmed as a gradual linear decrease with age, but instead follows an accelerated course causing a highly unbalanced age distribution in the overall study group. Infection spread moves in the opposite direction, with higher prevalence rates in the older groups because of worsening health condition with increasing age. There is a perception that ageing induces an overall healthcare cost increase [16, 17]. The example here, estimating overall healthcare costs of infectious disease management in ageing adults, indicates the opposite. It shows that overall cost may be lower when adjusted for non-homogeneity, compared with the overall cost of a study group when assumed to be fully homogeneous. This happens when the age structure is heavily unbalanced (fewer very old people), which imposes a lower absolute number of highly severe and costly treatments, despite having proportionally much more severe disease present in those older age-classes. This lower overall cost estimation may seem counterintuitive, but Fig. 6 shows it is possible. If, however, the age imbalance is marginal (Figs. 7 & 12), which is likely to be the development over time as the whole population lives longer, the problem of infectious diseases with more severe cases in the older groups could increase the overall cost above the homogeneously assumed estimates. Managing infectious disease in ageing people could therefore become a serious threat to tackle to help control healthcare cost increases over time.
Another surprising finding is that higher costs for treatment of more severe cases may not result in an obvious change in the cost difference between an assumed homogeneous evaluation and an adjusted non-homogeneity evaluation, as indicated in Fig. 10. The result could move in the opposite direction, with a higher negative cost difference with higher cost for the more severe cases, because the average cost in the homogeneous situation X1 also increases (Fig. 11). Also, changes in demographic composition may take time before a substantially higher cost is observed when substantially more people are living longer, as Fig. 12–13 indicates [18].
Having highlighted the issues of non-homogeneity with a hypothetical example using the ESAP, the results indicate the information that would be needed for developing more accurate estimates. Detailed demographic data are usually collected by national institutes of statistics at country level. However, infectious diseases are often neglected and not monitored or registered precisely and systematically across age groups. The recent COVID-19 pandemic highlighted the importance of measuring details of infection spread among different groups.
This analysis did not include all possible factors that could have an unbalanced spread across the population group, as it would have been too complex to model non-homogeneity if all known factors had been considered. However, the following additional elements could be considered that might influence the overall cost results with their unequally distributions across the group: sex; health condition expressed as the level of co-morbidities present in numbers and severity; frailty and disability; place of living (home, nursing homes, service flats); and hospitalisation. Regarding frailty, a recent review has shown the bi-directional movement of infection influencing the frailty condition of the individual and vice versa [19]. That may complicate the correct assessment of total infectious disease cost burden and the impact estimate of new interventions on that particular health condition like vaccination [20]. It is known and reported that frailty increases exponentially with aging which may justify the exploration here done of exponential graphs for infectious diseases [21]. The better knowledge about frailty that increases with time, allows for considering more appropriate and efficient prevention programs in ageing adults [22].
Regarding hospitalisation, infectious disease costs for hospitalisation are considerable and the cost differences by type of infection in hospital care could be large. This could potentially cause a high impact on the estimates of the overall infectious disease cost if these differences are not accounted for, and this was not evaluated in this analysis [23, 24]. Finally, another point not considered in this analysis is disease seasonality. This is particularly relevant for respiratory diseases, the most important type of infection in ageing adults [25]. Seasonal effect is an important non-homogeneous factor influencing good management of hospital beds across the year. It can severely impact the quality of care in hospital disease management, as reported for infectious diseases in children [26].
Non-homogeneous analysis of the data, such as that presented here, may indicate a different assessment of importance of the cost and the need for good management of the healthcare problem of infectious disease in older adults. Infectious diseases may spread beyond the initial cases and may harm many others during a considerable period. They become a serious threat when they accumulate, particularly in costly environments such as hospital settings. It is there that they cause most damage to society, although many infections could be avoided through prevention. Non-homogeneous analysis may capture more accurate and more detailed evaluations that better help understanding the costs of infectious disease, and consequently the potential benefits of preventive interventions.
The analysis presented here has some obvious limitations, as the objective of the simple hypothetical example used was to illustrate the use of the ESAP method to indicate potential impact on the differences between homogeneous and non-homogeneous analysis types. However, economic evaluations often work with simple models. A check against real-world data may give a more nuanced picture than indicated by the present simple data analysis. Future studies may go beyond this analysis, using more sophisticated models that could also capture the effect of seasonality and other variables not considered in this analysis, and the indirect effects of new interventions such as vaccination.
It is important to choose the most appropriate data analysis to obtain the most accurate estimates of the potential health and cost gains from preventive interventions. Ultimately, the big challenge concerns the next steps after evaluating the costs of disease management, including prevention strategies to support healthy ageing[27]. Understanding how non-homogeneity in variable categories may skew reported results could potentially help researchers to provide better modelling and evaluations of the healthcare cost data, closer to the real-world situation.