The potential value of person-centred statistical methods in ageing research Free

Key points

• Repeated latent class analysis (LCA) was used to identify the distinct latent ageing trajectories of older Australian women.

• The cohort underwent repeated assessments over a period of 20 years with surveys investigating measures of successful ageing.

• LCA identified a six-class model, with the classes characterised by the presence of disease and disability, and longevity.

• There were a large number of sociodemographic and lifestyle factors independently associated with class membership.

• LCA might allow targeted, cost-effective, intervention strategies to prevent disease and disability.

The process of human ageing in modern society is characterised by a significant inter-individual variability in clinical and demographic characteristics, disease burden, cognitive and functional capacity and social circumstances. The increasing availability of large population datasets has facilitated the epidemiological study of the factors that are associated with different ageing trajectories, such as ‘successful ageing,’ presence of chronic disability and premature mortality [1]. Issues with the definition of ‘successful ageing’ notwithstanding [2], conventional statistical methods including prediction models, typically assess the independent associations of individual variables with a predefined end-point. As such, they can be considered as ‘variable-centred’ approaches, with the focus being on the independent associations of each variable in the regression model on the outcome. This traditional approach contrasts with various clustering methods, including latent class analysis (LCA), that attempt to identify two or more subject groups, with each group having similar characteristics, e.g. patients with a similar (although not necessarily identical) set of comorbidities and physical function. This ‘patient-centric’ statistical approach, which allows identification of distinct patient groups, can be particularly useful in highly heterogeneous populations, and within healthcare, can help facilitate the development of targeted, cost-effective, intervention strategies to prevent disease and disability, particularly amongst those identified as highest risk. LCA is a ‘model-based’ clustering approach that relies on defined criteria such as the Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) to indicate the optimal number of clusters within the data. It is widespread within the social science disciplines where identification of different patterns of social behaviour is required. For example, latent classes of drinkers with differing alcohol related problems, latent classes of feared situations within social anxiety disorder and apathy subtypes in the older general population [3–5]. While other clustering methods, including for example k-means clustering can also be used, they suffer from the lack of any formal statistical measure to allow objective assessment of the optimal number of groups or classes. As such, the user may therefore quite arbitrarily define the final number of groups, with a greater potential for identifying spurious (i.e. chance) subgroups that cannot be validated in future datasets. LCA is, therefore, an efficient data reduction technique that allows the objective identification of underlying patterns within data, with each pattern defined by virtue of having a similar probability of class membership across a set of observed categorical variables [6]. In this issue of the journal, Byles et al. studied 12,432 participants of the 1921–1926 birth cohort of the Australian Longitudinal Study of Women’s Health, using survey data between 1996 (age 70–75 years at baseline) and 2016, with six waves of three-yearly surveys (1996–2011) and eight waves of 6-monthly surveys (2012–2016) [7]. Repeated measures LCA (RMLCA) was used to identify different ageing trajectories (latent classes), according to the presence of specific self-reported disease states, predefined physical functioning score thresholds, and measures of assistance with daily tasks, assessed during each survey. Following the creation of RMLCA models that included between one and eight latent classes, a six-class model was chosen based on the AIC and BIC scores, acceptable sample size for each class, and meaningful interpretation. The six classes included ‘successful agers’ (5.5%; relatively long survival without chronic disease, disability, or need for help until the age of 82–87 years), ‘managed agers long survivors’ (9.0%; relatively long survival with progressive increase in chronic disease and, to a lesser extent, disability), ‘usual agers long survivors’ (14.9%; relatively long survival with progressive increase in disability and, to a lesser extent, chronic disease), ‘missing surveys’ (18.3%; similar to usual agers and long survivors but with a higher rate of missing surveys after the age of 79–84 years), ‘usual agers’ (26.6%; sharp increase in chronic disease and disability after the baseline survey and a significant increase in mortality after the age of 82 years), and ‘early mortality’ (25.7%; about a quarter died by the age of 82 years) [7]. Multivariable multinomial logistic regression showed a number of associations between membership for each individual class and a range of baseline clinical, demographic, and social characteristics. In particular, using the ‘usual agers’ class as reference, ‘successful agers’ were less likely to be widowed (or divorced, separated, never married), have a lower education, be ex-smokers or current smokers, overweight or obese, and physically inactive and more likely to have social support. By contrast, the ‘early mortality’ class was more likely to experience difficulties with managing income, be ex-smokers or current smokers, be physically inactive, and less likely to have social support [7]. Elements of novelty in the study by Byles et al. are represented by the use of LCA to identify longitudinal ageing patterns in a large population dataset and the consideration of early mortality in class determination. Potential limitations include the lack of a comprehensive analysis of between-group differences in key biochemical parameters (e.g. C-reactive protein and glomerular filtration rate), disease states (e.g. dementia and depression) and prescribing of specific drug classes (e.g. anticholinergic and sedative drugs) that have shown significant associations with frailty, disease states, disability, and premature mortality [8–12]. Furthermore, the study of specific disease states based on patient self-assessment would benefit from corroborating clinical evidence. The latter could also help to determine the severity of a specific disease and whether the latter is controlled by means of pharmacological and/or non-pharmacological treatment. These issues notwithstanding, the study by Byles et al. provides useful additional knowledge regarding the applicability of patient-centred statistical approaches such as LCA for the identification, at the population level, of older age subgroups that are characterised by specific ageing trajectories. This information could be used to identify whether individual class membership is also predictive of important clinical end-points, in addition to morbidity and mortality, such as increasing access to specific health care services, hospital readmission rates, transition to intermediate or long-term residential care, and, potentially, patient-centred outcomes that are based on measures of physical and cognitive function, well-being, and social support [13]. In this context, further studies are required to investigate the role of patient-centred statistical approaches in facilitating the development of new, safe, and effective intervention strategies that increase the proportion of people in the ‘successful ageing’ trajectory.

Declaration of Conflicts of Interest:

None.

Declaration of Sources of Funding:

None.

References