2.1 Challenge 1: untying individual health forecasting from gradualist assumptions
A typical view of senescence is that individual health deteriorates with increasing age. Such a process can be described by mixed-effects regression methods through an intercept (i.e., baseline measure of health) and a slope (i.e., rate of health change over time; DiPrete and Eirich, 2006). Although these methods allow for individual variation in baseline levels (i.e., random intercepts) and in the rate of change (e.g., random slopes, quadratic terms, exponential functions), they operate under the assumption that health changes smoothly and gradually as individuals age. In doing so, forecasts from these methods do not align with empirical evidence demonstrating associations between cumulative disadvantage and late life outcomes, and this modeling approach may not fully capture within-person variability in health (Fig 1, survey data). In particular, the smoothing of mean health trajectories across social groups may not capture the effect of the social environment on individual health and consequent aging because this approach may underestimate the accumulation of poor health outcomes in socially disadvantaged persons (Engelman and Jackson 2019; Fig 1, mixed-effects).
In contrast to the gradualist assumptions of many models, empirical evidence suggests that human health can show periods of stability, slight deterioration, and then recovery from health insults (Bolano et al., 2019; Gill et al., 2010; Guilley et al., 2008; Keown, 2003). As life progresses through time, individuals may remain in the same health status (i.e., stasis) or transition among health states before death. That is, health does not decline gradually and homogeneously with no reversals. This highlights a crucial aspect of human health dynamics that accurate forecasting models must capture. In line with this, Engelman and Jackson (2019) proposed a new approach to health history forecasting by describing individual health trajectories as a punctuated equilibrium pattern where individuals experience periods of long-term stability interrupted by sudden changes in health status or mortality. The authors argue that while gradual approaches appeal to an intuitive reasoning about health change, such a modeling choice produces a mean health change that is not representative of the changes experienced by most of the individuals comprising the population.
One way to capture such health dynamics are multi-state models (Namboodiri and Suchindran, 2013; Schoen, 2006) and sequence analysis (Abbott and Tsay, 2000). These models describe discrete health states and transition probabilities among these states, and more accurately capture the health trajectories observed in real populations (Engelman and Jackson, 2019). We also argue that multi-state models describing changes across age as a Markov process may also provide a much-needed reconciliation between deterministic and stochastic processes when modeling individual health that would be otherwise obscured by the smoothing of health trajectories in gradualist approaches (Figure 1, multi-state).