Associations of stress and stress-related psychiatric disorders with GrimAge acceleration: review and suggestions for future work

The notion of “biological aging” as distinct from chronological aging has been of increasing interest in psychiatry, and many studies have explored associations of stress and psychiatric illness with accelerated biological aging. The “epigenetic clocks” are one avenue of this research, wherein “biological age” is estimated using DNA methylation data from specific CpG dinucleotide sites within the human genome. Many iterations of the epigenetic clocks have been developed, but the GrimAge clock continues to stand out for its ability to predict morbidity and mortality. Several studies have now explored associations of stress, PTSD, and MDD with GrimAge acceleration (GrimAA). While stress, PTSD, and MDD are distinct psychiatric entities, they may share common mechanisms underlying accelerated biological aging. Yet, no one has offered a review of the evidence on associations of stress and stress-related psychopathology with GrimAA. In this review, we identify nine publications on associations of stress, PTSD, and MDD with GrimAA. We find that results are mixed both within and across each of these exposures. However, we also find that analytic methods — and specifically, the choice of covariates — vary widely between studies. To address this, we draw upon popular methods from the field of clinical epidemiology to offer (1) a systematic framework for covariate selection, and (2) an approach to results reporting that facilitates analytic consensus. Although covariate selection will differ by the research question, we encourage researchers to consider adjustment for tobacco, alcohol use, physical activity, race, sex, adult socioeconomic status, medical comorbidity, and blood cell composition.

In the next step, we depict the behavioral mediators, based on the Oblak et al., 2021 review. This list of behavioral mediators is not intended to be exhaustive, but rather is a practical starting point based on current literature. Of note, the direction of the depicted relationships remains controversial, and some of variables may instead be confounders. See Supplementary Materials for more. In this simplified DAG, the distinction between total (all arrows), direct (blue arrow) and indirect (grey arrows) causal effects is made clear. However, because stress and psycho-pathology are complex exposures, researchers must still define what is meant by "direct" and "indirect" effects. Here, we consider effects mediated by known health-related behaviors to be indirect, while biological mechanisms are direct.
Comorbid medical illness is a challenging variable to address in analysis of stress/psychopathology and GrimAA. We depict it as a mediator, with suspected mechanisms including autonomic, neuroendocrine, and immune changes. However, medical illness is also a risk factor for stress and psychopathology, and therefore could be depicted as a confounder.
Studies of GrimAA commonly adjust for blood cell composition because the 2 nd -generation clocks are associated with changes in composition. However, changes in blood cell composition may also be a mechanism of clinical and scientific interest. Whether or not to adjust for blood cell composition depends entirely on the precise research question. Generally, providing both unadjusted and adjusted estimates of causal effect will be valuable.
In the next step, we consider the complex web of confounders that influence stress, risk of psychopathology, GrimAA, and the behavioral intermediates. For simplicity, here we show the five confounders considered without their relationships to other variables. There are also many other variables that could considered, such as environmental exposures and adverse childhood experiences.
As we add the causal relationships of the confounders to other variables, the DAG becomes complex. The mediators are now also part of confounding paths, as indicated by the pink rather than green lines. As DAG complexity increases, it becomes difficult to determine the covariates that must be adjusted for. To address this, dagitty.net provides a "minimally sufficient adjustment set" (MSAS) for both total and direct causal effects. A single DAG can have multiple MSAS options, and an MSAS can be selected based on the available data, most practical variables to measure, etc.

POSSIBLE FUTURE ADDITIONS
DAGs are evolving theory-based representations of a causal system, where the variables included and their relationships with one another change with new data and understanding. NOTE that this figure is hypothetical and is not intended to depict the current state of the field.
Research Question

Adjusting for Smoking Exposure
As noted in the main text, we depict smoking exposure as a mediator of the association between stress and stress-related psychopathology and GrimAA. As detailed in Ziedonis et al., the causal link of PTSD and MDD with smoking could be due to changes in HPA axis function following stress and trauma exposure, attempts to regulate negative emotional states associated with anxiety disorders, altered reward processing related to smoking, and increased vulnerability to social pressures related to smoking, among other mechanisms 1 . However, it is also possible that smoking is a confounder of the relationship between stress and psychopathology and GrimAA. In this model, smoking exposure could either directly increase the risk of stress and stress-related psychopathology -perhaps through changes in neurophysiology -or some other genetic or environmental factor predisposes to both stress/psychopathology and tobacco use. Both of these scenarios are depicted in Supplementary Figure  1.
Of note, both scenarios require adjustment for tobacco use, but for different reasons. In the scenario where tobacco use is a mediator, adjustment is required only to yield a direct causal effect estimate. This was detailed in the main text. In the scenario where tobacco use is a confounder, adjustment is necessary to yield any valid estimate of causal effect, including total causal effect. Tobacco use may be part of a confounding path caused by shared genetic risk factors for stress/stress-related psychopathology and tobacco use. Adjusting for either tobacco use or genetic risk will close the confounding path. C. Tobacco use may be part of a confounding path caused environmental risk factors for both stress/stress-related psychopathology and tobacco use. Again, adjusting for either tobacco use or environmental risk will close the confounding path. However, tobacco use is more practically quantifiable than either genetic or environmental risk.
On Comorbid Somatic Illness -Selection Bias and "Adjusting for a Collider": In a DAG, a "collider" is any variable that is a common effect of the exposure and the outcome. When colliders are adjusted for, they can induce a spurious statistical association between the exposure and outcome when no true causal association exists. In other words, controlling for a collider by statistical adjustment, stratification, or sample restriction introduces bias. Identifying colliders and their impacts on statistical associations is perhaps the most complex foundational principle in DAG use. They are easiest to identify when the collider is directly caused by both the exposure and the outcome. However, collider stratification bias can also occur when the collider and the exposure and/or outcome are associated by other means (e.g., a fourth variable causes both the collider and the exposure, and a fifth variable causes both the collider and the outcome Importantly, "collider stratification bias" is the DAG representation of a much more familiar conceptselection bias 4 . Selection bias occurs anytime that selection into a study is related to both the exposure and the outcome of interest. "Selection into the study" is the collider in this case, and is by necessity stratified upon when we analyze only the subset of the target population who enrolled in the study. Several biomedical "paradoxes" have been attributed to inadvertent collider stratification due to selection bias (e.g., obesity paradox 5 ).
In the studies reviewed here, selection bias can arise from several sources (Supplementary Figure 1). In studies that employed exposure-selective sampling (e.g., PTSD patients and PTSD-free controls), we can imagine that incentives to volunteer for a study may differ between participants with and without PTSD. For instance, those with PTSD might be highly motivated to enroll in a study to contribute to science despite other barriers to participation, such as medical illness or low socioeconomic status. Meanwhile, PTSD-free controls might lack the same intrinsic motivation to participate, and those who do may be more likely to be somatically well and able to attend a study. However, it may be difficult to predict how patients and controls may differ and in which direction the bias is likely to be.

Supplementary Figure 3: Directed Acyclic Graph Depicting Collider Stratification Bias.
A collider is any variable that is a common effect of both the exposure and the outcome. Variables can also be colliders when there is another variable that is a common cause of both the exposure and the collider, or the outcome and collider. In this model, 'Medical Illness' is a common cause of both 'Selection into Study' and 'GrimAA,' while 'Adult SES' is a common cause of both 'Stress & Stress-Related Psychopathology' and 'Selection into Study.' When a collider is adjusted for, it introduces bias into the analysis. In the case of selection bias, 'Selection into Study' is conditioned upon because investigators are only able to study the stratum of individuals who enrolled in the study and for whom data was collected. This type of selection bias is often difficult to detect without a DAG, and maybe be a challenge for observational studies relying on research volunteers.