Table 1 Illustrative listing of causal inference methodologies that could support deeper study of the association between adversity, stress, trauma, and accelerated biological aging.

Directed acyclic graphs (DAGs)

DAGs provide a visual illustration of causal assumptions and assist in identifying variables that need statistical control and situations where controlling for specific variables may worsen causal inferences [196].

A DAG representation of the potential confounding influences from early life adversity to accelerated aging in midlife is used to identify collider variables that artificially strengthen the association.

Propensity score analyses (PSA)

In some situations, stress exposures are random; in other situations, especially in cases of adversity, life event exposures are non-random. Propensity scores allow researchers to condition on a set of covariates that are used to predict exposure to the traumatic experience. The resulting propensity scores can then be used as a means of matching exposed vs. non-exposed groups, or by weighting participants in terms of their likelihood of being exposed to a traumatic experience [170, 197].

In a longitudinal study of combat-exposed veterans, propensity scores are used to examine the association between PTSD diagnoses and accelerated aging. Study participants are matched on their likelihood to experience PTSD. The results show that without propensity score matching (PSM), PTSD is associated with accelerated epigenetic aging. After PSM, the association between PSTD and accelerated epigenetic aging is not reliably different from zero.

Instrumental variable (IV) analysis

IV analysis is a method that helps account for unmeasured confounding in observational analyses. (Thus, whereas PSA methods account for measured confounders, IV methods account for unmeasured confounders.) IV estimation is based on a series of assumptions, the chief of which is that the instrument variable associated with the exposure is not associated with unmeasured confounds (see [198]). Increasingly, genetically-informed Mendelian randomization methods are used in health sciences as a form of IV analysis [199].

A group of investigators use Mendelian randomization methods to study the bidirectional association between PTSD and a methylation-based pace of aging outcome. The results reveal significant associations from the pace of aging to PSTD but not the reverse, which raises questions about the directionality of extant associations in the literature.

Natural experiments

Natural experiments occur when an event or condition affects a specific sub-population, creating a situation that is analogous to a randomized experiment and thus addresses many of the confounds inherent observational research [200]. There are many potential natural experiments in the stress-health literature, but the key is to ensure that the stress exposure is indeed random—e.g., comparing counties exposed to chronic stress of environmental contamination may seem like a natural experiment, but the question of whether such contaminations are random is open to debate.

To the extent that government policies, environmental emergencies, or crises vary randomly, a study can address whether people living in counties or cities exposed to such stressors (vs. those who are unexposed) show accelerated biological aging. In many situations, the naturally-occurring event can be considered an instrumental variable—e.g., see [201] for a study of the Vietnam lottery draft as a means of estimating the causal effect of veteran status. In this case, the participants are drafted randomly and the lottery status is an IV for veteran status.

Target trial emulation studies

Based on counterfactual theory, target trial emulation (TTE) methods attempt to recreate the conditions of a randomized clinical trial within observational data [202]. A key step in the TTE method is matching eligibility criteria within an observational data set, then following these people through some type of exposure to a pragmatic trial.

TTEs involve [1] identifying a causal question from observational data in the form of a hypothesized RCT protocol, and [2] emulating the components of the protocol. For example, in a longitudinal research study of perceived stress, match participants in a manner consistent with an RCT protocol, then evaluate if changes in perceive stress are associated with changes in accelerated aging (see [203] for details on implementation).

Sibling-comparison designs

Among a suite of quasi-causal within-family models, sibling-comparison designs offer a power test of causal life event exposure and other environmentally-mediated hypotheses [204]. Genes and environments are highly confounded—e.g., as a stress exposure, divorce is non-random, and thus studying its putative consequences (e.g., in the association between divorce exposure and biological aging) cannot rule-out the confounding role of genetics. Sibling-difference analyses represent a form of random genetic assignment during meiosis and can be a useful tool in assessing the environmental impact of within-family differences in stress exposures.

In a large longitudinal study that includes full biological siblings, researchers investigate whether exposure to chronic interpersonal stress and strain is associated with accelerated biological aging. The sibling comparison model involves creating a within-family average of interpersonal stress and strain, then centering each sibling’s score on this metric relative to the family-level average. The main hypothesis centers on the extent to which siblings with elevated stress and strain scores (relative the family average) show more accelerated epigenetic aging.

Cotwin control designs

Among a suite of quasi-causal within-family models, cotwin control designs provide another method that helps account for genetic confounding [191]. Cotwin control designs take advantage of the fact the monozygotic (MZ) twins share identical genotypes and when compared with dizygotic (DZ) twins, who share only 50% of their genetic material, this yields a biometric model that can help rule out genetic confounding in stress exposure to health outcome associations [192].

In a sample of MZ and DZ twin pairs, researchers study whether differential exposure to high levels of subjective psychological stress is associated with prospective changes in accelerated biological aging. Within this design, researchers examine whether, after accounting for genetic influences in the stress-aging association, within twin-pair differences in chronic psychological stress remains a significant predictor of accelerated biological aging.

G-Computation

The nonparametric g(general) formula for estimating causal effects has many applications and is based on estimating the probability of an outcome under different hypothesized conditions based on different sets of control variables [170, 172, 205]. G-computation is based on counterfactual theory and the conditional probabilities of the outcome under different exposure conditions.

Researchers are interested in the potential age acceleration consequences of caregiving stress, and outcome groups are created to classify participants into groups of people who do and do not show epigenetic age acceleration. Covariates include gender and a history of adverse childhood experiences. G-computation is completed on the marginal means to compare the age acceleration of those exposed vs. unexposed to caregiving stress under the different control conditions. The difference of these average probabilities is then calculated to create an estimate of the causal effect of caregiving on age acceleration.