Raw PROMIS scores were transformed to normalized t-scores (https://www.healthmeasures.net/). Scores representing “no pain” or “no fatigue” were replaced with zeros to evaluate these outcomes appropriately using a two-part statistical model described below. Structural and social stressors disproportionately occur among minoritized populations, inhibiting our ability to disentangle race and racism. Thus, race and ethnicity variables were not included in the main analyses to not violate statistical principles (e.g., collinearity). These efforts align with our goal to thoughtfully reconsider attributing statistical onus on race, versus what social and health system structural factors contribute to outcomes for PwMS [20, 21].

Aligning with previous work that evaluated lifetime stressors and physical outcomes [14], our analytic approach aimed to assess fit of increasingly complex models that include blocks of related predictors (i.e., collinear variables representing latent constructs) to determine if their collective contributions improve model fit. Of note, this type of analysis focuses on establishing whether there is a relationship between the latent constructs (e.g., blocks) and the outcomes, and not necessarily the change or individual variable contributions because they could be underestimated. Successive models were compared to prior models using likelihood ratio (LR) testing, and Akaike Information Criterion (AIC) as an index of relative model fit, with lower AIC indicating better fit. If a block of predictors did not significantly improve the model, it was removed from final analytic modeling. Table 1 shows the predictors and covariates in each of the three blocks, with each model nested within the next higher-level model.

The base model encompassed covariates to determine baseline contributions. Model fit of the base model was then compared to Model 2 which added childhood stressor predictors to assess if these contribute over and above the base model. For Model 3, adult stressor predictors were added to determine if they contributed additional predictive variance. After evaluating contributions of blocks of related predictors, contributions of individual predictors were examined. Yet, it must be noted that using related stressor variables is a strength of block modeling but may also cause an underestimation of individual stressor contributions.

The specific types of hierarchical regressions included Poisson regression for the count of psychiatric morbidities outcome. The two PROMIS outcomes, pain interference, and fatigue, represented a mixed distribution therefore two-part modeling was used. Part 1 included a dichotomous (yes/no) component of experiencing any pain interference or fatigue utilizing logistic regression, followed by part 2 a normal distribution characterizing the magnitude of pain interference or fatigue utilizing OLS linear regression.

Base model predictors contributed to a significant overall two-part model estimating the likelihood and magnitude of pain interference (logistic regression pseudo R² = 0.2219, P < 0.0001; OLS regression R² = 0.1831, P < 0.0001; model AIC = 3,751) (Table 3). The childhood stressor block of predictors in Model 2 improved predictions significantly over the base model (logistic regression pseudo R² = 0.2448, P < 0.0001; OLS regression R² = 0.2152, P < 0.0001; model AIC = 3,719, LR P < 0.0001). Similarly, the adult stressor predictors in Model 3 contributed significantly more information over the prior nested modes (logistic regression pseudo R² = 0.2578, P < 0.0001; OLS regression R² = 0.2667, P < 0.0001; model AIC = 3,685, LR P < 0.0001).

Regarding individual predictors, childhood stress severity was significantly associated with the higher magnitude of pain interference (b = 0.33, P = 0.005) in Model 2 but lost significance when adult stress was added for Model 3, suggesting shared variance among child and adult stressors. In the final model, age impacted both the likelihood (OR = 1.02, P < 0.03) and magnitude (b = –0.10, P < 0.001) of pain.

The base model of predictors contributed to a significant overall two-part model estimating the likelihood and magnitude of having fatigue (logistic regression pseudo R² = 0.074, P < 0.04; OLS regression R² = 0.086, P < 0.0001; model AIC = 4,192) (Table 4). The childhood stressor predictors in Model 2 contributed a significant amount of variance over and above the base model (logistic regression pseudo R² = 0.11, P < 0.01; OLS regression R² = 0.14, P < 0.0001; model AIC = 4,160, LR P < 0.0001). While adult stressor severity was independently significant for the magnitude of fatigue (b = 0.073, P = 0.001), overall, the adult stressor predictors in Model 3 did not significantly contribute and reduced model fit (logistic regression pseudo R² = 0.136, P = 0.004; OLS regression R² = 0.242, P < 0.0001; model AIC = 4,844, LR P = 1.0). Thus, adult stressors were removed from the final analytic model, as only base covariates and childhood stressors correlated with fatigue.

In the final model, childhood stress severity was significantly associated with 24% higher odds of experiencing any fatigue for each increase 1-unit increase in severity rating (OR = 1.24, P = 0.03), and with the magnitude of fatigue (b = 0.47, P < 0.001). Interpreting this in context of the average childhood stress severity (9.8), this translates to the average PwMS in this sample being 235% more likely to experience fatigue. Childhood stressor count (b = –1.23, P < 0.04) and age (b = –0.07, P = 0.019) were both negatively associated with the magnitude of fatigue.

The base model contributed significantly to estimating the risk of accumulating psychiatric morbidity (R² = 0.061, P < 0.0001; model AIC = 2,560) (Table 5). The childhood stressor predictors in Model 2 significantly improved over the base model (R² = 0.09, P < 0.0001; AIC = 2,485, LR P < 0.0001). Similarly, the adult stressor predictors in Model 3 contributed significantly more information over the prior nested model (R² = 0.116, P < 0.0001, AIC = 2,420, LR P < 0.0001). Therefore, childhood and adult stressors both correlate with psychiatric morbidity for PwMS.

Regarding individual predictors, adult stress severity was significantly associated with psychiatric morbidity (IRR = 1.01, P < 0.0001). As adult stressor severity increased by 1-unit, the risk of having an additional psychiatric diagnosis or symptom increased by 1%. Interpreting that within the context of the average adult stressor severity in this sample, 55.3 (30.8) this translates to a 55% increased risk for the average PwMS, with nearly 31% more risk just one standard deviation away. In Model 2, childhood stressor severity carried nearly five times that risk, with a 4.8% increased risk of psychiatric morbidity for each 1-unit increase in severity rating, however lost significance when adding the adult stressors for Model 3, again suggesting shared variance. In the final model, the risk of psychiatric morbidity decreased by 2% for each year since MS onset (IRR = 0.98, P < 0.001).

To our knowledge, this is the first study to use a lifetime approach to comprehensively measure the effect of child and adult stressors on three invisible issues in MS, fatigue, pain interference, and psychiatric morbidity. Use of hierarchical block modeling allowed us to determine the overall contribution of similar stressor variables, count and severity, to better assess the latent concept of stress. Cumulative childhood stressors correlated to all three outcomes, while adult stressors additionally associated with pain interference and psychiatric morbidity (Table 6). This work aligns with the few previous studies which associated only ACE-focused childhood stressors with adult fatigue [6], pain catastrophizing [8], and mental health outcomes [7, 9, 10] in PwMS or immune-mediated inflammatory diseases. Expanding beyond childhood stressor literature, this current study also aligns with evidence suggesting increased adult and lifetime stressors relate to worsening MS outcomes more broadly; although, this literature is largely focused on physical clinical outcomes (e.g., disease onset, relapses, progression) [14, 25, 26]. Specifically, since only childhood stressors related to fatigue in our study, our findings do not align with evidence that adult adversity (i.e., adverse life events in the last 60 days) is associated with MS fatigue [27]. This divergence may stem from measurement differences (e.g., adult vs. lifetime). The current findings suggest that stressor severity may individually carry a more significant impact relative to stressor count for some outcomes; therefore, relying solely on count-based measures (e.g., ACEs) is not ideal. Similarly, evidenced by multiple instances of shared variance or contributions in our analyses, examination of childhood stressors without consideration of adult stressors is not ideal; therefore, a lifetime approach should be used when possible.

Omitting race/ethnicity from analyses is not optimal for discerning unique racial/ethnic experiences or differences across outcomes, yet may be necessary due to included stressors (e.g., discrimination) to avoid statistical issues. Since MS samples are largely White, even when studies do include race/ethnicity in analyses, small individual cell sizes typically lead to collapsing multiple categories into a dichotomous variable to abide by ethical/IRB reporting standards to protect participant identity and to create less statistical error variance. A recent review revealed that race/ethnicity was not accounted for in a third of the studies assessing childhood stressors and MS risk or features [5], leaving much room for improvement. More diverse samples are needed to address health disparities and inequities in MS research and treatment; and race and ethnicity should be included in future analyses when possible [28].

Our finding that each incremental increase in childhood stressor severity increased the odds of experiencing fatigue by 24% is noteworthy. Fatigue is the most common symptom experienced by PwMS yet remains one of the most challenging symptoms to treat, in part because of its personalized nature that can be influenced by stress. Meta-analyses highlight that mindfulness-based stress reduction approaches may be effective for fatigue, with mixed results for pain for PwMS [29], however, symptoms often co-occur thus broader outcomes should be considered. A recent study by Braley and colleagues found that telephone-based version cognitive behavioral therapy (CBT) for fatigue performed similarly to pharmacological treatment with modafinil [30, 31] in terms of fatigue impact reduction; however, combination therapy with both was associated with more global benefits based on the Patient Global Impression of Change (PGIC) score. Although the PGIC was a secondary outcome in this trial, the findings suggest that an interdisciplinary approach may offer the most benefit when considering a person’s perception of global function. Further, benefits of interventions for invisible symptoms may be best captured by multifaceted instruments that capture overall activity, symptoms, mood, physical, and social function, which themselves can be influenced by stress, and targeted with psychotherapy. Future work is needed to implement and evaluate the benefits of adjunct therapies, in tandem with standard MS care, on invisible symptoms and broader measures.

Machine learning research has been burgeoning, especially for healthcare applications such as clinical decision making and treatment optimization [32, 33]. Our findings may inform future work such as using stressor history as one of the parameters in supervised machine learning to help determine whether those with high childhood adversity may respond better to cognitive or pharmacological treatment of fatigue. Similarly, integrating stress informed machine learning decision tools may help optimize treatment for other invisible and physical symptoms. Evidence suggests that interventions including stress reduction, coping/resilience skills, and smoking cessation, are useful for symptom management (e.g., pain, fatigue, depression, disability) [34–37]. However, such research progress and clinician acceptance of these strategies hinge on health system parameters and infrastructure [38]. Additional implementation studies focused on provision of new services (e.g., stress reduction, coping, smoking cessation clinics), and/or protocols that facilitate increased referrals (e.g., screening, therapy, smoking clinics) are sorely needed to demonstrate feasibility and acceptability throughout various neurological settings to promote clinical buy-in and adoption of translational change.

As PwMS aged, the risk of psychiatric morbidity decreased by 2%, suggesting that PwMS may be most vulnerable at diagnosis but may learn to cope or feel more in control of their disease over time. This is somewhat supported by the finding that while the presence of pain increased with age, the magnitude of interference in daily life decreased. Interestingly, as both the count of childhood stressors and age increased, the fatigue magnitude decreased, which may suggest that PwMS who experienced more stressors may have already received mental health support and similarly learned to cope better over time. Alternatively, this may also indicate that PwMS who experienced high childhood stressors and potentially high coping skills, may be faring better and participating in research more than their counterparts. Since fatigue was the only outcome to which adult stressors did not have a significant relationship, yet evidence of this relationship has previously been shown to be mediated by resilience [27], there may be additional factors like coping and resilience that may have different mediating impacts across the three outcomes. Other complex considerations may be differing genetic contributions and intergenerational transmission of trauma. As this emerging area of research grows, more prospective and mechanistic work is needed to determine how resilience, coping, and other complex factors may mediate, moderate, or otherwise impact relationships between stressors and health outcomes in PwMS.

Causal inference cannot be determined with cross sectional data. Yet, as the first study to assess many of these lifetime relationships with an MS focused sample, it fills an important gap. Those who responded may have increased ability to take an online survey (e.g., technology access, less disability). A response bias may be present due to the high number of PwMS on the NMSS listserv (approximately 80,000). Self-reported retrospective data has potential for recall bias, therefore, it is recommended to collect self-reported data using measures which have been validated using a test-retest approach and perform well over time (e.g., STRAIN) [12]. Sensitive information such as stressors could have a social desirability bias and be under-reported. However, the online format and anonymity may have facilitated more accurate reporting compared to other formats. Our sample was highly educated and may not represent all groups. However, this sample aligns with traditional US MS research samples including other studies that used the NMSS listserv and captures the widest geographical range and largest sample size in this emerging area and thus bolsters the internal and external validity [17, 22]. Additional strengths include a wide range of covariates compared to other work in this area [5]. While the PROMIS measures allowed us to compare two outcomes against a healthy group, we did not compare stressor experience. Future studies would be more robust by using a design that allows for a true comparison against other populations of interest (e.g., healthy controls, similar chronic diseases).

These findings support an association between childhood stressors and pain interference, fatigue, and psychiatric morbidity; as well as an association between adult stressors and pain interference, and psychiatric morbidity for PwMS. Additional studies are needed to assist clinical efforts of trauma informed precision medicine and intervention efforts to mitigate stressor impact on PwMS. While pediatric MS is far less common, investigating stressor experience across different developmental stages may be helpful in evaluating MS outcomes in this sub-population. Future research should replicate this work with more diverse MS samples and expand to include other positive aspects (e.g., coping, resiliency, social support) with mediation analyses and additional clinical features (e.g., sleep, cognition, substance use).