Generalized causal mediation and path analysis: Extensions and practical considerations

Mediation Analyses of the Mechanisms by Which Socioeconomic Status, Comorbidity, Stroke Severity, and Acute Care Influence Stroke Outcome

BACKGROUND AND OBJECTIVES

Low socioeconomic status (SES) is associated with increased risk of death and disability after stroke, but interventional targets to minimize disparities remain unclear. We aim to assess the extent to which SES-based disparities in the association between low SES and death and dependency at 3 months after stroke could be eliminated by offsetting differences in comorbidity, stroke severity, and acute care.

METHODS

This nationwide register-based cohort study included all 72 hospitals caring for patients with acute stroke in Sweden. All patients registered with an acute ischemic stroke in the Swedish Stroke Register in 2015-2016 who were independent in activities of daily living (ADL) during stroke were included. Data on survival and SES the year before stroke were retrieved by cross-linkage with other national registers. SES was defined by education and income and categorized into low, mid, and high. Causal mediation analysis was used to study the absolute risk of death and ADL dependency at 3 months depending on SES and to what extent hypothetical interventions on comorbidities, stroke severity, and acute care would equalize outcomes.

RESULTS

Of the 25,846 patients in the study, 6,798 (26.3%) were dead or ADL dependent 3 months after stroke. Adjusted for sex and age, low SES was associated with an increased absolute risk of 5.4% (95% CI 3.9%-6.9%; p < 0.001) compared with mid SES and 10.1% (95% CI 8.1%-12.2%; p < 0.001) compared with high SES. Intervening to shift the distribution of all mediators among patients with low SES to those of the more privileged groups would result in absolute reductions of these effects by 2.2% (95% CI 1.2%-3.2%; p < 0.001) and 4.0% (95% CI 2.6%-5.5%; p < 0.001), respectively, with the largest reduction accomplished by equalizing stroke severity.

DISCUSSION

Low SES patients have substantially increased risks of death and ADL dependency 3 months after stroke compared with more privileged patient groups. This study suggests that if we could intervene to equalize SES-related differences in the distributions of comorbidity, acute care, and stroke severity, up to 40 of every 1,000 patients with low SES could be prevented from dying or becoming ADL dependent.

Complete effect decomposition for an arbitrary number of multiple ordered mediators with time-varying confounders: A method for generalized causal multi-mediation analysis

Causal mediation analysis is advantageous for mechanism investigation. In settings with multiple causally ordered mediators, path-specific effects have been introduced to specify the effects of certain combinations of mediators. However, most path-specific effects are unidentifiable. An interventional analog of path-specific effects is adapted to address the non-identifiability problem. Moreover, previous studies only focused on cases with two or three mediators due to the complexity of the mediation formula in a large number of mediators. In this study, we provide a generalized definition of traditional path-specific effects and interventional path-specific effects with a recursive formula, along with the required assumptions for nonparametric identification. Subsequently, a general approach is developed with an arbitrary number of multiple ordered mediators and with time-varying confounders. All methods and software proposed in this study contribute to comprehensively decomposing a causal effect confirmed by data science and help disentangling causal mechanisms in the presence of complicated causal structures among multiple mediators.

Applied causal inference methods for sequential mediators

BACKGROUND

Mediation analysis aims at estimating to what extent the effect of an exposure on an outcome is explained by a set of mediators on the causal pathway between the exposure and the outcome. The total effect of the exposure on the outcome can be decomposed into an indirect effect, i.e. the effect explained by the mediators jointly, and a direct effect, i.e. the effect unexplained by the mediators. However finer decompositions are possible in presence of independent or sequential mediators.

METHODS

We review four statistical methods to analyse multiple sequential mediators, the inverse odds ratio weighting approach, the inverse probability weighting approach, the imputation approach and the extended imputation approach. These approaches are compared and implemented using a case-study with the aim to investigate the mediating role of adverse reproductive outcomes and infant respiratory infections in the effect of maternal pregnancy mental health on infant wheezing in the Ninfea birth cohort.

RESULTS

Using the inverse odds ratio weighting approach, the direct effect of maternal depression or anxiety in pregnancy is equal to a 59% (95% CI: 27%,94%) increased prevalence of infant wheezing and the mediated effect through adverse reproductive outcomes is equal to a 3% (95% CI: -6%,12%) increased prevalence of infant wheezing. When including infant lower respiratory infections in the mediation pathway, the direct effect decreases to 57% (95% CI: 25%,92%) and the indirect effect increases to 5% (95% CI: -5%,15%). The estimates of the effects obtained using the weighting and the imputation approaches are similar. The extended imputation approach suggests that the small joint indirect effect through adverse reproductive outcomes and lower respiratory infections is due entirely to the contribution of infant lower respiratory infections, and not to an increased prevalence of adverse reproductive outcomes.

CONCLUSIONS

The four methods revealed similar results of small mediating role of adverse reproductive outcomes and early respiratory tract infections in the effect of maternal pregnancy mental health on infant wheezing. The choice of the method depends on what is the effect of main interest, the type of the variables involved in the analysis (binary, categorical, count or continuous) and the confidence in specifying the models for the exposure, the mediators and the outcome.

The Chasm in Percutaneous Coronary Intervention and In-Hospital Mortality Rates Among Acute Myocardial Infarction Patients in Rural and Urban Hospitals in China: A Mediation Analysis

Objectives: To determine to what extent the inequality in the ability to provide percutaneous coronary intervention (PCI) translates into outcomes for AMI patients in China.

Methods: We identified 82,677 patients who had primary diagnoses of AMI and were hospitalized in Shanxi Province, China, between 2013 and 2017. We applied logistic regressions with inverse probability weighting based on propensity scores and mediation analyses to examine the association of hospital rurality with in-hospital mortality and the potential mediating effects of PCI.

Results: In multivariate models where PCI was not adjusted for, rural hospitals were associated with a significantly higher risk of in-hospital mortality (odds ratio [OR]: 1.19, 95% confidence interval [CI]: 1.03–1.37). However, this association was nullified (OR: 0.94, 95% CI: 0.81–1.08) when PCI was included as a covariate. Mediation analyses revealed that PCI significantly mediated 132.3% (95% CI: 104.1–256.6%) of the effect of hospital rurality on in-hospital mortality. The direct effect of hospital rurality on in-hospital mortality was insignificant.

Conclusion: The results highlight the need to improve rural hospitals’ infrastructure and address the inequalities of treatments and outcomes in rural and urban hospitals.

Applications of conceptual models from lifecourse epidemiology in ecology and evolutionary biology

In ecology and evolutionary biology (EEB), the study of developmental plasticity seeks to understand ontogenetic processes underlying the phenotypes upon which natural selection acts. A central challenge to this inquiry is ascertaining a causal effect of the exposure on the manifestation of later-life phenotype due to the time elapsed between the two events. The exposure is a potential cause of the outcome-i.e. an environmental stimulus or experience. The later phenotype might be a behaviour, physiological condition, morphology or life-history trait. The latency period between the exposure and outcome complicates causal inference due to the inevitable occurrence of additional events that may affect the relationship of interest. Here, we describe six distinct but non-mutually exclusive conceptual models from the field of lifecourse epidemiology and discuss their applications to EEB research. The models include Critical Period with No Later Modifiers, Critical Period with Later Modifiers, Accumulation of Risk with Independent Risk Exposures, Accumulation of Risk with Risk Clustering, Accumulation of Risk with Chains of Risk and Accumulation of Risk with Trigger Effect. These models, which have been widely used to test causal hypotheses regarding the early origins of adult-onset disease in humans, are directly relevant to research on developmental plasticity in EEB.

Post-Treatment Confounding in Causal Mediation Studies: A Cutting-Edge Problem and A Novel Solution via Sensitivity Analysis

In causal mediation studies that decompose an average treatment effect into indirect and direct effects, examples of post-treatment confounding are abundant. In the presence of treatment-by-mediator interactions, past research has generally considered it infeasible to adjust for a post-treatment confounder of the mediator-outcome relationship due to incomplete information: for any given individual, a post-treatment confounder is observed under the actual treatment condition while missing under the counterfactual treatment condition. This paper proposes a new sensitivity analysis strategy for handling post-treatment confounding and incorporates it into weighting-based causal mediation analysis. The key is to obtain the conditional distribution of the post-treatment confounder under the counterfactual treatment as a function of not just pretreatment covariates but also its counterpart under the actual treatment. The sensitivity analysis then generates a bound for the natural indirect effect and that for the natural direct effect over a plausible range of the conditional correlation between the post-treatment confounder under the actual and that under the counterfactual conditions. Implemented through either imputation or integration, the strategy is suitable for binary as well as continuous measures of post-treatment confounders. Simulation results demonstrate major strengths and potential limitations of this new solution. A re-analysis of the National Evaluation of Welfare-to-Work Strategies (NEWWS) Riverside data reveals that the initial analytic results are sensitive to omitted post-treatment confounding. This article is protected by copyright. All rights reserved.

Sex differences in the association of vital exhaustion with regional fat deposition and subclinical cardiovascular disease risk

OBJECTIVE

Vital exhaustion (VE) is more strongly associated with cardiovascular disease (CVD) risk for women than men. This study examined whether sex differences in associations of VE with CVD risk markers are accounted for by unique associations of VE with regional adiposity.

METHODS

The study enrolled 143 persons (18-55 years) without diagnosed conditions. VE was assessed by the Maastricht questionnaire. CVD indices were measured using the euglycemic-hyperinsulinemia clamp, resting blood pressure, and blood draws. Regional adiposity was measured using computed tomography and 2-D echocardiography. This cross-sectional study employed a path analysis approach, including relevant covariates.

RESULTS

Of the cohort, aged 38.7 ± 8.4 years, 65% were men, and 41% were obese. The final model had excellent fit (χ2(36) = 36.5, p = .45; RMSEA = 0.009, CFI = 0.999). For women, but not men, the model indicated paths from VE to: 1) lower insulin sensitivity (B = -0.10, p = .04), and higher total cholesterol to HDL ratio (B = 0.12, p = .09), triglycerides (B = 0.10, p = .08), and C-reactive protein (B = 0.08, p = .09) through visceral adiposity; 2) higher mean arterial pressure (B = 0.14, p = .04), lower insulin sensitivity (B = -0.09, p = .08), and higher C-reactive protein (B = 0.12, p = .07) through subcutaneous adiposity; 3) lower insulin sensitivity (B = -0.07, p = .08) and higher total cholesterol to HDL ratio (B = 0.16, p = .03) through liver adiposity; and 4) higher C-reactive protein (B = 0.08, p = .09) through epicardial adiposity.

CONCLUSION

Results extend prior evidence by showing that the association of VE with CVD risk in women is linked with specific regional adiposity elevation. Further study of adiposity-related mechanisms in women who experience early decline in vitality may inform clinical targets for CVD prevention.

Did the Socio-Economic Gradient in Depression in Later-Life Deteriorate or Weaken during the COVID-19 Pandemic? New Evidence from England Using Path Analysis

It is well established that there is a socioeconomic gradient in adult mental health. However, little is known about whether and how this gradient has been exacerbated or mitigated by the COVID-19 pandemic. This study aims to identify the modifiable pathways involved in the association between socioeconomic position (SEP) and mental health during the COVID-19 pandemic. The analysis included 5107 adults aged 50+ living in England and participating in the English Longitudinal Study of Ageing Wave nine (2018–2019) and the COVID-19 study (June 2020). Mental health was measured using a shortened version of the Centre for Epidemiologic Studies Depression scale. Path analysis with multiple mediator models was used to estimate the direct effect of SEP (measured by educational qualification and household wealth) on mental health (measured by depression), along with the indirect effects of SEP via three mediators: COVID-19 infection symptoms, service accessibility and social contact. The results show that the prevalence of depression for the same cohort increased from 12.6% pre-pandemic to 19.7% during the first wave of the pandemic. The risk of depression increased amongst older people who experienced COVID-19 infection, difficulties accessing services and less frequent social contact. The total effects of education and wealth on depression were negatively significant. Through mediators, wealth and education were indirectly associated with depression. Wealth also directly affected the outcome. The findings suggest that the socioeconomic gradient in depression among older people may have deteriorated during the initial phase of the pandemic and that this could in part be explained by increased financial hardship, difficulties in accessing services and reduced social contact.

Path-specific effects in the presence of a survival outcome and causally ordered multiple mediators with application to genomic data

Causal multimediation analysis (i.e. the causal mediation analysis with multiple mediators) is critical for understanding the effectiveness of interventions, especially in medical research. Deriving the path-specific effects of exposure on the outcome through a set of mediators can provide detail about the causal mechanism of interest However, existing models are usually restricted to partial decomposition, which can only be used to evaluate the cumulative effect of several paths. In genetics studies, partial decomposition fails to reflect the real causal effects mediated by genes, especially in complex gene regulatory networks. Moreover, because of the lack of a generalized identification procedure, the current multimediation analysis cannot be applied to the estimation of path-specific effects for any number of mediators. In this study, we derive the interventional analogs of path-specific effect for complete decomposition to address the difficulty of nonidentifiability. On the basis of two survival models of the outcome, we derive the generalized analytic forms for interventional analogs of path-specific effects by assuming the normal distributions of mediators. We apply the new methodology to investigate the causal mechanism of signature genes in lung cancer based on the cell cycle pathway, and the results clarify the gene pathway in cancer.

Locating and testing the healthy context paradox: examples from the INCLUSIVE trial

Background

The healthy context paradox, originally described with respect to school-level bullying interventions, refers to the generation of differences in mental wellbeing amongst those who continue to experience bullying even after interventions successfully reduce victimisation. Using data from the INCLUSIVE trial of restorative practice in schools, we relate this paradox to the need to theorise potential harms when developing interventions; formulate the healthy context paradox in a more general form defined by mediational relationships and cluster-level interventions; and propose two statistical models for testing the healthy context paradox informed by multilevel mediation methods, with relevance to structural and individual explanations for this paradox.

Methods

We estimated two multilevel mediation models with bullying victimisation as the mediator and mental wellbeing as the outcome: one with a school-level interaction between intervention assignment and the mediator; and one with a random slope component for the student-level mediator-outcome relationship predicted by school-level assignment. We relate each of these models to contextual or individual-level explanations for the healthy context paradox.

Results

Neither model suggested that the INCLUSIVE trial represented an example of the healthy context paradox. However, each model has different interpretations which relate to a multilevel understanding of the healthy context paradox.

Conclusions

Greater exploration of intervention harms, especially when those accrue to population subgroups, is an essential step in better understanding how interventions work and for whom. Our proposed tests for the presence of a healthy context paradox provide the analytic tools to better understand how to support development and implementation of interventions that work for all groups in a population.

Trial registration

Current Controlled Trials, ISRCTN10751359.

Varying Coefficient Mediation Model and Application to Analysis of Behavioral Economics Data

This article is concerned with causal mediation analysis with varying indirect and direct effects. We propose a varying coefficient mediation model, which can also be viewed as an extension of moderation analysis on a causal diagram. We develop a new estimation procedure for the direct and indirect effects based on B-splines. Under mild conditions, rates of convergence and asymptotic distributions of the resulting estimates are established. We further propose a F-type test for the direct effect. We conduct simulation study to examine the finite sample performance of the proposed methodology, and apply the new procedures for empirical analysis of behavioral economics data.

A biologist's guide to model selection and causal inference

A goal of many research programmes in biology is to extract meaningful insights from large, complex datasets. Researchers in ecology, evolution and behavior (EEB) often grapple with long-term, observational datasets from which they construct models to test causal hypotheses about biological processes. Similarly, epidemiologists analyse large, complex observational datasets to understand the distribution and determinants of human health. A key difference in the analytical workflows for these two distinct areas of biology is the delineation of data analysis tasks and explicit use of causal directed acyclic graphs (DAGs), widely adopted by epidemiologists. Here, we review the most recent causal inference literature and describe an analytical workflow that has direct applications for EEB. We start this commentary by defining four distinct analytical tasks (description, prediction, association, causal inference). The remainder of the text is dedicated to causal inference, specifically focusing on the use of DAGs to inform the modelling strategy. Given the increasing interest in causal inference and misperceptions regarding this task, we seek to facilitate an exchange of ideas between disciplinary silos and provide an analytical framework that is particularly relevant for making causal inference from observational data.

Nonlinear mediation analysis with high-dimensional mediators whose causal structure is unknown

With multiple possible mediators on the causal pathway from a treatment to an outcome, we consider the problem of decomposing the effects along multiple possible causal path(s) through each distinct mediator. Under a path-specific effects framework, such fine-grained decompositions necessitate stringent assumptions, such as correctly specifying the causal structure among the mediators, and no unobserved confounding among the mediators. In contrast, interventional direct and indirect effects for multiple mediators can be identified under much weaker conditions, while providing scientifically relevant causal interpretations. Nonetheless, current estimation approaches require (correctly) specifying a model for the joint mediator distribution, which can be difficult when there is a high-dimensional set of possibly continuous and noncontinuous mediators. In this article, we avoid the need to model this distribution, by developing a definition of interventional effects previously suggested for longitudinal mediation. We propose a novel estimation strategy that uses nonparametric estimates of the (counterfactual) mediator distributions. Noncontinuous outcomes can be accommodated using nonlinear outcome models. Estimation proceeds via Monte Carlo integration. The procedure is illustrated using publicly available genomic data to assess the causal effect of a microRNA expression on the 3-month mortality of brain cancer patients that is potentially mediated by expression values of multiple genes.

Estimation of Controlled Direct Effects in Longitudinal Mediation Analyses with Latent Variables in Randomized Studies

In a randomized study with longitudinal data on a mediator and outcome, estimating the direct effect of treatment on the outcome at a particular time requires adjusting for confounding of the association between the outcome and all preceding instances of the mediator. When the confounders are themselves affected by treatment, standard regression adjustment is prone to severe bias. In contrast, G-estimation requires less stringent assumptions than path analysis using SEM to unbiasedly estimate the direct effect even in linear settings. In this article, we propose a G-estimation method to estimate the controlled direct effect of treatment on the outcome, by adapting existing G-estimation methods for time-varying treatments without mediators. The proposed method can accommodate continuous and noncontinuous mediators, and requires no models for the confounders. Unbiased estimation only requires correctly specifying a mean model for either the mediator or the outcome. The method is further extended to settings where the mediator or outcome, or both, are latent, and generalizes existing methods for single measurement occasions of the mediator and outcome to longitudinal data on the mediator and outcome. The methods are utilized to assess the effects of an intervention on physical activity that is possibly mediated by motivation to exercise in a randomized study.

Effect decomposition through multiple causally nonordered mediators in the presence of exposure-induced mediator-outcome confounding

Avin et al (2005) showed that, in the presence of exposure-induced mediator-outcome confounding, decomposing the total causal effect (TCE) using standard conditional exchangeability assumptions is not possible even under a nonparametric structural equation model with all confounders observed. Subsequent research has investigated the assumptions required for such a decomposition to be identifiable and estimable from observed data. One approach was proposed by VanderWeele et al (2014). They decomposed the TCE under three different scenarios: (1) treating the mediator and the exposure-induced confounder as joint mediators; (2) generating path-specific effects albeit without distinguishing between multiple distinct paths through the exposure-induced confounder; and (3) using so-called randomised interventional analogues where sampling values from the distribution of the mediator within the levels of the exposure effectively marginalises over the exposure-induced confounder. In this paper, we extend their approach to the case where there are multiple mediators that do not influence each other directly but which are all influenced by an exposure-induced mediator-outcome confounder. We provide a motivating example and results from a simulation study based on from our work in dental epidemiology featuring the 1982 Pelotas Birth Cohort in Brazil.

Continuous-time causal mediation analysis

While causal mediation analysis has seen considerable recent development for a single measured mediator (M) and final outcome (Y), less attention has been given to repeatedly measured M and Y. Previous methods have typically involved discrete-time models that limit inference to the particular measurement times used and do not recognize the continuous nature of the mediation process over time. To overcome such limitations, we present a new continuous-time approach to causal mediation analysis that uses a differential equations model in a potential outcomes framework to describe the causal relationships among model variables over time. A connection between the differential equation models and standard repeated measures models is made to provide convenient model formulation and fitting. A continuous-time extension of the sequential ignorability assumption allows for identifiable natural direct and indirect effects as functions of time, with estimation based on a two-step approach to model fitting in conjunction with a continuous-time mediation formula. Novel features include a measure of an overall mediation effect based on the "area between the curves," and an approach for predicting the effects of new interventions. Simulation studies show good properties of estimators and the new methodology is applied to data from a cohort study to investigate sugary drink consumption as a mediator of the effect of socioeconomic status on dental caries in children.

Fusion Validity: Theory-Based Scale Assessment via Causal Structural Equation Modeling

Fusion validity assessments employ structural equation models to investigate whether an existing scale functions in accordance with theory. Fusion validity parallels criterion validity by depending on correlations with non-scale variables but differs from criterion validity because it requires at least one theorized effect of the scale, and because both the scale and scaled-items are included in the model. Fusion validity, like construct validity, will be most informative if the scale is embedded in as full a substantive context as theory permits. Appropriate scale functioning in a comprehensive theoretical context greatly enhances a scale's validity. Inappropriate scale functioning questions the scale but the scale's theoretical embedding encourages detailed diagnostic investigations potentially challenging specific items, the procedure used to calculate scale values, or aspects of the theory, but also possibly recommends incorporating additional items into the scale. The scaled items should have survived prior content and methodological assessments but the items may or may not reflect a common factor because items having diverse causal backgrounds can sometimes fuse to form a unidimensional entity. Though items reflecting a common cause can be assessed for fusion validity, we illustrate fusion validity in the more challenging context of a scale comprised of diverse items and embedded in a complicated theory. Specifically we consider the Leadership scale from the Alberta Context Tool with care aides working in Canadian long-term care homes.