Causal Mediation Analysis

Causal mediation analysis has gained increasing attention in recent years. This article guides empirical researchers through the concepts and challenges of causal mediation analysis. I first clarify the difference between traditional and causal mediation analysis and highlight the importance of adjusting for the treatment-by-mediator interaction and confounders of the treatment–mediator, treatment–outcome, and mediator–outcome relationships. I then introduce the definition of causal mediation effects under the potential outcomes framework and different methods for the identification and estimation of the effects. After that, I highlight the importance of conducting a sensitivity analysis to assess the sensitivity of analysis results to potential unmeasured confounding. I also list various statistical software that can conduct causal mediation analysis and sensitivity analysis and provide suggestions for writing a causal mediation analysis paper. Finally, I briefly introduce some extensions that I made with my colleagues, including power analysis, multisite causal mediation analysis, causal moderated mediation analysis, and relaxing the assumption of no post-treatment confounding.

The effect of socioeconomic status on adolescent substance abuse may be mediated by family socialization practices. However, traditional mediation analysis using a product or difference method is susceptible to bias when assumptions are not addressed. We aimed to use a potential outcomes framework to assess assumptions of exposure-mediator interaction and of no confounding of the results. We revisited a traditional mediation analysis with a multiple mediator causal mediation approach using data from 17,761 Norwegian young people (13-18 years), 51% female. Data were collected through a print questionnaire. Socioeconomic status was operationalized as parental education and employment status (employed or receiving welfare); drinking behavior as the frequency of alcohol consumption and frequency of intoxication in the past year; and socialization practices as general parenting measures, alcohol-related parental permissiveness, and parent drinking behavior. There was no consistent evidence of exposure-mediator interaction. Formal sensitivity analysis of mediator-outcome confounding was not possible in the multiple mediator model, and this analysis supported the hypothesis that socioeconomic status effects on adolescent substance abuse are fully mediated by family socialization practices, with apparently stronger effects in younger age groups observed in plots. We found that the effect of socioeconomic status on adolescent substance abuse was fully mediated by family socialization practices. While our analysis provides more rigorous support for causal inferences than past work, we could not completely rule out the possibility of unmeasured confounding.

BackgroundMediation analysis methodology underwent many advancements throughout the years, with the most recent and important advancement being the development of causal mediation analysis based on the counterfactual framework. However, a previous review showed that for experimental studies the uptake of causal mediation analysis remains low. The aim of this paper is to review the methodological characteristics of mediation analyses performed in observational epidemiologic studies published between 2015 and 2019 and to provide recommendations for the application of mediation analysis in future studies.MethodsWe searched the MEDLINE and EMBASE databases for observational epidemiologic studies published between 2015 and 2019 in which mediation analysis was applied as one of the primary analysis methods. Information was extracted on the characteristics of the mediation model and the applied mediation analysis method.ResultsWe included 174 studies, most of which applied traditional mediation analysis methods (n = 123, 70.7%). Causal mediation analysis was not often used to analyze more complicated mediation models, such as multiple mediator models. Most studies adjusted their analyses for measured confounders, but did not perform sensitivity analyses for unmeasured confounders and did not assess the presence of an exposure-mediator interaction.ConclusionsTo ensure a causal interpretation of the effect estimates in the mediation model, we recommend that researchers use causal mediation analysis and assess the plausibility of the causal assumptions. The uptake of causal mediation analysis can be enhanced through tutorial papers that demonstrate the application of causal mediation analysis, and through the development of software packages that facilitate the causal mediation analysis of relatively complicated mediation models.

Causal mediation analysis provides an attractive framework for integrating diverse types of exposure, genomic, and phenotype data. Recently, this field has seen a surge of interest, largely driven by the increasing need for causal mediation analyses in health and social sciences. This article aims to provide a review of recent developments in mediation analysis, encompassing mediation analysis of a single mediator and a large number of mediators, as well as mediation analysis with multiple exposures and mediators. Our review focuses on the recent advancements in statistical inference for causal mediation analysis, especially in the context of high-dimensional mediation analysis. We delve into the complexities of testing mediation effects, especially addressing the challenge of testing a large number of composite null hypotheses. Through extensive simulation studies, we compare the existing methods across a range of scenarios. We also include an analysis of data from the Normative Aging Study, which examines DNA methylation CpG sites as potential mediators of the effect of smoking status on lung function. We discuss the pros and cons of these methods and future research directions.

Mediation analysis is a useful tool in randomized trials for understanding how a treatment works, in particular how much of the treatment's effect on an outcome is explained by a mediator variable. The traditional approach to mediation analysis makes sequential ignorability assumption which precludes the existence of unobserved confounders between the mediator and outcome variables. Since the randomized experiment does not randomize the mediator, sequential ignorability may not be plausible. In this article, based on a statistical model termed sure outcomes of random events model, we propose an alternative approach to causal mediation analysis without relying on the sequential ignorability assumption for the case of binary treatment and mediator variables. When the outcome is also binary, we establish the identifiability of the average natural direct and indirect effects in the presence of an unobserved confounder between mediator and outcome variables. More importantly, if the identifiability conditions are violated, we provide new bounds that are narrower than those in the previous studies, and these bound results are extended to the case of an arbitrary bounded outcome. Simulation studies show good performance for the proposed estimators in finite samples. Finally, we use a job training intervention on the mental health study to illustrate our approach.

Frailty as a mediator of postoperative delirium in older adults undergoing orthopedic surgery: A causal mediation analysis

BackgroundHealth determinants such as income and employment significantly influence declarative memory in women with HIV (WWH). These social exposures may contribute to cognitive decline through complex biological mechanisms. We used high‐throughput metabolomics data and causal mediation analysis to identify multiple biological pathways linking adverse socioeconomic conditions and cognitive decline.MethodA total of 324 women (n = 225 WWH) from the New York sites of the Women’s Interagency HIV Study completed biennial cognitive testing, starting in 2009 and every two years, which included the Hopkins Verbal Learning Test (HVLT)‐R, a measure of verbal learning and memory. Metabolomic assessment was conducted from serum samples collected prior to or at first cognitive visit using liquid chromatography tandem mass spectrometry. We used multivariable generalized estimating equation models to link peripheral metabolomic profiles and cognitive trajectories in verbal learning and memory. Linear mixed effect models were used to predict subject‐specific trajectories for downstream mediation models. We used regression‐based multiple mediation analyses to estimate the joint effect of multiple metabolic mediators as pathways between income, employment, and cognitive trajectories. We considered a false discovery rate (FDR) of 5% for metabolomic discovery.ResultOur sample had average age of 43.1 years and were followed‐up for 11.9 years, 63% were non‐Hispanic Black, 76% reported earnings of <24K, and 65% were unemployed. We identified 16 FDR‐corrected metabolites associated with trajectories in verbal memory and 15 with verbal learning. Top metabolites (serotonin, taurine, adenosine, niacinamide, α‐glycerophosphocholine, ADMA, pseudouridine, and sphingomyelins) were common in both domains. Stratified models by serostatus indicated similar results in WWH and women without HIV. In causal mediation analysis, we found that 30% (95%CI: 4%‐100%) of the effect of employment and 18.2% (95%CI: ‐2.1% to 72.3%) of the effect of income on verbal learning was mediated through differences in metabolite levels. We also found that 35.6% (2.7%‐100%) of the effect of employment and 25.4% (95%CI: ‐1.7%‐100%) of the effect of income on verbal memory was mediated through altered metabolomic signatures.ConclusionHealth determinants influence biological process implicated in cognitive decline. Biosocial interventions are needed to address the conditions that perpetuate poor cognitive health in women.

Background:Surgical interventions allow for tailoring of treatment to individual patients and implementation may vary with surgeon and healthcare provider. In addition, in clinical trials assessing two competing surgical interventions, the treatments may be accompanied by co-interventions.Aims:This study explores the use of causal mediation analysis to (1) delineate the treatment effect that results directly from the surgical intervention under study and the indirect effect acting through a co-intervention and (2) to evaluate the benefit of the surgical intervention if either everybody in the trial population received the co-intervention or nobody received it.Methods:Within a counterfactual framework, relevant direct and indirect effects of a surgical intervention are estimated and adjusted for confounding via parametric regression models, for the situation where both mediator and outcome are binary, with baseline stratification factors included as fixed effects and surgeons as random intercepts. The causal difference in probability of a successful outcome (estimand of interest) is calculated using Monte Carlo simulation with bootstrapping for confidence intervals. Packages for estimation within standard statistical software are reviewed briefly. A step by step application of methods is illustrated using the Amaze randomised trial of ablation as an adjunct to cardiac surgery in patients with irregular heart rhythm, with a co-intervention (removal of the left atrial appendage) administered to a subset of participants at the surgeon’s discretion. The primary outcome was return to normal heart rhythm at one year post surgery.Results:In Amaze, 17% (95% confidence interval: 6%, 28%) more patients in the active arm had a successful outcome, but there was a large difference between active and control arms in the proportion of patients who received the co-intervention (55% and 30%, respectively). Causal mediation analysis suggested that around 1% of the treatment effect was attributable to the co-intervention (16% natural direct effect). The controlled direct effect ranged from 18% (6%, 30%) if the co-intervention were mandated, to 14% (2%, 25%) if it were prohibited. Including age as a moderator of the mediation effects showed that the natural direct effect of ablation appeared to decrease with age.Conclusions:Causal mediation analysis is a useful quantitative tool to explore mediating effects of co-interventions in surgical trials. In Amaze, investigators could be reassured that the effect of the active treatment, not explainable by differential use of the co-intervention, was significant across analyses.

In the disciplines of medicine, biology and social science, causal mediation analysis is widely used to examine the effect of an exposure on an outcome through different causal pathways. Recently, there is rapid development of mediation analysis with survival data. An early work of Lange and Hansen proposed to identify the direct and indirect causal effects under the additive hazards model with a single mediator. However, inaccurate measurements of mediators and confounders may lead to biased causal effect estimation. We propose a measurement error correction approach to tackle measurement error in the mediators and confounders under the additive hazards model where multiple mediators are present. We apply the proposed corrected method to the ACTG175 Study data set, and uncover interesting findings.

ABSTRACTIdentifying mediators in variable chains as part of a causal mediation analysis can shed light on issues of causation, assessment, and intervention. However, coefficients and effect sizes in a causal mediation analysis are nearly always small. This can lead those less familiar with the approach to reject the results of causal mediation analysis. The current paper highlights five factors that contribute to small path coefficients in mediation research: loss of information when measuring relationships across time, controlling for prior levels of a predicted variable, adding control variables to the analysis, ignoring measurement error in one’s variables, and using multiple mediators. It is argued that these issues are best handled by increasing the statistical power of the analysis, identifying the optimal temporal interval between variables, using bootstrapped confidence intervals to analyze the results, and finding alternate ways of assessing the meaningfulness of the indirect effect.

Causal mediation analysis is the study of mechanisms—variables measured between a treatment and an outcome that partially explain their causal relationship. The past decade has seen an explosion of research in causal mediation analysis, resulting in both conceptual and methodological advancements. However, many of these methods have been out of reach for applied quantitative researchers, due to their complexity and the difficulty of implementing them in standard statistical software distributions. The mediation package in R provides a set of simple commands that execute some of the newer causal mediation methods. This article will summarize some of the recent advances in mediation analysis, critically review the mediation package, and demonstrate, by example, some of its capabilities.

Background: People with psychosis have a reduced life expectancy by 15 years, mainly due to preventable physical illnesses for which obesity is a precursor. Obesity is three times more common in psychosis and antipsychotics are an important cause. Prediction could individualise obesity treatment but current models are not fully actionable for individuals.Aims: We test whether antipsychotic-induced weight increase at 1 year is causally mediated by weight change in the first 12 weeks of treatment, rather than the following 40 weeks. We then develop and internally validate a causal actionable prediction pathway to prevent antipsychotic-induced obesity.Methods: We use parallel causal mediation analysis to determine the natural direct and indirect, and total effects of antipsychotic choice on weight in a trial of olanzapine versus haloperidol in 97 participants. We develop a baseline causal actionable prediction model to predict weight gain at 12 weeks in 172 participants, and a 12-week model to predict obesity at 1 year in 97 of the participants, and then demonstrate counterfactual prediction.Results: Antipsychotic-induced weight gain at 1 year appears to be causally mediated by weight change during the first 12 weeks treatment – indirect effect 5.70 (95% CI 2.83 to 8.66). At internal validation, the discrimination c-statistic for the baseline causal actionable prediction model was 0.732 and the calibration slope was 0.768. For the 12-week model, the c-statistic was 0.914 and the calibration slope was 0.601. We use the models to predict the counterfactual outcomes of antipsychotic choice and 12-week weight change.Conclusions: Our study shows it may be early, rather than later weight change, which causally mediates antipsychotic-induced weight gain at 1 year. We also demonstrate the potential for causal prediction of counterfactuals targeting obesity for true precision medicine. Further research in larger samples is necessary before causal prediction models could be applied clinically.

SAS Macro for Causal Mediation Analysis with Survival Data

This study develops a joint modeling approach that incorporates latent traits into causal mediation analysis with multiple mediators and a survival outcome. A linear structural equation model is used to characterize the latent mediators with several highly correlated observable surrogates and depicts the relationships among multiple parallel or causally ordered mediators and the exposure. A proportional hazards model is used to derive the path-specific causal effects on the scale of hazard ratio under the counterfactual framework with a set of sequential ignorability assumptions. A Bayesian approach with Markov chain Monte Carlo algorithm is developed to perform efficient estimation of the causal effects. Posterior propriety theory is established for the proportional hazards model with latent variables. Empirical performance of the proposed method is verified through simulation studies. The proposed model is then applied to a study on the Alzheimer’s Disease Neuroimaging Initiative dataset to investigate the causal effects of APOE- allele on the disease progression, either directly or through potential mediators, such as hippocampus atrophy, ventricle expansion, and cognitive impairment.

BackgroundThere is evidence that disability acquisition causes a decline in mental health, but few studies have examined the causal mechanisms through which the effect operates. Our previous research examined three...