Chapter Four · failure evidence
What Causal Mediation Analysis got wrong, from 38 dissertations
The records document applications and methodological challenges of causal mediation analysis across diverse psychological, biological, and observational settings. Authors frequently abandon or adapt mediation techniques when specifications exhibit poor structural fit, violated temporal ordering, or unmet prerequisite associations between variables. These records come from PhD theses at 20 institutions, 2021 to 2026. Each links to its thesis. They were extracted by language models reading the full text, so treat each as a lead to read, not a verdict.
Complex structural equation models and moderated mediation configurations suffer from poor fit and convergence failures
Structural mediation models frequently display unacceptable global fit indices, including elevated RMSEA and suppressed CFI values across multigroup or continuous interaction configurations. Researchers also encounter bootstrapping convergence failures and non-significant likelihood ratio tests when adding interaction terms or separating variables.
Tried and failed
path analysis mediation model applied to cross-sectional behavioral survey data. Outcome: no signal. Reason: The specified structural mediation model exhibited very poor global fit and non-significant indirect mediation paths.
Tried and failed
continuous interaction modeling in structural equation models applied to longitudinal behavioral mediation analysis. Reason: produced unacceptable model fit indices including very high RMSEA
Tried and failed
moderated multigroup structural equation modeling applied to moderated mediation psychometric cascade pathways. Reason: models produced extremely poor fit indices across multigroup moderation configurations compared to direct mediation models
Mindfulness and its relationship to race-related stress, racial identity, age, gender, and ses across multiple racial minorities · Georgia Tech
Considered and rejected
Considered and rejected: Rejected moderated mediation models including interaction terms (dosage x working alliance and dosage x program satisfaction) due to poor model fit (CFI = .86, RMSEA = .10) and convergence failure in bootstrapping
Examining the Effects Between Teacher Qualifications and Program Implementation on Improving Social Skills among Students in K-3rd Grade · Scholars' Bank
Considered and rejected
Considered and rejected: Moderated mediation SEM models were rejected due to poor fit (RMSEA > .45) and non-significant likelihood ratio tests compared to mediation
Subjective social status, symbolic capital, and psychopathology in adolescence. · ResearchWorks
Considered and rejected
Considered and rejected: Mediation model separating intergenerational support into two observed variables (providing support and receiving support) was rejected due to poor fit (chi2(20) = 512.58, p < .001, CFI = .70, RMSEA = .14).
The impact of adult children's economic hardship on older parents' psychological well-being · Iowa State University Digital Repository
Mediation analysis is abandoned when prerequisite associations between variables fail to reach significance
Authors discard planned mediation workflows when requisite bivariate correlations between predictors, mediators, and outcomes do not reach statistical significance. Models are also omitted when treatment variables show no significant relation to the outcome or when main effects lack sufficient posterior confidence.
Considered and rejected
Considered and rejected: Rejected mediation analysis via Hayes Process Macro due to failure of requisite bivariate correlations between CTQ-SF total, BIS-11 total, and CWI4 to reach statistical significance
Exploring childhood adversity, impulsivity, and cognitive functioning within forensic populations · University of Nottingham Repository
Considered and rejected
Considered and rejected: Testing mediation of alliance on common factor-outcome relations in Study 2 was abandoned because common factors did not predict subsequent outcome.
Considered and rejected
Considered and rejected: Rejected planned mediation analyses because no treatment variables significantly correlated with competency restoration.
Juvenile Competency Restoration in Texas: Factors Associated with Remediation in Inpatient Settings · DSpace at SHSU
Considered and rejected
Considered and rejected: Omitted formal statistical tests of indirect effects (mediation) for research questions 3a, 4a, 5a, 6a, 7a, and 8a because main effects lacked strong posterior confidence.
Considered and rejected
Considered and rejected: Discarded mediation models for DA mediating conformity motives on alcohol quantity and Level III binge drinking due to violation of the assumption requiring significant X-Y relations.
Cross-sectional survey data invalidates causal mediation due to lack of temporal ordering and severe bias
Researchers reject formal mediation modeling on cross-sectional survey datasets because the designs lack temporal ordering and introduce severe methodological bias. Studies instead substitute experimental manipulations, indirect effect estimations, or sequential covariate adjustments to avoid these design limitations.
Considered and rejected
Considered and rejected: Rejected performing full mediation analysis on Study 1 cross-sectional data due to methodological invalidity.
Emotion Dynamics of Solitude: How Spending Time Alone Affects the Way We Experience and Manage Affect · Carleton University Institutional Repository
Considered and rejected
Considered and rejected: Rejected using classic mediation analysis for mental effort due to methodological failure concerns in experimental psychology (Lemmer & Gollwitzer, 2017), opting instead for experimental manipulation.
Leistungserwartungen und mentale Anstrengung bei Menschen mit depressiven Störungen · open_UMR Marburg DSpace 10.0
Considered and rejected
Considered and rejected: Rejected formal cross-sectional mediation modeling in favor of estimating indirect effects, citing severe methodological bias and lack of temporal ordering in cross-sectional designs
Considered and rejected
Considered and rejected: Decided against conducting a full mediation analysis due to the limiting nature of cross-sectional survey data, instead implementing sequential thematic covariate adjustment models.
Birth Matters: Examining Contextual Risk and Protective Factors for Maternal Postpartum Depression · Harvard
Proposed structural mediation models are beaten by simpler direct effect or unmoderated baselines
Chi-square difference tests show that direct-effect models and simple mediation structures frequently match or exceed the fit of complex moderated mediation specifications with lower complexity. Proposed structural paths also lose to basic measurement models without regression paths and to models without reversed mediator-outcome directions.
Lost to a baseline
Direct-effect SEM models were statistically preferable over indirect mediation models via chi-square difference tests (p = .059 to .859) due to equal fit with lower complexity.
Exploring Functional and Dysfunctional Features of Talent Development Environments in German Olympic and Paralympic Sports. Implications for the Holistic Development of Elite Youth Athletes · Publikationssystem UB Tuebingen
Lost to a baseline
A simple mediation model (PROCESS Model 4) outperformed the proposed doubly moderated mediation 3MP model in explaining Republican minority persuasion.
Bridging the Minority Influence Gap: The Roles of Social Identification and Prototypicality · unevada
Lost to a baseline
Measurement model without regression paths fit parent-report data significantly better than the proposed structural mediation model (chi-square difference = 33.70, p < .001).
Pathways to Parenting Stress and Challenging Behaviors for Preschoolers with Developmental Delays: The role of emotion regulation · Scholars' Bank
Considered and rejected
Considered and rejected: Rejected reversed mediation model (switching outcomes and mediators) due to worse model fit (CFI = .91 vs .95, RMSEA = .068 vs .048).
Summary statistic tests and joint significance methods face power loss and conservative p-values
Summary statistics-based mediation yields overly conservative p-values when confounder-mediator correlation is high, and tests like sMiST lose power under null direct effects by evaluating unnecessary variance components. In addition, joint significance testing fails to supply point estimates or standard errors, while simple mediation fails when high variable counts cannot be reduced into latent factors.
Lost to a baseline
When confounder-mediator correlation is high (0.4), summary statistics-based mediation p-values become over-conservative compared to individual-level MiST
Statistical Methods for Biomarker Informed Personalized Medicine: Discovery and Translation · ResearchWorks
Lost to a baseline
When there is no direct effect, sMiST can lose power compared to pure mediation tests (e.g. PrediXcan/TWAS) due to testing the additional variance component parameter with null effect
Statistical Methods for Biomarker Informed Personalized Medicine: Discovery and Translation · ResearchWorks
Considered and rejected
Considered and rejected: Rejected a simple mediation analysis because there were too many outcome and predictor variables that could not be combined into single latent factors.
Student Engagement & Social Disadvantage: The Importance of Supportive Relationships · Research Repository UCD
Considered and rejected
Considered and rejected: Rejected joint significance testing for mediation models because it fails to provide a single point estimate or standard error for the mediated effect.
Full mediation constraints fail statistical tests and are rejected in favor of partial mediation
Restricting direct paths to zero produces unacceptable fit compared to partial mediation alternatives, leaving substantial unaccounted variance in the outcome. Formal chi-square difference tests repeatedly reject full mediation in favor of models that retain direct paths between exposures and outcomes.
Tried and failed
Full mediation analysis of cognitive constructs applied to working memory and fluid intelligence relationship. Outcome: did not generalise. Reason: Attention control did not fully mediate the relationship, leaving significant unaccounted variance
Considered and rejected
Considered and rejected: In Study 3, a full mediation model fixing the direct path from MORE treatment to Grade 4 math to zero was rejected in favor of a partial mediation model based on a chi-square difference test (χ²(1) = 8.79, p < .01).
Considered and rejected
Considered and rejected: A full mediation model of OBSE was rejected because it did not provide an adequate fit to the data (χ² = 10.21, df = 4, RMSEA = .09, CFI = .90, TLI = .77) compared to the partial mediation model.
Exploring the Impact of Servant Leadership on Thriving at Work and Adaptive Performance · Carleton University Institutional Repository
Intermediate confounding and bidirectional feedback violate core causal assumptions
Researchers reject natural effects mediation when intermediate confounding between interrelated mediators creates bias, turning instead to interventional effect approaches. Models are also abandoned when measurement scales fail to separate actions from outcomes or when bidirectional reinforcement undermines the directed causal structure.
Considered and rejected
Considered and rejected: Rejected conducting temporal and causal mediation analysis on parent behaviors using the PSIS because the measure failed to capture distinct parent actions versus child outcomes.
Considered and rejected
Considered and rejected: Mediation models were considered based on prior literature but rejected in favor of moderation models due to insufficient causal justification from bidirectional reinforcement.
What did you expect? The role of alcohol outcome expectancies in the relationship between psychopathic traits and alcohol-related aggression · Carleton University Institutional Repository
Considered and rejected
Considered and rejected: Rejected using natural effects mediation analysis due to bias from intermediate confounding among interrelated school/family mediators, choosing an interventional effects approach instead.
Left open by the authors
Problems the authors named and did not get to.
Left open
Adapt mediational E-value sensitivity analysis methodology to evaluate unmeasured confounding in network causal models using simulations and public frameworks. Blocker: None
Networked resources and health in South Africa: Towards a Materialist Social Network Epidemiology · Harvard
Left open
Perform mediation analysis assessing whether gut microbiota-derived enterolignans mediate dietary lignan associations with gout risk. Blocker: Requires cohort biospecimen/metabolomics data (plasma or urinary enterolactone and enterodiol measurements)
Plant-Based Diets, Lignan Intake, Weight Change, and the Risk of Gout · Harvard
Left open
Develop causal mediation analysis methodology that jointly accounts for left-censoring and measurement error in viral persistence mediators. Blocker: None
Methods for causal mediation analysis with applications in HIV and cardiorespiratory fitness · OpenBU
Left open
Investigate mediation mechanisms linking ethnic enclaves and neighborhood poverty to adverse maternal and birth outcomes. Blocker: Requires access to restricted Texas vital statistics records and clear mediator variables not specified.
Left open
Perform Mendelian randomisation and mediation analysis to assess causal roles of identified metabolite levels in disease aetiology using GWAS summary statistics. Blocker: None
Common 'Inborn Errors' of Metabolism in the General Population · Cambridge
Left open
Perform mediation analysis to determine if inflammatory biomarkers mediate associations between prenatal PFAS exposure and pregnancy or child complications. Blocker: Requires access to restricted longitudinal pregnancy cohort data (MIREC/MIREC-ENDO)
Examining the impact of poly- and perfluoroalkyl substances (PFAS) exposure on metabolic health · Carleton University Institutional Repository
Left open
Identify specific causal mechanisms linking PM2.5 exposure to adult upward mobility using mediation analysis. Blocker: No specific datasets, targets, or mediation methodologies are defined.
ECONOMIC AND SOCIAL IMPACTS OF AIR POLLUTION: A CAUSAL INFERENCE APPROACH · Harvard
Left open
Extend pragmatic mediation to settings with unobserved confounders, soft or high-dimensional interventions, or partially known abstract feature sets. Blocker: The task is an open-ended research direction without concrete mathematical targets, formal formulations, or specific evaluation methods provided.
Casual and trustworthy machine learning: methods and applications · Oxford
Left open
Replicate neural mediation analyses of peer victimization and drinking motives on racially and culturally diverse fMRI and behavioral datasets. Blocker: Requires access to controlled-access neuroimaging datasets (e.g., ABCD study) with diverse cohorts and relevant behavioral phenotypes
Left open
Perform a mediation analysis evaluating how average institutional SAT scores mediate effects on the percentage of Pell Grant recipients. Blocker: None
Checking a claim in this area?
We can run the same search on any method or claim. If nothing turns up, we will say so, and that proves nothing on its own.