Chapter Four · failure evidence
What Moderation & Interaction Modeling got wrong, from 24 dissertations
The records document various methodological obstacles encountered when attempting to specify and estimate moderation and interaction models. Analysts frequently abandoned or simplified interaction terms because of severe multicollinearity, failure to improve upon simpler main-effects baselines, structural biases, and algorithmic non-convergence. These records come from PhD theses at 12 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.
Failure to improve explanatory or predictive power over simpler main effects baselines
Adding interaction terms failed to yield meaningful increments in explained variance or model fit relative to simpler additive specifications. Across linear, logistic, and mixed-effects models, researchers reverted to main-effects baselines because interaction terms increased computational cost or reduced predictive performance.
Considered and rejected
Considered and rejected: Reverted moderation regression models with sex interaction terms to main effects models only due to nonsignificant increments to R2.
Inside Their Mind: Examining the Relationship Between Visual Mental Imagery and Psychopathic Traits · Carleton University Institutional Repository
Considered and rejected
Considered and rejected: Rejected adding interaction terms to the final multinomial logistic regression model after exploratory analysis showed none significantly improved model fit
The influence of academic and social factors on degree attainment in BSc Sport and Exercise Science · University of Nottingham Repository
Tried and failed
adding two-way interaction terms to regression models applied to predicting beliefs from psychological features. Outcome: no signal. Reason: interaction terms failed to yield substantive improvements to model fit over simpler additive models
Representations of social and political attitudes, opinions, and facts in the mind and brain · Harvard
Tried and failed
hierarchical group-lasso for pairwise interactions applied to disease risk prediction from environmental exposures. Outcome: worse than baseline. Reason: interaction terms reduced predictive performance and increased computational cost compared to main effects alone
Lost to a baseline
Polynomial proxy with only interaction terms (mean R² = 0.9345) was beaten by the linear proxy with top two interactions (mean R² = 0.9409).
Considered and rejected
Considered and rejected: Discarded interaction terms in the main predictors regression model of story certainty in Study 1 after model comparison favored a simpler main-effects model
Considered and rejected
Considered and rejected: Rejected models with interaction terms for logistic regression modeling of P responsiveness, evaluating only simple main effects.
Soil fertility trials on trial for "noise": Statistical implications for soil fertility modeling · Iowa State
Tried and failed
adding demographic interaction terms to regression models applied to educator survey response prediction. Outcome: no signal. Reason: Interaction terms were statistically insignificant and did not improve model fit over main effects
Measuring Student Agency Conditions · Harvard
Considered and rejected
Considered and rejected: Rejected adding interaction terms between vowel type and vowel quality in linear mixed models for segmental epenthesis because they did not improve model fit.
Acoustic properties of underlying and derived contrasts in Bedouin Meccan Arabic · Texas Tech
Severe multicollinearity and variance inflation from interaction terms
Introducing interaction terms alongside correlated predictors generated severe multicollinearity that inflated standard errors and diluted main effects. Analysts were forced to drop interaction variables, evaluate them one at a time, or simplify models to prevent uninterpretable slope estimates.
Tried and failed
combining multiple correlated interaction terms in regression applied to moderation analysis of survey data. Outcome: no signal. Reason: high multicollinearity among correlated predictor interaction terms inflated standard errors, eliminating statistical significance
Religion's main and moderating effects on depressive symptoms for adults who have experienced stressful life events · Iowa State
Tried and failed
adding time interaction terms to linear models applied to longitudinal feature performance modeling. Outcome: no signal. Reason: Interaction terms were statistically non-significant and diluted the statistical significance of established main effects.
Modeling statistics ITAs’ speaking performances in a certification test · Iowa State
Considered and rejected
Considered and rejected: Rejected including interaction terms in the multiple linear regression model due to high multicollinearity among predictor variables (most ps < 0.001).
Tiered Approaches for Educational Equity: Modeling the Determinants of Special Education Disproportionality and Compliance · ResearchWorks
Considered and rejected
Considered and rejected: Decided against simultaneously estimating all latent interaction terms in LMS due to multicollinearity, opting to test one interaction model at a time
Considered and rejected
Considered and rejected: Rejected inclusion of 3 statistically significant interaction terms (Figures 6, 7, and 10) from the final modified model due to severe multicollinearity.
The impact of a mother's stressful employment conditions on her parenting practices and her adolescent child · Iowa State
Considered and rejected
Considered and rejected: Rejected the original complex regression model (Eq. 3 with simultaneously combined entropy, inverse SEn, and interaction terms) due to uninterpretable slopes and severe multicollinearity.
Measuring Dynamic Team Reorganization and Interdependency in Response to Uncertainty · Georgia Tech
Statistical bias and specification artifacts in nonlinear and latent interactions
Latent and nonlinear interaction structures introduced bias into model parameters and overall fit assessments. Researchers excluded latent interaction terms from measurement models to protect fit estimates and avoided nonlinear interactions that suffered from incidental parameter problems.
Considered and rejected
Considered and rejected: Logistic regression rejected for primary specification in favor of OLS due to incidental parameter problem with large fixed effects and bias in nonlinear interaction terms
Considered and rejected
Considered and rejected: Excluded latent interaction terms from baseline measurement model CFA because interactions bias model fit estimates.
Outcomes of Interpersonal Felt Distrust in the Workplace: An Identity Threat Perspective · YorkSpace
Algorithmic non-convergence and estimation instability from data constraints
Estimating complex interaction structures led to computational failures such as non-convergence during multiple imputation procedures. In other settings, constrained sample sizes left fully interacted choice models imprecisely estimated, which restricted practical interpretation.
Tried and failed
conditional logit with full interaction terms applied to discrete choice preference estimation. Outcome: data insufficient. Reason: interaction terms were imprecisely estimated with weak statistical significance, limiting interpretability
Considered and rejected
Considered and rejected: Rejected including interaction terms in multiple imputation equations due to non-convergence
Labour induction in Nova Scotia: What has contributed to rising rates? · DalSpace
Left open by the authors
Problems the authors named and did not get to.
Left open
Analyze interaction terms among restaurant factors contributing to average sentiment using regression models. Blocker: None
Left open
Regress PC loadings and alignment selections against interaction terms of institutional manager characteristics and macroeconomic factors. Blocker: None
Institutional Co-Holdings Geometries · Harvard
Left open
Estimate regression models including interaction terms between distinct management dimensions like goal setting and feedback to assess organizational reform heterogeneity. Blocker: Requires the proprietary organizational management survey and clinical performance dataset used in the dissertation
Left open
Estimate interaction terms for respondent occupation in the conditional logit model to evaluate its effect on willingness-to-pay for offshore wind co-location. Blocker: Requires the private survey response microdata containing respondent occupations.
Opportunity Between the Turbines: A Willingness-to-Pay Experiment Regarding Co-Location Activities with the Coastal Virginia Offshore Wind Farm · Virginia Tech
Left open
Perform a meta-analytic moderation analysis testing whether age or game type moderates the predictive validity of deliberate practice versus cognitive ability. Blocker: None
What Causes High Achievement? An Investigation of "Talent" and Its Alternatives · Penn
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.