Atlas
statminds
Interactive GLM (Conditional Effects Model)The underlying model family class (e.g. GLM, linear model, categorical matrix, log-linear).Parametric ReferenceStatistical methods that assume a specific probability distribution family (typically normal).12-stage workflow
Moderation Analysis
The blueprint for Context-Dependent Discovery. Moderation Analysis identifies if the strength or direction of a relationship depends on the level of a third variable, revealing the 'boundary conditions' of an effect.
G1
Interaction Detection
Determine if the effect of Factor A 'changes shape' across the levels of Moderator B.
G2
Boundary Mapping
Identify the specific thresholds where a treatment stops being effective or begins to peak.
G3
Contextual Precision
Quantify the 'Conditional Effect' of a predictor while accounting for complex human variability.
1
What is it?
Moderation Analysis evaluates if the relationship strength between X and Y changes (is moderated) depending on the value of a third moderator variable (M).
2
When to use it
- Contextual Impact: Evaluate if an effect is stronger in some conditions than others.
- Interaction Effects: Focus specifically on the X*M product term.
Interactive Sandbox
Moderation Slope Live Laboratory
Change the moderation interaction effect (X*M) to watch the three conditional slopes split.
Presets
Simple Slopes Space (X: -10 to 10; Y: 0 to 100)Low M (-1 SD): blue; Mean M: green; High M (+1 SD): amber
The 12-Stage Precision Workflow
01The Product Term
Hypotheses
We test if the 'Interaction Term' (X * Z) is significant—the only mathematical proof that moderation exists.
02Interaction Normality
Assumptions
Ensuring the residuals of the interactive model follow the bell curve, justifying the multivariable synergy math.
03Mean Centering
Diagnostics
Executing 'Center-to-Scale' logic to reduce the artificial multicollinearity caused by multiplying predictors—essential for interpreting main effects.
04focus
Does FlowMotion's effect on Pain depend on the patient's Baseline BMI? Testing if the intervention is 'Size-Dependent'.
05Multi-Group Pivot
Alternatives
Knowing when to switch to Multi-Group SEM if the moderator is a categorical grouping with potentially differing variance structures.
06Incremental R²
Significance
Testing if adding the interaction term meaningfully increases the 'Total Explained Variance' of the model.
07Conditional B
Effect Size
Interpreting the shift in 'B'—e.g., 'For every unit increase in the Moderator, the effect of the Predictor increases by Y.'
08Interaction Power
Sample Size
Accounting for the significant sample size required to detect moderation—which typically requires 4x the N of simple main effects.
09Simple Slopes
Reporting
Reporting the effect at -1 SD, Mean, and +1 SD of the moderator to visualize the 'Diverging Slopes' narrative.
10PROCESS Logic
Software
Executing the 'Interaction' or 'PROCESS' macro commands to automatically calculate the pick-a-point slope breakdown.
11focus
The fatal error of reporting main effects when a significant interaction exists, which dangerously oversimplifies context.
12focus
Tracing the model back to the conditional probability frameworks and the mid-century refinement by Aiken and West.
01Hypothesis test logic
Hypotheses
Pragmatic null and alternative hypotheses defined in mathematical notation.
We ask not just 'is there a link?', but 'how much does Y change for every unit of X?'
Logic CoreNull · H₀
H₀: β₃ = 0 (no interaction; effect of X on Y does not depend on Z)
Alternative · Hₐ
Hₐ: β₃ ≠ 0 (interaction exists; effect of X on Y varies across levels of Z)
Why it matters two-tailed
Model: Y = β₀ + β₁X + β₂Z + β₃(X×Z) + ε. The interaction coefficient β₃ tests moderation. Significant interaction requires probing simple slopes at different levels of the moderator.
02Model diagnostics
Assumptions
The core mathematical criteria needed to ensure that statistical testing remains unbiased and valid.
Linearity is a strong claim. Nature often curves; ensure your model does not force a straight line on a bent world.
Integrity Shield7
Assumptions
0
Critical / High Severity
How to check
Create scatterplots of X and Z against Y. Examine residual plots. Check component-residual plots. Note: Moderation itself represents a non-linear effect (multiplicative), but the underlying relationships should be linear.
If violated
Transform variables (log, polynomial). Consider non-linear moderation models. Use generalized additive models (GAM) for complex non-linear interactions.
How to check
Durbin-Watson test (values near 2.0 indicate independence). Check for temporal, spatial, or hierarchical structure in data. Plot residuals by sequence or cluster.
If violated
Use time series methods for temporal dependence. Apply multilevel/mixed models for clustered or hierarchical data. Add control variables or time lags.
How to check
Plot residuals vs. fitted values; look for funnel patterns. Breusch-Pagan or White test. Particularly important to check at different levels of the moderator.
If violated
Transform dependent variable. Use robust standard errors. Apply weighted least squares. Consider heteroscedasticity may vary by moderator level.
How to check
Q-Q plots of residuals. Shapiro-Wilk test (n<50) or Kolmogorov-Smirnov test (n>50). Histogram of residuals. Less critical with n>30-40 due to Central Limit Theorem.
If violated
Transform dependent variable (log, Box-Cox). Use bootstrapping. Remove outliers after investigation. Consider robust regression.
How to check
Calculate VIF for all predictors including interaction term. Before centering, VIF for X×Z is often very high (>10). After mean-centering X and Z, VIF should decrease substantially. VIF < 10 acceptable, <5 better.
If violated
Mean-center continuous predictors before creating interaction term. Standardize if needed. If multicollinearity persists, interpret with caution but proceed if interaction term is theoretically justified. Do not remove interaction due to VIF alone.
How to check
Verify that X and Z have been mean-centered before creating the interaction term X×Z. Check that means of centered variables are approximately zero. Note: Centering does not change the interaction effect or its significance.
If violated
Uncentered continuous predictors lead to: (1) high multicollinearity between main effects and interaction, (2) main effects represent effects when other variable = 0, which may be nonsensical (e.g., when 0 is outside observed range). Always center continuous predictors unless 0 has meaningful interpretation.
How to check
Quick
Examine standard deviations and ranges of X and Z. Check that ranges are not artificially truncated (e.g., by inclusion criteria, measurement ceiling/floor effects). Verify no extreme restriction of range (SD < 20% of theoretical range suggests problems). Create scatterplots of X vs Z to check for adequate spread and no clustering.
Rigorous
Calculate reliability-adjusted variance: true variance = observed variance × reliability coefficient. If reliability < .80, effective variance is reduced. Examine bivariate distributions: compute coefficient of variation (CV = SD/mean) for X and Z; CV < 0.15 indicates low variability. Test for restriction of range by comparing sample SD to population SD if available. Simulate power loss: restricted range reduces correlation by factor of √(1-u²) where u = ratio of restricted to unrestricted SD.
If violated
Restricted range/variance leads to: (1) severely reduced statistical power to detect interactions (power loss is multiplicative for interaction = power_X × power_Z, so worse than for main effects), (2) attenuated interaction effect sizes that underestimate true population effects, (3) increased Type II error rates (missing real interactions). Solutions: (1) Broaden inclusion criteria if ethically/practically feasible to increase range, (2) Use stratified sampling to ensure representation across full range of X and Z, (3) Over-sample extreme values of X and Z to maximize variance (but maintain representativeness for generalizability), (4) Consider latent variable models (SEM) that correct for measurement error and restriction, (5) Report restriction statistics and note limitations—do not claim 'no moderation' when power is compromised, (6) Meta-analyze across studies with varying ranges to estimate true effect.
03Residual Forensics
Diagnostics
Checking residual plots and indices to examine model deviations and ensure standard error integrity.
Trust, but verify. The outliers often hold more truth than the averages.
System HealthEssential checks
- Significance test of interaction term β₃ (t-test or F-test for X×Z coefficient)
- Simple slopes analysis: Test X→Y relationship at low (-1 SD), mean, and high (+1 SD) levels of Z
- Interaction plot: Visualize predicted Y across range of X at different Z levels (at least 3 levels)
- VIF check after mean-centering (all VIF < 10, preferably < 5)
- ΔR² and F-change from adding interaction term to main effects model
Recommended checks
- VIF comparison before vs. after centering (demonstrates centering benefit)
- Johnson-Neyman technique: Identify exact Z values where X→Y effect transitions from non-significant to significant
- Effect sizes: Cohen's f² for interaction = ΔR²/(1-R²_full), report with 95% CI
- Probing at multiple Z levels: Test simple slopes at quartiles or specific theory-driven values
- Confidence bands on interaction plot (visualize uncertainty in predicted values)
- Residual diagnostics for final model: Q-Q plot, residuals vs. fitted, scale-location, Cook's D
- Test interaction robustness: Bootstrap interaction coefficient (1000-2000 iterations) for stability
- Cross-validation: Split-sample or k-fold validation to assess interaction replicability
- Floodlight analysis: Plot simple slope and its p-value across full range of moderator
04Live Instances
Applied Minds
Review concrete study examples, data layout guidelines, and copy executable syntax scripts.
Theory is the map. Practice is the terrain. Simulation bridges the gap.
Applied WisdomExample 01
Baseline Severity Moderates Treatment Effect
Research question: Does baseline severity moderate the effect of treatment on outcomes? Specifically, is the treatment more (or less) effective for patients with higher baseline severity?
05Tactical Pivots
Alternatives
Structured fallback pathways for choosing alternative tests when normality or slopes requirements fail.
When the path is blocked, pivot. Rigor is not rigidity; it is the intelligent adaptation to reality.
Adaptive StrategyMeasurement Precision Ladder Ideal · Continuous interaction (X*Z)
Ratio x Ratio
Maintain Moderation logic. Captures the full multiplicative influence of X across the Z-spectrum.
Peak SynergyNominal x Ratio
Ideal for Categorical Interactions (e.g., Treatment x Baseline). Ensure 'Reference Coding' is transparent.
Standard PrecisionOrdinal Interaction
Pivot to Ordinal Logistic interaction models to prevent non-linear rank distortion.
Threshold LossNominal x Nominal
Abandon Moderated Regression. Use Two-Way ANOVA or Log-Linear Analysis to audit categorical synergy.
Identity OnlyTemporal Trajectory Audit Static Synergy Snapshot
Static Context
Single point audit.
Stay with Moderation. Map the drivers of context-dependent efficacy.
Time-Moderated
Trajectory shifts.
Pivot to Mixed ANOVA or LMM where 'Time' is the moderator—auditing the shift in slopes.
Adaptive Technical Safeguards · adaptive safeguards
multicollinearity inflation
- Mean-Centering Mandate — Center both predictors before creating the product term to neutralize artificial VIF inflation.
- Residual Centering — A more advanced alternative to isolate the interaction term from its constituents.
non linear moderation
- Generalized Additive Models (GAMs) — Model the interaction surface using two-dimensional splines.
- Polynomial Synergy — Include squared terms for the moderator to capture 'U-shaped' context-dependence.
low power detected
- Johnson-Neyman Strike — Shift from p-values to regions of significance to find where the effect 'Unlocked'.
06Adjusted Comparisons
Post-hoc
Group mean comparisons and correction controls (e.g. Tukey HSD, Bonferroni) to protect against Family-Wise Error Rates.
The omnibus test opens the door; post-hoc analysis explores the room.
Forensic DetailAdjusted Comparisons
- Center/standardize variables before creating interaction term
- Compare different moderator value choices
- Bootstrap confidence intervals for simple slopes
- Test three-way interactions if theoretically relevant
Interpretation Guidelines
Moderation analysis tests interaction effects. Post-hoc involves probing the interaction.
07Standardized scale impact
Effect Size
Understanding effect sizes (e.g., Cohen's d, Partial Eta-Squared) and clinical impact benchmarks.
Significance is noise. Magnitude is the signal. Measure the impact, not just the probability.
Impact MagnitudeCohen's f² for interaction measures local effect size: f² = ΔR²_interaction/(1-R²_full) where ΔR²_interaction is incremental R² from adding X×Z term beyond main effects. Interpretation: f² = 0.02 (small), 0.15 (medium), 0.35 (large). Note: Interaction effects typically have smaller f² than main effects, so even f² = 0.01-0.02 can be meaningful. Always report with 95% CI via bootstrapping (1000-2000 iterations).
ΔR² represents additional variance in Y explained by interaction beyond main effects (X, Z alone). Even small ΔR² (e.g., 0.01-0.03) can be theoretically and practically important if interaction reveals 'for whom' or 'under what conditions' X affects Y. Context matters: in applied fields (clinical, organizational), small ΔR² may have large practical impact. Report: ΔR² = .02, 95% CI [.005, .045], F(1,296) = 6.23, p = .013.
Simple slopes quantify X→Y relationship at specific Z levels (conventionally -1 SD, mean, +1 SD of Z). Each simple slope has its own significance test, standard error, and 95% CI. Report format: 'At low Z (-1 SD): b = 0.15, SE = 0.08, t = 1.88, p = .061, 95% CI [-0.01, 0.31]. At high Z (+1 SD): b = 0.52, SE = 0.09, t = 5.78, p < .001, 95% CI [0.34, 0.70].' Large differences between slopes indicate strong moderation. The interaction coefficient β₃ = (slope_high - slope_low) / 2.
Difference in simple slopes directly quantifies moderation effect. For continuous moderator: slope difference = β₃ × (2 × SD_Z). For categorical moderator (0/1): slope difference = β₃. Larger differences = stronger moderation. Example: 'The X→Y slope was 0.37 units stronger for high Z compared to low Z, representing a 246% increase in effect size.'
Advanced metric: Compute correlation between predicted slopes (from X×Z interaction) and observed slopes (via multilevel model or local regression) across Z range. R²_slope quantifies proportion of variability in X→Y relationship explained by Z. Useful when Z is continuous and you want single summary of 'how much does the effect vary?'.
Recommended Metric: Report all three: (1) ΔR² with 95% CI for overall interaction strength, (2) f² for standardized effect size, (3) Simple slopes with CIs to describe moderation pattern
Small
0.2
0.2
Medium
0.5
0.5
Large
0.8
0.8
0.50
Report all three: (1) ΔR² with 95% CI for overall interaction strength, (2) f² for standardized effect size, (3) Simple slopes with CIs to describe moderation pattern
Recommended Measure
5
Available Metrics
ReportUse Report all three: (1) ΔR² with 95% CI for overall interaction strength, (2) f² for standardized effect size, (3) Simple slopes with CIs to describe moderation pattern to represent clinical impact magnitude.
08Statistical Power
Sample Size
Guidelines for minimum sample requirements and power analysis parameters.
An underpowered study is an ethical failure. Respect the data by collecting enough of it.
Power ProtocolFloor Requirements
The 'Synergy Threshold': A minimum of 50-100 participants is essential for even simple moderation (X * Z). Detecting interactions typically requires 4x more power than detecting simple main effects.
| Effect Size | Parameters | Required n |
|---|---|---|
| Small Effect | ΔR² = .02 (Small) | n ≈ 600 total |
| Medium Effect | ΔR² = .10 (Medium) | n ≈ 110 total |
| Large Effect | ΔR² = .25 (Large) | n ≈ 45 total |
Key considerations
The 'Range-Restriction' Penalty: If your moderator (Z) has a very narrow range, your power will collapse. Ensure your recruitment strategy targets the 'Extremes' of the moderator to allow the 'Diverging Slopes' to reveal their true magnitude.
G*Power StrategyBenchmark: F-tests → Linear interaction (R² change). Parameters: Predictor (X), Moderator (Z), Interaction (X*Z), ΔR², α = .05, Power = .80. Note: Power is dictated by the magnitude of the 'Incremental Slope' change.
09APA narrative blueprint
Reporting
How to compile statistical results into publication prose matching APA and journal style guides.
The Beta coefficient is the currency of change. Interpret it in real-world units, not just standardized abstractions.
Narrative Arc10Exhibit Builder
Manuscript Lab
Copy standard summary tables and forensic reporting grids to outline analysis details.
Table 1: Moderation Analysis: Intervention Effect by Support Level
| Predictor | B | SE | t | p | 95% CI |
|---|---|---|---|---|---|
| Intervention (X) | 0.12 | 0.08 | 1.50 | .135 | [-0.04, 0.28] |
| Social Support (M) | 0.45 | 0.10 | 4.50 | < .001 | [0.25, 0.65] |
| X × M (Interaction) | 0.38 | 0.12 | 3.17 | .002 | [0.14, 0.62] |
Note. Outcome: Wellbeing. Predictor: Intervention. Moderator: Social Support. N = 240.
Interaction (p=.002)Confirms Moderation. You cannot claim the treatment 'works' without specifying the participant's social support environment.
Low Support (p=.406)A critical failure point. The intervention provides zero statistical benefit for isolated patients, requiring a design pivot for this subgroup.
Header glossary
The 'Context' Audit. Proves that the effectiveness of the intervention (X) is not universal but depends on the level of social support (M).
The 'Drill Down'. Reveals exactly WHERE the intervention works (e.g., it works for those with high support, but not for those with low support).
11Algorithmic Logic
Command Center
Syntax libraries and function parameters for executing calculations in stats packages.
Code your model to handle residuals. The errors tell you what your model missed.
Execution Engine# 1. Fit Interaction Model
model <- lm(wellbeing ~ intervention * support, data = df)
# 2. Execute Simple Slopes Audit
interactions::sim_slopes(model, pred = intervention, modx = support)
# 3. Visualize Moderation (Interaction Plot)
interactions::interact_plot(model, pred = intervention, modx = support)Library stack
R
interactionsinteractionsggplot2
Python
statsmodels
Elite Forensic Strike
Moderation is about 'Boundary Conditions'. Use Johnson-Neyman intervals to find the exact point on the moderator scale where the treatment stops being effective.
# Johnson-Neyman Significance Zone
interactions::johnson_neyman(model, pred = intervention, modx = support)12The Over-adjustment Trap
Common Mistakes
Analytical caveats and corrections to maintain modeling integrity.
Wisdom is learning from the failures of others. Anticipate the error before it occurs.
Defensive LogicWhy it's wrong
Uncentered predictors create non-essential multicollinearity between main effects and interaction term (often VIF > 10), making coefficients unstable and difficult to interpret. Main effects represent effects when the other variable equals zero, which may be meaningless if zero is outside the observed range.
The correction
Always mean-center continuous predictors (subtract the mean) before computing the interaction term. This reduces multicollinearity and makes main effects interpretable as average effects. For categorical predictors, use effect coding rather than dummy coding if interpretation of main effects is important.
Why it's wrong
A significant interaction coefficient only indicates that slopes differ across moderator levels—it doesn't reveal the nature or pattern of the moderation. Without simple slopes, you cannot determine at which levels of the moderator the predictor has significant effects.
The correction
Always conduct simple slopes analysis after finding a significant interaction. Test slopes at low (-1 SD), mean, and high (+1 SD) levels of the moderator. Report coefficients and significance tests for each level. Consider Johnson-Neyman technique to identify the exact moderator values where the effect transitions from non-significant to significant.
Why it's wrong
When an interaction is significant, main effects are conditional and only represent effects when the other variable equals zero (or its mean if centered). The main effect of X describes the X→Y relationship only at Z=0, not the overall or average effect of X.
The correction
When interaction is significant, focus interpretation on simple slopes at different moderator levels, not on main effects. Report that 'the effect of X on Y depends on Z' rather than interpreting main effects separately. Create interaction plots to visualize the conditional relationships.
Why it's wrong
Artificially dichotomizing continuous variables loses statistical power, distorts effect sizes, increases Type I and Type II errors, and assumes effects are uniform within each artificially created group. This practice is widely criticized in methodological literature.
The correction
Keep continuous moderators continuous. Test moderation using the interaction between continuous variables. If you must use categorical moderators for presentation, conduct the analysis continuously first, then use predicted values at specific levels for illustration only.
Why it's wrong
Interaction effects are typically smaller than main effects and harder to detect, especially with limited sample sizes. Non-significant results may reflect insufficient power rather than true absence of moderation. Failing to reject the null doesn't prove the null is true.
The correction
Consider statistical power for detecting interactions (often requires n > 200-400). Report effect sizes (ΔR², f²) even for non-significant interactions. Discuss whether sample size was adequate. Avoid strong claims about absence of moderation unless power is sufficient (≥.80) and effect sizes are negligible.
Why it's wrong
Interaction effects are inherently complex and difficult to understand from coefficients alone. Verbal descriptions often fail to convey the pattern clearly. Readers need visual representation to grasp the nature of the moderation.
The correction
Always create interaction plots showing predicted Y across range of X at different levels of moderator (typically -1 SD, mean, +1 SD). Include confidence bands if possible. Ensure axes are clearly labeled and the moderator levels are identified. Include these plots in manuscripts or presentations.
Why it's wrong
Testing many possible moderators (fishing expeditions) inflates Type I error rates. Interaction effects have lower base rates and smaller effect sizes, making false positives more likely when testing many interactions without adjustment or prior hypotheses. If you test 20 possible moderators at α=.05, expected number of false positives ≈ 1 even if no true moderation exists.
The correction
Specify which interactions are of theoretical interest a priori in pre-registration or introduction. Limit interaction tests to those with strong theoretical rationale. If exploratory, apply appropriate corrections (Bonferroni: α/k where k=number of tests; FDR control for large sets) or clearly label as exploratory hypothesis-generating analyses requiring replication. Report all interactions tested, not just significant ones. Use cross-validation or split-sample replication for exploratory findings.
Why it's wrong
Interactions are more sensitive to measurement error than main effects. Unreliable measurement of X or Z attenuates interaction effects more severely than main effects (multiplicative attenuation). With reliability < .70 for either variable, interaction power drops dramatically, leading to false negatives. This is especially problematic for continuous moderators measured with scales prone to error.
The correction
Report reliability (Cronbach's alpha, test-retest) for all continuous predictor and moderator scales. Aim for α ≥ .80. If reliability is low (.60-.75): (1) Acknowledge as major limitation, (2) Use latent variable SEM to model measurement error explicitly, (3) Apply correction for attenuation formulas to estimate true effect size, (4) Consider multi-item composites rather than single items. Never test moderation with single-item measures unless reliability is known to be high. Report both observed and disattenuated effect sizes when possible.
13Academic Lineage
References
Scholarly lineage and citation keys grounding the statistical framework.
We stand on the shoulders of giants. Honor the source of the method.
Academic Lineage[1]
Aiken, L. S., & West, S. G. (1991). Multiple regression: Testing and interpreting interactions. Sage Publications.
Classic and comprehensive text on testing and interpreting moderation/interactions in regression.
doi: 10.4135/9781412983587
[2]
Hayes, A. F. (2018). Introduction to mediation, moderation, and conditional process analysis: A regression-based approach (2nd ed.). Guilford Press.
Modern comprehensive treatment including PROCESS macro for moderation analysis.
[3]
Aguinis, H., Beaty, J. C., Boik, R. J., & Pierce, C. A. (2005). Effect size and power in assessing moderating effects of categorical variables using multiple regression: A 30-year review. Journal of Applied Psychology, 90(1), 94-107.
Critical review of sample size and power issues in moderation analysis.
doi: 10.1037/0021-9010.90.1.94
[4]
Dawson, J. F. (2014). Moderation in management research: What, why, when, and how. Journal of Business and Psychology, 29(1), 1-19.
Practical guide to conducting and interpreting moderation analyses with common pitfalls.
doi: 10.1007/s10869-013-9308-7
[5]
McClelland, G. H., & Judd, C. M. (1993). Statistical difficulties of detecting interactions and moderator effects. Psychological Bulletin, 114(2), 376-390.
Important paper on statistical challenges and power issues in detecting interactions.
doi: 10.1037/0033-2909.114.2.376
In a moderated world, there is no single truth. The relationship is a liquid that changes shape based on the context. Audit the interaction, or you are only seeing a shadow of the effect.
The Interpretive Rigor Directive
statminds · ModerationMind reference · v2.2 · updated 2026-01-1715 of 15 sections