Creativity Self-Efficacy Mediation Analysis

Purpose

This skill encodes expert methodological knowledge for analyzing the psychological mechanisms through which AI tools (ChatGPT, web search) affect human creativity. Specifically, it covers SEM-based mediation analysis with creative self-efficacy as a mediator, and moderation analysis using baseline creativity. A general-purpose programmer could run a mediation analysis package, but would not know why creative self-efficacy is the theoretically motivated mediator, how to measure it, what the RAT measures and why it is the appropriate baseline, or how to interpret the indirect effect in the context of creativity theory.

When to Use This Skill

Investigating why an intervention affects creativity (not just whether it does)
Testing whether creative self-efficacy mediates the effect of AI/tool use on creative output
Examining whether baseline creativity moderates the effect of AI assistance
Designing a study that needs both mediation and moderation analysis for creativity outcomes
Specifying SEM models for creativity research using lavaan (R)

Research Planning Protocol

Before executing the domain-specific steps below, you MUST:

State the research question — What mechanism or moderator is being tested?
Justify the method choice — Why SEM-based mediation (not Baron & Kenny, not PROCESS)? What alternatives were considered?
Declare expected outcomes — What pattern of indirect/direct effects would support vs. refute the hypothesis?
Note assumptions and limitations — What does SEM assume? Where could cross-sectional mediation mislead?
Present the plan to the user and WAIT for confirmation before proceeding.

For detailed methodology guidance, see the research-literacy skill.

⚠️ Verification Notice

This skill was generated by AI from academic literature. All parameters, thresholds, and citations require independent verification before use in research. If you find errors, please open an issue.

Theoretical Framework

Creative Self-Efficacy as Mediator

Creative self-efficacy (CSE) = an individual's belief in their ability to produce creative outcomes (Tierney & Farmer, 2002). It is grounded in Bandura's (1997) self-efficacy theory: people who believe they can be creative are more likely to attempt, persist at, and succeed in creative tasks.

Hypothesized causal chain (Lee & Chung, 2024):

AI tool use → ↓ Creative Self-Efficacy → ↓ Creative Output

Mechanism: Using AI to generate ideas may undermine the user's
belief in their own creative ability, leading to reduced creative
effort and output on subsequent tasks.

Baseline Creativity as Moderator

Baseline creativity moderates how much AI assistance affects creative output:

High-creativity individuals: May benefit less from AI (ceiling effect) or be harmed more (self-efficacy threat)
Low-creativity individuals: May benefit more from AI (scaffolding) or show less effect (floor effect)

Lee & Chung (2024) found that ChatGPT use disproportionately reduced creativity for individuals with higher baseline creativity (measured by RAT).

Measurement Instruments

Creative Self-Efficacy Scale (Tierney & Farmer, 2002)

3 items, 5-point Likert scale (1 = strongly disagree, 5 = strongly agree):

"I have confidence in my ability to solve problems creatively"
"I feel that I am good at generating novel ideas"
"I have a knack for further developing the ideas of others"

Property	Value	Source
Cronbach's alpha	0.83-0.89	Tierney & Farmer, 2002; Lee & Chung, 2024
Test-retest reliability	0.77	Tierney & Farmer, 2002
Scale score	Mean of 3 items	Tierney & Farmer, 2002
Administration time	<1 minute	—
Timing	Administer after the manipulation, before the creativity task	Lee & Chung, 2024

Critical timing note: CSE must be measured after the manipulation (e.g., after ChatGPT use) and before the outcome measure. Measuring CSE before the manipulation captures trait CSE, not the mediated state change.

Remote Associates Test — RAT (Mednick, 1962)

Used as the baseline creativity measure for moderation analysis.

Property	Value	Source
Items	15 three-word problems	Lee & Chung, 2024
Format	Each item presents 3 words; participant finds the common associate	Mednick, 1962
Time limit	30 seconds per item or untimed	Lee & Chung, 2024
Scoring	Number correct out of 15	Lee & Chung, 2024
What it measures	Convergent thinking — finding the single correct remote association	Mednick, 1962

Example item: FALLING / ACTOR / DUST → answer: STAR

Why RAT as baseline: RAT measures convergent thinking (a creativity component independent of divergent thinking), so it serves as a baseline creativity indicator without directly measuring the same construct as the AUT outcome (Lee & Chung, 2024).

SEM Mediation Model Specification

Model Structure (Lee & Chung, 2024)

 Creative Self-Efficacy (M)
 ↗ a b ↘
AI Condition (X) Creativity Score (Y)
 ————— c' —————→

Path a: Effect of AI condition on CSE
Path b: Effect of CSE on creativity, controlling for condition
Path c': Direct effect of condition on creativity, controlling for CSE
Indirect effect: a × b (the mediated portion)
Total effect: c = c' + a × b

lavaan Specification (R)

library(lavaan)

mediation_model <- '
 # Measurement model (if using latent variables)
 # CSE =~ cse1 + cse2 + cse3 # Uncomment for latent CSE

 # Structural model
 cse ~ a * condition # Path a: X → M
 creativity ~ b * cse + # Path b: M → Y
 cprime * condition # Path c': X → Y (direct)

 # Indirect and total effects
 indirect := a * b # Mediated effect
 total := cprime + a * b # Total effect
'

fit <- sem(mediation_model, data = df, se = "bootstrap", bootstrap = 5000)
summary(fit, ci = TRUE)

Key Specification Decisions

Decision	Recommendation	Rationale
SE estimation	Bootstrap (5000 samples)	Indirect effects are non-normal; bootstrap CIs are preferred over Sobel test (Preacher & Hayes, 2008)
CI type	Bias-corrected bootstrap	More accurate than percentile bootstrap for indirect effects (MacKinnon et al., 2004)
Estimator	ML (maximum likelihood)	Default for continuous outcomes; use MLR for non-normal data
Missing data	FIML (full information ML)	Handles missing data without listwise deletion
Significance	95% bootstrap CI excluding zero	Do NOT rely on p-values for indirect effects

Moderation Analysis

Baseline Creativity × Condition Interaction

Two approaches (Lee & Chung, 2024 used both):

Approach 1: Median Split (Descriptive)

Compute median RAT score across all participants
Split into high-creativity (above median) and low-creativity (below median) groups
Run separate ANOVAs or t-tests within each subgroup
Report condition effects separately for high vs low creativity

Limitation: Median split loses information and reduces power (MacCallum et al., 2002). Use for visualization/description; rely on continuous moderation for inference.

Approach 2: Continuous Moderation (Inferential)

# In lavaan or linear regression
moderation_model <- '
 creativity ~ b1 * condition +
 b2 * rat_score +
 b3 * condition:rat_score # Interaction term
'
# b3 = moderation effect
# Probe interaction at ±1 SD of RAT score (Aiken & West, 1991)

Interpreting Moderation Results

Pattern	Interpretation	Lee & Chung (2024) Finding
Significant interaction, negative b3	AI assistance is more harmful for high-creativity individuals	Confirmed: ChatGPT reduced creativity more for high-RAT participants
Significant interaction, positive b3	AI assistance benefits high-creativity individuals more	Not observed
No significant interaction	AI effect is similar across creativity levels	—

Simple Slopes Visualization

Plot creativity scores against condition, separately for high (+1 SD) and low (-1 SD) baseline creativity:

library(emmeans)
# For interaction probing
emtrends(model, ~ condition, var = "rat_score")
# Or Johnson-Neyman technique for regions of significance

Moderated Mediation (Full Model)

When both mediation and moderation are relevant, combine into a conditional indirect effect model:

Does the indirect effect (X → M → Y) depend on baseline creativity (W)?

Model:
 CSE ~ a1 * condition + a2 * rat + a3 * condition:rat
 creativity ~ b * cse + c' * condition

Conditional indirect effect at level w of RAT:
 (a1 + a3 * w) × b

Software: Use lavaan with bootstrap, or the mediation package in R, or PROCESS macro Model 7 (Hayes, 2022).

Common Pitfalls

Cross-sectional mediation as causal evidence: Mediation in a cross-sectional or single-session design cannot establish temporal causation. The X → M → Y sequence must be theoretically justified and, ideally, measured at different time points (Bullock et al., 2010). Lee & Chung (2024) addressed this by measuring CSE after manipulation but before the outcome task.
Interpreting non-significant direct effect as "full mediation": A non-significant c' does not prove full mediation — it may reflect insufficient power. Report both direct and indirect effects with CIs (Rucker et al., 2011).
Using the Sobel test instead of bootstrapping: The Sobel test assumes normality of the indirect effect, which is almost never met. Use bootstrap CIs exclusively (Preacher & Hayes, 2008).
Forgetting to measure the mediator at the right time: CSE must be measured after the manipulation and before the outcome. Measuring at the wrong time destroys the mediation logic.
Median split without continuous analysis: Dichotomizing a continuous moderator loses statistical power and can create spurious interactions. Always accompany median splits with continuous moderation analysis (MacCallum et al., 2002).
Ignoring measurement reliability: Low reliability of the CSE scale attenuates the mediated effect. Report Cronbach's alpha and consider latent variable SEM if reliability is below 0.80.
Not controlling for potential confounders: In online studies, prior AI experience, age, education, and task engagement may confound the condition-creativity relationship. Include as covariates or demonstrate randomization balance.

Minimum Reporting Checklist

Based on Lee & Chung (2024) and Preacher & Hayes (2008):

References

Aiken, L. S., & West, S. G. (1991). Multiple regression: Testing and interpreting interactions. Sage.
Bandura, A. (1997). Self-efficacy: The exercise of control. W.H. Freeman.
Bullock, J. G., Green, D. P., & Ha, S. E. (2010). Yes, but what's the mechanism? Journal of Personality and Social Psychology, 98(4), 550-558.
Fritz, M. S., & MacKinnon, D. P. (2007). Required sample size to detect the mediated effect. Psychological Science, 18(3), 233-239.
Hayes, A. F. (2022). Introduction to mediation, moderation, and conditional process analysis (3rd ed.). Guilford Press.
Lee, B. C., & Chung, J. (2024). An empirical investigation of the impact of ChatGPT on creativity. Nature Human Behaviour. https://doi.org/10.1038/s41562-024-01953-1
MacCallum, R. C., Zhang, S., Preacher, K. J., & Rucker, D. D. (2002). On the practice of dichotomization of quantitative variables. Psychological Methods, 7(1), 19-40.
MacKinnon, D. P., Lockwood, C. M., & Williams, J. (2004). Confidence limits for the indirect effect. Multivariate Behavioral Research, 39(1), 99-128.
Mednick, S. A. (1962). The associative basis of the creative process. Psychological Review, 69(3), 220-232.
Preacher, K. J., & Hayes, A. F. (2008). Asymptotic and resampling strategies for assessing and comparing indirect effects in multiple mediator models. Behavior Research Methods, 40(3), 879-891.
Rucker, D. D., Preacher, K. J., Tormala, Z. L., & Petty, R. E. (2011). Mediation analysis in social psychology: Current practices and new recommendations. Social and Personality Psychology Compass, 5(6), 359-371.
Tierney, P., & Farmer, S. M. (2002). Creative self-efficacy: Its potential antecedents and relationship to creative performance. Academy of Management Journal, 45(6), 1137-1148.

See references/lavaan-templates.md for complete model specification code.