26. SEMinR Lecture Series - Moderation Analysis in R

TL;DR

Moderation tests whether an independent construct’s effect on a dependent construct changes as a function of a moderator.

Briefing Cornell Notes

Briefing

Moderation analysis in R with semPLS hinges on a two-stage interaction workflow: estimate latent variable scores first, then build the interaction term from those scores and test whether it significantly changes the explained variance of the dependent construct. In the example used throughout, collaborative culture’s effect on organizational performance depends on role ambiguity—high role ambiguity weakens the relationship, while low role ambiguity strengthens it.

The lecture frames moderation as a situation where the relationship between two constructs is not constant but varies with a third variable (the moderator). Researchers may hypothesize that either a single path—or multiple paths—depend on the moderator, which determines the modeling strategy. For the two-stage approach, the process is split into Stage 1 (a main-effects model) and Stage 2 (the structural model including the interaction). Stage 1 estimates latent variable scores for the dependent construct, the independent construct, and the moderator using a model with only exogenous predictors. Those latent scores are then multiplied to create a single-item interaction term used in Stage 2.

A key practical constraint is handled carefully: measurement-model evaluation (reliability and validity checks) should be done before introducing the moderating effect. Once the interaction term is created as a single-item latent score, standard measurement-model quality criteria for that interaction become meaningless, so the interaction term is not assessed with the usual measurement-model steps. Instead, the workflow is: validate the measurement model for the constructs first, then add the interaction term to the structural model.

Statistical relevance of moderation is assessed using Cohen’s f² for the interaction effect, computed from R² values with and without the interaction term: f² = (R²_included − R²_excluded) / (1 − R²_included). The lecture notes common benchmarks for f² (small/medium/large often cited as 0.02/0.15/0.35), but also highlights that moderation effects in practice tend to be smaller; alternative guidance suggests more modest thresholds (0.005/0.01/0.025). In the worked example, the interaction produces a large f² (reported as about 0.09), indicating the moderator meaningfully contributes to explaining organizational performance.

After estimating the model, bootstrapping is used to test whether the interaction term is significant. The interaction coefficient is negative and statistically significant (t-value exceeding 1.96, with no zero in the confidence interval), meaning role ambiguity moderates the collaborative culture → organizational performance link in a weakening direction. Simple slopes and slope plots translate this into interpretable effects: at average role ambiguity, collaborative culture has a positive effect; at high role ambiguity (e.g., +1 SD), the effect decreases by the magnitude of the interaction; at low role ambiguity (−1 SD), the effect strengthens. The steepness of the plotted lines becomes the visual cue—steeper at low role ambiguity, flatter at high role ambiguity—summarizing how moderation changes the strength of the relationship.

Cornell Notes

Moderation analysis in semPLS (via semPLS in R) is implemented with a two-stage interaction workflow. First, a main-effects model estimates latent variable scores for the independent construct, the moderator, and the dependent construct. Second, the interaction term is created by multiplying the Stage 1 latent scores, then inserted into the structural model to test whether it changes the dependent construct’s explained variance. Measurement-model checks (reliability/validity) are performed before adding the interaction, because the interaction term is treated as a single-item score and standard measurement-model criteria don’t apply. Significance is tested with bootstrapping, and effect size is summarized using f² computed from R² with vs. without the interaction term. In the example, role ambiguity negatively moderates collaborative culture’s effect on organizational performance, weakening the relationship at high ambiguity.

Why does the two-stage approach matter for moderation in semPLS?

It separates estimation of latent scores from construction of the interaction term. Stage 1 fits a main-effects model with the independent construct and the moderator predicting the dependent construct, then extracts latent variable scores for the independent construct (X), the moderator (M), and the dependent construct (Y). Stage 2 multiplies the Stage 1 latent scores for the independent construct and moderator to form the interaction term used in the structural model. This avoids manually engineering interactions and supports better parameter recovery and statistical power compared with less robust alternatives.

What changes in measurement-model evaluation once the interaction term is added?

Measurement-model assessment (reliability and validity) should be completed for the constructs before moderation is introduced. After the interaction term is created, it is treated as a single-item latent score, so standard measurement-model quality criteria for that interaction become inappropriate. The interaction term is therefore not evaluated with the usual measurement-model steps; only the measurement model for the original constructs is assessed.

How is moderation effect size quantified using f²?

f² is computed from R² values of the endogenous construct with and without the interaction term: f² = (R²_included − R²_excluded) / (1 − R²_included). Here, R²_included comes from the model that includes the interaction term, while R²_excluded comes from the model where the interaction (and typically the moderator’s direct role, depending on the comparison) is removed. Larger f² indicates the interaction meaningfully increases explained variance for the dependent construct.

How does bootstrapping determine whether moderation is significant?

Bootstrapping generates sampling distributions for the path coefficients, including the interaction term coefficient. The interaction is deemed significant when the bootstrapped t-value exceeds the conventional threshold (e.g., 1.96 for a two-sided test) and the confidence interval does not cross zero. In the example, the interaction coefficient for collaborative culture × role ambiguity is negative and significant, indicating a weakening moderation effect.

How do simple slopes and slope plots translate the interaction into substantive meaning?

Simple slopes compute the effect of the independent construct on the dependent construct at specific moderator values (commonly −1 SD, mean, and +1 SD). With a negative interaction, the independent construct’s effect decreases as the moderator increases. The slope plot visualizes this through curve steepness: the line is steeper at low role ambiguity (stronger positive effect) and flatter at high role ambiguity (weaker effect).

Review Questions

In a two-stage moderation workflow, what exactly is multiplied to form the interaction term, and from which stage do those values come?
Why is measurement-model evaluation for the interaction term treated differently than for the original constructs?
How would you interpret a negative, significant interaction coefficient in terms of how the moderator changes the independent construct’s effect on the dependent construct?

Key Points

1
Moderation tests whether an independent construct’s effect on a dependent construct changes as a function of a moderator.
2
A two-stage approach estimates latent scores first (Stage 1) and then builds the interaction term by multiplying those scores (Stage 2).
3
Measurement-model reliability and validity checks should be completed before adding the interaction term; standard measurement-model criteria for the interaction term are not applied afterward.
4
Moderation relevance is quantified with f² using R² included vs. R² excluded models, where f² reflects the interaction’s contribution to explained variance.
5
Bootstrapping is used to test the interaction term’s significance; significance requires a t-value above threshold and a confidence interval that excludes zero.
6
A negative interaction coefficient means higher moderator values weaken the independent→dependent relationship; simple slopes and curve steepness make this interpretable.
7
In the example, role ambiguity weakens the positive relationship between collaborative culture and organizational performance, with stronger effects at low ambiguity and weaker effects at high ambiguity.

Highlights

The interaction term is created by multiplying Stage 1 latent variable scores, then inserted into the Stage 2 structural model.

Interaction-term measurement-model evaluation is skipped because the interaction is treated as a single-item latent score.

Moderation effect size uses f² = (R²_included − R²_excluded) / (1 − R²_included), tying moderation to explained-variance change.

In the worked case, the interaction is negative and significant: role ambiguity reduces the strength of collaborative culture’s impact on organizational performance.

Slope plots interpret moderation through curve steepness: low role ambiguity produces a steeper (stronger) effect than high role ambiguity.

Topics

Moderation Analysis
semPLS Two-Stage Interaction
f² Effect Size
Bootstrapping
Simple Slopes