Mediation, moderation & path analysis guide
The right method depends on what you’re asking: how one variable’s influence reaches another, when that influence changes, or whether several relationships belong in a single model.
Say a study-skills program raised students’ test scores. Why did it work? Maybe it boosted their motivation. Maybe it helped students who were already strong more than everyone else. Or maybe several factors were pushing on one another at once. Each of those stories is a different research question, and each calls for a different analysis.
Mediation looks at the process an effect travels through. Moderation looks at the conditions under which an effect gets stronger, weaker, or even changes direction. And when you need to examine several paths at once, path analysis or structural equation modeling (SEM) comes into play. This guide sorts out these four approaches and helps you match them to your research question.
Four questions to answer before you choose
Before you worry about technique names, get clear on your research question and how your variables relate.
Process or condition?
Are you explaining the pathway an effect travels along, or checking whether the effect changes under certain conditions?
Observed or latent?
Will you use variables you measure directly, like test scores, or latent concepts measured with several survey items?
Is the time order clear?
When was each variable measured? Was there random assignment, or a repeated-measures design?
One relationship or many paths?
Is this a single mediation or moderation relationship, or do you need to handle two or more outcomes and several paths at once?
Mediation, moderation, and SEM aren’t competing options. You can test a mediation effect inside an SEM, and a single research model can include both mediation and moderation.
A decision map for effect and path analysis
Start from your research question and narrow down what you need.
Treat this map as a way to tell the concepts apart, not as a rulebook. In practice, the right method also depends on how your variables are measured, how your data are structured, your sample size, the estimation method, and your study design.
When you want to know how an effect gets passed on
Put a mediator between the predictor and the outcome.
Mediation analysis
Does a study-skills program (X) raise students’ motivation (M), and is that motivation in turn related to their test scores (Y)? What you’re after is the indirect effect: the part of X’s influence that flows through M on its way to Y.
The indirect effect is usually expressed as the product of path coefficients (a × b), and its uncertainty is assessed with tools such as bootstrap confidence intervals. Keep in mind that finding an indirect effect doesn’t, on its own, prove that the process is causal.
| Question | Where to go |
|---|---|
| Does the effect work through another variable? | Mediation analysis |
| Are there two or more mediators? | Consider parallel or serial mediation models |
| Does the strength of the mediated path depend on conditions? | Consider moderated mediation |
When you want to know when an effect changes
Check whether a relationship holds the same way for everyone.
Moderation analysis
Does the link between study time (X) and test scores (Y) depend on grade level (W)? The usual tool is an interaction term (X × W), which captures whether the effect of X changes from one condition to another.
With a continuous moderator, you can look at the conditional effect of X at specific values of W; with a categorical moderator, you compare the effect group by group. If the interaction is significant, also check over what range of the moderator the difference actually shows up.
| Question | Where to go |
|---|---|
| Under what conditions does the effect grow or shrink? | Moderation analysis |
| Does the mediated effect itself differ across groups or conditions? | Moderated mediation |
| Do you want to compare several paths across groups at once? | Consider multigroup path analysis or multigroup SEM |
When several relationships belong in one model
Tell apart path analysis with observed variables and SEM with latent variables.
Path analysis
Path analysis estimates several direct and indirect paths among observed variables at the same time. For example, a single model can trace how academic support feeds into both motivation and study time, which in turn relate to test scores.
Structural equation modeling (SEM)
SEM earns its keep when your model includes latent variables—concepts such as motivation or satisfaction that you measure with several survey items—and you need to account for measurement error. It evaluates the measurement model and the structural relationships among variables in one go.
A complicated web of paths doesn’t automatically call for SEM. If your question is simple and you’re working only with observed variables, regression-based mediation and moderation analysis, or a path analysis, may be all you need.
If you use latent variables, make sure the items really measure the concepts they’re meant to. When the factor structure is already specified, checking the measurement model—with confirmatory factor analysis (CFA), for example—is essential.
What should you check before running the analysis?
Even a good model can’t make up for weaknesses in your data or study design.
| Check | What to look for |
|---|---|
| Study design | When each variable was measured, possible confounders, and the direction of cause and effect |
| Measurement | Reliability and validity of scales, measurement error, coding of categorical variables |
| Model specification | Missing or unnecessary paths, interaction terms, theoretical justification |
| Sample & missing data | Enough information for the number of paths; the pattern of missing data and how it’s handled |
| Estimation & diagnostics | Distributions, outliers, collinearity, estimation method, and model fit (where applicable) |
| Scope of interpretation | Not treating associations in observational data as proven causal effects |
In SEM, good fit indices don’t confirm the causal direction you assumed. Weigh alternative models and the theory behind your model as well.
Effect and path analysis methods at a glance
Pin down the purpose of your analysis first, then turn to each chapter for the details and settings.
These cards summarize the typical approach for each goal; they don’t map one-to-one onto the lessons available in the Academy.
Mediation is about process, moderation is about conditions, and SEM is about the scope of your model.
If you want to know how an effect gets from X to Y, think mediation first; if you want to know when it changes, think moderation. Path analysis estimates relationships among several observed variables at once, and SEM extends that to latent variables and measurement error. Let your research question, your measurement approach, and the structure of your data decide which one you need.
Don’t start with the most sophisticated technique—start with your research question. And when you analyze observational data, keep two claims separate: that a path is statistically significant, and that it supports a causal explanation.