Mediation and Moderation Analysis - Educational Administration

Much research in educational administration asks not only whether leadership is related to teacher outcomes, but how and under what conditions. Mediation analysis addresses the first question: it tests whether the association between a predictor (say, principals’ instructional leadership) and an outcome (teachers’ work engagement) runs through a third variable, such as teacher self-efficacy. Moderation analysis addresses the second: it tests whether the strength or direction of an association changes with the level of another variable, such as trust in the principal. Both are now standard in theses built on teacher and principal surveys, and both are frequently misreported. This guide sets out the decisions you face and what examiners and reviewers expect to see.

When to use it

Use mediation when your theory specifies a mechanism: a variable that is affected by the predictor and in turn affects the outcome. Use moderation when your theory says that a relationship depends on context, on a group, or on a characteristic of the respondent. Baron and Kenny (1986) drew the distinction that is still taught today. A mediator accounts for how or why a predictor relates to an outcome; a moderator changes the direction or strength of that relationship. The two are not interchangeable, and the choice should come from your conceptual framework, not from whichever analysis produced significant results.

When a mechanism itself is expected to vary with context, the two ideas combine in moderated mediation, also called conditional process analysis (Hayes, 2022). For example, the indirect path from instructional leadership to teaching practice through professional learning might be stronger where principals also show high levels of transformational leadership.

A caution belongs here at the outset. Mediation is a causal claim about a process that unfolds over time. Most thesis data are collected at a single time point from a single source. The statistical tools described below will produce indirect effects from such data, but they cannot turn correlational data into causal evidence.

Key concepts

Term	Meaning
a path	Association of the predictor (X) with the mediator (M)
b path	Association of M with the outcome (Y), controlling for X
Indirect effect (ab)	Product of a and b; the part of the X–Y association transmitted through M
Direct effect (c′)	Association of X with Y after controlling for M
Total effect (c)	Direct plus indirect effect (in simple linear models, c = c′ + ab)
Interaction term	Product of X and the moderator (W), testing whether the X–Y slope depends on W
Conditional effect	The slope of X on Y at a given value of W
Index of moderated mediation	A single test of whether an indirect effect depends linearly on a moderator (Hayes, 2015)

Step by step

Specify the model from theory

Draw the model before touching the data. State which variables are predictors, mediators, moderators and outcomes, and why. With several mediators, decide whether they operate in parallel (each carrying a separate indirect path) or in sequence (serial mediation). Preacher and Hayes (2008) showed how to estimate a total indirect effect alongside specific indirect effects for each mediator, and how to test contrasts between them.

Move beyond the causal-steps approach

Baron and Kenny (1986) proposed a sequence of regressions: X must predict Y, X must predict M, M must predict Y with X controlled, and the X–Y coefficient should shrink. Many theses still follow this logic, supplemented by the Sobel test for the indirect effect. Both have been largely superseded. Preacher and Hayes (2008), drawing on earlier simulation work showing that the causal-steps procedure has lower power than alternatives, judged it of little use with multiple mediators, and pointed out that the Sobel test assumes the sampling distribution of ab is normal, which holds only in large samples. Zhao et al. (2010) went further, arguing that a significant total effect is not a precondition for mediation and that the only requirement is a significant indirect effect. The current standard is to test ab directly with a bootstrap confidence interval: the data are resampled thousands of times, ab is estimated in each resample, and the interval is read from the resulting distribution. If the interval excludes zero, the indirect effect is treated as statistically significant.

Choose your software

Regression-based models with observed variables can be estimated with Hayes’s PROCESS, which is available for SPSS, SAS and R and documented in the third edition of his book (Hayes, 2022). PROCESS uses numbered model templates for common mediation, moderation and conditional process designs, and reports bootstrap intervals for indirect effects.

When your constructs are measured with multi-item scales and you want to model measurement error, use structural equation modelling (SEM). In lavaan for R (Rosseel, 2012), you label the a and b paths and define the indirect effect as a new parameter with the := operator; bootstrap standard errors are requested with se = "bootstrap". In Mplus, the MODEL INDIRECT command requests specific and total indirect effects, BOOTSTRAP in the ANALYSIS command sets the number of draws, and CINTERVAL(BOOTSTRAP) in the OUTPUT command produces bootstrap confidence intervals. With teachers nested in schools, consider whether a multilevel model is needed, since single-level analysis ignores the clustering.

Test and probe interactions

For moderation, enter X, W and their product in the model. A significant interaction coefficient tells you only that the slope of X varies with W; it does not tell you where the effect is present. Probe it. The simple slopes (pick-a-point) approach estimates the conditional effect of X at chosen values of W, commonly the mean and one standard deviation above and below it. The Johnson–Neyman technique instead identifies the range of W values over which the conditional effect is statistically significant, which avoids arbitrary cut points (Bauer & Curran, 2005). Always plot the interaction.

Test moderated mediation properly

In moderated mediation, report conditional indirect effects at several values of the moderator, each with a bootstrap interval. Showing that the indirect effect is significant at one value and not at another is not itself a test of moderation. Hayes’s (2015) index of moderated mediation, with its bootstrap interval, provides that formal test.

Confront the design limitations

Maxwell and Cole (2007) showed that when mediation is a longitudinal process, estimates from cross-sectional data can be substantially biased, even when the mediator fully accounts for the effect of X on Y. Their estimates of direct and indirect effects and of the proportion mediated were often seriously misleading. Maxwell et al. (2011) extended this to the more typical case of partial mediation and reached a similar conclusion. In practice, this means that a cross-sectional study can show that data are consistent with a mediation hypothesis, but cannot establish that mediation occurred. Write “is associated with … indirectly through”, not “affects”, “influences” or “leads to”.

Common method bias is a related threat. When the predictor, mediator and outcome are all teacher self-reports from one questionnaire, shared method variance can inflate or deflate the paths. Podsakoff et al. (2003) recommended procedural remedies first, such as obtaining measures from different sources or separating them in time, and judged Harman’s single-factor test to be insensitive and not a useful remedy on its own.

What to report

The theoretical model, drawn as a path diagram, with the role of each variable stated in advance.
Descriptive statistics, reliabilities and correlations for all variables; for SEM, measurement model fit.
Unstandardised (and, where helpful, standardised) a, b, c′ and c coefficients with standard errors.
Each indirect effect with its bootstrap confidence interval, the number of bootstrap resamples, and the type of interval (for example, percentile or bias-corrected).
For moderation: the interaction coefficient, conditional effects at stated values of the moderator or Johnson–Neyman regions, and a plot.
For moderated mediation: conditional indirect effects and the index of moderated mediation.
Software and version, how variables were centred (if at all), and how missing data were handled.
An explicit statement of design limitations, including cross-sectional data and single-source measurement.

Common mistakes

Using causal language (“effect”, “impact”, “leads to”) for cross-sectional survey data without qualification.
Relying on the causal-steps approach and abandoning mediation because the total effect was not significant.
Reporting only the Sobel test, or only p-values for the a and b paths, instead of a confidence interval for ab.
Treating “full” versus “partial” mediation as a substantive finding, when whether the direct effect reaches significance depends heavily on sample size.
Confusing mediators with moderators, or testing both on the same variable without a theoretical reason.
Interpreting a significant interaction without probing or plotting it.
Citing a non-significant Harman’s single-factor test as proof that common method bias is absent.
Ignoring the nesting of teachers within schools.

Examples from Turkish research

Turkish educational administration research has used mediation and moderation extensively to examine how school leadership is linked to teachers’ beliefs and practices, typically with teacher surveys and bootstrapped indirect effects. A recent doctoral study of 470 teachers, summarised in our research digest, found that most of the association between principals’ instructional leadership and teachers’ work engagement ran indirectly, through self-efficacy and collaboration. Published examples include:

Bellibaş et al. (2021), in Educational Administration Quarterly, surveyed 616 teachers in Turkish primary and secondary schools and used SEM with bootstrapping to test a moderated mediation model, finding that transformational leadership moderated the indirect association of instructional leadership with instructional practice through teacher professional learning.
Kılınç et al. (2024), in Educational Management Administration & Leadership, studied 611 teachers in 56 schools with multilevel SEM and Bayesian estimation; the indirect link from transformational leadership to innovative practice through teacher commitment depended on teachers’ trust in the principal (see also organisational trust).
Özdoğru et al. (2025), in Behavioral Sciences, used PROCESS with 5,000 bootstrap resamples on data from 357 high school teachers in Eskişehir and reported a significant indirect association between innovative school leadership and teacher professional learning through the principal–teacher relationship, while acknowledging that the cross-sectional design leaves the direction of the relationships open.

Key takeaways

A mediator explains how a predictor relates to an outcome; a moderator explains when or for whom. Choose on theoretical grounds.
Test indirect effects directly with bootstrap confidence intervals rather than the causal-steps approach or the Sobel test.
Probe every interaction with simple slopes or the Johnson–Neyman technique, and use the index of moderated mediation for conditional indirect effects.
Cross-sectional, single-source data can be consistent with mediation but cannot demonstrate it; avoid causal language and address common method bias in the design.

References

Baron, R. M., & Kenny, D. A. (1986). The moderator–mediator variable distinction in social psychological research: Conceptual, strategic, and statistical considerations. Journal of Personality and Social Psychology, 51(6), 1173–1182. https://doi.org/10.1037/0022-3514.51.6.1173
Bauer, D. J., & Curran, P. J. (2005). Probing interactions in fixed and multilevel regression: Inferential and graphical techniques. Multivariate Behavioral Research, 40(3), 373–400. https://doi.org/10.1207/s15327906mbr4003_5
Bellibaş, M. Ş., Kılınç, A. Ç., & Polatcan, M. (2021). The moderation role of transformational leadership in the effect of instructional leadership on teacher professional learning and instructional practice: An integrated leadership perspective. Educational Administration Quarterly, 57(5), 776–814. https://doi.org/10.1177/0013161X211035079
Hayes, A. F. (2015). An index and test of linear moderated mediation. Multivariate Behavioral Research, 50(1), 1–22. https://doi.org/10.1080/00273171.2014.962683
Hayes, A. F. (2022). Introduction to mediation, moderation, and conditional process analysis: A regression-based approach (3rd ed.). Guilford Press.
Kılınç, A. Ç., Polatcan, M., Savaş, G., & Er, E. (2024). How transformational leadership influences teachers’ commitment and innovative practices: Understanding the moderating role of trust in principal. Educational Management Administration & Leadership, 52(2), 455–474. https://doi.org/10.1177/17411432221082803
Maxwell, S. E., & Cole, D. A. (2007). Bias in cross-sectional analyses of longitudinal mediation. Psychological Methods, 12(1), 23–44. https://doi.org/10.1037/1082-989X.12.1.23
Maxwell, S. E., Cole, D. A., & Mitchell, M. A. (2011). Bias in cross-sectional analyses of longitudinal mediation: Partial and complete mediation under an autoregressive model. Multivariate Behavioral Research, 46(5), 816–841. https://doi.org/10.1080/00273171.2011.606716
Özdoğru, M., Doğuş, Y., & Akyürek, M. İ. (2025). The mediating role of the principal–teacher relationship in innovative school leadership and teacher professional learning according to Turkish teachers’ perceptions. Behavioral Sciences, 15(4), Article 450. https://doi.org/10.3390/bs15040450
Podsakoff, P. M., MacKenzie, S. B., Lee, J.-Y., & Podsakoff, N. P. (2003). Common method biases in behavioral research: A critical review of the literature and recommended remedies. Journal of Applied Psychology, 88(5), 879–903. https://doi.org/10.1037/0021-9010.88.5.879
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. https://doi.org/10.3758/BRM.40.3.879
Rosseel, Y. (2012). lavaan: An R package for structural equation modeling. Journal of Statistical Software, 48(2), 1–36. https://doi.org/10.18637/jss.v048.i02
Zhao, X., Lynch, J. G., Jr., & Chen, Q. (2010). Reconsidering Baron and Kenny: Myths and truths about mediation analysis. Journal of Consumer Research, 37(2), 197–206. https://doi.org/10.1086/651257

Last reviewed: October 2026