ACADEMIC WRITING GUIDE
How to Conduct and Report Mediation Analysis in a Dissertation (With Worked Examples)
A dissertation may establish that two variables are related without explaining why that relationship exists. For example, newly hired employees who find onboarding materials easier to use may also report better task coordination. Does understanding their role help explain that association?
That is the kind of question mediation analysis can investigate.
Mediation analysis examines whether an association or effect involving a predictor and outcome can be described partly through an intermediate variable. Used carefully, it can help researchers investigate plausible mechanisms. Used carelessly, it can encourage causal conclusions that their research designs cannot support.
This guide explains the main concepts, how to plan a mediation analysis, how to estimate and interpret indirect effects, and how to report findings in a dissertation. It includes worked numerical examples and a reporting template.
Important: All study scenarios, coefficients, confidence intervals and sample descriptions in the worked examples are hypothetical teaching illustrations. They are not findings from a real dataset or software analysis.
What Is Mediation Analysis?
Mediation analysis investigates the role of an intermediate variable, called a mediator, in a proposed relationship between a predictor and an outcome.
A simple mediation model includes three variables:
X — Predictor: The variable associated with, or hypothesized to influence, the outcome.
M — Mediator: The intermediate variable proposed to help explain the relationship.
Y — Outcome: The variable whose variation the research seeks to explain.
Consider a hypothetical workplace study:
X: Perceived usability of onboarding materials.
M: Role clarity.
Y: Self-reported task coordination.
The research question might be:
Is the association between perceived usability of onboarding materials and task coordination statistically mediated by role clarity?
The proposed model is:
Onboarding-material usability (X) → Role clarity (M) → Task coordination (Y)
The researcher estimates whether onboarding-material usability is associated with role clarity and whether role clarity is associated with task coordination after accounting for onboarding-material usability.
However, estimating this indirect association does not by itself demonstrate that onboarding-material usability causes changes in role clarity or task coordination.
Mediation Versus Moderation: What Is the Difference?
Mediation and moderation answer different research questions.
Mediation asks whether an intermediate variable helps account for the relationship between a predictor and outcome.
Moderation asks whether the strength or direction of a relationship differs according to another variable.
For example:
Mediation question: Does role clarity help explain the association between onboarding-material usability and task coordination?
Moderation question: Is the association between onboarding-material usability and task coordination stronger among employees working remotely than among those working on site?
A moderator changes how a relationship is modeled across conditions or levels. A mediator occupies an intermediate position in a proposed explanatory process.
A study can include both, but the researcher must specify what each variable is hypothesized to do and why the design can support that interpretation.
Understanding the Four Paths in Simple Mediation
A commonly used simple linear mediation model contains four quantities.
| Path | Meaning |
|---|---|
| a | Relationship between predictor X and mediator M |
| b | Relationship between mediator M and outcome Y, adjusting for X |
| c′ | Direct relationship between X and Y, adjusting for M |
| c | Total relationship between X and Y in the corresponding model without M |
The indirect effect is calculated as:
Indirect effect = a × b
In a compatible linear mediation model without exposure–mediator interaction and under the relevant modeling assumptions:
Total effect = Direct effect + Indirect effect
Or:
c = c′ + ab
Use the same analytic observations, coefficient scale and justified covariate set across the corresponding ordinary least-squares regressions, including intercepts. Changing the sample because of missing data, or adding different controls to different equations, can break the numerical comparison. With covariates C, the model can be written as:
M = iM + aX + gC + eM
Y = iY + c′X + bM + hC + eY
Y = iT + cX + kC + eTHere i denotes an intercept, g, h and k denote covariate coefficients, and e denotes a residual. The third equation estimates the total association without conditioning on M. The algebraic decomposition in a compatible additive linear model does not require the associations to be causal; causal interpretation requires additional assumptions.
These identities should not be applied automatically to nonlinear models, odds ratios or incompatible specifications. For a binary outcome modeled with logistic regression, coefficients from models with and without M are not directly comparable in the same way. With an X-by-M interaction, the indirect effect can depend on the value of X. Define the effect and its scale before choosing a suitable mediation estimator.
The official lavaan mediation tutorial demonstrates the basic path model and defines the indirect effect through the product of its estimated coefficients.
When Is Mediation Analysis Appropriate for a Dissertation?
Mediation analysis is most useful when the researcher has a meaningful explanatory hypothesis, suitable measurements and a design that can support the planned interpretation.
Before proceeding, answer four questions.
1. Is there a defensible theoretical explanation?
The mediator should not be chosen simply because it correlates with the predictor and outcome.
Suppose a researcher argues that usable onboarding materials may help new employees understand their responsibilities, and this role clarity may be related to better task coordination. Relevant theory and prior research should support why role clarity belongs between the other variables in the proposed model.
The theoretical argument should come before selecting statistically convenient mediators.
2. Is the proposed order of variables credible?
A mechanism normally implies an order: the predictor precedes the mediator, which precedes the outcome.
A cross-sectional questionnaire measuring all three variables simultaneously generally cannot establish that sequence.
Longitudinal and experimental designs can provide stronger evidence about temporal ordering, but even they do not automatically remove all mediator–outcome confounding.
3. Are the variables measured appropriately?
A researcher needs defensible operational definitions, scoring rules, reliability and validity evidence, and a clear understanding of what the instruments actually measure.
For example, self-reported task coordination cannot automatically be interpreted as observed team performance or independently recorded productivity.
Our guide to writing operational definitions of research variables explains how to connect a theoretical construct to an observable measure.
4. Can the analysis accommodate important alternative explanations?
Potential confounders must be considered in relation to the research design and the proposed causal structure.
For instance, relevant prior work experience may be associated with both role clarity and task coordination. Ignoring it could distort the estimated mediator–outcome relationship.
However, adjusting for every available variable is not necessarily appropriate either. Variables affected by the predictor, colliders or poorly measured controls may introduce additional problems.
Researchers should justify their adjustment strategy using substantive knowledge rather than automatically including every questionnaire item.
How to Conduct Mediation Analysis: Seven Practical Steps
Step 1. State the research question and hypothesis
Begin with a research question identifying the predictor, mediator, outcome and population.
For example:
Among newly hired employees at one organization, is role clarity a statistical mediator of the association between perceived usability of onboarding materials and self-reported task coordination?
A corresponding hypothesis is:
The indirect association between onboarding-material usability and task coordination through role clarity differs from zero.
This is an association-based hypothesis. It deliberately avoids assuming causality from the proposed mediation structure.
Step 2. Define your variables and research design
Document the instrument or data source for each variable, the participants, sampling strategy, measurement period and nature of the study.
Distinguish planned data collection from completed research.
Before collecting data, establish whether the design provides sufficiently informative measurements of X, M and Y. A mediation analysis cannot compensate for a research question that the data are incapable of answering.
Our dissertation methodology chapter guide explains how research questions, sampling, data collection, analysis and ethics should align.
Step 3. Prepare and examine the dataset
Before estimating paths, verify participant eligibility, duplicate records, variable coding, missing values and consistency of measurement scales.
Inspect descriptive statistics and relationships between variables. Check whether particular observations have disproportionate influence and whether model assumptions are appropriate for the chosen estimation method.
Document what you changed and why.
Do not replace missing responses with zero unless zero genuinely represents the observed value and the analysis plan supports that treatment.
The missing-data approach should be appropriate to the nature and extent of missingness and consistent with the models being estimated.
Step 4. Specify the mediation model
In a simple linear model, the mediator is predicted from X, while the outcome is predicted from X and M.
The first regression estimates path a.
The second regression estimates paths b and c′.
The indirect effect is the product ab.
Relevant covariates can be included when justified by theory and the design. State whether coefficients are standardized or unstandardized; do not combine the two casually in the same calculation.
The models should be identified and checked before interpreting the results.
Step 5. Estimate the indirect effect and its uncertainty
The product of two estimated coefficients can have a sampling distribution that is not well represented by a normal approximation.
For this reason, researchers frequently use bootstrap confidence intervals for indirect effects.
Bootstrapping repeatedly resamples observations from the available data, estimates the relevant model and constructs an interval for the indirect effect using a specified procedure.
Resample the complete observation rows together, refit both equations in each resample, and calculate a new product ab each time. Do not bootstrap a and b independently or multiply endpoints from their separate confidence intervals. For a percentile 95% interval, use the 2.5th and 97.5th percentiles of the bootstrap products. A request for 5,000 resamples is a practical illustration, not a guarantee of accuracy; inspect warnings, failed resamples and stability.
Report the interval method, requested and successfully used resamples, and random seed. Ordinary row resampling assumes independent observations. Employees nested within teams, repeated measurements and complex survey samples need a resampling scheme or estimator matched to that dependence; the simple row bootstrap is not automatically valid.
A 95% confidence interval that excludes zero provides evidence that the estimated indirect effect differs from zero under the analysis procedure and its assumptions.
It does not demonstrate that the mediation mechanism is causal.
The UCLA Statistical Consulting Group's mediation tutorial explains why uncertainty estimation matters for products of coefficients.
Step 6. Evaluate the model and interpret carefully
A statistical indirect effect can be present even if the total relationship is small or statistically uncertain.
Therefore, a statistically significant total effect is not always a necessary prerequisite for investigating an indirect effect.
Also avoid treating a nonsignificant direct effect as proof that the mediator explains everything.
The direct effect's uncertainty, the possibility of unmeasured confounding, the temporal ordering of measurements and the adequacy of the model all matter.
If the study is observational and cross-sectional, it may be more defensible to describe the results as consistent with a proposed statistical mediation model, rather than claiming a mechanism has been established.
For continuous-score regressions, inspect functional form, residual patterns, influential observations and collinearity, and justify the uncertainty method. Reliable composite scores still contain measurement error; a score-based model does not estimate an item-level measurement model. The usual three-variable path model with all a, b and c′ paths is just identified, with zero global degrees of freedom. Its perfect global fit cannot validate the proposed ordering or mechanism.
Step 7. Report the findings transparently
Your dissertation should explain:
The theoretical justification for the mediator.
The study design, sample and measurements.
The estimation method, software and relevant settings.
Any justified covariates and model checks.
Paths a, b and c′, plus the total relationship where applicable.
The indirect effect and its confidence interval.
The implications, limitations and boundaries of interpretation.
For additional guidance on presenting coefficients, uncertainty and statistical findings, consult our statistical-results reporting guide.
Worked Example 1: A Simple Mediation Model
Consider an entirely hypothetical cross-sectional survey of newly hired employees.
The researcher investigates the proposed relationship among onboarding-material usability, role clarity and task coordination.
Assume the same employees completed suitably designed questionnaires and the researcher estimated a compatible set of linear regression models.
The following coefficients are invented to demonstrate how the calculations work. They are not estimated from an actual sample. For this illustration, X, M and Y are separate mean composite scores on 1–7 scales, scored so that higher values indicate greater usability, clearer roles and better self-reported coordination. These are illustrative scoring definitions, not a claim that an instrument has been validated. All coefficients are unstandardized; the model has no covariates.
| Parameter | Hypothetical estimate | Interpretation |
|---|---|---|
| a | 0.40 | Higher onboarding-material usability is associated with higher role clarity |
| b | 0.50 | Higher role clarity is associated with higher task coordination, adjusting for onboarding-material usability |
| c′ | 0.10 | Direct adjusted association between onboarding-material usability and task coordination |
| ab | 0.20 | Estimated indirect association through role clarity |
| c | 0.30 | Total association under the stated linear model |
Calculate the indirect effect
Using the invented coefficients:
ab = 0.40 × 0.50 = 0.20
The estimated indirect association is 0.20 task-coordination scale points per one-point difference in onboarding-material usability through the modeled role-clarity pathway. Specifically, a = 0.40 role-clarity points per X point and b = 0.50 coordination points per M point, so the mediator units cancel in the product. This describes a model-based association, not a predicted causal improvement for an employee.
This calculation does not mean that 20% of task coordination is explained by role clarity. An indirect-effect coefficient is not automatically a percentage, variance-explained statistic or proportion mediated.
Calculate the total effect
Under the stated compatible linear model:
c = 0.10 + 0.20 = 0.30
The total estimated association is 0.30 coordination points per one-point difference in X. The direct coefficient of 0.10 is the adjusted X–Y association at a fixed M value in this model. No confidence intervals for a, b, c′ or c have been supplied, so the illustration cannot establish their individual statistical significance. A real dissertation should report their uncertainty as well as the interval for ab.
Assign an illustrative 95% percentile-bootstrap interval of [0.09, 0.32] to this teaching example. These endpoints are invented, not computed from the three path coefficients or an actual bootstrap run. Individual-level data and the complete estimation procedure would be needed to obtain a real bootstrap interval.
Because this hypothetical interval excludes zero, the researcher would report statistical evidence of a nonzero indirect association under the model.
But because the survey measured all variables at one time, the result would not establish whether role clarity actually preceded task coordination or whether unmeasured factors influenced both.
Example of cautious dissertation reporting
In this hypothetical cross-sectional mediation model, perceived usability of onboarding materials was positively associated with role clarity (a = 0.40). Role clarity was also positively associated with self-reported task coordination after adjustment for onboarding-material usability (b = 0.50). The estimated indirect association was 0.20, with an illustrative 95% bootstrap confidence interval of [0.09, 0.32]. The direct coefficient was 0.10, and the total coefficient was 0.30 under the specified linear model. Although these estimates are consistent with the proposed statistical mediation structure, the simultaneous measurement of variables prevents a strong conclusion about temporal order or causal mediation.
Notice that the example reports what the numbers represent and what they do not establish.
Worked Example 2: When the Indirect Effect Is Uncertain
Imagine a different hypothetical dissertation examining perceived supervisor support, research confidence and progress on a postgraduate research project.
Treat the following path estimates as unstandardized coefficients on consistently defined score scales:
a = 0.30
b = 0.20
c′ = 0.15
The indirect effect is:
ab = 0.30 × 0.20 = 0.06
The corresponding total association is c = 0.15 + 0.06 = 0.21 under the same additive linear-model conditions.
Assign an illustrative 95% percentile-bootstrap confidence interval of [−0.02, 0.15]. Its endpoints are invented teaching values, not software output.
Because this interval includes zero, the analysis does not provide sufficiently clear statistical evidence of a nonzero indirect effect using that interval-based criterion.
The appropriate interpretation is not necessarily that the mediator has no role. The estimate could be imprecise because of limited information, variability, measurement problems or other features of the design.
A careful dissertation statement might be:
The estimated indirect association through research confidence was 0.06, with an illustrative 95% bootstrap confidence interval of [−0.02, 0.15]. Because this interval includes zero, the analysis did not establish a statistically distinguishable indirect association under the specified model. The findings should not be interpreted as proof that research confidence plays no role.
Reporting an uncertain result accurately is preferable to changing models repeatedly until an apparently significant result appears.
How to Interpret Direct, Indirect and Total Effects
Mediation terminology can be misleading when treated as a simple classification exercise.
A significant indirect effect with a remaining direct effect
The estimated indirect relationship and the direct association may both be distinguishable from zero.
This can be consistent with multiple pathways or other explanations. It does not prove that the researcher has measured every relevant pathway.
A significant indirect effect with an uncertain direct effect
This does not automatically establish “full mediation.”
The confidence interval for the direct effect may still include values that matter substantively. Measurement error or unmeasured confounding may remain.
It is often more informative to describe the estimated direct and indirect effects separately than to assign a rigid full-versus-partial label.
Direct and indirect effects with different signs
The estimated direct and indirect effects can oppose each other.
For example, the direct coefficient may be positive while the indirect coefficient is negative.
Their combination can produce a small total effect, even though the component estimates are materially different.
For a separate numerical illustration, c′ = 0.25 and ab = −0.20 give c = 0.25 + (−0.20) = 0.05. This cancellation says nothing about statistical significance without uncertainty estimates. The ratio ab/c would be -4 here, so calling it a “percentage mediated” would be misleading.
This is one reason not to rely on the total effect alone when evaluating a theoretically motivated mediation hypothesis.
An indirect effect does not guarantee a causal mechanism
This is the central interpretation safeguard.
Statistical mediation models represent relationships among measured variables. Causal mediation analysis requires stronger identification assumptions, including consideration of confounding involving the predictor, mediator and outcome.
Imai, Keele and Tingley (2010) distinguish defining an effect, identifying it from a design, estimating it and assessing sensitivity. For a standard natural-effects interpretation, assumptions include adequate control of X–M, X–Y and M–Y confounding, appropriate temporal order, consistency of the exposure and mediator definitions, and sufficient overlap for relevant comparisons. The standard identification approach also rules out mediator–outcome confounders caused by X. Merely adding a post-X covariate does not solve that problem.
Randomizing X can address confounding of its assignment; it does not randomize M or automatically identify the mediated effect. Imai, Keele and Yamamoto (2010) develop identification and sensitivity analysis for these issues. For example, a researcher can examine how much unmeasured M–Y confounding would change the indirect-effect conclusion. Sensitivity analysis probes specified assumption violations; it does not certify the absence of bias. An ordinary adjusted coefficient c′ should not be labeled a natural or controlled direct causal effect without defining and identifying that estimand.
Mediation Analysis Using Statistical Software
Several established programs can estimate mediation models, including R, SPSS with appropriate procedures or extensions, Stata, Mplus and structural-equation-modeling software.
Software choice should follow the model and available data, not determine the substantive question.
A practical lavaan estimation template
The following template fits an observed-score model to your own complete, independent observations. Supply a data frame named dat with numeric columns X, M and Y, using the same rows for both equations. The template deliberately stops if values are missing or non-finite; it does not prescribe deletion of incomplete cases. Decide and justify a missing-data approach before fitting your study's model.
library(lavaan)
vars <- c("X", "M", "Y")
stopifnot(is.data.frame(dat), all(vars %in% names(dat)))
d <- dat[, vars, drop = FALSE]
stopifnot(all(vapply(d, is.numeric, logical(1))))
stopifnot(all(is.finite(as.matrix(d))))
model <- '
M ~ a*X
Y ~ b*M + cp*X
indirect := a*b
direct := cp
total := cp + a*b
'
set.seed(8613)
fit <- sem(model, data = d, estimator = "ML",
se = "bootstrap", bootstrap = 5000)
stopifnot(lavInspect(fit, "converged"))
pe <- parameterEstimates(fit, ci = TRUE, level = 0.95,
boot.ci.type = "perc",
standardized = FALSE)
pe[pe$op %in% c("~", ":="),
c("lhs", "op", "rhs", "est", "se", "ci.lower", "ci.upper")]
lavInspect(fit, "nobs")
packageVersion("lavaan")
sessionInfo()This is an adaptation template, not a run that generated the invented examples above. Check the installed package's help for your version. Read all bootstrap warnings and confirm usable resamples before reporting intervals; checking original-model convergence alone is insufficient. Retain output, code, the seed and software versions in your analysis record.
The official lavaan mediation guide explains labeled paths, defined parameters and bootstrap estimation. The package's parameterEstimates reference documents percentile intervals and notes that its displayed z-based p-values do not become bootstrap p-values. The template therefore extracts estimates and confidence intervals without presenting a z-test as the bootstrap test. These are intervals for unstandardized coefficients; they should not be attached to a standardized estimate.
For a justified baseline covariate C, add C to both regressions and include it in the prepared data. Clustered observations, categorical outcomes, latent variables, interactions and a missing-data estimator require an appropriately revised specification and uncertainty procedure; changing a single option is not a universal solution.
For dissertations that include latent constructs and more complex path models, our structural equation modeling guide explains broader model-specification and reporting considerations.
How Many Participants Do You Need for Mediation Analysis?
There is no universal minimum sample size that makes every mediation analysis reliable.
The information required depends on factors including the expected path magnitudes, measurement quality, model complexity, planned covariates, outcome distributions and desired precision or power.
A sample that provides adequate power for an ordinary regression coefficient may be insufficiently informative about a smaller indirect effect.
When planning a dissertation, justify the sample using a method matched to the proposed mediation model. Simulation-based planning may be appropriate when a simple formula cannot adequately represent the analysis. Fritz and MacKinnon (2007) compare sample-size requirements across path strengths and mediation tests. Their results support model-specific planning, not a universal minimum or a substitute for considering your own design.
Our power-analysis guide explains why sample-size planning must reflect the actual research question and model rather than a generic minimum.
A Dissertation Mediation-Analysis Reporting Template
Separate what was planned, what was done and what was found. The AGReMA reporting statement provides detailed guidance on reporting mediation analyses of trials and observational studies. Use it alongside your institution's requirements. The prompts below are an adaptation aid, not a claim of complete AGReMA compliance.
| Part | Information to supply |
|---|---|
| Rationale and design | Proposed pathway, target population, measurement occasions and whether the analysis was prespecified or exploratory |
| Data and measurements | Analytic sample size, exclusions, scale units, scoring and measurement evidence, missingness and its treatment |
| Specification | Equations, covariates and their justification, interactions, estimator, dependence structure and causal assumptions if invoked |
| Uncertainty | Confidence level and interval method, requested and usable bootstrap resamples, seed, software and version |
| Results | a, b, c′, ab and compatible c; coefficient scale, estimates and intervals; relevant diagnostics and failed or inadmissible solutions |
| Interpretation | Magnitude in meaningful units, uncertainty, plausible alternatives, sensitivity analyses and limits on causal claims |
Methods template — replace every bracketed field with what you actually did:
We examined whether [M] statistically mediated the association between [X] and [Y] in [design and population]. Measurements occurred at [times]. The analytic sample included [n] observations after [exclusions and missing-data handling]. We estimated [equations and estimator], adjusting for [justified covariates, or none], using [software and version]. Coefficients were [unstandardized, with units, or explicitly defined standardized estimates]. The indirect association was defined as ab, with a [confidence level and interval method] interval based on [requested and usable resamples] using seed [seed]. [Describe diagnostics, dependence handling and prespecification or exploratory changes.]
Results template — report estimates and their actual intervals:
The X–M coefficient was a = [estimate, interval]. The M–Y coefficient adjusted for X and [covariates] was b = [estimate, interval]. The indirect association was ab = [estimate, interval], using [interval method]. The direct coefficient was c′ = [estimate, interval], and the corresponding total coefficient was c = [estimate, interval]. In the reported units, [explain the magnitude]. [Describe uncertainty and substantive importance.] Because [design-specific limitations], these findings [state the defensible associational or explicitly justified causal interpretation].
Do not insert the article's teaching numbers into these templates as if they were your findings. When direct and indirect effects oppose each other or c is near zero, report the components clearly rather than presenting an unstable proportion mediated.
Common Mediation-Analysis Mistakes to Avoid
Choosing a mediator after seeing the results. A post-hoc mediator may produce an interesting exploratory model, but it must not be presented as a prespecified theoretical hypothesis.
Treating cross-sectional associations as causes. Simultaneous measurements cannot independently establish the sequence implied by a causal mechanism.
Ignoring confounding. An apparently meaningful indirect effect may be distorted by unmeasured factors associated with the mediator and outcome.
Relying only on the total effect. The absence of a statistically significant total association does not automatically rule out an indirect effect.
Declaring full mediation because c′ is nonsignificant. A nonsignificant coefficient does not prove that the underlying effect is exactly zero.
Reporting only p-values. Provide effect estimates and appropriate uncertainty intervals so readers can assess direction, precision and substantive importance.
Interpreting ab as a percentage. The product of coefficients is not automatically a percentage of an outcome or a proportion mediated.
Using standardized and unstandardized coefficients interchangeably. Keep the scale of reported coefficients explicit and calculate effects consistently.
Skipping measurement checks. Unreliable or invalid measures can make a theoretically appealing model difficult to interpret.
Treating bootstrap confidence intervals as a solution to bias. Bootstrapping addresses aspects of sampling uncertainty; it does not remove confounding or repair an inadequate research design.
Frequently Asked Questions
Does mediation analysis prove causation?
No. An estimated indirect effect does not by itself establish causation. Causal interpretation requires a defensible design and identification assumptions that the researcher can explain and evaluate.
Can I conduct mediation analysis using cross-sectional data?
Yes, a statistical mediation model can be estimated from cross-sectional data when appropriate. However, interpretation should generally be associational because the design cannot establish the required temporal sequence on its own.
Must the total effect be statistically significant before testing mediation?
Not necessarily. A theoretically justified indirect effect may be examined even when the total effect is uncertain, provided the analysis and interpretation are appropriate.
Should I use bootstrapping?
Bootstrap confidence intervals are widely used for indirect effects because the distribution of a product of coefficients can be asymmetric. The method must still be suitable for the sampling structure and model.
Can a mediator and moderator appear in the same model?
Yes. More complex models can examine indirect pathways that vary across groups or levels of another variable. Such analyses require clearly specified hypotheses and careful interpretation.
How do I report mediation results in APA style?
Report the study design, model and software, relevant path estimates, indirect effect and confidence interval, the method of estimating uncertainty, and appropriately cautious interpretation. Follow the APA guidance and institutional requirements applicable to your programme.
Final Checklist Before Submitting Your Mediation Analysis
Before finalizing the dissertation, confirm that:
The mediator is justified by theory or relevant prior evidence.
The research design and timing support the intended interpretation.
Predictor, mediator and outcome measurements are clearly defined.
The model, covariates and assumptions are explained.
Missing data and influential observations have been addressed appropriately.
The indirect effect and its uncertainty are reported.
Direct and total associations are distinguished where applicable.
Hypothetical, exploratory or post-hoc analyses are identified honestly.
Causal language does not exceed the evidence.
All tables, figures and reported values agree with the statistical output.
What a Defensible Mediation Analysis Should Show
Mediation analysis can help researchers investigate how a proposed relationship may operate through an intermediate variable. Its value lies not simply in obtaining a statistically significant indirect effect, but in building a coherent connection between theory, research design, measurement, estimation and interpretation.
A strong dissertation explains the proposed mediator, estimates the relevant paths transparently, reports uncertainty and acknowledges where the evidence does not justify a causal conclusion.
Need Research Support Tailored to Your Dissertation?
Acadelyra offers human-led research support, methodology guidance, statistical-results review, editing and dissertation support within our academic integrity standards. You remain responsible for your research decisions and submitted work; support must comply with your institution's rules.
If you need help understanding your supervisor's requirements, checking an analysis plan or improving how you explain your own research findings, explore Acadelyra's academic support services or request academic help.
Support is based on your brief, the agreed scope and applicable institutional requirements. Academic outcomes are not guaranteed.
Research Guidance and Further Reading
Imai, K., Keele, L., & Tingley, D. (2010). A general approach to causal mediation analysis. Psychological Methods, 15(4), 309–334. DOI: 10.1037/a0020761. Author page links to the full paper.
Imai, K., Keele, L., & Yamamoto, T. (2010). Identification, inference and sensitivity analysis for causal mediation effects. Statistical Science, 25(1), 51–71. DOI: 10.1214/10-STS321. Author page links to the full paper.
Lee, H., Cashin, A. G., Lamb, S. E., et al. (2021). A guideline for reporting mediation analyses of randomized trials and observational studies: The AGReMA statement. JAMA, 326(11), 1045–1056. DOI: 10.1001/jama.2021.14075. Linked author manuscript in PubMed Central.
Fritz, M. S., & MacKinnon, D. P. (2007). Required sample size to detect the mediated effect. Psychological Science, 18(3), 233–239. DOI: 10.1111/j.1467-9280.2007.01882.x. Linked PubMed record and abstract.
UCLA Statistical Consulting Group. Introduction to mediation analysis using Mplus. University tutorial on indirect effects, estimation and interpretation.
lavaan project. Mediation tutorial and parameterEstimates reference. Official software guidance for defined effects, bootstrap estimation and interval extraction.