Psychology dissertation mediation analysis examines whether an association or intervention effect may operate through an intermediate variable. It can connect theory to evidence, but only when the proposed pathway, time order, measurements, design, assumptions, and uncertainty are treated seriously.
This guide explains how to formulate a mediation question, distinguish a mediator from related third variables, plan the design, estimate and interpret indirect effects, use software responsibly, and report the analysis without claiming more than the data support. Examples use common psychology dissertation settings and internationally relevant methodological guidance.
What is a Psychology Dissertation Mediation Analysis?
Mediation analysis studies a proposed pathway from a predictor or exposure, usually written as X, through a mediator M, to an outcome Y. The indirect effect represents the part of the X to Y relationship associated with change in M under the specified model. The direct effect represents the remaining X to Y relationship not operating through that mediator as modelled.
For example, a student might ask whether academic stress is associated with sleep disturbance partly through rumination. Stress is X, rumination is M, and sleep disturbance is Y. The analysis can estimate a statistical indirect effect, but a cross-sectional survey cannot establish that stress occurred before rumination or that rumination caused later sleep problems.
Mediation is more than running several regressions. The model must follow psychological theory and the design must support its ordering. Review the broader psychology dissertation regression analysis guide if variable coding, covariates, model assumptions, or coefficient interpretation are still unclear.
Table of Contents
Distinguish mediators from moderators and confounders
Third variables can play very different roles. Labelling every variable between X and Y a mediator creates a model that may be statistically estimable but conceptually incoherent.
| Variable role | Question it addresses | Psychology example | Main planning issue |
|---|---|---|---|
| Mediator | Through what proposed pathway is X related to Y? | Rumination between stress and sleep disturbance | Temporal order and mediator-outcome confounding |
| Moderator | For whom, or under what conditions, does an association differ? | Social support changes the stress-sleep association | Interaction, scale, and adequate information across levels |
| Confounder | What prior factor may influence two variables and distort their association? | Prior depression affects rumination and later sleep | Measure before the variables it may confound |
| Covariate | What variable is included for a justified adjustment or precision purpose? | Prespecified baseline sleep score | Avoid automatic adjustment and post-treatment variables |
| Collider | What common consequence of two variables could create bias if conditioned on? | Service use affected by distress and access barriers | Do not control for it merely because it is available |
Moderation and mediation can occur in the same theoretical model, but moderated mediation is more demanding than simple mediation. It asks whether an indirect effect varies with another variable. Use it only when the conditional pathway is theoretically justified, prespecified, measured well, and supported by an informative sample.
Start with a psychological mechanism, not a software template
A credible model begins with a reason why X should influence M and why M should influence Y. Summarise the theoretical mechanism, relevant evidence, plausible alternative pathways, and expected timing. A numbered software model is an estimation template, not a theory.
Write the question before specifying equations. A cautious observational question might be: “Is the association between perceived discrimination and wellbeing statistically mediated by belonging, after adjustment for prespecified pre-exposure covariates?” A stronger causal question requires a design and assumptions capable of supporting causal effects.
Your psychology dissertation hypothesis should name X, M, Y, the proposed direction, population, and design. Separate confirmatory pathways from later exploratory alternatives.
Draw the causal structure before choosing covariates
Create a simple diagram showing X, M, Y, their timing, common causes, competing mediators, and variables affected by X. This exercise exposes assumptions that a regression output can hide.
In the familiar single-mediator notation:
- Path a estimates the association of X with M.
- Path b estimates the association of M with Y, conditional on X and included covariates.
- Path c is often used for the total X to Y effect under the selected framework.
- Path c-prime is the direct X to Y effect after accounting for M as specified.
- The indirect effect is commonly estimated as the product a × b in a linear model without an X by M interaction.
These coefficients describe a model. They become causal effects only under additional identification assumptions. Imai, Keele, and Tingley emphasise that causal mediation requires assumptions about the exposure and about unmeasured confounding of the mediator-outcome relationship. Randomising X does not usually randomise M.
Choose a design that supports the proposed order
Cross-sectional design
Measuring X, M, and Y at one occasion may identify an indirect association compatible with the proposed model. It cannot show whether the mediator preceded the outcome, exclude reverse direction, or demonstrate within-person change. Use language such as “statistical indirect effect” and discuss rival orderings.
Longitudinal design
Longitudinal measurement can strengthen temporal evidence when X is measured before M and M before Y. Include appropriate baseline measures when justified, choose intervals that match the psychological process, and address attrition and time-varying confounding. Three waves are not automatically causal, but they can be more informative than simultaneous measurement.
Experimental design
Randomising an intervention can support the causal effect of X on subsequent variables when implementation is sound. The mediator is still normally observed, not randomised, so mediator-outcome confounding remains possible. Plan manipulation fidelity, measurement timing, post-randomisation confounders, and sensitivity analysis. See the psychology dissertation experimental design guide.
Identify the assumptions your interpretation needs
No mediation method removes the need for assumptions. State which interpretation you seek and which conditions it requires.
- The model represents the relevant temporal and causal ordering.
- X, M, Y, and covariates are measured with adequate quality.
- There is no important unaddressed confounding for the relationships used to identify the effects.
- Covariates are not consequences of X unless the chosen method explicitly handles them.
- The functional forms and links are appropriate for the variable types.
- Observations are independent unless clustering or repeated measures are modelled.
- Missing data and selection processes do not invalidate the estimates under the stated method.
- There is sufficient overlap in exposure and covariate patterns for the intended comparisons.
A sensitivity analysis can show how strong an unmeasured mediator-outcome confounder would need to be to change a conclusion under a defined framework. It does not prove that confounding is absent. Report the assumptions and the limits of the sensitivity analysis.
Select and measure X, M, and Y carefully
Define each construct conceptually and operationally. Use instruments with evidence relevant to the population, language, administration mode, and intended score interpretation. Measurement error in a mediator can distort path estimates and indirect effects.
Do not select a mediator only because it was included in the questionnaire. Explain why it is modifiable or responsive to X, why it plausibly precedes Y, and why the interval permits change. If X is stable trait anxiety, M is momentary coping, and Y is long-term wellbeing, the proposed timing needs a defensible account.
Report scoring, reverse-coded items, missing-item rules, reliability evidence in the current sample where appropriate, and any adaptation or translation. The reliability and validity guide explains why one internal-consistency coefficient does not establish valid measurement.
Plan sample size for the indirect effect
Mediation sample size depends on the likely a and b paths, measurement quality, design, covariates, variable distributions, missingness, clustering, desired precision, and estimation method. A generic “ten participants per variable” rule is not a defensible calculation.
Plan around the primary indirect effect rather than the total effect or one individual path. Simulation is often useful for complex or nonlinear models because it can represent realistic distributions, missingness, reliability, and analysis choices. When prior evidence is uncertain, evaluate several plausible scenarios instead of selecting a convenient effect size.
Allow for exclusions and attrition transparently. If the achievable sample cannot estimate the proposed model with useful precision, simplify the model, collect better measures, narrow the question, or present the work as exploratory. Do not add parallel and serial mediators simply to make the model appear advanced.
Estimate the indirect effect directly
The older causal-steps approach asks whether several individual regression paths are statistically significant. Modern mediation analysis focuses on estimating the indirect effect and its uncertainty. A significant total X to Y effect is not a universal prerequisite for an indirect effect because direct and indirect pathways can differ in direction, and total effects can be imprecise.
For a simple linear model without X by M interaction, the product a × b is widely used. Its sampling distribution is often asymmetric. Bootstrap confidence intervals repeatedly resample observations, estimate the indirect effect in each sample, and use the resulting distribution for interval estimation. State the interval type, number of resamples, software, version, seed or reproducibility settings where relevant, and treatment of missing data.
Bootstrapping does not repair poor temporal order, confounding, unreliable measures, an inappropriate model, or a biased sample. It addresses uncertainty under the fitted procedure.
Match the model to the variables
| Situation | Planning consideration | Common mistake |
|---|---|---|
| Continuous M and Y | Check linearity, residual structure, influential observations, and scale | Assume normal predictors are required |
| Binary M or Y | Use an appropriate link and effect definition | Interpret logistic coefficients as ordinary mean differences |
| Count outcome | Consider Poisson or negative-binomial structure and overdispersion | Force a skewed count into linear regression without evaluation |
| Repeated measures | Model within-person dependence and time ordering | Treat repeated observations as independent participants |
| Clustered sample | Account for shared setting, class, clinic, or group | Use individual-level standard errors only |
| Latent constructs | Integrate the measurement and structural models | Treat fallible sum scores as error-free without discussion |
In nonlinear models, direct, indirect, and total effects may not combine in the same simple way as in ordinary linear regression. Define the effect scale and use a method appropriate to the outcome and mediator types rather than copying interpretations from a continuous-outcome example.
Handle covariates, missing data, and influential cases transparently
Select covariates from theory, design, and a causal diagram before examining which choices produce significance. A pre-exposure common cause of M and Y may need adjustment. A variable caused by X may introduce post-treatment bias if entered casually. Controlling every demographic field is not automatically safer.
Describe missingness for X, M, Y, and covariates. Complete-case analysis can change the target population and introduce bias when inclusion depends on relevant variables. Use a justified method compatible with the analysis, such as multiple imputation or full-information estimation when appropriate, and include sensitivity checks.
Inspect distributions, impossible values, leverage, residual patterns, and data coding. Do not delete an observation merely because it weakens the indirect effect. Define exclusion and influence decisions in advance where possible, report them, and compare reasonable analyses when conclusions depend on a few cases. The psychology dissertation data analysis guide provides a complete screening workflow.
Choose software after defining the model
Regression software, structural equation modelling programs, and specialised packages can all estimate mediation models. The correct choice depends on variable types, clustering, missing data, latent constructs, interactions, and the causal effects required.
- IBM SPSS Statistics: recent documentation includes mediation procedures for direct and indirect effects across several variable types.
- PROCESS: a regression-based tool for SPSS, SAS, and R that estimates direct, indirect, conditional, and multiple-mediator models. Use the current documentation and report the version.
- R: packages support regression-based, structural-equation, multilevel, and causal mediation approaches. Record package versions and code.
- Structural equation modelling: useful when the research requires latent variables, several simultaneous pathways, or an integrated measurement model.
Software cannot decide whether M is a mediator, whether a covariate is a confounder, or whether causal interpretation is justified. Save syntax or code, label every variable, and verify results against the planned equations. Our SPSS psychology dissertation workflow explains reproducible project organisation.
Worked psychology dissertation example
Suppose a student asks whether perceived social support statistically mediates the association between loneliness and depressive symptoms among postgraduate students.
- Theory: loneliness may be related to reduced perceived support, which is associated with depressive symptoms. Reverse and reciprocal pathways remain plausible.
- Design: all variables are measured in one online survey. The student therefore frames the result as cross-sectional statistical mediation, not a causal mechanism.
- Variables: validated scale scores are calculated using documented missing-item and reverse-scoring rules.
- Covariates: only prespecified variables with a defensible role are included. Models with and without them are clearly distinguished.
- Analysis: the student estimates a, b, direct, total, and indirect effects with confidence intervals, checks model conditions, and records syntax.
- Result: the indirect-effect interval is reported with its units and uncertainty. Individual paths and sample descriptives are also shown.
- Interpretation: the pattern is consistent with the proposed model, but simultaneous measurement, self-report error, selection, and unmeasured confounding prevent a causal conclusion.
A future three-wave or experimental study could test temporal ordering more directly. The dissertation should explain this next step rather than presenting one survey as proof that social support transmits the effect of loneliness.
Report mediation analysis clearly
The AGReMA statement provides consensus-based recommendations for primary and secondary mediation analyses in randomised trials and observational studies. Use it alongside disciplinary and institutional requirements.
| Dissertation section | Information to report |
|---|---|
| Introduction | Theory, prior evidence, pathway, competing explanations, and mediation hypothesis |
| Method | Design, timing, sample, measures, covariate rationale, model, assumptions, software, missing-data and sensitivity plans |
| Results | Participant flow, descriptives, path estimates, indirect and direct effects, confidence intervals, diagnostics, sensitivity and deviations |
| Discussion | Meaning on the reported scale, uncertainty, alternative pathways, design limits, measurement error, generalisability, and implications |
Include a path diagram that labels variables and estimates, but ensure the text and table remain understandable without it. Report exact estimates and intervals rather than only stars or “full” and “partial” mediation labels. The psychology dissertation results guide explains how to present statistics without turning the chapter into software output.
Common mediation mistakes and repairs
- Choosing a mediator after seeing correlations. Label exploration and seek independent confirmation.
- Using cross-sectional data to prove a process. Report an indirect association and discuss temporal ambiguity.
- Requiring a significant total effect. Estimate the prespecified indirect effect directly and explain the model.
- Testing each path separately. Report the indirect-effect estimate and interval.
- Calling every adjusted variable a confounder. Define roles from theory, timing, and a causal diagram.
- Ignoring mediator-outcome confounding. State the assumption, adjust appropriately, and use sensitivity analysis when possible.
- Adding many mediators to a small sample. Prioritise a parsimonious model with sufficient information.
- Deleting cases to obtain significance. Use transparent rules and sensitivity checks.
- Reporting only PROCESS model numbers. Describe the substantive paths, equations, options, and software version.
- Equating a bootstrap interval with causal proof. Separate statistical inference from design validity.
A practical mediation-analysis workflow
- Define X, M, Y, population, timing, and the intended interpretation.
- Justify the pathway with psychological theory and prior evidence.
- Draw a causal diagram with confounders, colliders, and alternative mediators.
- Select measures and intervals that match the proposed process.
- Choose the design and acknowledge what it cannot identify.
- Plan sample size for the primary indirect effect and expected missingness.
- Prespecify covariates, model form, missing-data rules, sensitivity checks, and software.
- Protect raw data, score measures transparently, and inspect data quality.
- Estimate indirect, direct, and relevant total effects with uncertainty.
- Check assumptions, influential cases, alternative specifications, and deviations.
- Report using AGReMA and appropriate psychology standards.
- Interpret at the level supported by the design.
Psychology dissertation mediation analysis checklist
- The mediator has a theoretical role and plausible temporal position.
- The research question distinguishes statistical from causal mediation.
- X, M, Y, covariates, and their measurement times are explicit.
- Competing pathways and mediator-outcome confounding are considered.
- The design and sample-size plan fit the primary indirect effect.
- Measure scoring, reliability, missingness, and exclusions are documented.
- Model form matches continuous, binary, count, clustered, or repeated data.
- The indirect effect is estimated directly with an interval.
- Software, version, options, code, and reproducibility settings are recorded.
- Sensitivity analyses and deviations are reported honestly.
- Tables and diagrams agree with the text.
- Causal language does not exceed the design and assumptions.
Frequently asked questions
Does X need to have a significant total effect on Y?
No universal rule requires a statistically significant total effect before estimating a prespecified indirect effect. Direct and indirect pathways can differ in direction, and the total estimate may be imprecise. Report all relevant effects and explain the theoretical model.
Can I conduct mediation analysis with cross-sectional data?
You can estimate a statistical indirect effect, but simultaneous measurement cannot establish temporal order or causal mediation. Use cautious language, discuss reverse pathways, and avoid describing the mediator as a proven mechanism.
Is bootstrapping required for mediation analysis?
Bootstrap intervals are common because the product of coefficients can have an asymmetric sampling distribution. The appropriate inference method depends on the model and data. Report and justify the chosen method; remember that bootstrapping does not solve confounding or poor design.
Can PROCESS prove mediation?
No. PROCESS estimates a specified regression-based model. Evidence for a causal mechanism also depends on theory, temporal order, measurement, confounding assumptions, sampling, and sensitivity to alternative explanations.
How many mediators should a dissertation include?
Include only pathways justified by theory and supported by the design and sample. One carefully specified mediator is often more defensible than several weakly measured parallel or serial mediators.
What should I report from a mediation model?
Report the design, sample, variable timing and scoring, covariates, path estimates, indirect and direct effects with intervals, model scale, software and options, missing-data handling, assumptions, sensitivity checks, and deviations. Use AGReMA where applicable.
Conclusion
Psychology dissertation mediation analysis can test whether evidence is consistent with a theoretically proposed pathway. Its value depends less on a software model number than on temporal logic, measurement quality, confounding control, appropriate effect definitions, direct estimation of the indirect effect, and transparent uncertainty.
If you need ethical academic support, Psychology Dissertation Help can review your own mediation diagram, analysis map, reporting table, or supervisor feedback. You remain responsible for the theory, data, analysis, interpretation, and authorship, and all support should comply with your institution’s rules.
Authoritative references
- EQUATOR Network: AGReMA reporting guideline
- Imai, Keele, and Tingley: A general approach to causal mediation analysis
- MacKinnon, Fairchild, and Fritz: Mediation analysis
- APA Style: quantitative research reporting standards
- Official PROCESS macro documentation site
- IBM SPSS Statistics: mediation analysis documentation
