Psychology dissertation Bayesian analysis offers a coherent way to update uncertainty with evidence, estimate effects, compare models, and make predictions. It is not a shortcut around careful design. A defensible dissertation still needs a clear question, an appropriate likelihood, justified priors, diagnostic checks, sensitivity analysis, and transparent reporting. This guide shows how to make those decisions without treating Bayesian software as a black box.
Bayesian methods can answer questions that students often try to force into a single p-value. You can estimate the probability that an effect exceeds a practically meaningful threshold, quantify evidence for a null model, or predict an outcome for a new participant. The method is valuable because it makes assumptions explicit and carries uncertainty through the analysis.
What Bayesian analysis changes
A Bayesian model combines a prior distribution with a likelihood. The prior represents plausible parameter values before observing the current data. The likelihood describes how the observed data are expected to arise for different parameter values. Their combination produces a posterior distribution, which expresses updated uncertainty after seeing the data.
This is a different logic from conventional null-hypothesis significance testing. A frequentist confidence interval is produced by a procedure that would cover the true value at a stated rate over repeated samples. A Bayesian credible interval, conditional on the model and data, assigns probability directly to a range of parameter values. The interpretations should not be interchanged.
Bayesian analysis does not make uncertainty disappear. Posterior conclusions depend on the model, measurement quality, sampling process, prior choices, and computational accuracy. The strongest dissertation explains those dependencies rather than presenting a posterior number as unquestionable truth.
| Research aim | Useful Bayesian output | Example psychology question |
|---|---|---|
| Estimate an effect | Posterior median or mean, credible interval, probability above a meaningful threshold | How much does sleep quality predict next-day anxiety? |
| Compare hypotheses or models | Bayes factor, posterior model probability when model priors are specified | Do the data favour no group difference or a non-zero difference? |
| Predict new observations | Posterior predictive distribution and predictive interval | What stress score is plausible for a new student with these characteristics? |
| Estimate variation across groups | Posterior distributions for population and group-level parameters | How much do intervention effects vary across clinics? |
Start with the research question, not the software
State what you want to learn before selecting a menu option or package. An estimation question might ask how strongly social support is associated with depressive symptoms. A model-comparison question might ask whether data support a null association, a positive association, or several competing mechanisms. A prediction question asks how accurately a model generalises to new cases.
Connect the question to a defensible psychology dissertation data analysis plan. Identify the outcome type, predictors, grouping structure, repeated observations, missingness, plausible interactions, and measurement limitations. A continuous approximately Gaussian outcome may suit a linear model, while a binary diagnosis requires a Bernoulli likelihood and logistic link. Repeated measurements may need a multilevel structure.
Define the estimand
The estimand is the quantity your analysis is meant to estimate. It might be a mean difference, regression coefficient, odds ratio, correlation, group-level variance, or predicted probability. Define its scale and direction. If a one-point change on a measure has little practical meaning, consider a standardised predictor or a threshold linked to theory or clinical practice.
Keep the estimand aligned with design. A cross-sectional association does not identify a causal treatment effect. A convenience sample does not automatically support population-wide claims. Bayesian computation can express uncertainty within a model, but it cannot repair selection bias, weak measurement, or an unidentified causal question.
Choose and justify priors
Priors should be plausible on the scale of the model. They can encode previous evidence, theoretical constraints, measurement limits, or weak regularisation that discourages implausibly large estimates. Calling every default “uninformative” is risky because a prior that seems broad on one scale may be influential after a nonlinear transformation.
A weakly informative prior usually allows a wide range of realistic effects while placing little probability on extreme values. An informative prior draws more strongly on relevant evidence. A sceptical prior concentrates near a null or small effect. The correct choice depends on the research purpose, population, measures, and model.
Translate subject knowledge into model units
Suppose the outcome is a 0 to 21 anxiety score and the predictor is weekly sleep duration in hours. A prior that puts substantial probability on a 30-point anxiety change per additional hour is impossible on the observed scale. Standardising variables can make generic priors easier to specify, but you must still explain how standardisation changes interpretation.
Use previous studies cautiously. A published estimate may come from a different age group, intervention, instrument, or model. If transportability is uncertain, widen the prior or centre it conservatively. Record the source, transformation, and rationale in the analysis plan. This makes your prior auditable rather than arbitrary.
Run prior predictive checks
Before fitting the model to observed outcomes, simulate data from the priors and likelihood. Ask whether the simulated scores, event rates, correlations, or group differences are plausible. Implausible predictions reveal conflicts between prior choices, parameterisation, and the measurement scale.
The Bayesian Analysis Reporting Guidelines recommend explaining the rationale for priors and using prior predictive assessment where appropriate. A prior predictive check is not a test that finds a uniquely correct prior. It is a practical way to reject choices that imply impossible or unreasonable observations.
| Prior strategy | When it can help | What to document |
|---|---|---|
| Weakly informative | Limited direct evidence, but realistic bounds are known | Scale, centre, and extreme outcomes the prior permits |
| Evidence-informed | Comparable studies and measures are available | Source studies, transformations, and population differences |
| Sceptical | Large effects are unlikely or false positives are costly | Meaning of “small,” width, and sensitivity to alternatives |
| Regularising hierarchical | Many groups or related coefficients require partial pooling | Group-level distribution and its effect on shrinkage |
Plan a psychology dissertation Bayesian analysis
A reproducible plan prevents the analysis from drifting toward whichever model gives the most attractive result. Use psychology dissertation preregistration to distinguish confirmatory decisions from later exploration. Specify the likelihood, predictors, contrasts, priors, sampling plan, diagnostic criteria, sensitivity analyses, exclusions, and reporting rules.
Plan sample size for the intended claim
Bayesian methods do not guarantee that a small sample is adequate. A weak dataset often produces a wide posterior or allows the prior to exert greater influence. Sample planning should reflect the desired precision, predictive performance, or probability of reaching a chosen evidence threshold under plausible data-generating conditions.
Simulation is often the clearest approach. Generate datasets under plausible parameter values, fit the proposed model, and examine interval width, bias, coverage in repeated simulations, or the frequency of decision outcomes. Compare several scenarios instead of planning around one optimistic effect. The site’s psychology dissertation power analysis guide explains the broader logic of prospective planning.
Separate confirmatory and exploratory analyses
Exploration is legitimate, but its status should be visible. If you revise a prior, add an interaction, transform an outcome, or exclude cases after seeing results, label the change and report the original specification when useful. A multiverse or sensitivity analysis can show whether the substantive conclusion depends on one defensible choice.
Fit the model and check computation
Many Bayesian models use Markov chain Monte Carlo sampling. The algorithm generates draws that approximate the posterior distribution. A numerical result is trustworthy only when the algorithm has explored the relevant posterior geometry adequately.
Run multiple chains from dispersed starting points. Inspect trace plots for mixing and stationarity. Review rank-normalised R-hat, effective sample size, Monte Carlo standard error, divergences, maximum tree-depth warnings, and energy diagnostics when relevant. The official Stan R-hat reference explains why R-hat should be close to one, while the Stan effective sample size documentation explains how autocorrelation affects precision.
Do not treat a single threshold as proof of convergence. A model can have an acceptable R-hat and still contain divergences or poor tail exploration. Conversely, a warning may reflect parameterisation rather than a scientific problem. Investigate warnings, consider rescaling predictors or using a non-centred parameterisation, and rerun the full diagnostic suite.
Use posterior predictive checks
Posterior predictive checks compare observed features with replicated data generated from the fitted model. The Stan guide to predictive checks distinguishes prior predictive simulation from posterior predictive assessment. The first evaluates what the model permits before seeing outcomes; the second evaluates whether a fitted model can reproduce important features of the observed data.
Choose checks that match the research problem. For symptom scores, compare the distribution, floor and ceiling rates, variance, and extreme values. For binary outcomes, compare prevalence across meaningful groups. For repeated observations, check within-person variability and time patterns. A model that predicts the mean but misses the tails or group structure may be inadequate for the intended claim.
Interpret posterior estimates responsibly
Report the full posterior summary rather than only whether an interval crosses zero. Provide a central estimate, interval level, units, direction, and practical context. For example, you might report a posterior median reduction of 2.1 points with a 95% credible interval from 0.4 to 3.8 points, conditional on the stated model and priors.
A posterior probability such as P(β > 0 | data, model) describes a model-conditional probability. It is not the probability that the entire theory is true. If using a region of practical equivalence, justify its bounds before viewing results and show how conclusions change under defensible alternatives. Link numerical interpretation to the principles in the psychology dissertation effect size guide.
Understand Bayes factors
A Bayes factor compares the predictive adequacy of two specified models. BF10 expresses evidence for model 1 relative to model 0, while BF01 is its reciprocal. The result depends on the hypotheses, prior distribution under each model, likelihood, and data. It does not by itself give posterior model odds unless prior model odds are also supplied.
A Bayes factor may support a null model, but weak evidence is not the same as evidence of absence. Avoid rigid adjective scales as substitutes for interpretation. Report the numerical value, models, priors, direction, and sensitivity. The peer-reviewed JASP guidelines for conducting and reporting Bayesian analysis provide practical recommendations covering planning, execution, interpretation, and reporting.
Conduct sensitivity analysis
Sensitivity analysis asks whether the conclusion changes under other defensible choices. Refit the model with narrower, wider, and differently centred priors. Examine alternative likelihoods for skewed or heavy-tailed outcomes. Test reasonable inclusion rules and missing-data assumptions. If a Bayes factor changes sharply under plausible prior widths, that dependence is a result to report.
Do not select the specification that produces the strongest evidence. Define the main analysis from theory and design, then use alternatives to reveal fragility. Sensitivity is especially important when data are limited, priors are informative, models are complex, or the substantive conclusion rests near a decision threshold.
Psychology-specific worked examples
Example 1: Bayesian regression for anxiety
A student investigates whether sleep quality predicts end-of-term anxiety after adjusting for baseline anxiety and workload. The outcome is continuous. The main estimand is the adjusted change in anxiety score associated with a one-standard-deviation improvement in sleep quality.
The student specifies regularising priors for coefficients, performs prior predictive checks against the scale range, fits four chains, and reviews convergence and divergences. Posterior predictive checks compare score distributions and residual patterns. The result is reported with a credible interval and the posterior probability that the association exceeds a pre-specified meaningful threshold. Prior sensitivity and a robust likelihood are included. For related model construction, see the psychology dissertation regression analysis guide.
Example 2: Bayes factor for a group comparison
Another student compares memory scores after two instructional conditions. The question includes whether the data support a negligible group difference, not merely whether a conventional test is non-significant. The student uses a Bayesian t-test with a pre-specified effect-size prior and reports BF10, the posterior effect estimate, and a credible interval.
The report states the prior scale and repeats the analysis with narrower and wider plausible priors. It avoids claiming that the null is true when evidence is weak. A conventional analysis can be discussed alongside the Bayesian result using the psychology dissertation t-test guide, provided the two inferential frameworks are interpreted correctly.
Select suitable software
Choose software according to model complexity, transparency, and supervisory support. JASP provides accessible interfaces for common tests and displays priors and Bayes factors. R packages can support reproducible scripts for regression, multilevel models, and model comparison. Stan offers flexible probabilistic programming and detailed diagnostics but requires stronger modelling and coding skills.
The official BayesFactor package reference documents common tests including t-tests, analysis of variance, regression, correlation, and contingency tables. Software defaults remain analytic choices. Record package versions, prior settings, sampler settings, random seeds, and code so another researcher can reproduce the analysis.
| Stage | Minimum evidence to retain | Common mistake |
|---|---|---|
| Model specification | Likelihood, link, formula, parameterisation, and estimand | Choosing a model only because it is available in a menu |
| Prior assessment | Prior source, scale, visualisation, and prior predictive simulations | Calling a default prior objective or non-informative |
| Computation | Chains, iterations, R-hat, effective sample sizes, warnings, trace plots | Reporting a posterior before resolving divergences |
| Model fit | Problem-specific posterior predictive checks | Checking only the average when tails or groups matter |
| Robustness | Defensible alternative priors and model choices | Presenting only the specification with the strongest result |
| Reporting | Posterior summaries, uncertainty, code, software, and limitations | Translating credible intervals as frequentist confidence intervals |
Report Bayesian analysis transparently
Readers should be able to reconstruct what was estimated and why. Report the research question, data structure, likelihood, link function, parameterisation, priors and their rationale, sampling algorithm, chain settings, convergence evidence, model checks, sensitivity analyses, and posterior summaries. Provide code and de-identified or simulated data when ethics and consent permit.
The Bayesian Analysis Reporting Guidelines offer a structured framework for reporting. Use them as a checklist, not as a replacement for disciplinary judgement. Explain deviations and distinguish pre-specified analyses from exploratory work.
Describe limitations that posterior intervals do not capture, including measurement error, unmeasured confounding, sampling bias, model misspecification, and limited generalisability. If a complex hierarchical or latent-variable model is used, connect the justification to the structural equation modelling guide or the relevant design resource.
Common errors to avoid
- Treating a default prior as neutral: inspect its implications on the outcome scale.
- Claiming small samples are automatically adequate: plan for precision, prediction, or evidence under realistic scenarios.
- Ignoring sampler warnings: resolve divergences and poor exploration before interpretation.
- Using Bayes-factor labels mechanically: report the value, compared models, priors, and uncertainty.
- Confusing probability statements: a parameter probability conditional on a model is not the probability that a broad theory is true.
- Skipping sensitivity checks: show whether reasonable priors or likelihoods change the conclusion.
- Overstating causal meaning: causal interpretation depends on design and identification, not on Bayesian computation alone.
Frequently asked questions
Is Bayesian analysis suitable for an undergraduate or master’s dissertation?
Yes, when the question, model, and supervision are appropriate. A well-justified Bayesian regression or group comparison is preferable to an unnecessarily complex model. Allow time to learn prior predictive checks, computation diagnostics, and reporting.
Can Bayesian methods prove the null hypothesis?
No method proves a statistical model in an absolute sense. A Bayes factor can show that observed data are more predictive under one specified model than another. The conclusion depends on those models and their priors.
Do I still need a sample-size calculation?
Yes. Plan sample size around posterior precision, predictive performance, decision accuracy, or expected evidence. Simulation can evaluate these targets under several plausible scenarios.
How should I choose a credible interval?
Choose the interval level in advance and state whether it is equal-tailed or highest-density when that distinction matters. A 95% interval is common, but the choice should support the decision context rather than follow habit alone.
What if my prior and data disagree?
Investigate measurement, coding, likelihood fit, prior relevance, and data quality. Prior-data conflict can reveal a population difference or model problem. Do not quietly replace the prior after seeing results; document revisions and compare specifications.
Should I report both p-values and Bayesian results?
Only when both answer a planned purpose and each is interpreted within its own framework. Adding a second method merely to search for a preferred conclusion increases ambiguity. Explain any differences in assumptions and estimands.
Which files should accompany the dissertation?
Where permitted, include analysis code, package versions, seeds, prior definitions, diagnostic plots, model summaries, and a data dictionary. Share de-identified data only when consent, ethics approval, and governance rules allow it.
Conclusion
A strong psychology dissertation Bayesian analysis begins with a precise estimand and ends with a transparent account of uncertainty. Justify priors on meaningful scales, simulate before fitting, diagnose computation, check predictions, test sensitivity, and interpret every probability within the stated model. These practices make the analysis more informative and easier to scrutinise.
If you need methodological support, seek guidance that helps you understand and defend every choice. Ethical dissertation assistance should strengthen your design, code, interpretation, and reporting while preserving your authorship and complying with your institution’s academic-integrity rules.
