Individual Project — To-do

Ten steps from question to presentation · mark each step as you go

The project is an individual project with data material selected by the student, presented orally and in writing, at roughly the complexity of a mixed model. Each step below has a status control (Not started / In progress / Done) that is saved in your browser and summarised on the dashboard. The steps link back to the module sections they rely on.

ImportantDates

Check Canvas for the official dates and format; this page gives study advice, not requirements.

• Mon 19 October 2026: data and research question settled (Announcement 3).
• Wed 28 October 2026: project deadline.
• Thu 29 October 2026: project presentations, 15–25 minutes each.

What counts as “Bayesian enough”?

From the teacher’s guidance on projects (Announcement 3):

• Do something in a Bayesian way. A coefficient in a linear model with weak priors on a large dataset is not enough: it is numerically a frequentist fit.
• The prior matters when the information per parameter is small: small groups, rare categories, few observations per unit, few studies. That is why a GLMM is the reference level (Module 5).
• Pragmatic use is fine: models that are hard to fit otherwise (Gamma, multinomial, ordinal, zero-inflated, hurdle, values below a detection limit), as long as the coefficients are interpreted with care.
• Large datasets are allowed, but MCMC gets slow: subsampling is fine.

TipDoes the prior matter with my data? Information per parameter

Parameter	Information	Does the prior matter?
A coefficient estimated from all rows	many observations	hardly: say so
An effect for a small group or cell	few observations	yes: partial pooling, hyperprior sensitivity
A between-group SD with few groups (or few studies)	as many units as groups	a lot
A quantity derived from several parameters	inherits all their uncertainty	the draws give it directly

A convincing check for the report: fit the model with and without informative priors (or by maximum likelihood) and show where the results differ.

The ten steps

Step 1 · Research question and estimand

• Write the question in one or two sentences, before fitting anything.
• Name the estimand(s): the number(s) computed from the posterior that answer the question (a difference, a probability, a ratio, a ranking).
• Say why the Bayesian part matters for this question (see above).

Relevant: 1.5, 4.11, 5.5.

Step 2 · Data: source, selection, cleaning, missing values

• Source and citation; conditions of use (keep restricted data out of public repositories).
• Selection, with the number of units at each exclusion step.
• Recoding (scales, centring, categories) written down.
• Missing data: how many, how handled (complete case, mi(), brm_multiple()), and could it bias the results?

Step 3 · Model: likelihood and structure, written out

• The model three ways: as distributions, in words, and as code (5.7).
• Justify the likelihood, the covariates, and random vs fixed effects (or what is shared between groups or studies).

Step 4 · Priors and a prior predictive check

• List every parameter that needs a prior (get_prior()).
• One sentence of justification per prior, on an interpretable scale (6.4).
• Prior predictive check (sample_prior = "only") with one plot for the report (6.2).

Step 5 · Fit and check convergence

• 4 chains, enough iterations, seed =, file = to save the fit.
• Report R-hat (< 1.01), bulk and tail ESS, divergent transitions, trace plots (3.8–3.10).
• Divergences: raise adapt_delta; with few groups, look at the group-level SD prior (funnel).

Step 6 · Posterior predictive checks

• Choose checks that match the question (4.3, 4.4): overall, by group, and by a variable not in the model (bayesplot::ppc_stat_grouped()).
• Ordinal or count outcomes: type = "bars" / "bars_grouped".
• State what the model misses.

Step 7 · Compare alternative models

• One to three plausible alternatives, compared with loo_compare() (4.9); check Pareto (k).
• Remember what LOO predicts (new units in existing groups), and when it says little (very few data points).

Step 8 · Sensitivity analysis

• Refit with different priors for the parameters the data inform least, e.g. group-level SDs (6.5).
• One debatable modelling or data choice, varied.
• Show how the estimands change.

Step 9 · Estimands from the draws: tables and figures

• Compute the estimands draw by draw (3.12, 4.11, 5.6, 5.11).
• One clear table (posterior median, 95% credible interval, probability of a relevant threshold) and one or two figures.
• Interpret on the outcome scale (probabilities, differences), not only log-odds.

Step 10 · Write-up and presentation

• Report: question and estimand; data; model and priors; computation (software versions, chains, diagnostics); checks and comparison; results in words; sensitivity; discussion of limitations and subjective choices (6.3, 6.8).
• Slides: one each for question, data, model, a check, the main result, sensitivity, limitations.
• Be ready to explain why each prior.

NoteCommon mistakes in Bayesian projects

• Interpreting the summary before checking diagnostics.
• Reporting parameters on the log-odds scale without translating them.
• Calling the 95% interval a confidence interval, or interpreting it as one.
• Leaving default priors without saying what they are.
• Showing only pp_check(fit) for a binary outcome: use stat and stat_grouped.
• Treating estimates that rest on few groups or studies as precise, without a prior sensitivity analysis.