Bayes Lab is a working statistics lab for media and mass communication research. Set a prior belief about an audience, hand it real survey data, and watch the exact posterior distribution, credible intervals, Bayes factors, and predictive probabilities compute themselves, live, from the underlying gamma and beta functions.
Four conjugate models, each rendered live from first principles. No approximation below the posterior, only closed-form updates.
A frequentist p-value can only tell you how surprising your data would be under a specific null. A Bayesian analysis answers the questions researchers actually ask: what should I now believe about my audience, and how confident am I? It combines theory with evidence, yields a full distribution of plausible effects (a credible interval you can interpret directly), and can even quantify evidence for a null hypothesis, something a p-value can never do.
Every section is interactive, mathematically exact, and grounded in a real media-research context.
What a prior actually is, Bayes' theorem in research vocabulary, conjugate priors, and why "no prior" is a myth.
Start hereSix conjugate models for the data shapes media research actually produces, from binary attitudes to Likert scales to exposure counts.
Open the labCultivation, parasocial interaction, third-person effect, agenda-setting, framing, spiral of silence, and more, with reference priors.
See your constructsBayes factors, ROPE, prior & posterior predictive checks, and prior-sensitivity analysis: evidence for and against effects.
Test an effectHeavy vs. light viewers, exposure groups, ad-vs-no-ad, posterior differences, Cohen's h, odds ratios, and probability of superiority.
Compare audiencesCombine effect sizes across studies with a random-effects model, the modern engine behind cultivation evidence reviews.
Combine studiesPrior elicitation you can defend to reviewers, and sample-size planning by posterior precision rather than long-run p-values.
Plan a studyThe vocabulary, precisely: credible vs. confidence intervals, HDI, KL divergence, ROPE, conjugate priors, with full citations.
Look it upEvery estimate is paired with the sentence a reviewer would accept, because a Bayesian result is only as defensible as the prior that seeded it and the model that digested it.
"p < .05" becomes "given this data, there is a 93% chance the effect exceeds zero."
A Bayes factor can support "no meaningful effect", essential when a media-effects null must be shown informative, not just underpowered.
The prior is written down before the data. A reviewer can evaluate it, criticise it, and see how much it changed the answer.
This is not toy math. The same models run here in exact closed form are the ones you'd fit in brms, Stan, or JASP for a real study.
# brms: Bayesian glm for a yes/no survey item brm(attitude_shift ~ 1, family = bernoulli(), prior = set_prior("beta(6, 14)", class = "Intercept"), data = survey)
Exact conjugate update: posterior = Beta(α₀ + x, β₀ + n − x).
# Stan: unknown mean, known-ish sampling SD data { int N; vector[N] y; real mu0; real s0; } parameters { real mu; } model { mu ~ normal(mu0, s0); y ~ normal(mu, sigma_est); }
Posterior precision = prior precision + n·(1/σ²): the conjugate normal-normal result shown live on the Models page.