In Bayesian versus frequentist inference, is the unknown parameter treated as random or as fixed?

## The one structural difference Both frameworks write down the same model for how data is generated: a likelihood `P(data | theta)` giving the probability of the observations for each candidate value of the parameter `theta`. They diverge on one question — is `theta` allowed to have a probability distribution? **Frequentist: no.** `theta` is a fixed unknown constant. It is a property of the world, not of your knowledge, and the world does not roll dice about it. Since `theta` cannot be random, every probability statement in the framework has to be about something that can be: the data. That is why frequentist reasoning is organised around *repeated sampling*. You imagine drawing the sample again and again from the same population, look at how the estimator `theta_hat` would scatter across those hypothetical draws, and evaluate procedures by that scatter — bias, variance, standard error. **Bayesian: yes.** `theta` gets a distribution. Before data, a prior `P(theta)`; after data, a posterior `P(theta | data)` obtained by multiplying prior by likelihood and renormalising. Here the data you actually collected is treated as fixed — you saw what you saw — and the uncertainty is carried by the parameter. So the randomness swaps sides. Frequentist: fixed parameter, random data. Bayesian: fixed (observed) data, uncertain parameter. ## What "random parameter" does and does not mean This phrase causes more confusion than any other part of the topic. Putting a distribution on `theta` is **not** a claim that the true conversion rate is jittering around from minute to minute. A Bayesian is perfectly free to believe the quantity is a single fixed number; the distribution describes their *state of knowledge* about that number, in the same way you can hold a distribution over the identity of a card that is already sitting sealed in an envelope. The card is not changing. Your information about it is incomplete. This follows directly from the degree-of-belief reading of probability. Once probability is allowed to describe knowledge rather than only repeatable physical setups, there is no obstacle to writing `P(theta)` for a fixed unknown `theta`. Reject that reading and the Bayesian machinery has nothing to attach itself to; accept it and the machinery is just conditional probability applied in the direction you actually want. ## The consequence you can actually use Because `theta` has a distribution, you can integrate over it and answer questions in the form decision-makers ask: - `P(theta > 0.05 | data)` — the probability the true rate clears a 5% bar. - `P(theta_B > theta_A | data)` — the probability one variant is genuinely better than another. - The expected loss of shipping versus not shipping, averaged over the posterior. A fixed-parameter framework cannot produce these directly, because `theta > 0.05` is not a random event within it — it is simply true or false, and its probability is 0 or 1 with no way to know which. The frequentist answers a different question: it constructs a *procedure* and characterises how that procedure behaves over repeated samples. That is a real and rigorous guarantee, but it is a statement about the procedure, not about this particular parameter. ## The cost The distribution over `theta` does not come free. Before seeing data you must state a prior, and someone can always challenge it — that is the standard objection to the Bayesian setup, and it has teeth when the prior is doing heavy lifting. How to choose priors, and how much they matter, is its own discussion; the point here is that the choice is unavoidable once you commit to treating the parameter as an uncertain quantity rather than a constant. The frequentist buys freedom from that choice at the price of not being able to say anything probabilistic about `theta` itself. ## How to answer in an interview Say the swap in one line — "frequentist: fixed parameter, random data; Bayesian: observed data fixed, parameter uncertain" — then immediately disarm the misreading by saying the Bayesian distribution encodes knowledge, not physical variation. Finish with the payoff: direct probability statements about the parameter, at the cost of stating a prior. Three sentences, in that order, is a complete senior-grade answer. A common stumble is to describe the difference as "Bayesians use priors and frequentists don't" and stop there. That is a symptom rather than the cause. The prior exists *because* the parameter was granted a distribution; get the ordering right and the rest of the framework follows from it.