Why is a posterior predictive interval for one new observation wider than the credible interval for the parameter?

## Two different questions A **credible interval** is a statement about a parameter: given the data, there is a 95% posterior probability that the conversion rate `p` lies between two numbers. A **posterior predictive interval** is a statement about data you have not seen: given the data, there is a 95% predictive probability that the next observation — or the count among the next 100 visitors — falls between two numbers. They are on different scales (a rate versus a count, a mean versus an individual measurement) and they cover different objects. Mixing them up produces forecasts that are confidently, reproducibly wrong. ## The variance decomposition The cleanest way to see the width difference is the **law of total variance**, conditioning on the parameter `theta` and averaging over its posterior: ``` Var(y_new | data) = E_theta[ Var(y_new | theta) ] + Var_theta( E[y_new | theta] ) = expected sampling noise + parameter uncertainty ``` **First term — sampling noise.** Even if someone told you the true parameter exactly, the next observation would still be random: a coin with a known bias still lands unpredictably. This term is a property of the likelihood, not of how much data you have. It does not shrink. Ever. **Second term — parameter uncertainty.** This is the spread of the posterior, pushed through onto the data scale. This is the *only* thing a credible interval measures, and it is the piece that shrinks roughly like `1/n` in variance as the sample grows. Because the predictive variance is the sum of a non-negative term and the parameter term, the predictive interval is always at least as wide as the parameter interval rescaled to the data scale, and in practice strictly wider. ## A worked feel for the two terms Take the conversion example: 12 conversions from 100 visitors, posterior for `p` centred near 0.127 with standard deviation about 0.033. Forecast the count among the next 100 visitors. - Parameter-uncertainty contribution on the count scale: `100 * 0.033`, about 3.3 conversions. - Sampling-noise contribution: about `sqrt(100 * 0.127 * 0.873)`, also about 3.3 conversions. - Combined (the two variances add, not the standard deviations): `sqrt(3.3^2 + 3.3^2)`, about 4.7 conversions. The near-equality is not a coincidence — you observed 100 visitors and are forecasting 100 more, so the two sources are comparable. Change the horizon and the balance changes sharply. Forecast **one** future visitor and sampling noise dominates: the parameter is known well enough that the only real question is a coin flip. Forecast **10,000** future visitors and parameter uncertainty dominates, because the sampling-noise variance grows linearly in the horizon while the parameter-uncertainty variance grows quadratically. That asymmetry is worth being able to state out loud; it is the reason long-horizon forecasts stay wide no matter how much history you have. ## The limiting behaviour As the observed sample size goes to infinity, the posterior collapses onto the true parameter, the second term vanishes, and the predictive distribution converges to the likelihood at that value. The credible interval shrinks toward zero width; the predictive interval converges to a **positive** floor set by the model's own noise. A candidate who claims 'with enough data the prediction interval also goes to zero' has confused a statement about a parameter with a statement about a random observation. ## How this shows up in practice A product manager asks 'how many conversions will we get next week?' and receives an interval that turns out to be a credible interval for the rate multiplied by traffic. That interval will be badly under-covered — actual weeks will land outside it far more than 5% of the time — and the miss will be blamed on the model rather than on the reporting. The fix is not a wider arbitrary fudge factor; it is generating replicate observations through the full two-step draw (parameter from the posterior, then data from the likelihood) and quoting quantiles of *those*. ## What to say in the interview One sentence for the framing (parameter versus observation), one for the decomposition (noise plus parameter uncertainty), one for the limit (the noise floor never disappears), and one for the consequence (never hand a parameter interval to someone who asked about future data). That answer covers the ground without hand-waving.