See what the model assumes before it sees the outcome

Put the model’s starting assumptions in business units

Name the modeled quantity, observation unit and exposure opportunity. Daily orders, orders per visitor and an underlying purchase probability are different objects. Document transformations, scaling and the actual predictor range used for the review. A coefficient that looks small on one scale may imply very large changes on another, so a parameter table alone is not an adequate acceptance artifact.

Ask the analyst which knowledge informed the prior and which uncertainty it intentionally retains. Separate a credible historical source, a structural constraint and a convenient software default. Record whether earlier data informing the prior will also enter the current analysis, so the specialist can evaluate reuse rather than inadvertently presenting the same evidence twice. Do not substitute management’s sales target for a distribution of plausible outcomes.

Request generated outcomes before conditioning on the result

The Stan User’s Guide describes prior predictive checking as generating data from the prior and the data model without conditioning on observed outcomes. It contrasts this with posterior predictive checking and explains that extreme implications can expose a poorly chosen prior. Its guidance also shows why an extremely broad prior need not be innocuous.

Have the analyst supply clearly labeled simulated examples under relevant campaign conditions. Include ordinary conditions and meaningful boundary cases, with the exposure or audience size held explicit. Review the generated observations, not only a smooth average line. Ask whether counts, dispersion, rare outcomes and changes across predictor values are plausible for the modeled process. These are simulations for diagnosis, not forecasts validated by a customer study.

Investigate implausibility without tuning to a desired winner

When generated outcomes look implausible, trace the issue with the analyst. The problem may involve the prior, predictor units, a transformation or the observation model rather than one parameter alone. A review that changes several ingredients without a record can make the final explanation impossible to reproduce. Keep each proposed revision and its rationale tied to the observation it addresses.

Use domain evidence to distinguish impossible values, highly unlikely values and merely unfamiliar values. Avoid narrowing the prior until the simulation reproduces the current observed answer or the preferred campaign result. If knowledge is genuinely weak, keep that uncertainty visible and ask what data could inform it. A defensible starting model need not predict the exact future; it needs assumptions whose implications the team can inspect and explain.

Keep the prior check separate from the fitted-model decision

After fitting, require the analyst’s normal computational diagnostics and a separate check of how well the model represents relevant observed patterns. Where prediction matters, also retain the appropriate out-of-sample evaluation. Passing the prior review cannot certify that the fit is correct, that a forecast will be accurate or that a modeled association has a causal interpretation.

Compare important conclusions under reasonable alternative starting assumptions when the analyst judges sensitivity material. Report whether the decision changes and which assumptions drive it rather than hiding that dependence behind one posterior number. Preserve the prior version, evidence dates, simulated examples and approved scope. When the outcome definition, predictor units or market context changes, revisit the implications rather than carrying an old default into a new problem unquestioned.

Make the reasoning concrete.

A fictional brand commissions a model for weekly trial purchases. A broad default prior produces simulated weeks with order volumes wildly out of scale with the stated audience opportunity. The supplier argues that a broad prior is automatically neutral and should not be discussed.

The brand requests a documented model revision and a fresh prior-predictive review. It keeps the simulated examples separate from forecasts and does not select assumptions to manufacture a favorable campaign effect.

Take the questions into your next brief.

Use these prompts in your own notes. Nothing is submitted from this page.

1. Quantity

   What observation, unit and exposure opportunity are modeled?

2. Rationale

   Which evidence or constraint informs each important prior?

3. Implications

   What generated outcomes appear under ordinary and boundary conditions?

4. Revision

   Which assumption changes and what evidence justifies the change?

5. Next gate

   What separate fit, prediction and causal checks remain?

Watch for the wrong conclusion.

• Assuming a very broad prior has no practical influence.
• Inspecting coefficient values without the outcome scale.
• Tuning the prior to reproduce the preferred observed result.

Sources & scope

Review workflows, worksheets and fictional marketing examples are original editorial analysis. Named documentation supports only the attributed statistical concepts. Qualified analysts must assess assumptions and implementation; no actual data analysis, causal outcome or human reviewer is claimed.

• Stan User’s Guide: Posterior and Prior Predictive Checks →Primary statistical software and methodological documentation · Supports: Supports the distinction between prior and posterior predictive generation and the use of simulated implications to diagnose unsuitable priors. No source code, sports example or numerical simulation result is reproduced.