Tensions in Practice: A simple update in tension with a more flexible model¶
Updating an uncertain event rate
A Bayesian update combines an initial probability model with new observations to produce a revised distribution of possibilities. Some specially matched models permit a direct algebraic update; that is convenient when their assumptions fit. A more flexible model can represent features that the simple one leaves out, but may require a numerical approximation and checks on its computation. Exact calculation within a model is different from the model fitting the situation.
Keep updating tractable
Obtain the revised distribution through a simple, reproducible calculation.
Represent relevant structure
Allow the model to express dependencies or variation that matter for the observations.
Why these aims pull against each other
Restricting the prior and observation model can make an exact update easy. Expanding what the model can express may remove that shortcut and add costly, fallible numerical computation.
Choose an arrangement to see what changes and what remains difficult.
Arrows are modeling and calculation dependencies. No posterior values, measured runtimes, or claim of guaranteed convergence are shown.
What this choice protects
What it costs
When it fits
Compare the arrangements
Use a matched simple model
Choose a prior and observation model from a compatible pair that permits a direct update.
- What it protects
- The posterior distribution can be computed without the particular approximation step shown in the other arrangement.
- What it costs
- The convenience can tempt an analyst to keep assumptions that do not fit the process being studied.
- When it fits
- Fits when the restricted model adequately represents the relevant observations. Computational simplicity alone is not evidence of fit.
Illustration note: The source lists Beta–Binomial and other matched families. This graph depicts their direct-update structure, not a worked statistical analysis or a claim that simple models are generally wrong.
Use a flexible model
Choose the prior and observation model for the structure needed, then approximate the posterior numerically when no usable direct solution is available. Examine two separate questions: whether the computation adequately approximates that posterior, and whether the selected model fits the situation.
- What it protects
- The model can represent relevant structure unavailable in the chosen simple family.
- What it costs
- Approximation and model assessment both take work. A concern about the numerical run may call for revising the computation; a concern about the assumptions may call for revising the model and starting computation again. Neither review necessarily settles the concern.
- When it fits
- Fits when the added structure matters and the approximation can be assessed well enough for the intended use. Flexibility does not ensure adequacy.
Illustration note: The source names fallible approximate methods and varying priors or likelihoods. These two review routes are an editorial illustration of different obligations, not a diagnostic algorithm. Not every flexible model lacks a direct solution, and a failed check need not identify a unique defect.
What this illustration does—and does not—establish
Bayesian Updating: Conjugacy and tractability vs model realism supplies the matched-family shortcut versus model-flexibility tension; Bayesian Updating: Sequential coherence vs computational tractability and Bayesian Updating: Model specification burden vs inferential rigor support the distinct burdens of approximation and assumption review. The second graph makes their targets different: revising the run changes the computation, while revising the model changes what is being computed. Neither arrangement is universally preferable.
- Both outputs are probability distributions, not single scores or automatic decisions.
- A correct numerical calculation can still use a misspecified prior or observation model. A good model can also be undermined by poor computation.
- “Exact” is conditional on the selected model. “Flexible” does not mean every real process is captured, and no algorithmic convergence guarantee is asserted.
- The two assessment branches are not a complete diagnostic procedure. Passing a numerical check does not validate the model; an apparent model discrepancy need not uniquely locate the error. Either assessment may remain inconclusive.
Source entries
Bayesian Updating
This source passage supplies the contextual tension. The concrete arrangements and schematic examples are editorial illustrations, not measured findings.
Conjugacy and tractability vs model realism
T6 — Conjugacy and tractability vs model realism. Conjugate prior-likelihood pairs (Beta-Binomial, Normal-Normal, Gamma-Poisson) admit closed-form posterior solutions, making Bayesian inference computationally simple; but real data-generating processes rarely conform to these restricted families. Non-conjugate models require approximate inference (MCMC, variational, particle filters), adding computational complexity but permitting more realistic modeling. The tension is between computational tractability (favoring conjugate families) and model adequacy (favoring flexible non-conjugate specifications).
The source operation
(1) Bayesian updating is the systematic process of revising a probability distribution over possibilities — the *prior* — by combining it with the likelihood of new evidence given each possibility, producing a revised *posterior* distribution.
A correct update can inherit a bad model
- Not automatically well-calibrated — posterior calibration depends on the prior and likelihood being approximately correct; misspecified models produce miscalibrated posteriors.
Approximation can remain inaccurate
T2 — Sequential coherence vs computational tractability. The mathematical elegance of Bayesian updating — posteriors become priors for the next observation — is a theoretical ideal that in practice often requires approximate methods (MCMC, variational inference, sequential Monte Carlo) whose convergence and accuracy are not always guaranteed. High-dimensional or complex-likelihood problems can stretch computational resources dramatically.
Varying the assumed model
Contemporary practice addresses this through sensitivity analysis (varying priors and likelihoods to assess robustness) and weakly-informative default priors that reduce the burden while preserving inferential transparency.