Why Mathematics? | Deviance Information Criterion for Bayesian Reliability Models: Posterior Fit, Effective Parameters, MCMC and Model Limits

Start with the posterior object. Hierarchical shrinkage and latent structure mean fitted parameters can contribute less than one full degree of freedom each. A fleet model contains many unit effects but pulls them toward a shared population distribution. This step matters because posterior fit, effective complexity and Bayesian reliability judgement require the mathematical object, prediction target and reliability question to stay aligned. Counting every latent effect as one can exaggerate practical flexibility. Estimate effective complexity under the declared posterior.

A reproducible audit should record variables, units, observation boundaries, censoring codes, likelihood factorisation and software settings. Keep the posterior draws, observed-data deviance definition and effective-complexity convention together. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to support a model comparison.

Try a small learning experiment: recalculate D-bar, D at theta-bar, p_D and DIC, then sketch a bimodal posterior to explain why a componentwise mean can fail. Use why bayesian complexity is not always a raw parameter count as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Did You Know? DIC can change under a nonlinear reparameterisation because deviance at a posterior mean is not generally invariant. That is why a careful report includes the construction, not only the final scalar.

Contents · Next chapter

Now separate fit from flexibility. Deviance is commonly defined as minus twice the log likelihood of the observed data, up to a constant shared across compared models. A failure contributes log f(t), while a right-censored unit contributes log S(c). This step matters because posterior fit, effective complexity and Bayesian reliability judgement require the mathematical object, prediction target and reliability question to stay aligned. Using complete-data likelihood in one model and observed-data likelihood in another breaks comparison. Name the likelihood and constant convention.

A reproducible audit should record variables, units, observation boundaries, censoring codes, likelihood factorisation and software settings. A negative or implausible effective parameter count is evidence about posterior geometry, not a discount coupon. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to support a model comparison.

Try a small learning experiment: recalculate D-bar, D at theta-bar, p_D and DIC, then sketch a bimodal posterior to explain why a componentwise mean can fail. Use deviance from the observed-data likelihood as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

The reliability interpretation comes next. Average D(theta) over retained posterior draws to describe typical fitted lack of fit under posterior uncertainty. A chain yields an average deviance of 248. This step matters because posterior fit, effective complexity and Bayesian reliability judgement require the mathematical object, prediction target and reliability question to stay aligned. Poor MCMC mixing can make the average unreliable. Check chains before computing DIC.

A reproducible audit should record variables, units, observation boundaries, censoring codes, likelihood factorisation and software settings. Compare conditional and marginal DIC only after stating which future unit the model is meant to predict. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to support a model comparison.

Try a small learning experiment: recalculate D-bar, D at theta-bar, p_D and DIC, then sketch a bimodal posterior to explain why a componentwise mean can fail. Use posterior mean deviance as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

Start with the posterior object. Evaluate the deviance at the componentwise posterior mean theta-bar under the same likelihood. The worked model gives D(theta-bar)=240. This step matters because posterior fit, effective complexity and Bayesian reliability judgement require the mathematical object, prediction target and reliability question to stay aligned. The posterior mean can lie in a low-density or implausible region for skewed or multimodal posteriors. Inspect posterior geometry before trusting this plug-in point.

A reproducible audit should record variables, units, observation boundaries, censoring codes, likelihood factorisation and software settings. Reparameterisation sensitivity deserves a worked sensitivity result whenever the posterior mean is not representative. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to support a model comparison.

Try a small learning experiment: recalculate D-bar, D at theta-bar, p_D and DIC, then sketch a bimodal posterior to explain why a componentwise mean can fail. Use deviance at the posterior mean as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Primary research: https://doi.org/10.1111/1467-9868.00353. Read the authors’ definition and conditions before carrying the formula into a censored reliability model.

Contents · Previous chapter · Next chapter

Now separate fit from flexibility. The original effective-complexity measure p_D is posterior mean deviance minus deviance at the posterior mean. 248-240 gives p_D=8. This step matters because posterior fit, effective complexity and Bayesian reliability judgement require the mathematical object, prediction target and reliability question to stay aligned. Negative or unexpectedly large p_D is a diagnostic signal, not a value to hide. Report both deviance terms and p_D.

A reproducible audit should record variables, units, observation boundaries, censoring codes, likelihood factorisation and software settings. DIC is a fitted-data summary; posterior predictive checks and later data answer different questions. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to support a model comparison.

Try a small learning experiment: recalculate D-bar, D at theta-bar, p_D and DIC, then sketch a bimodal posterior to explain why a componentwise mean can fail. Use effective parameters p_d as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

The reliability interpretation comes next. Add p_D to posterior mean deviance, or equivalently add twice p_D to deviance at the posterior mean. 248+8=256 and 240+16=256. This step matters because posterior fit, effective complexity and Bayesian reliability judgement require the mathematical object, prediction target and reliability question to stay aligned. Rounding each MCMC summary early can create disagreement between the two routes. Keep full precision and verify both identities.

A reproducible audit should record variables, units, observation boundaries, censoring codes, likelihood factorisation and software settings. Keep the posterior draws, observed-data deviance definition and effective-complexity convention together. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to support a model comparison.

Try a small learning experiment: recalculate D-bar, D at theta-bar, p_D and DIC, then sketch a bimodal posterior to explain why a componentwise mean can fail. Use completing dic 256 as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Did You Know? DIC can change under a nonlinear reparameterisation because deviance at a posterior mean is not generally invariant. That is why a careful report includes the construction, not only the final scalar.

Contents · Previous chapter · Next chapter

Start with the posterior object. DIC ranks a declared comparable candidate set and has no universal pass threshold. Values 256 and 260 give a difference of four. This step matters because posterior fit, effective complexity and Bayesian reliability judgement require the mathematical object, prediction target and reliability question to stay aligned. The smaller value can still belong to an inadequate model. Pair DIC with posterior predictive checks.

A reproducible audit should record variables, units, observation boundaries, censoring codes, likelihood factorisation and software settings. A negative or implausible effective parameter count is evidence about posterior geometry, not a discount coupon. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to support a model comparison.

Try a small learning experiment: recalculate D-bar, D at theta-bar, p_D and DIC, then sketch a bimodal posterior to explain why a componentwise mean can fail. Use lower is relative, not absolute as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

Now separate fit from flexibility. Every candidate must use the same observations, censoring definitions and response scale. All models fit the same 120 component records. This step matters because posterior fit, effective complexity and Bayesian reliability judgement require the mathematical object, prediction target and reliability question to stay aligned. Dropping records with missing covariates changes the target comparison. Freeze the analysis set and missing-data model.

A reproducible audit should record variables, units, observation boundaries, censoring codes, likelihood factorisation and software settings. Compare conditional and marginal DIC only after stating which future unit the model is meant to predict. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to support a model comparison.

Try a small learning experiment: recalculate D-bar, D at theta-bar, p_D and DIC, then sketch a bimodal posterior to explain why a componentwise mean can fail. Use same-data and same-likelihood rule as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

The reliability interpretation comes next. Because D(theta-bar) uses a posterior mean, nonlinear reparameterisation can change p_D and DIC. A mean on the log-scale transformed back is not the mean on the natural scale. This step matters because posterior fit, effective complexity and Bayesian reliability judgement require the mathematical object, prediction target and reliability question to stay aligned. Equivalent probability models can receive different DIC values under careless computation. Use a stable scientifically meaningful parameterisation.

A reproducible audit should record variables, units, observation boundaries, censoring codes, likelihood factorisation and software settings. Reparameterisation sensitivity deserves a worked sensitivity result whenever the posterior mean is not representative. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to support a model comparison.

Try a small learning experiment: recalculate D-bar, D at theta-bar, p_D and DIC, then sketch a bimodal posterior to explain why a componentwise mean can fail. Use parameterisation sensitivity as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

Start with the posterior object. A variance-based alternative uses one half of the posterior variance of the deviance. If Var(D)=18, p_V=9. This step matters because posterior fit, effective complexity and Bayesian reliability judgement require the mathematical object, prediction target and reliability question to stay aligned. Different definitions need not agree in finite or irregular posteriors. Label the exact DIC convention.

A reproducible audit should record variables, units, observation boundaries, censoring codes, likelihood factorisation and software settings. DIC is a fitted-data summary; posterior predictive checks and later data answer different questions. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to support a model comparison.

Try a small learning experiment: recalculate D-bar, D at theta-bar, p_D and DIC, then sketch a bimodal posterior to explain why a componentwise mean can fail. Use alternative effective complexity p_v as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

Now separate fit from flexibility. DIC was motivated partly by models where the number of effective parameters is ambiguous. Batch-level frailties share information across manufacturing lots. This step matters because posterior fit, effective complexity and Bayesian reliability judgement require the mathematical object, prediction target and reliability question to stay aligned. The choice of focus—conditional or marginal prediction—changes the relevant likelihood. Match DIC to the intended predictive unit.

A reproducible audit should record variables, units, observation boundaries, censoring codes, likelihood factorisation and software settings. Keep the posterior draws, observed-data deviance definition and effective-complexity convention together. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to support a model comparison.

Try a small learning experiment: recalculate D-bar, D at theta-bar, p_D and DIC, then sketch a bimodal posterior to explain why a componentwise mean can fail. Use hierarchical reliability models as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Did You Know? DIC can change under a nonlinear reparameterisation because deviance at a posterior mean is not generally invariant. That is why a careful report includes the construction, not only the final scalar.

Contents · Previous chapter · Next chapter

The reliability interpretation comes next. Conditional DIC may assess predictions for existing units, while marginal likelihood formulations target new units after integrating latent effects. A machine-specific frailty is known for an existing machine but unknown for a new one. This step matters because posterior fit, effective complexity and Bayesian reliability judgement require the mathematical object, prediction target and reliability question to stay aligned. Mixing these focuses creates misleading rankings. State whether prediction is conditional or population-level.

A reproducible audit should record variables, units, observation boundaries, censoring codes, likelihood factorisation and software settings. A negative or implausible effective parameter count is evidence about posterior geometry, not a discount coupon. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to support a model comparison.

Try a small learning experiment: recalculate D-bar, D at theta-bar, p_D and DIC, then sketch a bimodal posterior to explain why a componentwise mean can fail. Use latent variables and focus as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Second authoritative source: https://doi.org/10.1007/s11222-016-9696-4. Compare its prediction target, diagnostics and finite-sample cautions with this worked example.

Contents · Previous chapter · Next chapter

Start with the posterior object. DIC is a posterior summary and inherits every failure of the sampling algorithm. One chain remains trapped in a high-deviance mode. This step matters because posterior fit, effective complexity and Bayesian reliability judgement require the mathematical object, prediction target and reliability question to stay aligned. A precise printed DIC can conceal nonconvergence. Check rank plots, effective sample size and chain agreement.

A reproducible audit should record variables, units, observation boundaries, censoring codes, likelihood factorisation and software settings. Compare conditional and marginal DIC only after stating which future unit the model is meant to predict. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to support a model comparison.

Try a small learning experiment: recalculate D-bar, D at theta-bar, p_D and DIC, then sketch a bimodal posterior to explain why a componentwise mean can fail. Use mcmc convergence as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

Now separate fit from flexibility. Finite posterior draws make D-bar, p_D and DIC estimates noisy. Two independent runs differ by 0.8 DIC units. This step matters because posterior fit, effective complexity and Bayesian reliability judgement require the mathematical object, prediction target and reliability question to stay aligned. Treating a tiny difference as decisive ignores simulation error. Repeat computation or estimate Monte Carlo uncertainty.

A reproducible audit should record variables, units, observation boundaries, censoring codes, likelihood factorisation and software settings. Reparameterisation sensitivity deserves a worked sensitivity result whenever the posterior mean is not representative. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to support a model comparison.

Try a small learning experiment: recalculate D-bar, D at theta-bar, p_D and DIC, then sketch a bimodal posterior to explain why a componentwise mean can fail. Use monte carlo error as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

The reliability interpretation comes next. A strongly skewed parameter posterior makes the componentwise mean a poor representative point. A weakly identified shape parameter has a long right tail. This step matters because posterior fit, effective complexity and Bayesian reliability judgement require the mathematical object, prediction target and reliability question to stay aligned. D(theta-bar) can overstate or understate typical fit. Plot the posterior and consider predictive alternatives.

A reproducible audit should record variables, units, observation boundaries, censoring codes, likelihood factorisation and software settings. DIC is a fitted-data summary; posterior predictive checks and later data answer different questions. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to support a model comparison.

Try a small learning experiment: recalculate D-bar, D at theta-bar, p_D and DIC, then sketch a bimodal posterior to explain why a componentwise mean can fail. Use skewed posterior distributions as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

Start with the posterior object. When several separated modes explain the data, theta-bar may average incompatible solutions. Two latent-class labelings create symmetric modes. This step matters because posterior fit, effective complexity and Bayesian reliability judgement require the mathematical object, prediction target and reliability question to stay aligned. DIC can produce unstable or nonsensical effective complexity. Resolve identifiability or use LOO-based evaluation.

A reproducible audit should record variables, units, observation boundaries, censoring codes, likelihood factorisation and software settings. Keep the posterior draws, observed-data deviance definition and effective-complexity convention together. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to support a model comparison.

Try a small learning experiment: recalculate D-bar, D at theta-bar, p_D and DIC, then sketch a bimodal posterior to explain why a componentwise mean can fail. Use multimodality as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Did You Know? DIC can change under a nonlinear reparameterisation because deviance at a posterior mean is not generally invariant. That is why a careful report includes the construction, not only the final scalar.

Contents · Previous chapter · Next chapter

Now separate fit from flexibility. Mixtures and redundant latent structures can violate regular approximations behind simple complexity summaries. Two mixture components become indistinguishable. This step matters because posterior fit, effective complexity and Bayesian reliability judgement require the mathematical object, prediction target and reliability question to stay aligned. A single p_D may not capture the geometry. Use methods designed for singular models.

A reproducible audit should record variables, units, observation boundaries, censoring codes, likelihood factorisation and software settings. A negative or implausible effective parameter count is evidence about posterior geometry, not a discount coupon. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to support a model comparison.

Try a small learning experiment: recalculate D-bar, D at theta-bar, p_D and DIC, then sketch a bimodal posterior to explain why a componentwise mean can fail. Use singular models as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

The reliability interpretation comes next. Heavy censoring can leave tail and shape parameters weakly informed even with many total units. Only six failures occur before the study closes. This step matters because posterior fit, effective complexity and Bayesian reliability judgement require the mathematical object, prediction target and reliability question to stay aligned. DIC cannot manufacture information about the late-life hazard. Inspect profiles, priors and posterior predictive tails.

A reproducible audit should record variables, units, observation boundaries, censoring codes, likelihood factorisation and software settings. Compare conditional and marginal DIC only after stating which future unit the model is meant to predict. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to support a model comparison.

Try a small learning experiment: recalculate D-bar, D at theta-bar, p_D and DIC, then sketch a bimodal posterior to explain why a componentwise mean can fail. Use sparse failures and weak identification as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

Start with the posterior object. The posterior and therefore DIC depend on the prior, especially when reliability evidence is sparse. A regularising prior stabilises a shape parameter. This step matters because posterior fit, effective complexity and Bayesian reliability judgement require the mathematical object, prediction target and reliability question to stay aligned. A ranking that changes across defensible priors is not robust. Publish a prior sensitivity table.

A reproducible audit should record variables, units, observation boundaries, censoring codes, likelihood factorisation and software settings. Reparameterisation sensitivity deserves a worked sensitivity result whenever the posterior mean is not representative. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to support a model comparison.

Try a small learning experiment: recalculate D-bar, D at theta-bar, p_D and DIC, then sketch a bimodal posterior to explain why a componentwise mean can fail. Use prior sensitivity as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

Now separate fit from flexibility. WAIC uses pointwise log predictive densities and a pointwise variance penalty, targeting predictive accuracy more directly. Two models have similar DIC but different WAIC. This step matters because posterior fit, effective complexity and Bayesian reliability judgement require the mathematical object, prediction target and reliability question to stay aligned. Asymptotic or finite-sample behaviour differs. Do not treat the acronyms as interchangeable.

A reproducible audit should record variables, units, observation boundaries, censoring codes, likelihood factorisation and software settings. DIC is a fitted-data summary; posterior predictive checks and later data answer different questions. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to support a model comparison.

Try a small learning experiment: recalculate D-bar, D at theta-bar, p_D and DIC, then sketch a bimodal posterior to explain why a componentwise mean can fail. Use dic versus waic as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

The reliability interpretation comes next. PSIS-LOO estimates leave-one-out predictive performance and supplies influence diagnostics. One component has a large Pareto-k diagnostic. This step matters because posterior fit, effective complexity and Bayesian reliability judgement require the mathematical object, prediction target and reliability question to stay aligned. DIC provides no equivalent pointwise reliability warning. Use predictive tools when the goal is out-of-sample performance.

A reproducible audit should record variables, units, observation boundaries, censoring codes, likelihood factorisation and software settings. Keep the posterior draws, observed-data deviance definition and effective-complexity convention together. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to support a model comparison.

Try a small learning experiment: recalculate D-bar, D at theta-bar, p_D and DIC, then sketch a bimodal posterior to explain why a componentwise mean can fail. Use dic versus psis-loo as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Did You Know? DIC can change under a nonlinear reparameterisation because deviance at a posterior mean is not generally invariant. That is why a careful report includes the construction, not only the final scalar.

Contents · Previous chapter · Next chapter

Start with the posterior object. Simulate replicated lifetimes and censoring-relevant summaries from the fitted model. Replicated early-failure counts miss the observed cluster. This step matters because posterior fit, effective complexity and Bayesian reliability judgement require the mathematical object, prediction target and reliability question to stay aligned. A lower DIC does not excuse a visible predictive failure. Check mechanisms and time regions separately.

A reproducible audit should record variables, units, observation boundaries, censoring codes, likelihood factorisation and software settings. A negative or implausible effective parameter count is evidence about posterior geometry, not a discount coupon. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to support a model comparison.

Try a small learning experiment: recalculate D-bar, D at theta-bar, p_D and DIC, then sketch a bimodal posterior to explain why a componentwise mean can fail. Use posterior predictive checks as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

Now separate fit from flexibility. Freeze the Bayesian model and test predictive distributions on later units or another site. A development fleet differs from field deployment. This step matters because posterior fit, effective complexity and Bayesian reliability judgement require the mathematical object, prediction target and reliability question to stay aligned. DIC uses the original data and does not measure transportability. Report calibration, log score and operational outcomes externally.

A reproducible audit should record variables, units, observation boundaries, censoring codes, likelihood factorisation and software settings. Compare conditional and marginal DIC only after stating which future unit the model is meant to predict. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to support a model comparison.

Try a small learning experiment: recalculate D-bar, D at theta-bar, p_D and DIC, then sketch a bimodal posterior to explain why a componentwise mean can fail. Use external validation as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Contents · Previous chapter · Next chapter

The reliability interpretation comes next. DIC joins logarithms, averages, variance, Bayesian updating and model comparison. Students can verify 248-240=8 and 248+8=256. This step matters because posterior fit, effective complexity and Bayesian reliability judgement require the mathematical object, prediction target and reliability question to stay aligned. Arithmetic is only the first layer; the representative posterior point matters. Draw a posterior with one mode and then two modes.

A reproducible audit should record variables, units, observation boundaries, censoring codes, likelihood factorisation and software settings. Reparameterisation sensitivity deserves a worked sensitivity result whenever the posterior mean is not representative. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to support a model comparison.

Try a small learning experiment: recalculate D-bar, D at theta-bar, p_D and DIC, then sketch a bimodal posterior to explain why a componentwise mean can fail. Use parent and student learning route as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Useful next reading: Mathematics Learning Hub, eduKate Mathematics Learning System, approach to learning mathematics, Bukit Timah Mathematics curriculum overview, cross-subject learning approach, nested likelihood-ratio guide. Choose the destination that matches the next question rather than treating links as a checklist.

Contents · Previous chapter · Next chapter

Start with the posterior object. DIC is not a Bayes factor, not a probability that one model is correct and not always invariant to reparameterisation. A negative p_D is a warning rather than a complexity bonus. This step matters because posterior fit, effective complexity and Bayesian reliability judgement require the mathematical object, prediction target and reliability question to stay aligned. Predictive validation remains necessary. Read Spiegelhalter and colleagues, then compare the result with modern LOO guidance.

A reproducible audit should record variables, units, observation boundaries, censoring codes, likelihood factorisation and software settings. DIC is a fitted-data summary; posterior predictive checks and later data answer different questions. Rebuild the displayed number from saved inputs before interpreting it. A result that cannot be reproduced is not ready to support a model comparison.

Try a small learning experiment: recalculate D-bar, D at theta-bar, p_D and DIC, then sketch a bimodal posterior to explain why a componentwise mean can fail. Use common misconceptions and next reading as the focus. Predict the direction of change before using software, alter one defensible input, and finish with two sentences: what the calculation supports, and what remains uncertain. This habit turns mathematics education into transferable problem-solving rather than formula recall.

Quick answers

What must be fixed before comparing models? Use the same observations, censoring mechanism, time scale, likelihood convention and predictive unit.

Does a preferred value prove a model is true? No. The comparison is conditional on its candidates, data and assumptions.

Which mathematics matters? Algebra organises identities, logarithms stabilise products, probability represents uncertainty, variance measures instability and graphs expose influential cases.

How should a student practise? Preserve intermediate quantities, predict directions before calculating and change one input at a time.

What follows the comparison? Diagnostics, uncertainty analysis, posterior or predictive checks, later-data validation and a separate decision analysis when costs or safety matter.

Contents · Previous chapter · Mathematics Learning Hub