Choose your reading route
eduKateSG · Why Mathematics? · Bayesian Reliability: Priors, Likelihoods, Posterior Failure Rates and Predictive Decisions
Start with the question, audit the worked example, or jump to the assumptions and practice. Every chapter stays open for uninterrupted reading.
Understand the question
Define the quantity and follow the mechanism.Route 2 · Chapters 8–14
Reproduce the example
Audit the numbers one step at a time.Route 3 · Chapters 15–21
Check assumptions
See where design and calibration matter.Route 4 · Chapters 22–28
Explore applications
Connect the method to a concrete investigation.Route 5 · Chapters 29–60
Study and verify
Practise, inspect failure cases and read further.
Full chapter index · Practice and checking · Mathematics Learning Hub
Why is mathematics important in Bayesian reliability? It combines earlier evidence and current observations into a posterior distribution, then turns that uncertainty into transparent predictions and decisions.
Did You Know? One divided by the posterior mean failure rate is generally not the same as the posterior mean MTBF.
Full chapter index
Chapters 1–7 · Understand the question
Chapters 8–14 · Reproduce the example
Chapters 15–21 · Check assumptions
Chapters 22–28 · Explore applications
Chapters 29–60 · Study and verify
- Applications and careers
- Common misconceptions
- Connection to recent mathematics articles
- Frequently asked questions
- A reporting checklist
- A final audit
- Useful next reading
- Beta–binomial reliability
- Posterior probability of a requirement
- Robust priors
- Commensurate evidence
- Overdispersion
- Repairable systems
- Survival prediction
- Decision thresholds
- Optional stopping
- Communication with frequencies
- Reproducibility
- Age-appropriate learning path
- Ethical use
- Practice questions
- Prior-data conflict
- Posterior predictive counts
- Transformations and summaries
- Calibration over repeated decisions
- Sensitivity to exposure quality
- Model comparison
- Communicating prior influence
- Governance of updates
- Posterior robustness report
- Exposure stratification
- Independent verification
Reliability evidence often arrives in pieces: earlier tests, current exposure, expert knowledge and field data. Bayesian mathematics represents starting uncertainty with a prior, combines it with a likelihood and produces a posterior distribution. It can answer predictive questions directly. The result is transparent only when prior relevance, model fit and sensitivity are shown.
The prior describes parameter uncertainty before current data. The likelihood describes how probable the observed data are for each parameter value. Bayes’ rule multiplies them and normalises to form the posterior. A posterior is conditional on all three ingredients: prior, likelihood and actual observations. It is not created by data alone.
NIST explains Bayesian reliability as combining old information or judgement with current data through Bayes’ formula, using conjugate priors where useful. Read NIST on Bayesian reliability methodology. NIST also notes advantages and disadvantages, including the challenge of validating a chosen prior.
For a constant failure rate λ with exponential lifetimes or a homogeneous Poisson process, a gamma prior is conjugate. Using shape a and rate b, prior density is proportional to λ^(a−1)exp(−bλ). After r failures in total exposure T, the posterior is Gamma(a+r,b+T). State whether b is a rate parameter; software uses competing conventions.
Take a prior Gamma(shape 2, rate 1000 hours). Observe one failure in 500 operating hours. The posterior is Gamma(shape 3, rate 1500 hours). Prior mean and posterior mean failure rate are both 0.002 per hour in this illustration because the new failure rate matches the prior mean. Other data would move it.
For Gamma(shape a, rate b), E[λ]=a/b. The posterior mean is 3/1500=0.002 per hour. A point summary hides distribution width, so report credible intervals or quantiles. Mean rate is not the only reliability decision quantity.
| Quantity | Value |
|---|---|
| Prior shape / rate | 2 / 1,000 hours |
| New failures / exposure | 1 / 500 hours |
| Posterior shape / rate | 3 / 1,500 hours |
| Posterior mean rate | 0.002/hour |
| Predictive P(no failure in 100h) | about 0.824 |
The reciprocal of posterior mean rate is 500 hours, but posterior mean of MTBF=E[1/λ] is b/(a−1)=750 hours when a>1. These differ because expectation and reciprocal do not commute. State whether you report reciprocal mean rate, median MTBF or mean MTBF. Ambiguous “Bayesian MTBF” invites mistakes.
For a future exposure τ under the gamma–Poisson model, the probability of zero failures is (b/(b+τ))^a using posterior parameters. With a=3, b=1500 and τ=100, this is (1500/1600)^3≈0.824. This prediction averages over parameter uncertainty instead of plugging in one rate.
NIST’s gamma/exponential reliability model assumes an exponential or homogeneous Poisson process and explains how prior gamma parameters may come from previous data or structured judgement. Read NIST on Bayesian reliability assumptions. Constant rate and exchangeable exposure must be checked.
NIST describes posterior parameters a′=a+r and b′=b+T and estimation of failure rate or MTBF from the resulting gamma model. Read NIST on Bayesian gamma reliability estimation. Use the source’s parameterisation carefully when translating formulas into software.
Historical failures and exposure can form prior pseudo-counts when the earlier system, mission and observation process are relevant. Similarity is an engineering judgement supported by evidence. Discounting can weaken borrowing, but the discount itself needs justification. Never count the same historical data in both prior and current likelihood.
Experts may express quantiles, plausible ranges or probabilities that can be fitted to a prior. Use multiple perspectives, calibration questions and documented reasoning. Separate aleatory variability from epistemic uncertainty. A consensus meeting can suppress disagreement; retain sensitivity to defensible alternatives.
A diffuse prior is not automatically neutral and can behave unexpectedly under transformations. A strong prior can dominate sparse data. Show effective information, prior predictive implications and posterior movement. If reasonable priors lead to different decisions, the correct conclusion is sensitivity, not one convenient answer.
| Stage | Question |
|---|---|
| Prior | What did we believe before current data? |
| Likelihood | How does the model connect parameters to observations? |
| Posterior | How did current data update the distribution? |
| Predictive | What may happen in future exposure? |
| Sensitivity | What changes under another defensible prior? |
Before seeing current data, simulate observations from the prior predictive distribution. If it assigns most probability to impossible failure counts or exposure behaviour, revise the prior or model. Prior predictive checks translate abstract parameters into observable implications that engineers can inspect.
A gamma–Poisson update is inappropriate when failure rate changes with age, repairs change the system, units differ strongly or exposure is measured poorly. Weibull, repairable-system, hierarchical or degradation models may be needed. Conjugacy saves arithmetic; it does not validate constant rate.
Right-censored lifetimes still contribute exposure under an exponential likelihood. Record entry, exit and failure status. Informative censoring can bias the model. A terminal test without failure is evidence, but it does not prove an infinite lifetime.
Related components, sites or fleets can share information through a hierarchical model. Partial pooling shrinks noisy estimates toward a group distribution while allowing differences. Exchangeability must be plausible. A unique design revision or environment may deserve a separate level or no borrowing.
Simulate future failure counts or lifetimes from the posterior and compare relevant summaries with observed patterns. Poor fit can reveal overdispersion, changing rate or unmodelled heterogeneity. A posterior can be mathematically correct for a bad likelihood; checking remains essential.
A 90% Bayesian credible interval contains 90% posterior probability for the parameter under the model and prior. It is not a frequentist confidence interval. Report parameter, transformation and interval type. For MTBF, transforming rate quantiles reverses endpoints; do not transform a symmetric interval mechanically.
Probabilities do not choose an action without values. Define actions and losses, then minimise posterior expected loss subject to legal, safety or operational constraints. A release, test or redesign decision may have asymmetric consequences. Do not invent monetary values simply to force an optimum.
Expected value of information asks whether additional testing could change a decision enough to justify its cost and delay. It requires a decision model, not only uncertainty. A highly uncertain parameter may have low value if no action depends on it; a moderately uncertain threshold parameter may be valuable.
Bayesian updating can occur as data arrive, provided the model and stopping process are handled transparently. The posterior after one batch becomes the prior for the next. Preserve data lineage and do not repeatedly reinterpret old evidence as new. Changes in design or mission may require a model boundary.
When several plausible lifetime models remain, Bayesian model averaging can combine predictions using model probabilities. Results depend on candidate set and priors. Sensitivity across models may be easier to communicate than one averaged answer. Averaging cannot include a missing mechanism.
Simple conjugate models can be calculated directly. Complex models may need Markov chain Monte Carlo or other numerical methods. Check convergence, effective sample size, Monte Carlo error and posterior geometry. Reproduce known special cases. A colourful trace plot is not a substitute for model validation.
For the gamma–Poisson example, enter prior shape and rate, failures and exposure. Add to obtain posterior parameters, compute mean rate and predictive zero-failure probability. Label rate units and parameterisation. Use trusted software for quantiles. Add a prior-sensitivity table rather than hiding one selected prior.
Use explicit distribution constructors with named shape and rate or scale. Unit-test zero exposure, zero failures and known conjugate updates. Verify predictive probabilities by simulation. Store prior rationale, data version and model code. Prevent double-counting. Generate a reproducible report with sensitivity across priors.
Use a beta–binomial coin example first, then the gamma–Poisson reliability example. Plot prior and posterior, calculate predictive probabilities and change prior strength. Explain why a posterior probability differs from a confidence statement. Keep the system fictional and focus on conditional reasoning.
Ask the student what evidence entered before and after the current test. Have them explain the likelihood without saying it is the probability of the parameter. Reproduce the update and predictive calculation. Parents can reinforce that judgement is not hidden when it is documented and tested for sensitivity.
Bayesian reliability appears in aerospace, manufacturing, energy, transport, medicine and software systems. Work combines probability, computing, engineering evidence and decisions. Mathematics supports pathways but does not guarantee employment or income. Students should build foundations and consult current official programme and professional requirements.
A prior is not automatically bias, and a diffuse prior is not automatically neutral. Posterior probability is conditional on the model and prior. Credible intervals are not confidence intervals. Reciprocal mean rate is not mean MTBF. Conjugacy does not prove constant failure rate. Bayesian updating cannot repair irrelevant historical data.
Read Reliability Demonstration Testing for frequentist evidence plans, Weibull Reliability for lifetime models and Warranty Analysis for field data. The Mathematics Learning Hub connects probability and inference.
### Is the prior made up? It must be justified from evidence or elicitation and tested for sensitivity.
### Why use a gamma prior? It is conjugate to constant-rate exposure models and keeps updating transparent.
### Is posterior mean MTBF just one over mean rate? No. Nonlinear transformation changes the expectation.
### Does Bayesian mean small samples are enough? No. Sparse data remain sensitive to prior and model choices.
State parameter, units, prior family and rationale, likelihood, data, censoring and posterior. Name rate versus scale parameterisation. Report intervals, predictive quantities, prior and model sensitivity, checks and decision rule. Preserve provenance. Avoid hiding judgement or presenting posterior results as universal facts.
Recalculate a+r and b+T, units, mean rate and predictive probability. Check MTBF transformation and interval direction. Verify historical data relevance and no double-counting. Test alternative priors and likelihoods. Review convergence if simulation was used. Confirm NIST sources. Remove any sentence implying the posterior is assumption-free.
Continue with NIST Bayesian methodology, assumptions and gamma-model estimation, then explore hierarchical reliability and Bayesian decision analysis. Compare posterior prediction with frequentist demonstration bounds. Use the Mathematics Learning Hub for Bayes’ rule, gamma distributions, expectations and simulation.
For binary missions, a beta prior Beta(a,b) is conjugate to binomial successes and failures. After s successes and f failures, the posterior is Beta(a+s,b+f). This supports posterior probability that mission reliability exceeds a target. It is a different model from gamma–Poisson exposure; do not mix their parameters.
A Bayesian demonstration may calculate P(R≥R0|data). This is a direct posterior probability under the prior and likelihood. State the target, prior and mission. It is not the same as a frequentist lower confidence bound. Compare frameworks by the decision they support rather than relabelling one as the other.
A robust prior can mix a historical informative component with a vague component, allowing current conflict to reduce borrowing. The mixture weight still needs justification. Show results under several weights. Robustness protects against some prior-data conflict; it cannot detect a new failure mechanism absent from every model.
Historical systems may differ in design, manufacturing, environment or reporting. A commensurate or hierarchical model estimates how much to borrow rather than assuming complete equivalence. Record the dimensions of similarity. More sophisticated borrowing does not eliminate the need for domain judgement.
If failure counts vary more than a Poisson model allows, unit heterogeneity or time variation may be present. Posterior predictive checks can reveal this. A negative-binomial or hierarchical model may fit better. Simply widening a prior on λ does not always reproduce the dependence structure.
A homogeneous Poisson model assumes a constant rate and repair that restores the process in a defined way. Reliability growth or deterioration needs a nonhomogeneous process. Bayesian methods can fit those models too, but the conjugate update changes. Do not call every event stream gamma–Poisson.
For nonrepairable items, posterior predictive survival averages R(t|θ) over the posterior distribution of θ. This incorporates parameter uncertainty. It still conditions on the lifetime family and covariates. Report predictive survival at relevant times rather than a single universal MTBF.
A rule might release when posterior probability of meeting a requirement exceeds a threshold. The threshold reflects losses and governance; it is not supplied by Bayes’ rule. Examine false release and false rejection under plausible truths. Prespecify the rule before data.
Bayesian updating remains coherent under many stopping rules when the likelihood and decision process are properly specified, but operational and regulatory plans may require prespecification. Data-dependent testing can still create cost, selection and reporting problems. Preserve all looks and stopping reasons.
Translate posterior predictions into a reference class: under the model, what range of failures might occur in the next exposure? Avoid saying the parameter “has changed” when only knowledge changed. Distinguish process uncertainty from randomness in future outcomes.
Store raw exposure and failures, prior version, code, random seeds and software environment. A posterior cannot be reconstructed from a chart alone. Version the configuration boundary. If prior evidence is later corrected, rerun the full chain and preserve the earlier result.
Students can begin with conditional probability and beta–binomial updating. Advanced learners can study gamma distributions, predictive mixtures and decisions. Use fictional systems. The key habit is naming what is uncertain and what evidence changes the distribution.
Bayesian models can make expert judgement visible, but authority and group dynamics can still bias elicitation. Include diverse expertise, document disagreements and avoid laundering preference into a technical prior. High-stakes use requires competent review and applicable regulation.
1. Update Gamma(2,1000) with one failure and 500 hours. 2. Find mean rate. 3. Compare two MTBF summaries. 4. Predict zero failures in 100 hours. 5. Weaken the prior. 6. Explain prior predictive checks. 7. Identify double-counting. 8. Distinguish credible and confidence intervals. 9. Define a decision loss. 10. Write a transparent conclusion.
When current data are surprising under the prior predictive distribution, investigate configuration, exposure, reporting and failure mechanisms. Do not simply average conflict away. Report posterior sensitivity to weakened or excluded historical evidence. A conflict can be valuable evidence that the population changed.
The gamma–Poisson model gives a negative-binomial posterior predictive distribution for future counts. Report a range of plausible counts, not only the probability of zero. Compare predictions with operational capacity and consequences. Predictive dispersion includes uncertainty in λ as well as future random variation.
Failure rate, MTBF and mission reliability are nonlinear transformations. A posterior median transforms directly under a monotone map, but means generally do not. Compute summaries on the scale stakeholders need and label them. Avoid moving an interval or mean across transformations by intuition.
Across many comparable systems, posterior predictive probabilities can be assessed for calibration. One outcome cannot validate a probability. Track prediction bins and realised counts while accounting for changing populations. Recalibrate transparently and retain the original model version.
Operating hours may be estimated, rounded or missing. Treat uncertain exposure explicitly or analyse plausible ranges. A precise posterior based on poor denominators is misleading. Align clock time, active time and mission cycles with the failure process. State which exposure the rate uses.
Compare exponential, Weibull, growth or mixture models using predictive checks and scientific reasoning. Bayes factors or information criteria can assist but depend on priors and candidate models. A selected best model is not proven true. Keep a sensitivity view when decisions change across families.
Show prior and posterior curves, effective prior exposure and the fraction of posterior information attributable to current data where meaningful. Avoid defensive claims that the data overwhelmed the prior unless demonstrated. Readers should see exactly what changed.
Define who can change priors, include new data or alter the likelihood. Record approvals and reasons. Automated dashboards should flag model drift and conflict rather than silently updating a decision threshold. Bayesian arithmetic can be continuous while authorisation remains controlled.
Publish a compact table of posterior summaries under the primary prior, a weaker prior and a sceptical alternative. Add predictive checks and the decision under each case. If the decision changes, collect more evidence or use a robust action. Sensitivity is a result, not a flaw to hide.
When environments differ, model rates by stratum or include covariates rather than summing all hours. The posterior for an average rate may be useless for a high-stress subgroup. State how predictions are weighted to the future use profile. Historical weights may not match a new deployment.
Have a second analyst reproduce the conjugate calculations and a sample of numerical posterior results. Verify rate-versus-scale settings and units. For simulation, compare posterior moments with analytic values. Independent checking is especially important when a prior or transformation makes the result look counterintuitive.
Extended practice laboratory
Practice case 1: Gamma prior
Setup: Use failure-rate prior Gamma(shape 2, rate 1000 hours).
Task: Find its mean failure rate.
Checking target: Obtain 0.002 per hour.
Practice case 2: Posterior update
Setup: Observe r=1 failure in T=500 hours.
Task: Add r to shape and T to rate.
Checking target: Obtain Gamma(shape 3, rate 1500 hours).
Practice case 3: Posterior mean rate
Setup: Use shape divided by rate.
Task: Calculate 3/1500.
Checking target: Obtain 0.002 per hour.
Practice case 4: MTBF distinction
Setup: Compare reciprocal posterior mean rate with posterior mean of 1/lambda.
Task: Calculate both for shape 3, rate 1500.
Checking target: Obtain 500 hours versus 750 hours; they answer different summaries.
Practice case 5: Predict zero failures
Setup: Use posterior Gamma(3,1500) for a 100-hour future exposure.
Task: Calculate (1500/1600)^3.
Checking target: Obtain about 0.824.
Practice case 6: Prior sensitivity
Setup: Replace the prior with a weaker alternative.
Task: Recalculate posterior summaries.
Checking target: Show how the current evidence and prior share influence.
Practice case 7: Prior-data relevance
Setup: Audit similarity of historical systems and missions.
Task: Decide what can be borrowed.
Checking target: Do not count irrelevant hours as equivalent evidence.
Practice case 8: Posterior predictive check
Setup: Simulate future counts from the posterior predictive model.
Task: Compare with observed patterns.
Checking target: Model fit still matters after conjugate updating.
Practice case 9: Credible interval
Setup: Calculate a posterior interval for the rate.
Task: State parameterisation and probability meaning.
Checking target: Do not call it a frequentist confidence interval.
Practice case 10: Decision loss
Setup: Compare actions under explicit costs or constraints.
Task: Calculate posterior expected loss.
Checking target: Probabilities do not choose values automatically.
Practice case 11: Hierarchical extension
Setup: Let related fleets share information partially.
Task: Explain shrinkage and exchangeability.
Checking target: Do not pool unlike systems completely.
Practice case 12: Write a conclusion
Setup: Report prior, likelihood, posterior and sensitivity.
Task: Separate inference from authorisation.
Checking target: Avoid hiding judgement inside a convenient prior.