Choose your reading route
eduKateSG · Why Mathematics? · Burn-In and Reliability Screening: Infant Mortality, Truncation, Yield and Selection Bias
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 29–44
Study and verify
Practise, inspect failure cases and read further.
Full chapter index · Practice and checking · Mathematics Learning Hub
Why is mathematics important in burn-in and reliability screening? It tracks how an early-stress screen changes a population, catches weak units, rejects some good units and creates selection bias in later field data. It helps teams compare screening with fixing the manufacturing cause.
Did You Know? A screened population can look much more reliable even when the production process has not improved, because the weakest units were removed before customers ever saw them.
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 29–44 · Study and verify
- Guidance for students and parents
- Common misconceptions
- Connection to recent mathematics articles
- Practice questions
- Frequently asked questions
- A reporting checklist
- A final audit
- Useful next reading
- Mixture models for weak and normal units
- Optimising screen time
- Accelerated stress and equivalence
- Lot-to-lot monitoring
- Competing risks in screening
- Verifying the screen itself
- Student investigation
- Economic comparison without false precision
Burn-in and reliability screening expose units to time or stress before shipment so that some early weaknesses appear in a controlled setting. The mathematics tracks early-failure behaviour, detection, false rejection, yield and the post-screen population. Screening can remove weak units, but it does not necessarily remove the process that created them. A useful analysis separates defect discovery, selection and true process improvement instead of claiming that exposure itself manufactures reliability.
NIST describes the early region of the familiar bathtub curve as a period with high but decreasing failure rate, often called infant mortality. This shape is an empirical pattern, not a law that every product must follow. A mixed population of weak and normal units can create a decreasing aggregate hazard even when individual hazards differ. Look for evidence in failure times and mechanisms before choosing a burn-in policy.
For a Weibull lifetime model, shape beta below one gives a decreasing hazard. Beta equal to one gives a constant exponential hazard; beta above one gives increasing wear-out behaviour. Fitting beta to screened data can be biased because early failures have already been removed. Use unscreened development or process data when possible and include censoring. A fitted curve summarises a population under stated conditions; it does not identify the defect cause by itself.
A screen applies a profile and a pass rule. Weak units may fail and be removed, normal units may pass, and some normal units may be falsely rejected. Some weak units may escape. Define the screened population, stresses, duration, measurements, repair policy and disposition. A screen that detects one mechanism may miss another. Its value depends on detection and consequence, not merely on how many units it rejects.
Imagine 1,000 units with an expected 4% weak subgroup: 40 weak and 960 normal. If a screen catches 80% of weak units, it removes 32. If it falsely rejects 1% of normal units, it rejects another 9.6 in expectation. Expected values can be fractional even though actual counts are integers. The example makes both benefits and losses visible before any decision is made.
After the illustrative screen, eight weak units and 950.4 normal units pass in expectation, for 958.4 shipped. The expected weak fraction is about 8/958.4, or 0.835%. This is lower than the original 4% because of selection. It does not mean the manufacturing process improved. The post-screen population is different from the unscreened population, so comparisons need the selection rule.
| Group | Before | Rejected | Shipped |
|---|---|---|---|
| Weak subgroup | 40 expected | 32 true removals | 8 |
| Normal subgroup | 960 expected | 9.6 false rejects | 950.4 |
| Total | 1,000 | 41.6 expected | 958.4 expected |
Screen sensitivity is the probability a weak or target-defect unit is rejected. Specificity is the probability a normal unit passes. These require a reference classification, which may be difficult when latent defects are not directly observable. Estimates from tiny, enriched studies may not transfer to routine prevalence. Report uncertainty and the defect definition. A screen can have high sensitivity yet poor shipped quality if the base population or escape mechanisms differ.
Yield is the fraction that passes the screen, but high yield is not automatically good if weak units escape. Low yield can reflect excellent detection, a damaged process or an overly harsh screen. Separate true defect removals, false rejects, rework and scrap. Include the cost and risk of diagnostic confirmation. A dashboard that shows only yield invites teams to weaken the screen without understanding the reliability consequence.
Longer burn-in gives weak units more opportunity to fail but consumes time, energy and useful life. It may also expose normal units to damage. Choose duration from the time distribution of the target mechanism, not tradition. Compare detection probability and false-reject or ageing cost across candidate durations. A flat benefit curve suggests little value from extra exposure. A duration is a decision variable, not a badge of thoroughness.
Temperature, voltage, vibration, cycling or software workload can accelerate failure mechanisms. Excessive stress may create failures that normal use would never see, while mild stress may miss latent defects. Accelerated stress needs physical understanding and, when translated to use conditions, a validated acceleration model. Preserve failure-mode evidence. A high reject count under an unrealistic stress does not prove better field screening.
Reliability screening is sometimes described as precipitating latent defects and detecting the resulting failures. These are distinct stages. A stress can make a weakness observable; measurements and pass criteria must then recognise it. Poor detection leaves precipitated defects in the shipped population. Conversely, a sensitive measurement can flag harmless variation. Map each stress to a mechanism and each measurement to a decision threshold.
NIST’s reliability handbook explains the early-failure period, Weibull hazard shapes, censoring and high-stress projection. Read NIST on the bathtub curve, censoring and projection to use conditions. These sources support modelling principles, not a universal burn-in recipe.
Screening sorts the output; process improvement reduces the creation of weak units. Root-cause correction can be more durable and less wasteful than permanent screening. Track defect modes back to material, assembly, software, contamination or supplier causes. A falling screen reject rate can signal improvement, a weakened screen or a changing mix. Confirm the mechanism before claiming success.
| Measure | Question |
|---|---|
| Sensitivity | How many weak units are caught? |
| Specificity | How many normal units pass? |
| Yield | How many units ship? |
| Escape rate | How many weak units pass? |
| Post-screen reliability | How does the selected population perform? |
Post-screen field data describe units that survived the screen. Comparing them with unscreened prototypes without adjustment confounds design and selection. A strong screen can make field reliability appear excellent while factory yield deteriorates. Preserve pre-screen identifiers, stress histories and failures. When estimating the original production distribution, include screened-out units. When estimating customer performance, analyse the shipped selected population and say so.
Burn-in can create left truncation because field observation begins after units have survived an initial period. Units that fail before entry never appear in the field cohort. During the screen, units that complete without failure are right-censored at the duration. These mechanisms differ and need correct likelihood treatment. Discarding screened failures and analysing only survivors produces a conditional lifetime distribution, not the original population life.
A failed screen unit may be scrapped, reworked or repaired and screened again. A successful re-screen does not erase the original defect or prove as-good-as-new condition. Track the repair action and configuration. Reworked units may form a separate population with different risk. Define whether yield counts first-pass, final-pass or shipment. The denominator can change the story dramatically.
A supplier lot or assembly shift can create a cluster of weak units. Pooling all production into one weak fraction hides this variation. Stratify by lot and use hierarchical or overdispersed models when appropriate. Sampling only a few units per lot may miss a local problem. A screen can protect customers while a lot investigation identifies the process cause. Keep both objectives visible.
A burn-in chamber, fixture or software script can damage many normal units together. Outcomes are then dependent and specificity estimates from another setup may not apply. Monitor chamber uniformity, calibration, power events and operator interventions. Include control units or redundant measurements where useful. A sudden reject cluster can be a product problem or a test-system problem; investigate before disposition.
Many screens reject units when a measured parameter crosses a limit. Measurement error near the threshold creates false rejects and escapes. Estimate repeatability, bias and drift, then choose guard bands or confirmatory tests. Do not tune thresholds solely to hit a yield target. The limit should connect to functional risk or a validated degradation relationship. Record raw values, not only pass or fail.
Stress consumes some fraction of life for mechanisms that accumulate damage. A screen intended to remove early defects can move normal units closer to wear-out. Model cumulative exposure when the mechanism warrants it and verify with follow-up tests. If stress activates a different mechanism, a simple equivalent-age conversion is invalid. The benefit must be compared with this potential cost.
Burn-in uses energy, equipment time and materials, and false rejects create waste. Measure these outcomes alongside field returns and customer disruption. A process correction can reduce both defects and screening burden. If screening remains necessary, optimise profile, sampling and confirmation. Do not collapse environmental and safety consequences into one invented dollar figure when stakeholders need to see them separately.
Weak prevalence, sensitivity, specificity and post-screen failure rates are estimated. Propagate uncertainty rather than multiplying point estimates as if known. Beta-binomial or Bayesian models can represent limited validation data, while sensitivity analysis shows which assumption drives shipped risk. A precise expected weak fraction from uncertain inputs should not be reported to four decimals. Use ranges and decision thresholds honestly.
A screen may use stages: a quick measurement, confirmatory test and extended stress only for borderline units. This can reduce cost while preserving detection. Adaptive rules must be prespecified and validated to avoid inconsistent treatment. Track conditional paths and final disposition. A unit receiving extra stress differs from one passing immediately, so downstream analysis should retain route information.
Link screen history to warranty or field outcomes. Compare failure modes among first-pass, reworked and borderline units, adjusting for exposure. A weak association can mean the screen is ineffective, the field mechanism differs or data linkage is poor. Avoid changing the screen after every small fluctuation. Use control charts, confidence intervals and engineering review to separate signal from noise.
Create 1,000 virtual units with hidden weak or normal labels. Give the screen different sensitivities and specificities, then count rejects, escapes and shipped weak fraction. Change prevalence and observe positive predictive value. Add a cost for false rejection and field escape. Students learn conditional probability, expected counts and selection bias. The exercise remains fictional and does not define a real production screen.
Use rows for true class and columns for screen outcome. Multiply population counts by sensitivity and false-positive rates, then calculate yield, escape fraction and positive predictive value. Add uncertainty scenarios and duration choices. Keep expected values labelled so fractional counts are not mistaken for observed units. A separate event sheet should store serial number, lot, stress, time, raw measurements, failure mode and disposition.
Simulate mixture populations, failure times, stress response and measurement error. Unit-test perfect sensitivity, zero prevalence and no-screen cases. Fit censored lifetime models to screen data and conditional models to shipped units. Version thresholds and chamber profiles. Return uncertainty distributions for yield and escapes. Code review should verify that screened-out units are not silently dropped from process-quality analysis.
Burn-in and screening appear in semiconductors, electronics, batteries, aerospace, automotive systems and other manufactured products. Roles combine reliability, materials, statistics, testing, manufacturing and quality. Mathematics supports these careers but does not guarantee employment or income. Students can explore conditional probability and lifetime models before choosing engineering, statistics, computing or technical education routes based on current official information.
Students should draw the before-and-after population before using equations. Ask which units were weak, which were caught and which normal units were lost. Parents can use the example to discuss why selection changes a population. Reproduce the expected counts, then test other sensitivities in a spreadsheet. The transferable idea is that a filter can improve the selected group without fixing the source process.
Burn-in does not create reliability by magic. Every product does not follow a bathtub curve. Longer and harsher screens are not automatically better. High yield is not proof of low escapes. A post-screen field population is not the original population. Rework does not erase failure history. False rejects matter. Accelerated stress cannot be translated to use without a mechanism and model.
Read Weibull Reliability for hazard shape, Accelerated Life Testing for stress projection and Statistical Process Control for production monitoring. Visit the Mathematics Learning Hub for probability and statistics foundations.
1. Calculate weak units in 1,000 at 4%. 2. Find true removals at 80% sensitivity. 3. Find false rejects at 99% specificity. 4. Calculate shipped weak fraction. 5. Explain selection bias. 6. Distinguish censoring and truncation. 7. Give one screen-induced damage mechanism. 8. Explain why yield alone is insufficient. 9. Design a confirmatory stage. 10. State a cautious field conclusion.
### Does burn-in improve every unit? It selects and may precipitate defects; it does not necessarily improve surviving units.
### Is a decreasing hazard proof of infant mortality? It supports a pattern under a model but needs mechanism evidence.
### Why can normal units be rejected? Measurement error and overlapping response distributions create false positives.
### Should screening continue forever? The policy should be reviewed as process capability and failure mechanisms change.
Define population, target defect, stress profile, duration, measurements and thresholds. Report sensitivity, specificity, yield, false rejects, escapes, rework and uncertainty. Keep lot, configuration and raw failure-time data. Distinguish original and post-screen populations. Check stress realism and induced ageing. Link primary sources accurately. Separate screening protection from process correction and avoid universal recipes.
Recalculate the 40, 960, 32, 9.6, 8 and 950.4 expected counts and the shipped weak fraction. Confirm actual counts remain integers. Verify hazard-shape language and censoring treatment. Inspect threshold uncertainty, chamber effects and rework histories. Confirm source claims. Remove any claim that a screen proves field reliability or that all early failures share one mechanism.
Continue with NIST sections on the bathtub curve, Weibull models, censoring and accelerated projection. Compare burn-in with reliability demonstration, process capability and reliability growth. Use the Mathematics Learning Hub to review conditional probability, mixtures, survival functions and measurement error.
A screened population can be described as a mixture of latent subgroups with different early-life hazards. Even a simple two-group model explains why the overall hazard may fall after weak units fail or are removed. Estimate the subgroup fraction cautiously and test whether the model predicts holdout data. A mixture is a modelling device, not proof that every unit belongs to a clear biological-style category. Manufacturing variables and continuous variation may be more realistic.
Screen duration balances expected defect removal, false rejection, capacity, energy and the risk of consuming useful life. Write an objective function with explicit costs or constraints, then perform sensitivity analysis rather than declaring one universal optimum. If detection probability rises quickly and then levels off, additional hours may add little benefit. If the stress damages normal units, the apparent optimum may shift sharply. The chosen duration should be validated against failure mechanisms and updated when the process changes.
Temperature or voltage acceleration may shorten observation time, but only if the accelerated failure mechanism relates to use conditions. State the acceleration model, parameter estimates and range of validation. Extrapolating far beyond tested stress can create confident but fragile predictions. A screen intended to reveal workmanship defects is different from a life test intended to estimate long-term wear-out. Keep the question, stress, failure mode and inference aligned.
A screen can hide process drift if the team studies only shipped units. Track incoming lot, supplier, date, equipment and failure mode before and after screening. Use control charts or hierarchical models to distinguish ordinary variation from a systematic shift. A rising reject rate is information about production, even if field returns remain low. The aim is to remove the cause, not merely increase screen severity until the output looks stable.
Units may fail during screening for several reasons. Treating every event as one generic early failure can mislead the duration and stress choice. Classify failure modes with evidence, record exposure time and use competing-risk methods where appropriate. A stress that improves detection of one mechanism may suppress or create another. Report cause-specific findings and uncertainty, including unresolved failures, instead of forcing every event into a convenient category.
Screens need measurement-system analysis, fixture checks, calibration and repeatability studies. A noisy pass/fail boundary produces both escapes and false rejects. Challenge the screen with known conditions, blind replicates and boundary cases. Monitor software versions and operator effects. Validation should cover detection, classification and restoration after the test. A screen that cannot reproduce its own result cannot support a precise reliability claim.
Create a spreadsheet with 1,000 simulated units, a weak-subgroup indicator and two pass probabilities. Apply a screen with chosen sensitivity and specificity, then compare the pre-screen and post-screen mixtures. Repeat the simulation many times to see variability around the expected values. Change prevalence and specificity to see how yield and escapes respond. Conclude with a paragraph explaining selection bias and why post-screen reliability does not describe an unscreened population.
CHAPTER 44 OF 44 · Study and verify
44. Economic comparison without false precision
A screening decision should compare more than the cost of failed units. Include equipment time, energy, handling, diagnosis, false rejects, escapes, delayed delivery and the value of learning about the process. Express uncertain inputs as ranges and test which assumptions change the preferred option. Separate accounting cost from safety or regulatory constraints that cannot simply be traded away. Compare screening with upstream process control, supplier improvement and design change. A low-cost screen may be unattractive if it conceals a worsening process; an expensive screen may be justified temporarily while a root cause is removed. Report expected values together with tail scenarios and practical capacity. Do not assign invented monetary values merely to make the spreadsheet complete. The analysis should clarify the decision, expose missing evidence and define what will be monitored after implementation.
Extended practice laboratory
Practice case 1: Map the population
Setup: Use 1,000 units with an expected 4% weak subgroup.
Task: Calculate expected weak and normal counts.
Checking target: Obtain 40 weak and 960 normal units.
Practice case 2: Screen sensitivity
Setup: Assume the screen catches 80% of weak units.
Task: Calculate expected true removals.
Checking target: Obtain 32 expected weak-unit removals.
Practice case 3: False rejects
Setup: Assume 1% of normal units are rejected.
Task: Calculate expected false rejects.
Checking target: Obtain 9.6 as an expectation, not a literal fraction of a unit.
Practice case 4: Post-screen mixture
Setup: Track weak units that pass and normal units that pass.
Task: Calculate the expected shipped weak fraction.
Checking target: Obtain about 0.835% in the illustrative model.
Practice case 5: Hazard shape
Setup: Use a Weibull beta below one.
Task: Interpret the decreasing hazard.
Checking target: Recognise a possible early-failure pattern, not proof of a universal bathtub curve.
Practice case 6: Screen duration
Setup: Compare two burn-in durations.
Task: Track weak detection and good-unit ageing.
Checking target: Longer is not automatically better.
Practice case 7: Stress choice
Setup: List temperature, voltage or load stresses.
Task: Identify which failure mechanisms each stress may activate.
Checking target: Avoid creating non-use mechanisms.
Practice case 8: Censoring
Setup: Record units that complete the screen.
Task: Keep exposure and failure times.
Checking target: Do not analyse only failed units.
Practice case 9: Selection bias
Setup: Compare pre-screen and post-screen populations.
Task: Explain why their failure rates differ.
Checking target: Do not generalise post-screen survival to unscreened production.
Practice case 10: Yield audit
Setup: Separate true defect removal from false rejection.
Task: Calculate screening yield and escape risk.
Checking target: Use both quality and reliability measures.
Practice case 11: Corrective action
Setup: Trace early failures to causes.
Task: Choose process correction or screening.
Checking target: Prefer eliminating causes when feasible.
Practice case 12: Write a decision
Setup: State population, screen, mechanism and limits.
Task: Report yield, escapes and uncertainty.
Checking target: Avoid claiming burn-in creates reliability by itself.