Choose your reading route
eduKateSG · Why Mathematics? · Cox Proportional Hazards Reliability: Covariates, Hazard Ratios, Censoring and Diagnostics
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 11–20
Reproduce the example
Audit the numbers one step at a time.Route 3 · Chapters 21–30
Check assumptions
See where design and calibration matter.Route 4 · Chapters 31–40
Explore applications
Connect the method to a concrete investigation.Route 5 · Chapters 41–50
Study and verify
Practise, inspect failure cases and read further.
Full chapter index · Practice and checking · Mathematics Learning Hub
Why is mathematics important in Cox proportional hazards reliability? Components often operate under different temperatures, loads, manufacturers or maintenance histories. The Cox model relates covariates to failure hazard while leaving the baseline hazard unspecified, allowing censored lifetime data to support relative-risk estimates.
This guide develops risk sets, partial likelihood, hazard ratios and diagnostics with checked examples. The examples are educational. Real engineering decisions require accurate exposure definitions, representative data, proportional-hazards checks, uncertainty and domain review.
Full chapter index
Chapters 1–10 · Understand the question
Chapters 11–20 · Reproduce the example
Chapters 21–30 · Check assumptions
Chapters 31–40 · Explore applications
Chapters 41–50 · Study and verify
CHAPTER 1 OF 50 · Understand the question
1. From one survival curve to covariate effects
A single reliability curve averages over different operating conditions. Regression asks how recorded covariates relate to failure time. The Cox model compares instantaneous hazards among units still at risk. It is semiparametric because the covariate effects are parameterised while the baseline hazard is left flexible.
The hazard h(t) is an instantaneous failure rate conditional on survival to time t. It is not a probability and carries units of inverse time. A higher hazard tends to shorten lifetime, but the relation to mission reliability depends on the integrated hazard over the full interval.
For covariate vector x, the Cox model is h(t|x)=h_0(t)exp(βᵀx). The baseline h_0(t) applies when x=0. Each coefficient changes hazard multiplicatively. Coding and centring determine what x=0 means, so document them.
Comparing x_1 with x_0 gives hazard ratio exp[βᵀ(x_1-x_0)] if proportional hazards holds. The unknown baseline cancels. A hazard ratio of two means twice the instantaneous conditional hazard, not twice the probability of failing by a fixed time.
Suppose β=ln(2) for high versus standard load. The hazard ratio is exp[ln(2)]=2. If the standard-load hazard at one moment is 0.003/hour, the model gives 0.006/hour for high load at that moment. Both hazards may vary over time while their ratio stays two.
If the standard cumulative hazard to 100 hours is H_0=0.2, its survival is exp(-0.2)=0.8187. With hazard ratio two, cumulative hazard is 0.4 and survival is exp(-0.4)=0.6703. This calculation needs the estimated baseline cumulative hazard in addition to β.
| Quantity | Reference | High load |
|---|---|---|
| Hazard ratio | 1 | 2 |
| Cumulative hazard | 0.2 | 0.4 |
| Survival | 0.8187 | 0.6703 |
At each observed failure time, the risk set contains units still under observation and capable of failing just before that time. Right-censored units contribute until censoring. Incorrect entry or exit times change the comparisons and can bias the coefficient.
The partial likelihood compares the covariate value of the failed unit with values in its risk set. Baseline hazard terms cancel, allowing β estimation without choosing a lifetime distribution. This elegant cancellation does not remove assumptions about censoring, covariates or proportionality.
If two units are at risk and their linear predictors differ by ln(2), the model assigns relative failure weights 2 and 1. Conditional on one failing at that time, the higher-risk unit has model probability 2/(2+1)=2/3. The full partial likelihood multiplies such contributions across failure times.
Rounded clocks or batch inspections can produce ties. Breslow, Efron and exact methods handle tied partial-likelihood contributions differently. Efron often improves on Breslow when ties are moderate. State the method and inspect whether ties reflect common events rather than rounding.
A unit removed or still functioning at study end contributes survival exposure but no failure at censoring. Censoring should be conditionally independent of failure given modelled information. If high-risk units leave because of warning signs, ordinary analysis can be biased.
A unit enters observation only after surviving to its entry age. Include it in risk sets after entry, not from time zero. Ignoring delayed entry overrepresents survivors and can create immortal-time bias. Record operating history and eligibility rules.
Calendar time, age, operating hours and cycles answer different questions. Choose the clock linked to exposure and failure mechanism. Using age as time while treating calendar temperature as a covariate differs from using calendar time with age as a covariate.
A one-unit coefficient assumes a log-linear effect unless transformed. For temperature, vibration or load, check functional form with splines, plots or residual-based tests. A hazard ratio per one degree may be less interpretable than per ten degrees.
Choose a reference category and avoid sparse levels. Hazard ratios compare each category with the reference under other covariates. Pooling categories should follow mechanism and evidence, not only significance. Report counts and failures by level.
An interaction allows the effect of one covariate to change with another. High load may matter more at high temperature. Include interactions when scientifically plausible and supported by data. Main-effect interpretations then depend on the other covariate’s reference value.
Some covariates change during follow-up, such as load, maintenance state or environment. Split records into intervals with correct values. Avoid using future measurements to predict earlier failures. Internal covariates affected by prior health need careful causal interpretation.
The basic model assumes each hazard ratio is constant over time. Baseline hazard can vary freely, but relative effects cannot. A crossing survival pattern or changing residual trend can violate proportionality. Add time interactions, stratify or choose another model when needed.
Schoenfeld residuals compare observed covariates at each failure with risk-set expectations. Scaled residual plots against time help detect changing effects. A flat pattern supports proportionality; uncertainty bands and global tests matter. One nonsignificant test does not prove the assumption.
| Candidate | Relative weight |
|---|---|
| Exposed unit A | 2 |
| Exposed unit B | 2 |
| Unexposed unit | 1 |
| Exposed share | 4/5 |
Martingale residuals compare observed event counts with model-expected cumulative hazards. They can reveal nonlinear functional form and influential observations, though their distribution is skewed. Deviance residuals symmetrise them for plotting.
Delta-beta or dfbeta measures show how coefficients change when an observation is removed. Investigate influential units, data errors and unusual exposure. Do not delete valid extremes merely to stabilise estimates; report sensitivity.
After estimating β, Breslow-type estimators recover cumulative baseline hazard. Then S(t|x)=exp[-H_0(t)exp(βᵀx)]. Baseline estimates are step functions at event times and uncertain in the tail where few units remain.
Hazard ratios can be transportable while absolute reliability differs across populations because the baseline changes. Decisions need survival probabilities at mission times, not only relative effects. Report both when the evidence supports them.
The Weibull reliability guide chooses a parametric baseline. Weibull proportional hazards is one special form. Cox regression sacrifices some parametric efficiency for baseline flexibility.
CHAPTER 25 OF 50 · Check assumptions
25. Connection to accelerated failure time models
A proportional-hazards coefficient scales hazard; an accelerated failure time coefficient scales time. Both can fit Weibull data under specific parameterisations, but their interpretations differ. The companion AFT article develops time ratios.
If several causes stop observation, a cause-specific Cox model treats other causes as censored for the cause-specific hazard. This does not directly yield cumulative incidence without combining hazards. Use the competing-risks guide for that distinction.
Multiple components from one system or lot may share environment. Robust sandwich standard errors can address some within-cluster correlation for coefficients, while frailty models represent random effects. Few clusters require caution.
Noisy covariates usually bias and blur effects, sometimes toward zero. Calibration error can also alter functional form. Validate instruments and consider measurement-error models. More lifetime records do not repair systematically wrong load values.
Complete-case analysis can lose information and bias results when missingness is informative. Multiple imputation should include outcome and censoring information under a defensible missingness model. Report missing patterns and sensitivity.
Sparse failures limit stable estimation even with many censored units. Penalisation or simpler models may help. Avoid adding many interactions and spline terms solely because software permits them. Report effective information and uncertainty.
Use mechanism, prespecification, penalisation or validated predictive performance. Stepwise p-value hunting creates unstable effects. Retain variables needed for confounding control or design even when individually uncertain.
Evaluate calibration of predicted survival at mission times, discrimination such as time-dependent concordance and performance on later or held-out units. A high concordance score does not guarantee calibrated reliability probabilities.
Report coefficient intervals, hazard-ratio intervals and confidence bands for survival or cumulative hazard. Bootstrap whole sampling units when clustering exists. Include model-form and proportionality sensitivity, not only standard errors.
Confirm hazard-ratio and survival calculations, risk-set membership, tied-event method and table row counts. Fit a simulated exponential baseline where the truth is known. Check that shifting a centred covariate changes the baseline interpretation but not fitted hazards.
Compare predicted failures across covariate groups and time periods. Inspect residuals, calibration and later cohorts. Review whether operating conditions and maintenance policy match the study population. A mathematically correct model can be nontransportable.
Use β=ln(2), H_0(100)=0.2 and reproduce survival 0.8187 versus 0.6703. Draw a risk set with one censored unit and calculate a 2/3 partial-likelihood contribution. Explain what proportional hazards means.
Ask the student what changes and what stays constant when hazard ratio is two. Encourage a timeline with risk sets before formulas. Discuss why a censored working unit still contributes useful information.
CHAPTER 38 OF 50 · Explore applications
38. Common misconception: hazard ratio is a probability ratio
Hazard is conditional instantaneous rate. A hazard ratio of two does not mean 80% reliability becomes 40% or failure probability doubles. Convert through cumulative hazard and baseline survival for an absolute mission estimate.
CHAPTER 39 OF 50 · Explore applications
39. Common misconception: unspecified baseline means assumption free
The model still assumes proportional hazards, log-linear covariate effects, appropriate censoring and correct risk sets. Semiparametric does not mean assumption free.
Cox regression estimates relative effects without choosing a baseline lifetime distribution because risk-set comparisons cancel that nuisance term. The baseline returns later when absolute reliability is needed.
Define hazard, risk set and partial likelihood. Reproduce the 2/3 contribution and 100-hour survivals. Explain censoring, left truncation, ties and Schoenfeld residuals. Distinguish relative and absolute risk.
Avoid the basic form when proportionality is untenable, event times are defined poorly or covariate histories are unavailable. Time-varying effects, AFT models, parametric models or state models may fit the mechanism better.
No. They describe conditional associations under the model. Causal claims require design, confounding control and stronger assumptions. Reliability intervention decisions should also use mechanism evidence.
Report outcome, clock, entry and censoring, risk sets, covariate coding, missingness, ties, interactions, baseline estimator, proportionality and functional-form diagnostics, clustering, uncertainty, calibration, validation and configuration.
Kumar and Klefsjö review the proportional-hazards model in reliability. Hess reviews graphical checks of proportional hazards.
Mathematics matters because operating conditions change failure risk. Cox regression uses risk sets and partial likelihood to estimate covariate effects while leaving baseline hazard flexible. Responsible use checks proportionality, converts relative effects to mission reliability and validates on relevant systems.
A stratified Cox model allows separate baseline hazards for groups while keeping selected covariate effects common. It handles nonproportional nuisance factors but does not estimate their hazard ratios. Check enough events within strata.
Ridge, lasso and elastic net can stabilise many covariates. Penalty choice needs cross-validation or prespecification, and coefficient shrinkage changes inference. Validate absolute survival after penalisation.
Cox models appear in reliability, medicine, insurance and operations. Students practise exponentials, conditional probability, optimisation and diagnostics. These skills support pathways, while professional conclusions need domain evidence and ethical use.
A student should build a risk set, interpret a hazard ratio, convert cumulative hazard to survival and diagnose proportionality. A strong answer names censoring, baseline uncertainty and the difference between prediction and causation.
Extended practice laboratory
Practice case 1: Hazard ratio
Setup: Use beta=ln(2).
Task: Calculate exp(beta).
Checking target: Obtain a hazard ratio of 2.
Method note: For each case, define the event, time origin, risk set and censoring rule before calculating. Write the linear predictor and explain the reference profile. Check proportionality, functional form, influential observations, ties and calibration. Then name one omitted common cause or operational change that could alter the association, and preserve the model version and uncertainty interval.
Reflection: Explain why the checking target follows, identify the assumption that carries the most consequence, and describe one verification or validation check. Finish with a cautious sentence that distinguishes the illustrative calculation from a real operational approval.
Worked reasoning for Hazard ratio: Start from this stated setup—Use beta=ln(2). Make the requested move explicitly: Calculate exp(beta). The calculation is complete only when it reaches this check—Obtain a hazard ratio of 2. Now change one assumption, unit or observation rule and repeat the reasoning. Explain whether the numerical answer, its interpretation, or both would change. This turns a short exercise into a transferable method rather than a number to memorise.
Audit note for Hazard ratio: Record the formula or algorithm, substitutions, units, rounding rule and result in that order. Add the evidence a real analysis would need, such as failure-time records, censoring indicators, inspection histories, cost definitions or cluster labels. Name one diagnostic that could reject the model and one independent check that could catch a transcription error. Finish by separating the mathematical illustration from an engineering, safety or maintenance decision.
Practice case 2: Survival translation
Setup: Use baseline cumulative hazard 0.2 and hazard ratio 2.
Task: Calculate both survivals.
Checking target: Obtain about 0.8187 and 0.6703.
Method note: For each case, define the event, time origin, risk set and censoring rule before calculating. Write the linear predictor and explain the reference profile. Check proportionality, functional form, influential observations, ties and calibration. Then name one omitted common cause or operational change that could alter the association, and preserve the model version and uncertainty interval.
Reflection: Explain why the checking target follows, identify the assumption that carries the most consequence, and describe one verification or validation check. Finish with a cautious sentence that distinguishes the illustrative calculation from a real operational approval.
Worked reasoning for Survival translation: Start from this stated setup—Use baseline cumulative hazard 0.2 and hazard ratio 2. Make the requested move explicitly: Calculate both survivals. The calculation is complete only when it reaches this check—Obtain about 0.8187 and 0.6703. Now change one assumption, unit or observation rule and repeat the reasoning. Explain whether the numerical answer, its interpretation, or both would change. This turns a short exercise into a transferable method rather than a number to memorise.
Audit note for Survival translation: Record the formula or algorithm, substitutions, units, rounding rule and result in that order. Add the evidence a real analysis would need, such as failure-time records, censoring indicators, inspection histories, cost definitions or cluster labels. Name one diagnostic that could reject the model and one independent check that could catch a transcription error. Finish by separating the mathematical illustration from an engineering, safety or maintenance decision.
Practice case 3: Risk set
Setup: Use two exposed and one unexposed candidate at an event time.
Task: Calculate the exposed weight share.
Checking target: Obtain 2/(2+1).
Method note: For each case, define the event, time origin, risk set and censoring rule before calculating. Write the linear predictor and explain the reference profile. Check proportionality, functional form, influential observations, ties and calibration. Then name one omitted common cause or operational change that could alter the association, and preserve the model version and uncertainty interval.
Reflection: Explain why the checking target follows, identify the assumption that carries the most consequence, and describe one verification or validation check. Finish with a cautious sentence that distinguishes the illustrative calculation from a real operational approval.
Worked reasoning for Risk set: Start from this stated setup—Use two exposed and one unexposed candidate at an event time. Make the requested move explicitly: Calculate the exposed weight share. The calculation is complete only when it reaches this check—Obtain 2/(2+1). Now change one assumption, unit or observation rule and repeat the reasoning. Explain whether the numerical answer, its interpretation, or both would change. This turns a short exercise into a transferable method rather than a number to memorise.
Audit note for Risk set: Record the formula or algorithm, substitutions, units, rounding rule and result in that order. Add the evidence a real analysis would need, such as failure-time records, censoring indicators, inspection histories, cost definitions or cluster labels. Name one diagnostic that could reject the model and one independent check that could catch a transcription error. Finish by separating the mathematical illustration from an engineering, safety or maintenance decision.
Practice case 4: Censoring
Setup: Right-censor one unit at 600 hours.
Task: Place it in each earlier risk set.
Checking target: Do not count censoring as failure.
Method note: For each case, define the event, time origin, risk set and censoring rule before calculating. Write the linear predictor and explain the reference profile. Check proportionality, functional form, influential observations, ties and calibration. Then name one omitted common cause or operational change that could alter the association, and preserve the model version and uncertainty interval.
Reflection: Explain why the checking target follows, identify the assumption that carries the most consequence, and describe one verification or validation check. Finish with a cautious sentence that distinguishes the illustrative calculation from a real operational approval.
Worked reasoning for Censoring: Start from this stated setup—Right-censor one unit at 600 hours. Make the requested move explicitly: Place it in each earlier risk set. The calculation is complete only when it reaches this check—Do not count censoring as failure. Now change one assumption, unit or observation rule and repeat the reasoning. Explain whether the numerical answer, its interpretation, or both would change. This turns a short exercise into a transferable method rather than a number to memorise.
Audit note for Censoring: Record the formula or algorithm, substitutions, units, rounding rule and result in that order. Add the evidence a real analysis would need, such as failure-time records, censoring indicators, inspection histories, cost definitions or cluster labels. Name one diagnostic that could reject the model and one independent check that could catch a transcription error. Finish by separating the mathematical illustration from an engineering, safety or maintenance decision.
Practice case 5: Ties
Setup: Give three failures the same recorded time.
Task: Choose Breslow, Efron or exact handling.
Checking target: Report the convention.
Method note: For each case, define the event, time origin, risk set and censoring rule before calculating. Write the linear predictor and explain the reference profile. Check proportionality, functional form, influential observations, ties and calibration. Then name one omitted common cause or operational change that could alter the association, and preserve the model version and uncertainty interval.
Reflection: Explain why the checking target follows, identify the assumption that carries the most consequence, and describe one verification or validation check. Finish with a cautious sentence that distinguishes the illustrative calculation from a real operational approval.
Worked reasoning for Ties: Start from this stated setup—Give three failures the same recorded time. Make the requested move explicitly: Choose Breslow, Efron or exact handling. The calculation is complete only when it reaches this check—Report the convention. Now change one assumption, unit or observation rule and repeat the reasoning. Explain whether the numerical answer, its interpretation, or both would change. This turns a short exercise into a transferable method rather than a number to memorise.
Audit note for Ties: Record the formula or algorithm, substitutions, units, rounding rule and result in that order. Add the evidence a real analysis would need, such as failure-time records, censoring indicators, inspection histories, cost definitions or cluster labels. Name one diagnostic that could reject the model and one independent check that could catch a transcription error. Finish by separating the mathematical illustration from an engineering, safety or maintenance decision.
Practice case 6: Schoenfeld check
Setup: Plot residuals against time.
Task: Look for a systematic trend.
Checking target: A trend challenges proportional hazards.
Method note: For each case, define the event, time origin, risk set and censoring rule before calculating. Write the linear predictor and explain the reference profile. Check proportionality, functional form, influential observations, ties and calibration. Then name one omitted common cause or operational change that could alter the association, and preserve the model version and uncertainty interval.
Reflection: Explain why the checking target follows, identify the assumption that carries the most consequence, and describe one verification or validation check. Finish with a cautious sentence that distinguishes the illustrative calculation from a real operational approval.
Worked reasoning for Schoenfeld check: Start from this stated setup—Plot residuals against time. Make the requested move explicitly: Look for a systematic trend. The calculation is complete only when it reaches this check—A trend challenges proportional hazards. Now change one assumption, unit or observation rule and repeat the reasoning. Explain whether the numerical answer, its interpretation, or both would change. This turns a short exercise into a transferable method rather than a number to memorise.
Audit note for Schoenfeld check: Record the formula or algorithm, substitutions, units, rounding rule and result in that order. Add the evidence a real analysis would need, such as failure-time records, censoring indicators, inspection histories, cost definitions or cluster labels. Name one diagnostic that could reject the model and one independent check that could catch a transcription error. Finish by separating the mathematical illustration from an engineering, safety or maintenance decision.
Practice case 7: Time interaction
Setup: Add x times log(t).
Task: Interpret the coefficient.
Checking target: The effect may vary with time.
Method note: For each case, define the event, time origin, risk set and censoring rule before calculating. Write the linear predictor and explain the reference profile. Check proportionality, functional form, influential observations, ties and calibration. Then name one omitted common cause or operational change that could alter the association, and preserve the model version and uncertainty interval.
Reflection: Explain why the checking target follows, identify the assumption that carries the most consequence, and describe one verification or validation check. Finish with a cautious sentence that distinguishes the illustrative calculation from a real operational approval.
Worked reasoning for Time interaction: Start from this stated setup—Add x times log(t). Make the requested move explicitly: Interpret the coefficient. The calculation is complete only when it reaches this check—The effect may vary with time. Now change one assumption, unit or observation rule and repeat the reasoning. Explain whether the numerical answer, its interpretation, or both would change. This turns a short exercise into a transferable method rather than a number to memorise.
Audit note for Time interaction: Record the formula or algorithm, substitutions, units, rounding rule and result in that order. Add the evidence a real analysis would need, such as failure-time records, censoring indicators, inspection histories, cost definitions or cluster labels. Name one diagnostic that could reject the model and one independent check that could catch a transcription error. Finish by separating the mathematical illustration from an engineering, safety or maintenance decision.
Practice case 8: Stratification
Setup: Use different baseline hazards by site.
Task: Keep a common coefficient.
Checking target: Do not report a site hazard ratio without a coefficient.
Method note: For each case, define the event, time origin, risk set and censoring rule before calculating. Write the linear predictor and explain the reference profile. Check proportionality, functional form, influential observations, ties and calibration. Then name one omitted common cause or operational change that could alter the association, and preserve the model version and uncertainty interval.
Reflection: Explain why the checking target follows, identify the assumption that carries the most consequence, and describe one verification or validation check. Finish with a cautious sentence that distinguishes the illustrative calculation from a real operational approval.
Worked reasoning for Stratification: Start from this stated setup—Use different baseline hazards by site. Make the requested move explicitly: Keep a common coefficient. The calculation is complete only when it reaches this check—Do not report a site hazard ratio without a coefficient. Now change one assumption, unit or observation rule and repeat the reasoning. Explain whether the numerical answer, its interpretation, or both would change. This turns a short exercise into a transferable method rather than a number to memorise.
Audit note for Stratification: Record the formula or algorithm, substitutions, units, rounding rule and result in that order. Add the evidence a real analysis would need, such as failure-time records, censoring indicators, inspection histories, cost definitions or cluster labels. Name one diagnostic that could reject the model and one independent check that could catch a transcription error. Finish by separating the mathematical illustration from an engineering, safety or maintenance decision.
Practice case 9: Nonlinearity
Setup: Model temperature with a spline.
Task: Compare with a straight line.
Checking target: Inspect shape and uncertainty.
Method note: For each case, define the event, time origin, risk set and censoring rule before calculating. Write the linear predictor and explain the reference profile. Check proportionality, functional form, influential observations, ties and calibration. Then name one omitted common cause or operational change that could alter the association, and preserve the model version and uncertainty interval.
Reflection: Explain why the checking target follows, identify the assumption that carries the most consequence, and describe one verification or validation check. Finish with a cautious sentence that distinguishes the illustrative calculation from a real operational approval.
Worked reasoning for Nonlinearity: Start from this stated setup—Model temperature with a spline. Make the requested move explicitly: Compare with a straight line. The calculation is complete only when it reaches this check—Inspect shape and uncertainty. Now change one assumption, unit or observation rule and repeat the reasoning. Explain whether the numerical answer, its interpretation, or both would change. This turns a short exercise into a transferable method rather than a number to memorise.
Audit note for Nonlinearity: Record the formula or algorithm, substitutions, units, rounding rule and result in that order. Add the evidence a real analysis would need, such as failure-time records, censoring indicators, inspection histories, cost definitions or cluster labels. Name one diagnostic that could reject the model and one independent check that could catch a transcription error. Finish by separating the mathematical illustration from an engineering, safety or maintenance decision.
Practice case 10: Prediction
Setup: Estimate a survival curve for one profile.
Task: Combine baseline and covariates.
Checking target: State the reference profile and horizon.
Method note: For each case, define the event, time origin, risk set and censoring rule before calculating. Write the linear predictor and explain the reference profile. Check proportionality, functional form, influential observations, ties and calibration. Then name one omitted common cause or operational change that could alter the association, and preserve the model version and uncertainty interval.
Reflection: Explain why the checking target follows, identify the assumption that carries the most consequence, and describe one verification or validation check. Finish with a cautious sentence that distinguishes the illustrative calculation from a real operational approval.
Worked reasoning for Prediction: Start from this stated setup—Estimate a survival curve for one profile. Make the requested move explicitly: Combine baseline and covariates. The calculation is complete only when it reaches this check—State the reference profile and horizon. Now change one assumption, unit or observation rule and repeat the reasoning. Explain whether the numerical answer, its interpretation, or both would change. This turns a short exercise into a transferable method rather than a number to memorise.
Audit note for Prediction: Record the formula or algorithm, substitutions, units, rounding rule and result in that order. Add the evidence a real analysis would need, such as failure-time records, censoring indicators, inspection histories, cost definitions or cluster labels. Name one diagnostic that could reject the model and one independent check that could catch a transcription error. Finish by separating the mathematical illustration from an engineering, safety or maintenance decision.
Practice case 11: Validation
Setup: Hold out later units.
Task: Check calibration and discrimination.
Checking target: Do not rely on concordance alone.
Method note: For each case, define the event, time origin, risk set and censoring rule before calculating. Write the linear predictor and explain the reference profile. Check proportionality, functional form, influential observations, ties and calibration. Then name one omitted common cause or operational change that could alter the association, and preserve the model version and uncertainty interval.
Reflection: Explain why the checking target follows, identify the assumption that carries the most consequence, and describe one verification or validation check. Finish with a cautious sentence that distinguishes the illustrative calculation from a real operational approval.
Worked reasoning for Validation: Start from this stated setup—Hold out later units. Make the requested move explicitly: Check calibration and discrimination. The calculation is complete only when it reaches this check—Do not rely on concordance alone. Now change one assumption, unit or observation rule and repeat the reasoning. Explain whether the numerical answer, its interpretation, or both would change. This turns a short exercise into a transferable method rather than a number to memorise.
Audit note for Validation: Record the formula or algorithm, substitutions, units, rounding rule and result in that order. Add the evidence a real analysis would need, such as failure-time records, censoring indicators, inspection histories, cost definitions or cluster labels. Name one diagnostic that could reject the model and one independent check that could catch a transcription error. Finish by separating the mathematical illustration from an engineering, safety or maintenance decision.
Practice case 12: Conclusion
Setup: State endpoint, time origin and covariates.
Task: Report diagnostics and uncertainty.
Checking target: Separate association from causation.
Method note: For each case, define the event, time origin, risk set and censoring rule before calculating. Write the linear predictor and explain the reference profile. Check proportionality, functional form, influential observations, ties and calibration. Then name one omitted common cause or operational change that could alter the association, and preserve the model version and uncertainty interval.
Reflection: Explain why the checking target follows, identify the assumption that carries the most consequence, and describe one verification or validation check. Finish with a cautious sentence that distinguishes the illustrative calculation from a real operational approval.
Worked reasoning for Conclusion: Start from this stated setup—State endpoint, time origin and covariates. Make the requested move explicitly: Report diagnostics and uncertainty. The calculation is complete only when it reaches this check—Separate association from causation. Now change one assumption, unit or observation rule and repeat the reasoning. Explain whether the numerical answer, its interpretation, or both would change. This turns a short exercise into a transferable method rather than a number to memorise.
Audit note for Conclusion: Record the formula or algorithm, substitutions, units, rounding rule and result in that order. Add the evidence a real analysis would need, such as failure-time records, censoring indicators, inspection histories, cost definitions or cluster labels. Name one diagnostic that could reject the model and one independent check that could catch a transcription error. Finish by separating the mathematical illustration from an engineering, safety or maintenance decision.