Choose your reading route
eduKateSG · Why Mathematics? · Multi-Task Logistic Regression for Reliability: Time Grids, Dependent Regressors, Censoring and Survival Curves
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 22–27
Explore applications
Connect the method to a concrete investigation.Route 5 · Chapters 28–31
Study and verify
Practise, inspect failure cases and read further.
Full chapter index · Practice and checking · Mathematics Learning Hub
Why is mathematics important in multi-task logistic regression for reliability? Failure risk unfolds across time, so one binary classifier at one horizon cannot describe an entire survival distribution. MTLR treats a sequence of time boundaries as related prediction tasks and constrains their joint patterns so cumulative failure remains coherent. This guide develops time grids, dependent regressors, censoring likelihoods, regularisation, worked probabilities and validation without turning a flexible curve into false certainty.
Full chapter index
Chapters 1–7 · Understand the question
Chapters 8–14 · Reproduce the example
Chapters 15–21 · Check assumptions
Chapters 22–27 · Explore applications
Chapters 28–31 · Study and verify
CHAPTER 1 OF 31 · Understand the question
1. Why one classifier is not a survival curve
A model for failure by 100 hours says nothing directly about 200 or 500 hours. Fitting independent classifiers can produce impossible results, such as lower cumulative failure at a later time. MTLR links horizon-specific regressors through a joint distribution over ordered event-time intervals. The output is a coherent probability mass function and survival curve.
Select boundaries τ1<τ2<…<τm that divide follow-up into intervals. Quantile-based grids can balance events; equal-width grids preserve clock spacing; operational boundaries align with inspections. Too many intervals create sparse tasks, while too few hide timing. Report the grid and test sensitivity.
For event time T, define binary labels indicating whether failure has occurred by each boundary. Valid sequences contain zeros before failure and ones afterwards. Patterns such as 1,0,1 are impossible. MTLR models only ordered sequences, which is how it enforces a non-increasing survival function.
Each time boundary has a linear predictor, but the sequence probability uses their cumulative scores inside one normalising denominator. The tasks share the requirement that event time belongs to exactly one interval. Regularisation can also smooth neighbouring coefficient vectors so effects change gradually over time.
The model assigns probability to failure in each interval and to survival beyond the final boundary. These masses are nonnegative and sum to one. Survival at a boundary is the sum of masses in later intervals. This construction produces monotone survival without post-hoc clipping.
Suppose interval failure masses are 0.10, 0.15, 0.20 and beyond-grid survival mass 0.55. Survival after boundary one is 0.90; after two, 0.75; after three, 0.55. Cumulative failure is 0.10, 0.25 and 0.45. Check sums and monotonicity before interpreting covariates.
| Interval | Failure mass | Survival after interval |
|---|---|---|
| 1 | 0.10 | 0.90 |
| 2 | 0.15 | 0.75 |
| 3 | 0.20 | 0.55 |
| Beyond grid | 0.55 | — |
Conditional hazard in interval two is interval-two failure mass divided by survival entering that interval: 0.15/0.90≈0.1667. Interval three hazard is 0.20/0.75≈0.2667. Probability mass, cumulative risk and conditional hazard are related but not interchangeable. Label the output in every table.
If an asset is censored in an interval, its likelihood sums probabilities for event times consistent with surviving to the censoring time. It does not choose the censoring interval as failure. Boundary conventions matter when censoring occurs exactly at τj. Verify the package likelihood on a tiny case.
MTLR usually localises failure to an interval, not an exact continuous time. Two events within the same interval receive the same time category. A coarse grid trades precision for stability. Interpolation may be added for presentation, but it does not create observed timing information.
A predictor can have different coefficients at different boundaries, allowing time-varying effects. Plot coefficient trajectories with uncertainty and regularisation context. Sharp changes may reflect sparse events or grid choices. Predictive coefficients are not causal effects.
Penalising differences between neighbouring coefficient vectors encourages gradual change across time. The penalty strength balances flexibility and stability. Tune it inside training data and compare coefficient paths plus calibration. Excessive smoothing hides real transitions; weak smoothing follows event noise.
An L1 penalty can select features, but correlated sensors may be chosen inconsistently across folds. Group penalties can retain related signals together. Feature selection must be repeated inside every validation split. Report stability, not only the final nonzero list.
Replace linear predictors with neural or kernel functions while keeping the MTLR sequence structure. This captures interactions but increases tuning and explanation complexity. Compare with the linear model under identical grids and validation. The structured likelihood, not the network alone, maintains coherent time probabilities.
A zero or reference feature vector has its own survival distribution; each asset shifts interval scores according to covariates. Plot individual curves against the marginal benchmark. Avoid choosing a visually dramatic asset after seeing outcomes. Representative cases and distribution summaries are more honest.
The predicted median is the earliest time where survival falls to or below 0.5. If the curve never crosses within the grid, the median is beyond follow-up rather than equal to the final boundary. Quantiles inherit grid resolution. Report interval or interpolation rules.
Expected time can be approximated from the survival curve, but a finite grid leaves an unknown tail beyond the last boundary. Assuming zero survival immediately afterwards biases low; extending the last mass requires a tail model. Prefer restricted mean survival within a supported horizon.
Restricted mean survival time to τ is the area under S(t) up to τ. On a discrete grid, sum interval widths times survival within each interval under a stated convention. It gives an interpretable average event-free time within support and avoids unsupported tail extrapolation.
With three 100-hour intervals and survival at their starts 1.00, 0.90 and 0.75, a left-step approximation to 300 hours is 100(1+0.90+0.75)=265 hours. A different interpolation changes the number. State the convention and do not call it unrestricted mean lifetime.
| Term | Value |
|---|---|
| Interval width | 100 hours |
| Start survival | 1.00, 0.90, 0.75 |
| Left-step total | 265 hours |
Separate tasks can represent event type and time, but masses across causes must remain coherent. A naive independent model can make total cumulative incidence exceed one. Use a validated competing-risks formulation and compare cause-specific calibration. Label the first-event target.
Late intervals often contain few observed events because of censoring and attrition. Loss contributions and regularisation can be dominated by early tasks. Show counts at risk and events per interval. Do not claim precise late predictions when support is weak.
Evaluate survival or cumulative risk at meaningful boundaries with censoring-aware calibration curves and proper scores. A structured model guarantees monotonicity, not empirical honesty. Recalibration must preserve ordering across time or be performed with a coherent survival method.
Time-dependent AUC can show whether ranking changes by boundary. MTLR may identify different risk factors early and late. Report the case-control definition and censoring adjustment. Strong ranking at one time does not validate the whole distribution.
Refit with fewer, shifted or quantile-based boundaries. Compare Brier curves, calibration and restricted mean predictions. If results change drastically, time resolution exceeds information. Choose a stable grid before final assessment and preserve it with the model version.
Time grid, penalties, feature selection and nonlinear architecture belong inside the training loop. Outer folds estimate the selected pipeline. Group by asset and hold out later periods where possible. Reusing final outcomes to choose boundaries is subtle leakage.
A different fleet can change both covariate effects and baseline timing. Monitor interval event rates, curve calibration and prediction mass near the final boundary. A model that pushes many assets into the last interval may be extrapolating rather than expressing reliable long survival.
CHAPTER 26 OF 31 · Explore applications
26. Common misconception: each task is independent
The horizon regressors are coupled by the joint normalisation and ordered label structure. Training separate logistic models discards that coherence. MTLR is multi-task because time-specific components cooperate to define one distribution.
CHAPTER 27 OF 31 · Explore applications
27. Common misconception: monotone means calibrated
A curve can decrease smoothly from one to zero and still be wrong for every horizon. Structural validity checks probabilities; calibration checks frequencies. Use both, plus a benchmark and external validation.
Start with masses 0.10, 0.15, 0.20 and 0.55. Build survival, cumulative failure and interval hazards. Change the second mass while keeping the sum one and observe every later survival value. Then compute a restricted mean on equal-width intervals.
Fractions, cumulative sums, conditional probability and graphs are the best entry points. Ask the learner to distinguish probability of failing in an interval from probability of having failed by then. The clarity transfers to finance, medicine, engineering and AI.
MTLR turns survival analysis into a sequence of linked classification tasks while still producing one coherent event-time distribution. The idea connects school-level cumulative probability with modern machine learning.
Read Yu et al.'s original paper at https://papers.nips.cc/paper_files/paper/2011/hash/1019c8091693ef5c5f55970346633f92-Abstract.html. Connect it to https://edukatesg.com/2026/11/03/why-mathematics-deep-survival-learning-reliability-neural-hazards-censoring-regularisation-validation/ and https://edukatesg.com/2026/11/03/why-mathematics-competing-risks-reliability-cause-specific-hazards-cumulative-incidence-first-failure/. MTLR matters because mathematics keeps many horizon predictions inside one valid probability distribution.
Extended practice laboratory
Build the curve from the probability-mass ledger, not from independent horizon guesses. Ordered label patterns, joint normalisation and regularisation create coherence; calibration and external validation determine whether the curve is honest. Inspect event support per interval, grid sensitivity and the tail beyond the final boundary. Compare linear and nonlinear versions against simple survival baselines. Keep masses, hazards, cumulative risks, quantiles and restricted means clearly labelled. The mathematical achievement is coordination: many time-specific regressors cooperate to describe one distribution rather than issuing contradictory forecasts.
Students can master MTLR through cumulative sums and conditional fractions. Parents can ask whether later failure probability can ever be lower than earlier cumulative failure. The answer builds practical probability literacy.
A capstone probability ledger should list every interval mass, survival value, cumulative incidence, conditional hazard and restricted-mean contribution for one asset. Reconstruct the values independently from stored coefficients or network outputs and check that probability sums to one. Repeat under a shifted grid and later cohort. Pair structural checks with empirical calibration, because a coherent curve may still be wrong. Add a simple Kaplan–Meier and Cox benchmark. The final report should make time support visible, especially mass beyond the last boundary, so a polished personalised curve does not imply knowledge of an unobserved tail.
One distribution from linked tasks
MTLR divides follow-up into ordered intervals and learns related regressors. The tasks are not independent classifiers because event time has one direction: an asset can remain event-free and then fail, but cannot return to event-free. The parameterisation assigns probability only to valid ordered patterns, inducing masses for interval of failure plus survival beyond the grid. Neighbouring parameters may be regularised together. Every reported survival value should therefore be recoverable from non-negative masses that sum to one.
A four-mass ledger
Suppose fitted masses are 0.10, 0.15 and 0.20 for three 100-hour intervals, with 0.55 beyond the grid. Survival is 0.90, 0.75 and 0.55; cumulative failure is 0.10, 0.25 and 0.45. Conditional hazards are 0.15/0.90=0.1667 in interval two and 0.20/0.75=0.2667 in interval three. A left-step restricted mean through 300 hours is 100(1.00+0.90+0.75)=265 hours. The 0.55 tail mass prevents an unrestricted mean without extra assumptions.
Censoring contributes a tail sum
If an asset survives through interval two and is then censored, its exact future interval is unknown. Its likelihood contribution is the sum of all masses consistent with survival to that point, not a guessed failure class and not an automatic beyond-grid label. Exact boundary times need a documented convention. Verify software with tiny hand cases: one event per interval, one censoring time at each boundary and one beyond-grid survivor. Compare each likelihood term with the probability ledger.
The grid is part of the model
Narrow intervals add detail but create sparse tasks and weak censoring support. Wide intervals reduce resolution and can hide decision-relevant changes. Candidate grids can use engineering milestones, balanced event counts or a hybrid. Choose them within training data and reserve the test set. Report sensitivity to moving boundaries. A grid ending at 1,000 hours does not support mean lifetime to infinity. Extending it simply to force a median replaces missing evidence with an arbitrary cut.
Regularisation and feature paths
A feature can have different coefficients across intervals. Smoothness discourages jagged paths, while shrinkage controls complexity. Tune both with nested asset-grouped or temporal validation. Plot coefficient paths with resampling uncertainty and compare a constant-effect model. A sign change may represent time variation, sparse data or correlated predictors; it is not automatically causal. Store standardisation, missing-data rules and the final interval definition with the coefficient matrix so another analyst can reproduce the curve.
Monotonicity is not calibration
A survival curve that never rises passes a structural check yet may be overconfident. Evaluate calibration and proper scores at prespecified horizons, with censoring-aware estimators and support warnings. Compare Kaplan–Meier, a simple Cox model and a constant-risk baseline. Test a later period or unseen fleet and inspect groups separately. Deployment monitoring should reconstruct the curve from stored outputs, confirm probability sums, watch covariate drift and reject requests beyond the supported grid.
Student probability lab
Use cards labelled 0.10, 0.15, 0.20 and 0.55. Students build survival by subtracting cumulative masses, then compute conditional hazards using only those still at risk. Move one boundary and discuss which card changes meaning. Censor an asset after interval two and circle every compatible future card. Parents can ask why the 0.55 tail is not failure at exactly 300 hours. The activity connects fractions, cumulative sums, conditional probability, functions and honest limits on extrapolation.
Decision-ready reporting
For a maintenance horizon, report the predicted risk by that horizon, calibration evidence, supported follow-up and an alternative simple model. Do not present an entire smooth-looking curve as equally certain at every time. Show how thresholds translate into alerts and missed events, then keep intervention benefits separate. If the model is recalibrated, verify that cross-time coherence remains. A versioned deployment package contains grid boundaries, feature order, transformations, parameters, censoring convention and a test record with expected masses.
Grid and model sensitivity dossier
A rigorous dossier repeats the analysis under at least two defensible grids. For each, show event counts, censoring support and calibration at decision horizons. Compare weak and strong parameter smoothing, and explain which differences matter operationally. A horizon that moves from 18% to 32% risk when one nearby boundary shifts deserves investigation even if the average score barely changes. Bootstrap by asset to preserve within-asset dependence and display uncertainty for probability masses, survival and restricted mean. Check the final tail mass by cohort and flag assets whose requested horizon exceeds training support. Finally, reconstruct one prediction from raw features through standardisation, interval logits, normalising constant and cumulative sums. The reconstruction should match the production result within a declared tolerance. This exercise catches boundary conventions, feature-order errors and numerical overflow that aggregate validation can miss.
Misconceptions to correct
MTLR does not guarantee accurate probabilities merely because its curves are coherent. It does not turn interval coefficients into causal effects, and it does not identify an unrestricted mean when follow-up ends early. More intervals do not automatically create more information. A smooth curve between coarse boundaries is a presentation choice unless the model defines that interpolation. Students should label failure mass, cumulative failure, survival and conditional hazard separately, then check every quantity against the same probability ledger. This vocabulary discipline prevents many otherwise sophisticated reporting errors.
A reproducibility handoff
Give a second analyst one frozen feature row, the ordered interval boundaries, coefficient matrix, intercepts and normalisation rule. They should reproduce every mass and survival value without consulting the training notebook. Include a censored training example and explain its allowed future patterns. Then repeat the handoff after exporting the model to production code. Differences often expose reversed interval order, omitted intercepts or unstable exponentiation. Record numeric precision and the rule used when a requested time falls inside an interval. A reliable survival model is a complete computational contract, not only a fitted object.
Final self-check
Before accepting a result, confirm that all masses are non-negative and total one, survival never rises, cumulative failure never falls and conditional hazards use the correct at-risk denominator. Verify that reported horizons fall within the grid and later-cohort follow-up supports their evaluation. These simple checks do not prove accuracy, but they catch contradictions before a curve reaches a family, technician or decision-maker.
Practice case 1: Probability ledger
Setup: Use four interval masses.
Task: Build survival and cumulative failure.
Checking target: Masses must sum to one.
Start with interval masses 0.10, 0.15, 0.20 and 0.55. Build cumulative failure 0.10, 0.25, 0.45 and survival 0.90, 0.75, 0.55. Verify all masses are nonnegative and sum to one. Move 0.05 from the final mass to interval two and trace every later survival value. This simple ledger is the foundation of a coherent event-time distribution.
Practice case 2: Conditional hazard
Setup: Convert masses to hazards.
Task: Calculate interval risks.
Checking target: Label conditional and cumulative quantities.
Calculate interval hazards from probability masses: 0.15/0.90=0.1667 and 0.20/0.75=0.2667. Explain why a mass is unconditional probability of failing in the interval while hazard is conditional on entering it. Label cumulative risk separately. Misidentifying these three quantities can produce correct arithmetic and wrong decisions, so every API and chart needs an explicit output name.
Practice case 3: Ordered patterns
Setup: Encode one failure time.
Task: List valid label sequences.
Checking target: Never allow failed-to-alive transitions.
Encode failure in the second interval as an ordered zero-then-one label sequence. Write valid patterns for all intervals and one impossible pattern such as 1,0,1. Explain how the joint normalisation gives probability only to valid event-time positions. Repeat for censoring after interval one and list all event intervals consistent with the observation instead of assigning a false failure.
Practice case 4: Smoothness
Setup: Penalise neighbouring coefficients.
Task: Compare weak and strong smoothing.
Checking target: Tune inside training data.
Fit or imagine coefficient paths over time for one predictor. Apply a smoothness penalty to neighbouring differences and compare weak versus strong penalties. Strong smoothing produces stable paths but can hide a sudden mechanism change; weak smoothing follows sparse interval noise. Tune inside training data and inspect both prediction calibration and coefficient stability. The penalty is a modelling decision, not a discovered physical law.
Practice case 5: Restricted mean
Setup: Use three 100-hour intervals.
Task: Calculate supported mean time.
Checking target: Do not extrapolate the tail.
Use three 100-hour intervals with survival at their starts 1.00, 0.90 and 0.75. Compute left-step restricted mean 265 hours. Compare with trapezoidal interpolation and explain why results differ. Keep the restriction horizon in the name. If substantial mass remains beyond the grid, unrestricted mean lifetime is unsupported unless a tail model is added.
Practice case 6: Grid sensitivity
Setup: Compare three boundary schemes.
Task: Recode and refit.
Checking target: Time resolution must match evidence.
Create equal-width, event-quantile and operational time grids. Recode the same events and compare counts per interval, curve calibration and restricted means. A fine grid may create empty late tasks; a coarse grid hides timing. Choose the grid without inspecting final test performance. Preserve boundary rules for events and censoring exactly at cut points.
Practice case 7: Nested validation
Setup: Tune grid and penalties.
Task: Evaluate outer grouped folds.
Checking target: Avoid horizon-selection leakage.
Run nested validation where grid, penalties, feature selection and any neural extension are chosen inside outer training folds. Group by asset and hold out later cohorts. Compare with independently fitted horizon classifiers that can violate monotonicity. MTLR should be judged on calibration, proper scores and curve coherence, not merely the attractiveness of a smooth plot.
Practice case 8: Deployment package
Setup: Version grid and normaliser.
Task: Reconstruct one curve.
Checking target: Monitor final-interval mass.
Package time boundaries, feature scaler, coefficient matrix or network, normaliser, censoring convention and calibration checks together. Reconstruct one asset's interval masses from stored values and verify they sum to one. Add monitors for mass accumulating at the final interval, which may signal unsupported long survival or changing follow-up. Version the grid whenever retraining changes it.
Practice case 9: Ordered labels
Setup: List valid sequences for four intervals.
Task: Reject impossible patterns.
Checking target: The model distribution must respect time order.
Write 0000 for survival beyond the grid and sequences that switch once from zero to one for failures. Count m+1 valid patterns rather than all 2^m binary strings. This restriction is how dependent tasks become one event-time variable. Confirm the implementation never assigns mass to a sequence that returns from failed to alive.
Practice case 10: Censoring sum
Setup: Censor after interval two.
Task: Sum consistent future masses.
Checking target: Do not select one unknown failure interval.
If the asset survives through two intervals, its likelihood includes failure in any later interval or survival beyond the grid. Add those masses and compare with survival at boundary two. Censoring provides a tail probability, not a class label. Boundary rules should be tested with exact cut-point times.
Practice case 11: Grid tail
Setup: Place large mass beyond the final boundary.
Task: Assess median and mean reporting.
Checking target: Do not collapse unknown tail to the last time.
When final survival mass is high, the median may be beyond observed follow-up and unrestricted expected lifetime is unidentified. Report that limitation. Compute restricted mean only to a supported horizon. Extend the grid only when data support it, not to force a numerical median for every asset.
Practice case 12: Feature path
Setup: Plot one coefficient across intervals.
Task: Bootstrap its trajectory.
Checking target: Time-varying effects need stability evidence.
A coefficient can change sign across horizons. Repeat the fit across grouped resamples and show uncertainty. Sparse intervals may create dramatic but unstable changes. Compare with a constant-effect model. Predictive coefficient paths are descriptions of the fitted distribution, not causal instructions.
Practice case 13: Calibration matrix
Setup: Evaluate several horizons and groups.
Task: Show probability honesty.
Checking target: Monotonicity is only a structural check.
Construct calibration-in-the-large and Brier values at each boundary for the full cohort and predeclared groups. Flag cells with weak follow-up. A curve can be perfectly monotone while systematically too optimistic. Recalibration should preserve cross-time coherence and be tested on new data.
Practice case 14: Competing mass
Setup: Split interval mass between two event types.
Task: Check total probability.
Checking target: Cause probabilities must share one simplex.
For each interval, allocate mass to failure modes while ensuring all cause-time masses plus survival beyond the grid sum to one. Compute cumulative incidence by cause. Independently fitted cause models can exceed total probability one. Use a validated competing-risks extension and compare cause-specific calibration.