Survival Analysis

Parent: Data Analysis · Published reference · snapshot 2026-09-08 · skill da-analytical-methods/references/da-24-survival-analysis.md

↓ Facts as markdown all context files

Modeling the time until an event happens when some observations are incomplete (censored or truncated). This is its own discipline because ordinary regression cannot use a row that says "this customer

These notes link each claim to its source. Some citations identify archived research reports that are no longer hosted on this site. A published reference means the content is available; it does not certify independent review or accuracy.Read the editorial policy and follow the sources before relying on a claim.

Survival Analysis / Time-to-Event Analysis

Modeling the time until an event happens when some observations are incomplete (censored or truncated). This is its own discipline because ordinary regression cannot use a row that says "this customer had not churned yet when we stopped looking" - survival methods extract information from exactly those incomplete rows. Canonical textbooks: Klein & Moeschberger Survival Analysis: Techniques for Censored and Truncated Data (2nd ed, 2003); Therneau & Grambsch Modeling Survival Data (2000). Primary Python tooling: lifelines and scikit-survival; R: survival + survminer. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#survival-analysis-time-to-event-analysis]

When to use this skill

The outcome is a duration until an event: death, machine failure, churn, loan default, conversion, hospital readmission. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#when-to-use-this-skill]
Some subjects have not experienced the event by end of observation (censoring), or only entered observation partway through (truncation). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#when-to-use-this-skill]
You need a survival curve, hazard ratio, median time-to-event, or cumulative incidence. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#when-to-use-this-skill]

When NOT to use this skill

Forecasting a numeric series over calendar time → da-15-forecasting [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#when-not-to-use-this-skill]
Regression/classification with fully observed outcomes → da-6 / da-7 [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#when-not-to-use-this-skill]
Causal/experiment analysis with no time component → da-12 [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#when-not-to-use-this-skill]
A descriptive cohort retention table (no estimator, no model) → da-21-product-analytics [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#when-not-to-use-this-skill]
Computing a CLV dollar figure with BG/NBD, Pareto/NBD, or Gamma-Gamma spend models → da-23-customer-lifetime-value (this skill covers only the time-to-churn / survival-curve half) [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#when-not-to-use-this-skill]

1. Censoring and truncation — the defining feature

The reason survival analysis exists. Get this wrong and every downstream estimate is biased. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#1-censoring-and-truncation-the-defining-feature]
Key distinction: censoring keeps the subject but loses event-time detail; truncation removes the subject from the sample entirely (Stats Ox lecture notes, 2020; NJIT Math 659 Ch.3, 2011; GeeksforGeeks, 2024). The standard estimators assume censoring is non-informative (independent of the event process). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#1-censoring-and-truncation-the-defining-feature]

2. The survival, hazard, and cumulative-hazard functions

Three interchangeable views of the same distribution; pick whichever the audience reads best. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#2-the-survival-hazard-and-cumulative-hazard-functions]
- Survival function S(t) = P(T > t) - probability of surviving past t. Monotone non-increasing from 1. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#2-the-survival-hazard-and-cumulative-hazard-functions]
- Hazard function h(t) = lim Δ→0 P(t ≤ T < t+Δ | T ≥ t)/Δ - instantaneous event rate given survival so far. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#2-the-survival-hazard-and-cumulative-hazard-functions]
- Cumulative hazard H(t) = ∫₀ᵗ h(u)du, with the bridge identity S(t) = exp(−H(t)). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#2-the-survival-hazard-and-cumulative-hazard-functions]
The hazard is the modeling target for most methods (lifelines Quickstart v0.30, 2025; Klein & Moeschberger Ch. 2, 2003). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#2-the-survival-hazard-and-cumulative-hazard-functions]

3. Kaplan-Meier & Nelson-Aalen (non-parametric estimators)

The first thing to compute on any survival dataset - assumption-free descriptive curves. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#3-kaplan-meier-nelson-aalen-non-parametric-estimators]
- Kaplan-Meier (product-limit) estimator of S(t): at each event time multiply by (1 − dᵢ/nᵢ). Step function; censored subjects drop out of the risk set without a step. Report median survival and confidence bands. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#3-kaplan-meier-nelson-aalen-non-parametric-estimators]
- Nelson-Aalen estimator of H(t): sum of dᵢ/nᵢ. Estimates cumulative hazard under independent right-censoring and left-truncation (lifelines NelsonAalenFitter docs, 2025). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#3-kaplan-meier-nelson-aalen-non-parametric-estimators]
Sources: lifelines Quickstart (2025); Klein & Moeschberger Ch. 4 (2003); CPSC 330 Survival lecture (2023). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#3-kaplan-meier-nelson-aalen-non-parametric-estimators]

4. The log-rank test (comparing groups)

Compares two-or-more KM curves; null = equal survival across groups. Chi-square test accumulating observed-minus-expected events at each event time; weights all time points equally (Wilcoxon/Tarone-Ware variants weight early times more). Gives a p-value, not an effect size - for an effect size use Cox. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#4-the-log-rank-test-comparing-groups]
Sources: lifelines.statistics (2025); STHDA (2018); Klein & Moeschberger Ch. 7 (2003). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#4-the-log-rank-test-comparing-groups]

5. Cox proportional-hazards model (the workhorse)

Semi-parametric: h(t|x) = h₀(t) · exp(βᵀx). Baseline hazard h₀(t) is unspecified; β estimated via partial likelihood (Cox 1972). exp(βⱼ) is the hazard ratio - multiplicative, time-constant. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#5-cox-proportional-hazards-model-the-workhorse]
Tie handling: Efron (default) or Breslow. Report HRs with CIs. Sources: lifelines CoxPHFitter (2025); Therneau & Grambsch (2000); Researchers' Guide (2021). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#5-cox-proportional-hazards-model-the-workhorse]

6. The proportional-hazards assumption & diagnostics

Cox is only valid if hazard ratios are constant over time. Always check. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#6-the-proportional-hazards-assumption-diagnostics]
- Scaled Schoenfeld residuals: zero slope against (a function of) time under PH. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#6-the-proportional-hazards-assumption-diagnostics]
- Grambsch-Therneau test (cox.zph in R, cph.check_assumptions() / proportional_hazard_test in lifelines): null = PH holds; small p-value flags a violation. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#6-the-proportional-hazards-assumption-diagnostics]
- Graphical: ggcoxzph() (survminer) - LOESS smooth should be flat. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#6-the-proportional-hazards-assumption-diagnostics]
Fixes when violated: stratify (strata=), add a covariate×time interaction, split follow-up into intervals, or switch to AFT. Sources: UCLA OARC (2021); Stata stcox (2015); STHDA (2018). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#6-the-proportional-hazards-assumption-diagnostics]

7. Parametric models: exponential, Weibull, and AFT

For a smooth curve, extrapolation, or a generative model. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#7-parametric-models-exponential-weibull-and-aft]
- Exponential: constant hazard h(t)=λ. Memoryless baseline. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#7-parametric-models-exponential-weibull-and-aft]
- Weibull: monotone increasing (ρ>1) or decreasing (ρ<1) hazard. The default parametric choice. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#7-parametric-models-exponential-weibull-and-aft]
- AFT: log(T) = βᵀx + error; covariates accelerate/decelerate time-to-event (exp(β) = time ratio). More interpretable for "this doubles the expected lifetime." [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#7-parametric-models-exponential-weibull-and-aft]
- Weibull is the only distribution expressible as both PH and AFT. Log-logistic / log-normal AFT allow non-monotone hazards. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#7-parametric-models-exponential-weibull-and-aft]
Sources: AFT model - Wikipedia (2025); CRAN eha (2024); AFT vs Cox PMC4645729 (2015). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#7-parametric-models-exponential-weibull-and-aft]

8. Competing risks (cause-specific vs. Fine-Gray)

When a subject can fail from mutually exclusive causes, naïve KM/Cox on one cause over-estimates its incidence by treating competing events as censored. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#8-competing-risks-cause-specific-vs-fine-gray]
- Cause-specific hazard (Cox per cause): rate of cause k among those still at risk. Best for etiology. Censor competing events. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#8-competing-risks-cause-specific-vs-fine-gray]
- Fine-Gray subdistribution hazard: links covariates to the cumulative incidence function (CIF) - the actual probability of cause k, accounting for competing events. Best for prediction / risk communication (sHR). Competing-event subjects stay in the risk set with decaying weights. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#8-competing-risks-cause-specific-vs-fine-gray]
Caveats: separate Fine-Gray per cause → CIFs can sum > 1; avoid multiple Fine-Gray models - prefer cause-specific for multi-event questions. For causal effects, Fine-Gray is discouraged. Sources: Austin & Fine, Stat Med (2017); Austin et al. (2021); Statistical Horizons (2023). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#8-competing-risks-cause-specific-vs-fine-gray]

9. Time-varying covariates

When a predictor changes during follow-up, a single baseline value is wrong. Use long (counting-process) format: one row per subject per interval (id, start, stop, event, covariates). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#9-time-varying-covariates]
Also the standard fix for a time-varying coefficient (a PH violation) - though that needs a covariate×time interaction. Sources: lifelines Time-varying regression (2025); CoxTimeVaryingFitter docs (2025); Therneau & Grambsch Ch. 3 (2000). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#9-time-varying-covariates]

10. Discrete-time survival & churn / retention / CLV

When time is naturally binned and many events tie at the same bin, discrete-time survival beats continuous Cox. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#10-discrete-time-survival-churn-retention-clv]
- Method: expand to person-period rows, fit ordinary logistic regression with the period (or a flexible function of it) as predictor. Fitted per-period probabilities are the discrete hazards; chain into a survival/retention curve. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#10-discrete-time-survival-churn-retention-clv]
- Churn / retention: tenure = duration, churn = event, active customers = right-censored. KM gives the retention curve; Cox/AFT give "what drives churn timing"; integrating S(t) gives expected lifetime, the backbone of CLV (CLV ≈ Σ margin·S(t)·discount). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#10-discrete-time-survival-churn-retention-clv]
Survival beats a static churn classifier: it answers when, uses censored customers correctly, and yields retention curves and CLV directly. Sources: SAS Survival Data Mining (2012); SAS CLV (2003); Springer churn prediction (2025). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#10-discrete-time-survival-churn-retention-clv]

11. Machine-learning survival models

When effects are nonlinear/interacting/high-dimensional and accuracy beats interpretability. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#11-machine-learning-survival-models]
- Random Survival Forests (RSF): survival trees split on the log-rank statistic; ensemble cumulative-hazard estimate; handles nonlinearities, right-censoring, variable importance. Ishwaran et al. (2008). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#11-machine-learning-survival-models]
- Gradient-boosted survival: boosts weak learners against a survival loss; often the strongest tabular baseline. scikit-survival GradientBoostingSurvivalAnalysis ~0.75 C-index on the standard example. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#11-machine-learning-survival-models]
- DeepSurv (Katzman et al., 2018): deep net optimizing the Cox partial-likelihood loss; nonlinear Cox for personalized risk. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#11-machine-learning-survival-models]
- Evaluation: Harrell's concordance index (C-index) (0.5 random, 1.0 perfect); time-dependent AUC; integrated Brier score. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#11-machine-learning-survival-models]
Sources: Ishwaran et al., Ann. Appl. Stat. 2(3):841-860 (2008); scikit-survival RSF & boosting guides (2025); Katzman et al., DeepSurv, BMC Med Res Methodol / arXiv 1606.00931 (2018). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#11-machine-learning-survival-models]

Methodology (default workflow)

Define the timeline: t=0 origin, the event, the censoring rule; check for left truncation / delayed entry. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#methodology-default-workflow]
Describe: KM curve + median survival; Nelson-Aalen for cumulative hazard; stratify by key groups. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#methodology-default-workflow]
Compare groups: log-rank (effect size deferred to Cox). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#methodology-default-workflow]
Model effects: Cox PH first; parametric/AFT for extrapolation or a smooth curve. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#methodology-default-workflow]
Check assumptions: Schoenfeld residuals / cox.zph; repair PH violations. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#methodology-default-workflow]
Handle structure: competing risks → cause-specific or Fine-Gray; changing covariates → time-varying; binned time → discrete-time logistic. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#methodology-default-workflow]
Predict at scale: RSF / gradient boosting / DeepSurv. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#methodology-default-workflow]
Validate: C-index, time-dependent AUC, integrated Brier, calibration; never plain accuracy. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#methodology-default-workflow]

Practical patterns

Always plot KM first - reveals crossing curves (PH violation), plateaus (cured fraction), data problems. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#practical-patterns]
Encode the outcome as a pair (event_indicator, time) - scikit-survival needs a structured array; lifelines takes two columns. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#practical-patterns]
Report hazard ratios with CIs and translate: "HR 1.4 → 40% higher instantaneous churn rate." [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#practical-patterns]
Use the right time origin (calendar / age / time-since-enrollment); left-truncate on delayed entry. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#practical-patterns]
For churn/CLV, integrate the survival curve for expected lifetime instead of averaging completed tenures (which ignores censored = still-active customers). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#practical-patterns]

Anti-patterns

Dropping censored rows - the cardinal sin; discards most information and badly biases estimates. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#anti-patterns]
Treating time-to-event as an OLS regression target - censoring makes the target undefined for survivors. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#anti-patterns]
Treating competing events as plain censoring when estimating one cause's incidence - over-states it; use CIF / Fine-Gray. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#anti-patterns]
Fitting Cox without checking PH - silently corrupts every hazard ratio. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#anti-patterns]
Reporting only a log-rank p-value with no effect size or curve. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#anti-patterns]
Evaluating an ML survival model with accuracy/AUC on a binarized label instead of C-index / Brier. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#anti-patterns]
One Fine-Gray model per cause read together - CIFs can sum past 1; prefer cause-specific for multi-event questions. [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#anti-patterns]

References

Klein & Moeschberger, Survival Analysis: Techniques for Censored and Truncated Data, 2nd ed., Springer (2003). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#references]
Therneau & Grambsch, Modeling Survival Data: Extending the Cox Model, Springer (2000). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#references]
lifelines docs - https://lifelines.readthedocs.io/en/latest/ (v0.30, 2025). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#references]
scikit-survival user guide - https://scikit-survival.readthedocs.io/en/stable/ (2025). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#references]
Censoring & truncation - https://www.stats.ox.ac.uk/~mlunn/lecturenotes1.pdf (2020); https://web.njit.edu/~wguo/Math%20659_2011/Math659_Chapter3.pdf (2011). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#references]
PH test - https://stats.oarc.ucla.edu/other/examples/asa2/testing-the-proportional-hazard-assumption-in-cox-models/ (2021); https://www.stata.com/manuals14/ststcoxph-assumptiontests.pdf (2015). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#references]
AFT - https://en.wikipedia.org/wiki/Accelerated_failure_time_model (2025); https://pmc.ncbi.nlm.nih.gov/articles/PMC4645729/ (2015). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#references]
Fine-Gray - https://onlinelibrary.wiley.com/doi/10.1002/sim.7501 (2017); https://onlinelibrary.wiley.com/doi/full/10.1002/sim.9023 (2021); https://statisticalhorizons.com/for-causal-analysis-of-competing-risks/ (2023). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#references]
Discrete-time / churn / CLV - https://support.sas.com/resources/papers/proceedings12/132-2012.pdf (2012); https://support.sas.com/resources/papers/proceedings/proceedings/sugi28/120-28.pdf (2003); https://link.springer.com/article/10.1057/s41270-025-00450-2 (2025). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#references]
ML survival - Ishwaran et al. RSF https://ishwaran.org/papers/IKBL.AOAS.pdf (2008); Katzman et al. DeepSurv https://link.springer.com/article/10.1186/s12874-018-0482-1 / https://arxiv.org/abs/1606.00931 (2018). [archived source: https://llms-explorer.com/sources/mdb-context-hub/da-24-survival-analysis/#references]

Where this helps

Modeling customer churn where "hasn't churned yet" is a censored observation, not a missing data point ordinary regression can use as-is. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Comparing time-to-event curves between two groups (e.g. two onboarding flows) with the log-rank test before fitting a full regression model. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Estimating a hazard ratio for a covariate (like plan tier) on churn timing with a Cox proportional-hazards model, when the exact shape of the baseline hazard doesn't need to be specified. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Modeling cause-specific attrition (e.g. downgrade vs. cancellation as competing risks) where naive Kaplan-Meier on one cause alone would over-estimate its incidence. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.

Project ideas

Build a churn survival-curve dashboard using Kaplan-Meier estimation on customer tenure data, with active customers correctly right-censored. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Fit a Cox proportional-hazards model on subscription data to identify which covariates raise or lower the instantaneous churn rate, and check the proportional-hazards assumption with scaled Schoenfeld residuals. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Implement a discrete-time survival model (person-period expansion plus logistic regression) for a naturally-binned retention problem like monthly subscription churn. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Build a customer lifetime value estimate by integrating a fitted survival function, using the standard survival-analysis backbone for CLV. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.

Common mistakes

Dropping censored rows instead of keeping them — this discards most of the informative data and badly biases every downstream estimate. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Fitting ordinary least-squares regression on time-to-event data, which is undefined for subjects who haven't experienced the event yet. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Treating competing events as plain censoring when estimating one cause's incidence, which over-states that cause's probability — use the cumulative incidence function or Fine-Gray instead. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Fitting a Cox model without checking the proportional-hazards assumption, silently invalidating the resulting hazard ratios if the assumption is violated. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.

Known issues

The Cox model's hazard ratios are only valid under the proportional-hazards assumption; a violation, checked via Grambsch-Therneau or scaled Schoenfeld residuals, requires either a time-varying-covariate fix or a different model entirely. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Competing-risks analysis is easy to get subtly wrong — cause-specific hazards answer an etiology question while Fine-Gray's subdistribution hazard answers a prediction question, and using the wrong one for the intended audience misleads. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Machine-learning survival models (Random Survival Forests, gradient-boosted survival) trade interpretability for accuracy on nonlinear or high-dimensional effects, so the choice against a Cox model is a real interpretability-vs-accuracy tradeoff, not a strict upgrade. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Time-varying covariates require reshaping data into long, counting-process format — a modeling decision that changes the data pipeline, not just the model call, and is easy to skip under time pressure. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.

Where this helps

Modeling customer churn where "hasn't churned yet" is a censored observation, not a missing data point ordinary regression can use as-is. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Comparing time-to-event curves between two groups (e.g. two onboarding flows) with the log-rank test before fitting a full regression model. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Estimating a hazard ratio for a covariate (like plan tier) on churn timing with a Cox proportional-hazards model, when the exact shape of the baseline hazard doesn't need to be specified. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Modeling cause-specific attrition (e.g. downgrade vs. cancellation as competing risks) where naive Kaplan-Meier on one cause alone would over-estimate its incidence. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.

Project ideas

Build a churn survival-curve dashboard using Kaplan-Meier estimation on customer tenure data, with active customers correctly right-censored. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Fit a Cox proportional-hazards model on subscription data to identify which covariates raise or lower the instantaneous churn rate, and check the proportional-hazards assumption with scaled Schoenfeld residuals. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Implement a discrete-time survival model (person-period expansion plus logistic regression) for a naturally-binned retention problem like monthly subscription churn. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Build a customer lifetime value estimate by integrating a fitted survival function, using the standard survival-analysis backbone for CLV. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.

Common mistakes

Dropping censored rows instead of keeping them — this discards most of the informative data and badly biases every downstream estimate. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Fitting ordinary least-squares regression on time-to-event data, which is undefined for subjects who haven't experienced the event yet. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Treating competing events as plain censoring when estimating one cause's incidence, which over-states that cause's probability — use the cumulative incidence function or Fine-Gray instead. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Fitting a Cox model without checking the proportional-hazards assumption, silently invalidating the resulting hazard ratios if the assumption is violated. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.

Known issues

The Cox model's hazard ratios are only valid under the proportional-hazards assumption; a violation, checked via Grambsch-Therneau or scaled Schoenfeld residuals, requires either a time-varying-covariate fix or a different model entirely. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Competing-risks analysis is easy to get subtly wrong — cause-specific hazards answer an etiology question while Fine-Gray's subdistribution hazard answers a prediction question, and using the wrong one for the intended audience misleads. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Machine-learning survival models (Random Survival Forests, gradient-boosted survival) trade interpretability for accuracy on nonlinear or high-dimensional effects, so the choice against a Cox model is a real interpretability-vs-accuracy tradeoff, not a strict upgrade. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.
Time-varying covariates require reshaping data into long, counting-process format — a modeling decision that changes the data pipeline, not just the model call, and is easy to skip under time pressure. [source] — AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.

Children

Censoring and Truncation (frontier)
Survival and Hazard Functions (frontier)
Kaplan-Meier and Nelson-Aalen Estimators (frontier)
Log-Rank Test (frontier)
Cox Proportional-Hazards Model (frontier)
Proportional-Hazards Assumption and Diagnostics (frontier)
Parametric and Accelerated Failure Time Models (frontier)
Competing Risks (Cause-Specific and Fine-Gray) (frontier)
Time-Varying Covariates (frontier)
Discrete-Time Survival and Churn/CLV (frontier)
Machine-Learning Survival Models (frontier)

Frontier under this node: Censoring and Truncation, Competing Risks (Cause-Specific and Fine-Gray), Cox Proportional-Hazards Model, Discrete-Time Survival and Churn/CLV, Kaplan-Meier and Nelson-Aalen Estimators, Log-Rank Test, Machine-Learning Survival Models, Parametric and Accelerated Failure Time Models, Proportional-Hazards Assumption and Diagnostics, Survival and Hazard Functions, Time-Varying Covariates

← the whole tree · 3D view· how to read this page