Survival Analysis

The defining feature of survival analysis is its ability to handle incomplete observations. In customer analytics, most users haven't churned at the time of analysis.

• Right Censoring: The most common type—a customer is still active when the study ends, so their true lifetime is unknown but bounded below
• Left Censoring: A customer's start date precedes data collection, obscuring their true origin
• Interval Censoring: Churn is known to have occurred between two observation points, but the exact moment is unknown

Ignoring censored data by treating active customers as non-churners systematically underestimates churn risk and inflates CLV projections.

The hazard function h(t) represents the instantaneous risk of the event occurring at time t, conditional on survival up to that moment. It's the core modeling target in survival analysis.

• Interpretation: The probability density of churning in the next infinitesimal interval, given the customer is still active
• Shape possibilities: Constant (memoryless), increasing (wear-out), decreasing (burn-in), or bathtub-shaped
• Relationship to survival: The survival function S(t) = exp(-∫h(u)du) is derived directly from the cumulative hazard

Unlike a simple churn probability score, the hazard function reveals when risk peaks—enabling time-sensitive retention interventions.

The Kaplan-Meier estimator is the foundational non-parametric method for estimating the survival function from observed lifetime data with censoring.

• Step function: Survival probability drops only at observed event times, creating a staircase curve
• Product-limit formula: Ŝ(t) = ∏(1 - dᵢ/nᵢ) where dᵢ are events and nᵢ are at-risk subjects at time tᵢ
• Key assumption: Censoring must be non-informative—the reason for censoring is unrelated to the event risk

In CLV contexts, Kaplan-Meier curves visualize cohort retention decay and provide empirical benchmarks against which parametric models are validated.

The Cox model is a semi-parametric regression that assesses how multiple covariates influence the hazard rate without requiring specification of the baseline hazard shape.

• Core equation: h(t|X) = h₀(t) × exp(β₁X₁ + β₂X₂ + ... + βₚXₚ)
• Proportional hazards assumption: The effect of a covariate is constant over time—the hazard ratio between two subjects remains fixed
• Partial likelihood: Estimation uses only the ordering of event times, not their exact values, making it robust to outliers

For churn modeling, coefficients directly quantify risk factors: a coefficient of 0.7 means a unit increase in the predictor multiplies the hazard by exp(0.7) ≈ 2.01, doubling churn risk.

Standard survival models assume fixed predictor values, but customer behavior evolves. Time-varying covariates extend the framework to incorporate dynamic features.

• Extended Cox model: h(t|X(t)) = h₀(t) × exp(β₁X₁(t) + β₂X₂ + ...) where X₁(t) changes over time
• Counting process format: Each customer contributes multiple rows, one per time interval with updated covariate values
• Examples: Recent purchase frequency, support ticket volume, session recency, or engagement score that updates weekly

This captures the reality that a customer's churn risk isn't static—a sudden drop in engagement dramatically elevates the hazard, enabling real-time intervention triggers.

Customers can exit for multiple reasons—voluntary churn, involuntary churn, or death of the relationship. Competing risks analysis models these distinct event types simultaneously.

• Cause-specific hazard: The instantaneous risk of a specific event type, treating other events as censoring
• Cumulative incidence function (CIF): The probability of experiencing a particular event by time t, accounting for the fact that other events preclude it
• Fine-Gray subdistribution hazard: Models the direct effect of covariates on the CIF without requiring independence of competing events

For CLV, distinguishing between a customer lost to a competitor versus one lost to category disinterest informs fundamentally different retention strategies.

The instantaneous potential per unit of time for the event of interest (e.g., churn) to occur at a specific time point, given that the individual has survived up to that point. Unlike simple probability, the hazard rate captures dynamic risk over time.

• Interpretation: A hazard rate of 0.05 at month 12 means 5% of remaining customers are expected to churn in that month
• Shapes: Can be constant, increasing (wear-out), decreasing (burn-in), or bathtub-shaped
• Relationship: Directly linked to the survival function via the cumulative hazard: S(t) = exp(-H(t))

A semi-parametric regression model that assesses the effect of multiple covariates on the hazard rate without requiring specification of the baseline hazard distribution.

• Key Assumption: Proportional hazards — the ratio of hazards between two individuals is constant over time
• Output: Hazard ratios (HR); HR > 1 indicates increased risk, HR < 1 indicates protective effect
• Use Case: Identifying which customer attributes (tenure, support tickets, usage frequency) most strongly predict churn risk
• Extension: Time-varying covariates can be incorporated to model changing risk factors

A non-parametric statistic used to estimate the survival function from lifetime data, producing a step function that decreases only at observed event times.

• Censoring Handling: Naturally accommodates right-censored observations by adjusting the risk set at each time point
• Visualization: Generates the iconic 'survival curve' showing the proportion of customers retained over time
• Comparison: The log-rank test statistically compares Kaplan-Meier curves between groups (e.g., high vs. low engagement cohorts)
• Limitation: Cannot directly adjust for multiple confounding variables

The handling of incomplete observations where the exact event time is unknown, a defining feature that distinguishes survival analysis from standard regression.

• Right Censoring: Most common in CLV — customer has not churned by end of observation window
• Left Censoring: Customer entered the study after their initial acquisition event
• Interval Censoring: Churn is known to have occurred within a time interval but exact date is unknown
• Informative Censoring: When censoring is related to the event risk (e.g., high-value customers are less likely to be lost to follow-up), requiring specialized methods

A parametric alternative to the Cox model that directly models the logarithm of survival time as a linear function of covariates, effectively 'accelerating' or 'decelerating' the time scale.

• Distributions: Assumes a specific distribution for survival times — Weibull, log-normal, log-logistic, or exponential
• Interpretation: Coefficients represent time ratios; a coefficient of 0.5 means the expected survival time is halved
• Advantage: Produces directly interpretable predictions of expected remaining lifetime
• When to Use: Preferred when the proportional hazards assumption is violated or when predicting absolute survival times is the primary goal

An extension of survival analysis that handles multiple mutually exclusive event types, where the occurrence of one event precludes observation of others.

• Cause-Specific Hazard: The instantaneous rate of a specific event type in the presence of competing risks
• Cumulative Incidence Function (CIF): Estimates the probability of experiencing a specific event by time t, accounting for competing events
• Application: Distinguishing between voluntary churn, involuntary churn (failed payment), and account dormancy as distinct competing outcomes
• Fine-Gray Model: A subdistribution hazard regression that directly models the CIF for a specific event type