Kaplan-Meier Survival Curves

How long do customers remain subscribed before cancelling? Kaplan-Meier estimates the chance of remaining subscribed over time, using each customer's observed duration and whether cancellation occurred. Customers still subscribed at the observation cutoff contribute censored durations rather than being counted as cancellations.

What is the Kaplan-Meier Estimator?

The Kaplan-Meier estimator is a non-parametric statistical method that calculates the probability of survival past specific time points when some observations are incomplete or censored. Named after Edward Kaplan and Paul Meier who published the technique in 1958 (Kaplan and Meier, 1958), it is the standard nonparametric estimate of a survival curve in medicine, engineering and customer analytics.

To estimate how long until your own customers cancel, counting the ones who have not yet, try the free survival analysis tool.

Unlike parametric survival methods that assume a specific distribution (exponential, Weibull, etc.), the Kaplan-Meier estimator makes no assumptions about the shape of the survival curve. This flexibility makes it robust and widely applicable to diverse business scenarios, from customer retention analysis to equipment maintenance planning.

The estimator calculates survival probabilities at each event time using the number at risk and the number of events. For a censored observation, we know the subject remained event-free until its recorded duration but do not know the subsequent event time. The method adjusts the risk set to retain that partial information.

When to Use the Kaplan-Meier Estimator

The Kaplan-Meier estimator excels in scenarios where you need to understand time-to-event patterns and some observations haven't experienced the event by the end of your study period. Here are specific situations where this technique delivers actionable insights:

Customer Analytics

• Churn analysis: Estimate the probability customers remain active over time, accounting for those still subscribed at analysis time (see the simulated worked example below)
• Time to first purchase: Analyze how long it takes prospects to convert, even when some haven't converted yet
• Retention cohorts: Compare survival curves across different customer segments or acquisition channels
• Subscription renewals: Model renewal timing and identify critical drop-off periods

Product and Operations

• Equipment reliability: Estimate failure rates and plan maintenance schedules based on survival probabilities
• Product lifespan: Analyze warranty claims and product durability when not all items have failed
• Time to defect: Study manufacturing quality by tracking when defects emerge
• Feature adoption: Measure how long it takes users to adopt new features after release

Financial Services

• Loan default analysis: Model time until default while accounting for loans still performing
• Employee retention: Estimate turnover probabilities and identify high-risk tenure periods
• Investment holding periods: Analyze how long investors hold positions before selling

Use the Kaplan-Meier estimator when your primary goal is to estimate and visualize the survival function itself. If you need to model the effect of multiple covariates on survival, consider Cox proportional hazards regression as a next step after establishing baseline survival patterns.

Worked Example (Simulated): Two Customer Groups

The numbers below come from simulated data, so they show how the method behaves, not how your customers will behave. Two groups of 200 customers each are generated. Each group has a constant monthly hazard rate, meaning the rate at which subscribers cancel: 0.10 per month (about 10% a month) for group A and 0.06 per month (about 6%) for group B (random seed 42). The study ends at month 24, so anyone still subscribed then is censored, meaning we know they lasted at least that long but not when they will cancel. Because the true rates are known, the estimates can be checked against the truth: the true 12-month survival is exp(-1.2) = 0.30 for A and exp(-0.72) = 0.49 for B.

The Script

Everything is computed with numpy and scipy: the Kaplan-Meier product, the Greenwood standard error from Step 4, and the log-rank statistic.

import numpy as np
from scipy.stats import chi2

rng = np.random.default_rng(42)
n = 200

# Simulated: monthly churn rate 0.10 (group A)
# and 0.06 (group B); the study ends at month 24
def simulate(rate):
    t = rng.exponential(1 / rate, n)
    event = (t <= 24).astype(int)
    return np.minimum(t, 24), event

tA, eA = simulate(0.10)
tB, eB = simulate(0.06)

def km(t, e):
    times = np.unique(t[e == 1])
    S, var_sum, rows = 1.0, 0.0, []
    for u in times:
        n_i = (t >= u).sum()
        d_i = ((t == u) & (e == 1)).sum()
        S *= (n_i - d_i) / n_i
        if n_i > d_i:
            var_sum += d_i / (n_i * (n_i - d_i))
        # Boundary S=0 needs a separate interval method; do not report zero SE as certainty.
        se = S * np.sqrt(var_sum) if n_i > d_i else float("nan")
        rows.append((u, n_i, d_i, S, se))
    return np.array(rows).reshape(-1, 5)

def at(rows, month):
    r = rows[rows[:, 0] <= month]
    if len(r) == 0:
        return 1.0, 0.0
    return r[-1, 3], r[-1, 4]

def median(rows):
    r = rows[rows[:, 3] <= 0.5]
    return r[0, 0] if len(r) else float("nan")

def logrank(t1, e1, t2, e2):
    both = np.concatenate([t1[e1 == 1], t2[e2 == 1]])
    O1 = E1 = V = 0.0
    for u in np.unique(both):
        n1, n2 = (t1 >= u).sum(), (t2 >= u).sum()
        d1 = ((t1 == u) & (e1 == 1)).sum()
        d2 = ((t2 == u) & (e2 == 1)).sum()
        n, d = n1 + n2, d1 + d2
        O1 += d1
        E1 += d * n1 / n
        if n > 1:
            share = n1 / n
            V += d * share * (1 - share) * (n - d) / (n - 1)
    if V <= 0:
        return O1, E1, float("nan"), float("nan")
    stat = (O1 - E1) ** 2 / V
    return O1, E1, stat, chi2.sf(stat, 1)

kA, kB = km(tA, eA), km(tB, eB)
for name, k, e in (("A", kA, eA), ("B", kB, eB)):
    s12, se12 = at(k, 12)
    lo, hi = s12 - 1.96 * se12, s12 + 1.96 * se12
    print(name, "events", int(e.sum()),
          "censored", int(n - e.sum()),
          "median", round(median(k), 2))
    print("  S(12)", round(s12, 3),
          "95% CI", round(lo, 3), round(hi, 3))
O1, E1, stat, p = logrank(tA, eA, tB, eB)
print("logrank observed A", int(O1),
      "expected A", round(E1, 1),
      "chi2", round(stat, 2), "p", p)

What It Returns

Group	Churned	Censored	Median survival	12-month survival (95% interval)
A (rate 0.10)	185	15	6.8 months	0.300 (0.236 to 0.364)
B (rate 0.06)	154	46	11.4 months	0.485 (0.416 to 0.554)

The log-rank test finds 185 churn events in group A against 139.6 expected if the two groups shared one curve, giving a chi-square of 25.6 on one degree of freedom and p below 0.001. Both 12-month intervals contain the true values (0.30 and 0.49), and the median estimates sit near the true medians of 6.9 and 11.6 months.

What the Numbers Do and Do Not Say

• The groups differ, and by how much: the 12-month intervals do not overlap, and the gap is about 18.5 points of survival. A log-rank p-value alone would not tell you that.
• Censoring is doing real work: group B has 46 customers still subscribed at month 24. Dropping them, or counting them as churned, would bias the curve.
• The step from the curve to a decision is yours: a gap of this size is only worth acting on if the cost of the groups' acquisition and the revenue per month make it matter. The estimator does not know either.

Sources

• Kaplan E, Meier P (1958). Nonparametric Estimation from Incomplete Observations. Journal of the American Statistical Association. doi:10.1080/01621459.1958.10501452. Covers: Kaplan-Meier.

How the Kaplan-Meier Estimator Works

The Kaplan-Meier estimator calculates survival probability as the product of conditional survival probabilities at each event time. This multiplicative approach allows the method to handle censored observations gracefully while building a complete survival curve.

The Mathematical Foundation

At each distinct event time t, the estimator calculates:

S(t) = ∏(1 - d_i/n_i)

Where:

• S(t) is the survival probability at time t
• d_i is the number of events at time i
• n_i is the number of subjects at risk just before time i
• The product is taken over all event times up to and including t

The "at risk" count n_i includes all subjects who have neither experienced the event nor been censored before time i. Risk sets shrink as events and censoring occur. This reduces late follow-up information, but absolute confidence-interval width also depends on the survival estimate and need not increase monotonically.

Handling Censored Observations

Censored observations contribute to the risk set up until their censoring time, then are removed from subsequent calculations. This assumes that censoring is non-informative: subjects who are censored have the same underlying risk as those who remain in the study. This assumption is critical and should be checked in your context.

For example, a cancellation is the event, not censoring. But if customers whose accounts the company closes for non-payment, or who move to another plan, are recorded as censored, those customers are probably at higher risk than the ones who remain, so the censoring is informative and violates the assumption. If the study end date is the censoring mechanism, this is typically non-informative.

Step-by-Step Implementation Process

Applying the Kaplan-Meier estimator effectively requires careful attention to data preparation, execution, and validation. Follow these actionable steps to ensure reliable results.

Step 1: Prepare Your Data Structure

Your dataset needs exactly two pieces of information for each subject:

1. Time: Duration from entry to event or censoring (days, months, etc.)
2. Event status: Binary indicator (1 = event occurred, 0 = censored)

Validate your data quality before proceeding:

• Investigate negative durations against the time origin. Zero durations may reflect events at entry or rounding; check their meaning rather than automatically discarding them.
• Verify that censoring times reflect actual information cutoff, not arbitrary assignment
• Ensure time units are consistent across all observations
• Identify and document any tied event times (multiple events at the same time)

Example data structure:

customer_id    time_months    churned
1              12             1
2              8              0
3              24             1
4              6              0
5              18             1

Step 2: Calculate Risk Sets and Event Counts

For each unique event time, determine:

• Number of subjects at risk (survived and not censored before this time)
• Number of events at this exact time
• Number of censored observations at this time

Handle ties (events and censoring at the same time) by the standard convention: count events first, then remove censored observations, so a customer censored at the same time as an event is treated as still at risk when it happens.

Step 3: Compute Survival Probabilities

Start with survival 1 before any event, including events recorded at time zero. Multiply by the conditional survival probability at each event time:

1. Calculate conditional survival: (n_i - d_i) / n_i
2. Multiply by the previous survival probability
3. Carry forward survival probability until the next event

The survival curve remains flat between event times. It only drops when events occur. This creates the characteristic step function appearance.

Step 4: Calculate Confidence Intervals

Confidence intervals quantify estimation uncertainty and are essential for decision-making. The Greenwood formula is the standard approach for calculating variance:

Var[S(t)] = S(t)² × ∑[d_i / (n_i × (n_i - d_i))]

A pointwise 95% normal-approximation interval, suitable only away from problematic boundaries, is:

S(t) ± 1.96 × √Var[S(t)]

This interval can extend outside 0 to 1. For interior estimates, a log-log transformation can keep bounds within the probability range. At S(t)=0 or 1, use the endpoint handling provided by the chosen method; clipping a normal interval does not repair its coverage. If all remaining subjects experience the event at once (n_i=d_i), the displayed Greenwood sum divides by zero: do not evaluate it directly or interpret a degenerate estimate as certainty.

Absolute interval width need not increase with time; it depends on both the risk set and survival estimate. When few subjects remain, report that limited information and avoid strong tail conclusions.

Step 5: Create Actionable Visualizations

The survival curve is your primary communication tool. Effective visualizations include:

• Step function: Shows exact survival probabilities with clear drops at event times
• Confidence bands: Shaded regions showing estimation uncertainty
• Censoring marks: Small vertical ticks indicating censored observations
• Risk table: Shows number at risk at key time points below the x-axis
• Median survival line: Horizontal line at 50% survival probability

When comparing groups, plot multiple curves on the same axes with clear color coding. Add a log-rank test p-value to quantify whether differences are statistically significant.

Interpreting Kaplan-Meier Results for Business Decisions

Raw survival curves contain valuable information, but extracting actionable next steps requires systematic interpretation focused on business context.

Key Metrics to Extract

1. Median Survival Time

The estimated median is the first time the survival estimate is at or below 0.5, including an exact value of 0.5. If it stays above 0.5 throughout follow-up, report the median as not reached rather than extrapolating.

Action step: Compare median survival times across segments to prioritize retention efforts. For example, a segment with a median of 6 months against 18 months elsewhere is worth investigating first, provided the intervals do not overlap and the segment is valuable enough for the gap to matter.

2. Survival Probability at Key Milestones

Read survival probability at business-relevant time points: end of 14-day window, annual renewal, warranty expiration, etc.

Action step: If the curve drops sharply around day 30 (for example, right after a free period ends), design interventions for the days just before that drop. If 80% of customers survive to 90 days but only 50% to 180 days, investigate what changes between months 3-6.

3. Hazard Rate Changes

When many subjects are still at risk, a steep drop means the hazard in that stretch is high. Flat regions mean few events. Late in follow-up, a drop of the same size can come from one or two events among a few remaining subjects, so check the number at risk before reading a late step as a risk phase.

Action step: Deploy targeted interventions just before high-risk periods. If equipment failures spike at 24 months, schedule preventive maintenance at 22 months.

Comparing Groups with Statistical Rigor

When comparing survival curves between groups (treatment vs. control, segment A vs. B), don't rely on visual assessment alone. Use the log-rank test to determine if differences are statistically significant:

• Null hypothesis: Survival curves are identical across groups
• Interpretation: P-value < 0.05 suggests statistically significant difference
• Limitation: Proportional hazards is not required for null validity of the log-rank test. It influences power under alternatives: crossing or otherwise nonproportional hazards can reduce power. A hazard ratio compares event rates, not cumulative risks.

Action step: If groups show statistically significant differences, investigate what drives the divergence. Is it an intervention effect? Underlying population differences? Random variation? Translate statistical significance into practical magnitude: a 5% difference in 12-month survival may be statistically significant but not operationally meaningful.

Best Practices for Reliable Analysis

Following these best practices ensures your Kaplan-Meier analysis produces trustworthy, actionable results.

Data Quality Standards

• Information available: Report group sizes, observed events and numbers at risk. No universal minimum guarantees precision for every survival question.
• Follow-up duration: Ensure sufficient follow-up time to observe meaningful events. Analyzing customer churn with only 3 months of data will miss long-term patterns.
• Censoring check: Document why observations are censored and assess whether that reason could be related to risk. Interview stakeholders to understand it; the data alone cannot confirm that censoring is non-informative.
• Time zero definition: Clearly define the starting point. For customer analysis, is it signup date? First purchase? End of the free period? Inconsistent definitions invalidate comparisons.

Statistical Considerations

• Report confidence intervals: Always show uncertainty bands. Decision-makers need to know if a 60% survival estimate might actually be 45-75%.
• Address small risk sets: When only a handful of subjects remain at risk, note this limitation explicitly. Don't over-interpret tail behavior.
• Handle ties appropriately: For Kaplan-Meier, use the risk set just before the event time, counting tied events before removing censoring at that time. Breslow and Efron approximations concern Cox regression, not this product-limit update.
• Choose the comparison: Inspect possible nonproportional hazards and prespecify a comparison suited to the question. Crossing curves can reduce log-rank power; a proportional-hazards model has additional assumptions.

Presentation and Communication

• Show risk tables: Include number at risk at regular intervals below your survival curve. This builds trust and highlights when estimates become uncertain.
• Mark censored observations: Visual tick marks show stakeholders how much data is incomplete versus complete.
• Translate probabilities: Instead of "72% survival at 12 months," say "we expect 28 out of 100 customers to churn within the first year."
• Connect to business metrics: Link survival probabilities to revenue, cost, or strategic goals. For example, work out what a 10-point gain in 12-month survival would be worth in annual recurring revenue at your own price and customer count.

Common Pitfalls and How to Avoid Them

Pitfall 1: Informative Censoring

Problem: Censored subjects have different risk profiles than uncensored subjects, biasing survival estimates.

Example: Analyzing equipment failure where machines are removed from service (censored) when they show early warning signs. These machines would have failed sooner than those running normally.

Solution: Investigate censoring mechanisms. If censoring is informative, results are biased. Consider sensitivity analyses or competing risks models.

Pitfall 2: Ignoring Confidence Intervals

Problem: Focusing only on point estimates without acknowledging uncertainty.

Example: Reporting median survival of 14 months when the 95% CI spans 8-22 months. Decisions based on 14 months may be inappropriate.

Solution: Always visualize and report confidence intervals. Make decisions based on the lower bound for conservative planning.

Pitfall 3: Small Sample Extrapolation

Problem: Trusting survival estimates when very few subjects remain at risk.

Example: Using survival probability at 36 months when only 3 subjects remain at risk out of an initial 200.

Solution: Before reading the curves, choose the smallest number at risk you will report from, state it, and stop reporting (or clearly flag) estimates past that point. There is no universal cut-off: the right one depends on how wide an interval your decision can tolerate, which the Greenwood interval at each time shows.

Pitfall 4: Comparing Without Statistical Tests

Problem: Concluding groups differ based on visual curve separation alone.

Example: Two curves appear different, but random variation could explain the gap.

Solution: Use log-rank or other formal tests. Report p-values and effect sizes. Visual differences without statistical significance should be interpreted cautiously.

Pitfall 5: Competing Risks Ignorance

Problem: Multiple mutually exclusive events can occur, but only one is analyzed.

Example: Analyzing customer churn (cancellation) without accounting for upgrades to enterprise contracts. Customers who upgrade can't churn via cancellation, but standard Kaplan-Meier treats them as censored.

Solution: Use competing risks analysis (cumulative incidence functions) when multiple end states exist. This gives a more accurate picture of event-specific probabilities.

Related Survival Analysis Techniques

The Kaplan-Meier estimator is a foundational tool, but related methods extend its capabilities for different analytical needs.

Nelson-Aalen Estimator

An alternative non-parametric estimator that directly estimates the cumulative hazard function rather than the survival function. The Nelson-Aalen estimator is the natural choice when the cumulative hazard, rather than the survival probability, is the quantity you want to report.

Use when: Your focus is on hazard rates (risk at specific times) rather than overall survival probabilities.

Cox Proportional Hazards Regression

Cox regression extends survival analysis by modeling the effect of multiple covariates on survival while accounting for censoring. It produces hazard ratios showing how covariates influence event risk.

Use when: You need to understand which factors affect survival (customer demographics, product features, etc.) or want to predict survival for new subjects based on their characteristics.

Parametric Survival Models

Methods like exponential, Weibull, or log-normal models assume survival times follow a specific distribution. These enable extrapolation beyond observed data but rely on distributional assumptions.

Use when: You have strong theoretical reasons to believe a particular distribution fits your data, or you need to extrapolate survival probabilities beyond your observation window.

Log-Rank and Other Comparison Tests

Statistical tests for comparing survival curves across groups. The log-rank test is most common, while Wilcoxon and Tarone-Ware tests weight early and late differences differently.

Use when: You need formal hypothesis testing to determine if survival differs across groups, beyond visual assessment.

Implementing Your Next Steps

Moving from theory to practice requires a structured implementation plan. Here's your step-by-step methodology to apply the Kaplan-Meier estimator in your organization.

Phase 1: Define Your Business Question (Week 1)

1. Identify the event of interest: What specific outcome are you analyzing?
2. Define the time origin: When does the clock start for each subject?
3. Specify the study population: Who is included and excluded?
4. Determine grouping variables: What segments will you compare?
5. Establish success criteria: What findings would change decisions?

Phase 2: Prepare Your Data (Week 2)

1. Extract time-to-event data from your data warehouse or CRM
2. Create the event indicator (1/0) based on your outcome definition
3. Review data quality: investigate negative durations, rounding, entry-time events and unusual observations
4. Document censoring mechanisms and assess if they're non-informative
5. Create grouping variables (acquisition channel, product tier, etc.)
6. Calculate descriptive statistics: total subjects, total events, median follow-up time

Phase 3: Execute Analysis (Week 3)

1. Run Kaplan-Meier estimator overall and by group
2. Calculate confidence intervals using Greenwood formula
3. Perform log-rank tests for group comparisons
4. Generate survival curves with risk tables and censoring marks
5. Extract key metrics: median survival, milestone probabilities
6. Check estimator and comparison assumptions, including censoring and observation independence; proportional hazards is an additional issue for models or power, not a requirement for Kaplan-Meier estimation

Phase 4: Interpret and Communicate (Week 4)

1. Identify high-risk time periods where survival drops sharply
2. Quantify differences between groups with effect sizes
3. Translate statistical findings to business language
4. Create executive summary with visualizations and key takeaways
5. Present confidence intervals and limitations prominently

Phase 5: Design Interventions (Week 5-6)

1. Map high-risk periods to potential causes
2. Design targeted interventions for critical time windows
3. Estimate expected impact: if intervention reduces risk by X%, what's the business value?
4. Create implementation plan with owners and timelines
5. Establish monitoring metrics to track intervention effectiveness

Phase 6: Monitor and Iterate (Ongoing)

1. Re-run analysis quarterly or after major interventions
2. Compare new cohorts to baseline survival curves
3. Track whether high-risk periods improve over time
4. Refine interventions based on what moves the curve against a holdout group
5. Expand analysis to new segments or deeper stratification

Conclusion: From Data to Decisions

The Kaplan-Meier estimator turns incomplete time-to-event data into survival estimates you can use in business decisions. Its non-parametric flexibility, ability to handle censored observations, and intuitive visual outputs make it a common starting point for survival analysis.

The key to extracting value lies not in the mathematics, but in the systematic application: careful data preparation, rigorous statistical execution, thoughtful interpretation, and translation to actionable next steps. By following the step-by-step methodology outlined in this guide, you can move from raw data to targeted interventions, and then test whether they change survival.

Remember that the Kaplan-Meier estimator is a tool for exploration and communication. When you need to control for multiple factors simultaneously or predict individual-level risk, complement it with Cox proportional hazards regression or other advanced techniques. When analyzing competing events, consider cumulative incidence functions. When you need to extrapolate beyond your observation period, parametric models may be more appropriate.

Start simple: apply the Kaplan-Meier estimator to your most pressing time-to-event question. Visualize the survival curve. Identify the inflection points. Quantify the differences across segments. Then design interventions specifically for the high-risk periods your analysis revealed. Re-run the analysis quarterly to see whether the curve moved, and compare against a holdout group to tell whether the interventions, rather than other changes, moved it.

Your next step is clear: gather your time-to-event data, structure it with event indicators and time variables, and generate your first Kaplan-Meier curve.

Your turn

Bring your own data and the question you actually need answered.

MCP AnalyticsData science by email Send us your data and your question, and we will reply. ds@mcpanalytics.ai

Frequently Asked Questions

What is the Kaplan-Meier estimator and when should I use it?

The Kaplan-Meier estimator is a non-parametric method for estimating survival probabilities over time when some observations are censored. Use it when you need to analyze time-to-event data where not all subjects have experienced the event by the end of the study, such as customer churn, equipment failure, or patient survival. It's ideal when you want to understand how survival probability changes over time without making assumptions about the underlying distribution.

How do I interpret a Kaplan-Meier survival curve?

A Kaplan-Meier survival curve shows the probability of survival (not experiencing the event) on the y-axis against time on the x-axis. The curve steps down at each event time, with the size of each drop proportional to the number of events. Censored observations are typically marked with tick marks. A steeper decline indicates higher event rates, while a flatter curve suggests better survival. Confidence intervals help assess uncertainty. Later risk sets contain less information, but absolute interval widths need not increase monotonically.

What are the key assumptions of the Kaplan-Meier estimator?

The Kaplan-Meier estimator makes three key assumptions: (1) Censoring is independent of the event: subjects who are censored have the same risk as those who remain in the study; (2) Survival probabilities are the same for subjects recruited early and late in the study; (3) Events happen at the times specified. Violations of these assumptions, especially non-independent censoring, can bias your results. Always check how censoring arises and whether it could be related to event risk; the data alone cannot confirm that it is not.

How is the Kaplan-Meier estimator different from other survival analysis methods?

The Kaplan-Meier estimator is non-parametric and makes no assumptions about the distribution of survival times, making it flexible and widely applicable. In contrast, parametric methods like exponential or Weibull models assume a specific distribution. Cox proportional hazards regression allows you to model the effect of covariates on survival, while Kaplan-Meier focuses on estimating the overall survival curve. For simple survival curve estimation without covariates, Kaplan-Meier is the standard choice. When you need to understand the impact of multiple factors, consider Cox regression.

What are common pitfalls when applying the Kaplan-Meier estimator?

Common pitfalls include: (1) Informative censoring, when censored subjects have different risk profiles than those who remain, biasing results; (2) Small sample sizes at later time points, leading to wide confidence intervals and unreliable estimates; (3) Ignoring competing risks when multiple types of events can occur; (4) Comparing groups without proper statistical tests like the log-rank test; (5) Misinterpreting the median survival time when the survival curve doesn't reach 50%. Always check sample sizes over time, visualize confidence intervals, and check censoring assumptions.