What Is the Poisson Model and Kaplan-Meier Survival Analysis, and How Do You Use Them Together?

When you need to analyze how often events occur over time—or how long subjects survive without an event—two classic statistical tools come into play: the Poisson model for event counts and the Kaplan-Meier (KM) survival estimate for time-to-event data. This article explains what each one is, how they work, and how they complement each other in quality and reliability engineering.

What It Is

The Poisson model is a probability distribution used to describe the number of events occurring in a fixed interval of time or space, assuming events happen independently and at a constant average rate. It is widely used for defect counts, failure frequencies, or customer complaints per unit time.

The Kaplan-Meier survival estimate (also called the product-limit estimator) is a non-parametric method for estimating the survival function from lifetime data. It calculates the probability that an individual survives beyond a given time, even when some data are censored (e.g., a unit has not failed by the end of the study).

Together, they give you two complementary views: the Poisson model answers "how many events per unit time?", while Kaplan-Meier answers "what proportion of items survive past time t?"

How It Works / Formula or Steps

### Poisson Model

The Poisson probability of observing exactly k events is:

\[
P(k) = \frac{e^{-\lambda} \lambda^k}{k!}
\]

where:

• \(\lambda\) = average rate of events per interval (e.g., failures per month)
• k = number of events observed
• e ≈ 2.71828 (Euler's number)

The mean and variance of a Poisson distribution are both equal to \(\lambda\). A common quality application is the Poisson control chart (c-chart or u-chart) for monitoring defect counts.

### Kaplan-Meier Survival Estimate

Given ordered distinct event times \(t_1 < t_2 < \dots < t_m\), the KM estimator at time t is:

\[
\hat{S}(t) = \prod_{t_i \le t} \left(1 - \frac{d_i}{n_i}\right)
\]

where:

• \(d_i\) = number of events (failures) at time \(t_i\)
• \(n_i\) = number of subjects at risk just before time \(t_i\)
• The product is taken over all event times up to t

Steps to compute KM:

1. Sort all observed event times (ignore censored times for the product, but keep subjects at risk).
2. At each event time, compute the proportion surviving that instant: \(1 - d_i/n_i\).
3. Multiply these proportions cumulatively from the earliest time onward.
4. Censored observations reduce \(n_i\) at the next event time but do not contribute to \(d_i\).

A Worked Illustrative Example

Example data (illustrative only). Suppose you track 10 machines for failure time (in weeks). Three machines are still running at the end of the study (censored at week 10). Failure times: 2, 4, 6, 8, and 10 (one each); censored at 10 for three machines.

Poisson rate: You observe 5 failures over 10 weeks for 10 machines, i.e., 50 machine-weeks of exposure. The average failure rate is \(\lambda = 5/50 = 0.1\) failures per machine-week. If you want the probability of exactly 0 failures in one machine-week: \(P(0) = e^{-0.1} \approx 0.9048\).

Kaplan-Meier estimate:

• At t=2: \(n_1 = 10\), \(d_1 = 1\), survival = \(1 - 1/10 = 0.9\)
• At t=4: \(n_2 = 9\), \(d_2 = 1\), survival = \(0.9 \times (1 - 1/9) = 0.9 \times 0.8889 = 0.8\)
• At t=6: \(n_3 = 8\), \(d_3 = 1\), survival = \(0.8 \times (1 - 1/8) = 0.8 \times 0.875 = 0.7\)
• At t=8: \(n_4 = 7\), \(d_4 = 1\), survival = \(0.7 \times (1 - 1/7) = 0.7 \times 0.8571 = 0.6\)
• At t=10: \(n_5 = 6\) (three censored at 10 are not yet removed before the event), \(d_5 = 1\), survival = \(0.6 \times (1 - 1/6) = 0.6 \times 0.8333 = 0.5\)

So the KM estimate of the probability of surviving past 10 weeks is 0.5 (50%).

Common Pitfalls

• Confusing the two tools: Poisson gives the rate of events, not the probability of survival. Use KM when you have time-to-event data with censoring.
• Ignoring censoring in KM: Removing censored subjects from the denominator too early biases the survival estimate downward.
• Assuming constant rate for Poisson: If the event rate changes over time, the Poisson assumption is violated—consider a piecewise or time-dependent model.
• Small sample sizes: KM estimates become step-like and imprecise with few events; Poisson confidence intervals widen dramatically.

---

To compute Poisson probabilities or Kaplan-Meier survival curves quickly, try the free tool at https://www.6sq.com/tools/poisson_km/—it handles both calculations in one place, saving you time and reducing manual errors.