Stochastic Modeling

Chapter 6
Renewal, Reliability, and Survival Models

Graduate Program in Business Analytics
Stochastic Modeling • Graduate

Renewal, Reliability, and Survival Models

Introduction

Organizations manage systems whose components fail, recover, renew, or leave populations over time. Examples include machines requiring replacement, employees departing, customers churning, software services experiencing outages, and insurance policies terminating after claims or death. These processes are stochastic because the timing of events is uncertain. Renewal theory, reliability analysis, and survival analysis provide related but distinct frameworks for understanding such event histories.

Renewal theory focuses on repeated events and the long-run behavior of cycles. Reliability theory studies the probability that an asset or system performs without failure for a specified period. Survival analysis models the time until an event, while allowing for censoring and covariates. Together, these tools support maintenance planning, warranty pricing, workforce forecasting, customer-lifecycle management, and service-level design.

Learning Objectives

By the end of this chapter, readers should be able to:

Formulate renewal processes from interarrival or lifetime distributions.
Distinguish reliability, survival, hazard, and cumulative hazard functions.
Apply renewal-reward theory to long-run cost, output, and availability.
Analyze preventive and corrective maintenance policies.
Explain series, parallel, and k-out-of-n reliability systems.
Interpret censored survival data and estimate survival quantities.
Evaluate model assumptions, governance risks, and managerial implications.

1. Renewal Processes and Interarrival Times

Let X₁, X₂, ... denote successive cycle lengths, where X_i ≥ 0 is the time between renewal i−1 and renewal i. The renewal epochs are

S_n = X₁ + X₂ + ··· + X_n, n = 1, 2, …

where S_n is the time of the nth renewal. The number of renewals observed by time t is

N(t) = max{n : S_n ≤ t}.

A renewal process generally assumes that interrenewal times are independent and identically distributed (i.i.d.) with distribution function F(x) and finite mean

μ = E[X] = ∫₀^∞ x dF(x).

The i.i.d. assumption means that the process probabilistically restarts after each renewal. It is appropriate for identical replacement cycles, but less appropriate when equipment ages, learning occurs, or environmental conditions change.

The renewal function is

m(t) = E[N(t)].

It satisfies the renewal equation

m(t) = F(t) + ∫₀^t m(t − x) dF(x).

The first term represents a renewal occurring by time t. The integral accounts for the expected renewals after the first renewal at time x. In general, this equation is solved numerically or by exploiting special distributions.

Important special cases

If interarrival times are exponential with rate λ, the process is a Poisson process and

m(t) = λt.

For a general renewal process with finite mean, the elementary renewal theorem states

lim_t→∞ m(t)/t = 1/μ.

Thus, the long-run renewal rate is 1/μ renewals per unit time. This result does not require exponential lifetimes, although it describes asymptotic rather than short-run behavior.

2. Renewal-Reward Theory

Suppose cycle i has length X_i and generates reward R_i. Rewards may represent revenue, production, operating cost, claims, or customer value. Under i.i.d. cycles with finite means and E[X] > 0, the renewal-reward theorem gives

lim_t→∞ Total reward by time t / t = E[R] / E[X]

with probability one under standard regularity conditions. The ratio is a long-run average reward rate, measured in reward units per time unit.

This theorem is powerful because it converts a time-dependent stochastic process into an expected reward per cycle divided by expected cycle length. It is used in replacement analysis, inventory cycles, call-center workloads, and customer lifecycles.

Example: replacement cost

A component is replaced whenever it fails. Its lifetime has mean 800 operating hours. Each replacement costs $1,200, including parts and labor. Ignoring downtime and discounting, the long-run replacement cost rate is

Cost rate = $1,200 / 800 hours = $1.50 per operating hour.

This calculation is valid when replacement cycles are independent and have the same lifetime distribution, and when every replacement restores the component to an equivalent as-good-as-new condition.

If planned preventive replacement occurs at age T, while random failure may occur earlier, the cycle length is min(X,T). If failure costs C_f and planned replacement costs C_p, then one common cost-rate model is

C(T) = [C_f P(X ≤ T) + C_p P(X > T)] / E[min(X,T)].

This expression assumes immediate replacement at failure or at age T, no overlap between cycles, and costs incurred at the renewal epoch. More detailed models may include downtime, inspection, labor, salvage, and discounting.

3. Lifetime, Reliability, and Survival Functions

Let T be a nonnegative lifetime or event time. The cumulative distribution function is

F(t) = P(T ≤ t).

The survival or reliability function is

S(t) = R(t) = P(T > t) = 1 − F(t).

In reliability engineering, R(t) usually means the probability that an item performs its required function without failure through time t. In survival analysis, S(t) is the probability of surviving beyond t. The same mathematical object has different disciplinary language.

If a density exists, it is f(t) = dF(t)/dt, and

S(t) = ∫_t^∞ f(u) du.

The mean lifetime, or mean time to failure (MTTF), is

E[T] = ∫₀^∞ S(t) dt,

provided the integral is finite. The units of E[T] are time.

Hazard function

The hazard rate describes the instantaneous event risk conditional on survival to time t:

h(t) = lim_∆t→0 P(t ≤ T < t + ∆t | T ≥ t) / ∆t.

When a density exists,

h(t) = f(t) / S(t).

The hazard is a rate, not a probability. Its units are inverse time. The cumulative hazard is

H(t) = ∫₀^t h(u) du,

and the survival function can be recovered as

S(t) = exp[−H(t)].

A high hazard at a particular time does not necessarily imply a high unconditional failure probability at that time, because few units may remain at risk.

Common lifetime distributions

Distribution	Survival function	Hazard behavior	Typical interpretation
Exponential	S(t)=e^−λt	Constant λ	Memoryless failures
Weibull	S(t)=e^{−(t/η)^β}	Increasing if β>1, decreasing if β<1	Aging, early failures, or mixed mechanisms
Lognormal	S(t)=1−Φ((ln t−μ)/σ)	Typically nonmonotone	Multiplicative degradation
Gamma	Depends on shape and rate	Flexible	Accumulated stages or shocks

For the Weibull model, β is the shape parameter and η is the scale parameter, both positive. Its hazard is

h(t) = (β/η)(t/η)^β−1.

The Weibull mean is η Γ(1+1/β), where Γ is the gamma function. The exponential distribution is the Weibull special case β=1.

4. Conditional Lifetime and Residual Life

For an item known to have survived to age a, the probability of surviving an additional duration x is

P(T > a+x | T > a) = S(a+x) / S(a).

The expected residual life at age a is

e(a) = E[T−a | T>a] = ∫₀^∞ S(a+x)/S(a) dx.

For an exponential lifetime, the residual-life distribution does not depend on a; this is the memoryless property. For an aging Weibull model with β>1, residual life generally decreases as age increases. This distinction is central to maintenance decisions: replacing an item because it is old is not justified by age alone if the hazard is constant.

5. Availability and Maintainability

Reliability concerns failure-free operation. Availability concerns whether an asset is operational at a specified time or over a period. For alternating operating and repair periods, steady-state availability is

A = MTBF / (MTBF + MTTR),

where MTBF is mean time between failures and MTTR is mean time to repair. This formula assumes a regenerative cycle consisting of one operating period and one repair period, stable averages, and negligible startup complications.

For a repairable system with failure rate λ and repair rate μ_r, both constant, the two-state continuous-time Markov model has steady-state availability

A = μ_r / (λ + μ_r).

Since MTBF=1/λ and MTTR=1/μ_r, the two formulas agree.

Maintainability is the probability that a failed item is restored within time t. If repair time D has distribution function G, then

M(t) = P(D ≤ t) = G(t).

If repair time is exponential with rate μ_r, then M(t)=1−e^{−μ_r t}. Maintainability analysis helps organizations set staffing levels, spare-parts inventories, and service-level commitments.

6. Reliability of Systems

For independent component states, system reliability depends on system structure.

Series systems

A series system works only if every component works. For components with reliabilities R₁(t), ..., R_n(t),

R_series(t) = ∏_i=1ⁿ R_i(t).

If all components have reliability R(t), this becomes R(t)ⁿ. Series systems become less reliable as components are added.

Parallel systems

A parallel system works if at least one component works. Therefore,

R_parallel(t) = 1 − ∏_i=1ⁿ [1−R_i(t)].

This assumes independent failures and that one functioning component is sufficient. Common-cause failures violate this assumption and can make the formula overoptimistic.

k-out-of-n systems

A k-out-of-n system works when at least k of n components work. For identical independent components with reliability R(t),

R_k|n(t) = ∑_j=kⁿ C(n,j) [R(t)]^j [1−R(t)]^n−j.

Here C(n,j) is the binomial coefficient. If component failures are dependent, a joint distribution or fault-tree model is required.

7. Survival Analysis with Observational Data

Survival datasets frequently contain right-censored observations. For a customer study, the event may be churn. If a customer remains active when observation ends, the exact churn time is unknown but exceeds the observed follow-up time. Treating such customers as if they had churned at the observation boundary creates bias.

The Kaplan–Meier estimator is a nonparametric estimate of survival. At each distinct event time t_j, let d_j be the number of events and n_j the number at risk immediately before that time:

Ŝ(t) = ∏_{t_j ≤ t} (1 − d_j/n_j).

Censored observations reduce the later risk sets but do not directly produce a failure factor. The estimator assumes noninformative censoring: conditional on the measured information, censoring is unrelated to the future event process.

A semiparametric alternative is the Cox proportional-hazards model:

h(t|z) = h₀(t) exp(β₁z₁ + ··· + β_pz_p).

Here h₀(t) is an unspecified baseline hazard, z is a vector of covariates, and β contains estimated effects. For a one-unit increase in covariate z_j, the hazard ratio is e^β_j, holding other covariates constant. The proportional-hazards assumption means that hazard ratios do not vary with time unless the model includes time interactions or stratification.

Survival curves and hazard ratios answer different questions. A survival curve estimates the probability of remaining event-free over time; a hazard ratio compares instantaneous event rates conditional on survival. Neither automatically measures causal impact. Causal interpretation requires an appropriate design, confounding control, and defensible assumptions.

8. Applied Case: Predictive Maintenance for a Wind-Turbine Fleet

The following is an illustrative case.

Background

A wind-energy operator manages 240 gearboxes. Historical records contain installation dates, failure dates, repairs, operating hours, temperature, vibration, and load. Corrective replacement costs $85,000, while planned replacement costs $42,000. Downtime costs an estimated $6,000 per day.

Problem

Managers currently replace gearboxes every six years. Some units fail early, while others operate well beyond six years. The fixed policy may replace healthy equipment unnecessarily and still miss high-risk failures.

Analysis

The analytics team first defines the event as gearbox failure and records right-censoring for units still operating at the study end. A Weibull model is fitted to estimate baseline aging. A Cox model adds vibration and load as covariates. Model diagnostics show that vibration has a time-varying effect, so the team includes a vibration-by-age interaction rather than assuming a constant hazard ratio.

The team then compares policies using renewal-reward logic. For each policy, simulation generates lifetimes, inspection intervals, repair delays, planned replacement costs, corrective costs, and downtime. The performance metric is expected total cost per operating day, supplemented by availability and the 95th percentile of annual downtime.

Approach and solution

The recommended policy is condition-based: inspect vibration monthly, flag units whose risk exceeds a specified threshold, and schedule replacement during a low-wind maintenance window. The threshold is not selected solely to maximize predictive accuracy. It incorporates false alarms, missed failures, production losses, worker safety, and available maintenance capacity.

Outcome

In the simulation, the condition-based policy reduces emergency failures but increases planned interventions. Its adoption is justified only if the reduction in downtime and catastrophic risk exceeds the additional inspection and replacement costs. The result also depends on whether vibration sensors remain calibrated and whether failure mechanisms remain stable across seasons.

Learning points

Survival models support risk estimation, but decisions require an economic objective.
Censoring must be represented correctly.
Predictive accuracy does not by itself establish causal mechanisms.
Dependence among components and common environmental shocks can invalidate simple system formulas.
Maintenance policies should be stress-tested under alternative failure and repair assumptions.

9. Customer, Workforce, and Warranty Applications

The same framework applies outside engineering. In customer analytics, T may be time to churn, upgrade, renewal, or complaint. In workforce analytics, it may represent time to resignation, promotion, or absence. In warranty analysis, it represents time to claim, with claims often affected by usage intensity and operating environment.

A warranty provider can estimate expected claims over a coverage horizon using the failure distribution. If each claim costs c and the probability of at least one failure by time w is F(w), a simple expected claim cost for a single non-repairable item is cF(w). This is incomplete when multiple claims, deductibles, repairable products, inflation, and discounting are relevant.

Customer-lifetime value can also be expressed through survival probabilities. If expected net contribution at time t is g(t) and the discount rate is r, a continuous-time expected value representation is

CLV = ∫₀^∞ e^−rt S(t) g(t) dt.

This formulation assumes g(t) is the contribution conditional on still being active and that survival adequately represents continuation. Acquisition costs, competing risks, time-varying behavior, and dependence between purchasing and churn may require extensions.

10. Limitations, Governance, and Emerging Practice

Model limitations

Renewal models can fail when cycles are not identically distributed, when repairs do not restore the system to an as-good-as-new state, or when failures are dependent. Reliability formulas based on independence can be seriously misleading under common-cause failures. Parametric survival models may be sensitive to distributional misspecification, while flexible models may overfit limited event data.

Data quality

Event definitions must be precise. “Failure,” “churn,” and “repair complete” can each have multiple operational meanings. Timestamp errors, delayed reporting, left truncation, interval censoring, and missing covariates affect inference. Analysts should document risk-set construction, censoring rules, exclusion criteria, and time origins.

Ethics and governance

Workforce survival models can create unfair monitoring or penalize employees who exhibit characteristics correlated with protected attributes. Customer churn models may trigger unequal retention offers. Governance should include purpose limitation, transparency, access controls, bias assessment, model monitoring, and human review. Predictions should not be treated as inevitable outcomes or used to deny opportunities without due process.

Technology and AI

Machine-learning survival models, recurrent event models, gradient boosting, and neural survival methods can capture nonlinearities and time-varying effects. However, interpretability, calibration, temporal validation, and dataset shift remain essential. A model that ranks risk well may still produce poorly calibrated probabilities. AI systems should therefore be evaluated with discrimination measures, calibration curves, decision-curve analysis, and operational cost metrics rather than accuracy alone.

Best practice is to combine statistical survival models with domain engineering, causal reasoning, simulation, and controlled implementation. Predictions should be monitored after deployment because maintenance regimes, product design, customer behavior, and economic conditions evolve.

Chapter Summary

Renewal theory models repeated stochastic cycles through interrenewal times, renewal epochs, and renewal counts. Its central long-run result is that the renewal rate equals the reciprocal of the mean cycle length. Renewal-reward theory extends this result to long-run costs, revenues, output, and service measures.

Reliability and survival functions describe the probability of remaining event-free. The hazard function gives the instantaneous conditional event rate, while cumulative hazard links hazard to survival. Exponential, Weibull, lognormal, and gamma distributions offer different assumptions about aging and failure behavior.

Availability combines operating and repair durations, and maintainability describes the probability of completing repair within a specified time. Series, parallel, and k-out-of-n formulas describe system reliability under independence, but common-cause failures require more advanced models. Kaplan–Meier estimation accommodates right censoring, while Cox regression relates covariates to hazard under a proportional-hazards assumption.

Applications span assets, warranties, customers, and employees. Sound decisions require correct event definitions, appropriate censoring treatment, economic evaluation, sensitivity analysis, ethical governance, and continuous monitoring of deployed models.

Key Terms

Renewal process: A stochastic process recording repeated events separated by cycle times.

Interrenewal time: The duration between successive renewals.

Renewal epoch: The time at which a renewal occurs.

Renewal function: The expected number of renewals by a specified time.

Renewal-reward theorem: A result equating long-run reward rate with expected reward per cycle divided by expected cycle length.

Lifetime: Time from origin until failure or another defined event.

Reliability function: Probability that an item performs without failure through time t.

Survival function: Probability that an event time exceeds t.

Hazard rate: Instantaneous event rate conditional on survival to a given time.

Cumulative hazard: Integral of the hazard rate over time.

MTTF: Mean time to failure for a non-repairable item.

MTBF: Mean time between failures for a repairable item.

MTTR: Mean time to repair.

Availability: Probability or proportion of time that a system is operational.

Maintainability: Probability that repair is completed within a specified time.

Right censoring: Knowing that an event has not occurred by the observation end, without knowing its eventual time.

Kaplan–Meier estimator: A nonparametric product-limit estimator of survival.

Cox proportional-hazards model: A semiparametric model relating covariates to hazard through a multiplicative form.

Competing risks: Multiple mutually exclusive event types that can occur first.

Common-cause failure: A shared event or condition causing several components to fail.

Review Questions

Short Answer Questions

What is the difference between a renewal epoch and an interrenewal time?
Define the reliability function and hazard function.
Why is the exponential distribution called memoryless?
What does right censoring mean in survival analysis?
State the assumptions behind the basic MTBF–MTTR availability formula.

Long Answer Questions

Explain the renewal-reward theorem and discuss two business applications.
Compare exponential and Weibull lifetime models, including their hazard behavior.
Derive the reliability of independent series and parallel systems and discuss when the formulas fail.
Explain how the Kaplan–Meier estimator handles censored observations.
Discuss the managerial, ethical, and governance issues in deploying predictive survival models.

Analytical / Application Questions

A component has mean lifetime 1,000 hours and replacement cost $2,500. Under an as-good-as-new replacement assumption, calculate and interpret the long-run replacement cost rate.
A repairable asset has MTBF of 900 hours and MTTR of 100 hours. Calculate steady-state availability and explain one operational method for improving it.
A two-component parallel system has independent component reliabilities of 0.90 and 0.80 over one year. Calculate system reliability and explain how dependence or common-cause failure would alter the conclusion.

Multiple Choice Questions with Answers

Question 1

A renewal process has mean interrenewal time 5 hours. What is its long-run renewal rate?

A. 0.05 renewals per hour

B. 0.20 renewals per hour

C. 5 renewals per hour

D. 25 renewals per hour

Correct Answer: B. 0.20 renewals per hour

Question 2

Which quantity is defined as P(T > t)?

A. Density function

B. Hazard function

C. Survival function

D. Cumulative hazard

Correct Answer: C. Survival function

Question 3

For an exponential lifetime with rate λ, the hazard rate is:

A. Increasing over time

B. Decreasing over time

C. Constant at λ

D. Equal to λt

Correct Answer: C. Constant at λ

Question 4

For independent components in a series system, system reliability is:

A. The sum of component reliabilities

B. The product of component reliabilities

C. One minus the product of reliabilities

D. The maximum component reliability

Correct Answer: B. The product of component reliabilities

Question 5

A repairable asset has MTBF = 900 hours and MTTR = 100 hours. Its steady-state availability is:

A. 0.10

B. 0.50

C. 0.90

D. 0.99

Correct Answer: C. 0.90

Question 6

In right-censored survival data, the analyst knows that:

A. The event occurred at the censoring time

B. The event will never occur

C. The event time exceeds the observed follow-up time

D. The event occurred before observation began

Correct Answer: C. The event time exceeds the observed follow-up time

Question 7

In a Weibull model, β > 1 generally indicates:

A. A decreasing hazard

B. A constant hazard

C. An increasing hazard

D. No possibility of failure

Correct Answer: C. An increasing hazard

Question 8

The Kaplan–Meier estimator is primarily used to estimate:

A. A product price

B. A survival function

C. A repair cost

D. A component count

Correct Answer: B. A survival function

Question 9

For independent parallel components, the system works when:

A. All components fail

B. Exactly half the components fail

C. At least one component works

D. The component with the lowest reliability works

Correct Answer: C. At least one component works

Question 10

In a Cox proportional-hazards model, exp(β_j) represents the:

A. Mean lifetime

B. Hazard ratio for a one-unit covariate increase

C. Cumulative hazard at time zero

D. Probability of censoring

Correct Answer: B. Hazard ratio for a one-unit covariate increase

Stochastic Modeling
Chapter 6 of 15

Chapter 6 Renewal, Reliability, and Survival Models