Hazard Function, Failure Rates, Censoring and Life Testing
CSIR NET statistics: reliability and hazard functions, IFR and DFR classes, Weibull and exponential lives, censoring types, Kaplan-Meier and life-test MLEs.
- 23 min
On this page · 15 sections
- Prerequisites
- The Hook: Ten Bulbs, Four Failures, One Estimate
- Lifetime, Reliability and Hazard
- Standard Lifetime Distributions
- The Bathtub Curve and Ageing Classes
- Censoring
- Life Testing under the Exponential Model
- Random Censoring: Kaplan-Meier and Nelson-Aalen
- Memory Tricks
- Don’t Confuse
- Common Mistakes
- Practice
- Quick Revision Table
- Key Takeaways
- What’s Next
Prerequisites
- Sample Space, Independence and Bayes: the hazard is a conditional probability per unit time, and every survival statement on this page is a statement about conditional probabilities.
- Riemann and Lebesgue Integration: the mean life is the improper integral \( \int_0^\infty R(t)\,dt \), and the cumulative hazard is an integral of the hazard.
- Standard distributions: exponential, gamma, Weibull and the chi-square distribution with even degrees of freedom.
- The chapter map is on Reliability and Operations Research.
The Hook: Ten Bulbs, Four Failures, One Estimate
Lifetime, Reliability and Hazard
Let \( T \ge 0 \) be the lifetime of a unit, with distribution function \( F(t) = P(T \le t) \) and density \( f(t) = F'(t) \).
| Quantity | Definition | Meaning |
|---|---|---|
| Reliability (survival) function | \( R(t) = P(T > t) = 1 - F(t) \) | probability the unit is still working at time \( t \) |
| Hazard (failure rate) function | \( h(t) = \lim\limits_{\Delta \to 0^+} \dfrac{P(t < T \le t + \Delta \mid T > t)}{\Delta} = \dfrac{f(t)}{R(t)} \) | instantaneous failure rate of a unit that has survived to \( t \) |
| Cumulative hazard | \( H(t) = \int_0^t h(u)\,du \) | total “risk” accumulated by time \( t \) |
| Mean time to failure (MTTF) | \( E[T] = \int_0^\infty R(t)\,dt \) | expected life |
| Conditional reliability | \( R(x \mid t) = \dfrac{R(t + x)}{R(t)} \) | probability of surviving a further \( x \) given survival to \( t \) |
| Mean residual life | \( m(t) = E[T - t \mid T > t] = \dfrac{\int_t^\infty R(u)\,du}{R(t)} \) | expected remaining life at age \( t \) |
The one identity that drives everything: since \( h(t) = \frac{f(t)}{R(t)} = -\frac{d}{dt}\ln R(t) \) and \( R(0) = 1 \),
$$\boxed{R(t) = \exp\!\left(-\int_0^t h(u)\,du\right) = e^{-H(t)}, \qquad f(t) = h(t)\,e^{-H(t)}.}$$So any one of \( F, R, f, h, H \) determines all the others. A function \( h \ge 0 \) is the hazard of a proper lifetime distribution exactly when \( \int_0^\infty h(u)\,du = \infty \) (so that \( R(t) \to 0 \)).
MTTF as an integral of \( R \). For \( T \ge 0 \), \( E[T] = \int_0^\infty P(T > t)\,dt \) (integrate \( T = \int_0^\infty \mathbf{1}\{T > t\}\,dt \) and swap expectation and integral). This is usually easier than \( \int t f(t)\,dt \), and it is the formula that makes system MTTFs on the next page one-line computations.
Average failure rate and life-table data
Over an interval \( [t_1, t_2] \) the average failure rate is
$$\bar h(t_1, t_2) = \frac{H(t_2) - H(t_1)}{t_2 - t_1} = \frac{\ln R(t_1) - \ln R(t_2)}{t_2 - t_1}.$$If \( R(t_1) = 0.9 \) and \( R(t_2) = 0.8 \) with \( t_2 - t_1 = 100 \) hours, \( \bar h = \frac{\ln(0.9/0.8)}{100} \approx 1.18 \times 10^{-3} \) per hour.
From grouped test data, the empirical hazard in an interval is estimated by
$$\hat h = \frac{\text{number failing in the interval}}{(\text{number surviving at the start}) \times (\text{interval length})}.$$In discrete time (lifetime in cycles \( 1, 2, 3, \dots \)) the hazard is \( h(k) = P(T = k \mid T \ge k) \), a genuine probability; it is constant exactly for the geometric distribution, the discrete analogue of the exponential.
Standard Lifetime Distributions
| Distribution | \( R(t) \) | \( h(t) \) | Shape of hazard | Mean |
|---|---|---|---|---|
| Exponential, mean \( \theta \) | \( e^{-t/\theta} \) | \( \frac{1}{\theta} = \lambda \) | constant | \( \theta \) |
| Weibull, shape \( \beta \), scale \( \eta \) | \( e^{-(t/\eta)^\beta} \) | \( \frac{\beta}{\eta}\left(\frac{t}{\eta}\right)^{\beta - 1} \) | decreasing if \( \beta < 1 \), constant if \( \beta = 1 \), increasing if \( \beta > 1 \) | \( \eta\,\Gamma\!\left(1 + \frac{1}{\beta}\right) \) |
| Rayleigh (Weibull, \( \beta = 2 \)) | \( e^{-ct^2/2} \) | \( ct \) | linear increasing | \( \sqrt{\frac{\pi}{2c}} \) |
| Gamma, shape \( k \), rate \( \lambda \) | (no closed form unless \( k \) is an integer) | tends to \( \lambda \) as \( t \to \infty \) | increasing if \( k > 1 \), decreasing if \( k < 1 \) | \( \frac{k}{\lambda} \) |
| Erlang (gamma, \( k = 2 \)), \( \lambda = 1 \) | \( (1 + t)e^{-t} \) | \( \frac{t}{1 + t} \) | increasing to 1 | 2 |
| Uniform on \( (0, b) \) | \( 1 - \frac{t}{b} \) | \( \frac{1}{b - t} \) | increasing, unbounded | \( \frac{b}{2} \) |
| Lomax (Pareto II) | \( \left(1 + \frac{t}{\sigma}\right)^{-\alpha} \) | \( \frac{\alpha}{\sigma + t} \) | decreasing | \( \frac{\sigma}{\alpha - 1} \) (\( \alpha > 1 \)) |
| Gompertz | \( \exp\!\left(-\frac{a}{b}(e^{bt} - 1)\right) \) | \( a e^{bt} \) | exponentially increasing | (no simple form) |
| Lognormal | \( 1 - \Phi\!\left(\frac{\ln t - \mu}{\sigma}\right) \) | rises then falls | non-monotone | \( e^{\mu + \sigma^2/2} \) |
The exponential is the only continuous lifetime with constant hazard, and this is equivalent to the memoryless property
$$P(T > s + t \mid T > s) = P(T > t) \quad \text{for all } s, t \ge 0.$$A used exponential unit is as good as new: its conditional reliability and its mean residual life (\( m(t) = \theta \)) do not depend on its age.
The Bathtub Curve and Ageing Classes
Many populations of manufactured units show a hazard of three phases:
graph LR
A["Infant mortality
decreasing hazard (DFR)
manufacturing defects"] --> B["Useful life
roughly constant hazard
random shocks"]
B --> C["Wear-out
increasing hazard (IFR)
fatigue, corrosion"]Burn-in (running units for a while before release) is worthwhile only in a DFR phase: it removes the weak units and lowers the hazard of the survivors. For an IFR unit, burn-in only uses up life.
IFR, DFR and their weaker cousins
| Class | Definition | Equivalent / intuition |
|---|---|---|
| IFR (increasing failure rate) | \( h(t) \) non-decreasing | \( \ln R \) concave; \( R(x \mid t) \) non-increasing in \( t \) for each \( x \) |
| DFR (decreasing failure rate) | \( h(t) \) non-increasing | \( \ln R \) convex |
| IFRA (IFR on average) | \( \frac{H(t)}{t} \) non-decreasing | \( R(t)^{1/t} \) non-increasing |
| NBU (new better than used) | \( R(s + t) \le R(s) R(t) \) | a used unit survives a further \( t \) with probability at most that of a new one |
| NBUE (new better than used in expectation) | \( m(t) \le m(0) = \mu \) | expected remaining life never exceeds the new-unit mean |
| DMRL (decreasing mean residual life) | \( m(t) \) non-increasing |
The implications run one way:
$$\text{IFR} \Rightarrow \text{IFRA} \Rightarrow \text{NBU} \Rightarrow \text{NBUE}, \qquad \text{IFR} \Rightarrow \text{DMRL} \Rightarrow \text{NBUE}.$$The exponential is the boundary case: it is both IFR and DFR (in the weak, “non-decreasing / non-increasing” sense), and the same for every class above. The dual classes (DFR, DFRA, NWU, NWUE, IMRL) reverse every inequality.
Censoring
A life test rarely runs until every unit fails. An observation is censored when we know only a bound on the lifetime.
| Type | What happens | What is random |
|---|---|---|
| Type I (time censoring) | \( n \) units on test; stop at a fixed time \( t_0 \) | the number of failures \( r \) |
| Type II (failure censoring) | \( n \) units on test; stop at the \( r \)-th failure, \( r \) fixed in advance | the duration \( t_{(r)} \) |
| Random (right) censoring | unit \( i \) has a censoring time \( C_i \) independent of \( T_i \); we see \( Y_i = \min(T_i, C_i) \) and \( \delta_i = \mathbf{1}\{T_i \le C_i\} \) | both |
| Left censoring | we know only that \( T_i \le c_i \) (failed before the first inspection) | |
| Interval censoring | we know only \( a_i < T_i \le b_i \) (failed between two inspections) |
Type I and Type II are both right censoring, the kind CSIR NET asks about. A test may also be run with replacement (each failed unit is immediately replaced by a new one, so \( n \) units are always on test) or without replacement.
The censored likelihood
With non-informative right censoring, a failure at \( t_i \) contributes \( f(t_i) \) and a unit censored at \( t_i \) contributes \( R(t_i) \):
$$L = \prod_{i=1}^n f(t_i)^{\delta_i}\, R(t_i)^{1 - \delta_i} = \prod_{i=1}^n h(t_i)^{\delta_i}\, R(t_i).$$For the exponential with mean \( \theta \) this collapses to
$$L(\theta) = \theta^{-r} \exp\!\left(-\frac{T}{\theta}\right), \qquad r = \sum \delta_i, \quad T = \sum_{i=1}^n t_i,$$where \( T \) is the total time on test (TTT): the sum of all observed times, failed or censored. Maximising,
$$\boxed{\hat\theta = \frac{T}{r} = \frac{\text{total time on test}}{\text{number of failures}}, \qquad \hat\lambda = \frac{r}{T}.}$$This one formula covers Type I, Type II and random censoring; only the way \( T \) is computed changes. If \( r = 0 \), the likelihood increases without bound in \( \theta \) and there is no finite MLE.
Life Testing under the Exponential Model
Let the ordered failure times be \( t_{(1)} \le t_{(2)} \le \dots \).
| Test plan | Total time on test \( T \) | Exact distribution |
|---|---|---|
| Type II, without replacement | \( \sum_{i=1}^r t_{(i)} + (n - r)\,t_{(r)} \) | \( \frac{2T}{\theta} \sim \chi^2_{2r} \) |
| Type II, with replacement | \( n\,t_{(r)} \) | \( \frac{2T}{\theta} \sim \chi^2_{2r} \) |
| Type I, without replacement | \( \sum_{i=1}^r t_{(i)} + (n - r)\,t_0 \) | \( r \) random; no simple exact law |
| Type I, with replacement | \( n\,t_0 \) | number of failures \( r \sim \text{Poisson}\!\left(\frac{n t_0}{\theta}\right) \) |
Why \( \chi^2_{2r} \) for Type II. For an exponential sample, the normalised spacings
$$D_i = (n - i + 1)\left(t_{(i)} - t_{(i-1)}\right), \qquad t_{(0)} = 0,$$are independent exponential variables with mean \( \theta \) (by memorylessness: after the \( (i-1) \)-th failure, \( n - i + 1 \) fresh-as-new units race, and the minimum of them is exponential with mean \( \frac{\theta}{n - i + 1} \)). A telescoping sum gives \( T = D_1 + \dots + D_r \), so \( T \sim \text{Gamma}(r, \theta) \) and \( \frac{2T}{\theta} \sim \chi^2_{2r} \). With replacement, failures form a Poisson process of rate \( \frac{n}{\theta} \), and \( n t_{(r)} \) is again a sum of \( r \) independent exponentials with mean \( \theta \).
Consequences for Type II (either plan):
- \( \hat\theta = \frac{T}{r} \) is unbiased, with \( \text{Var}(\hat\theta) = \frac{\theta^2}{r} \). It is also the UMVUE. The precision depends on \( r \), the number of failures, not on \( n \).
- A \( 100(1 - \alpha)\% \) two-sided confidence interval for \( \theta \) is
where \( \chi^2_{\nu, q} \) is the \( q \)-quantile.
- Expected test duration. Without replacement, \( E[t_{(r)}] = \theta \sum_{i=1}^{r} \frac{1}{n - i + 1} \); with replacement, \( E[t_{(r)}] = \frac{r\theta}{n} \).
Random Censoring: Kaplan-Meier and Nelson-Aalen
When no parametric model is assumed, the reliability function is estimated from the distinct failure times \( t_1 < t_2 < \dots \), with \( d_j \) failures at \( t_j \) and \( n_j \) units at risk just before \( t_j \) (still under observation: not yet failed and not yet censored).
$$\hat R(t) = \prod_{t_j \le t} \left(1 - \frac{d_j}{n_j}\right) \quad \text{(Kaplan-Meier, product-limit)}, \qquad \hat H(t) = \sum_{t_j \le t} \frac{d_j}{n_j} \quad \text{(Nelson-Aalen)}.$$Conventions that MCQs test:
- A unit censored at time \( c \) is counted in the risk set for failures at times \( \le c \) and removed after. If a failure and a censoring are tied, the failure is taken to come first.
- \( \hat R \) is a step function that jumps only at failure times; censored times change the later risk sets, not \( \hat R \) itself.
- With no censoring, \( \hat R \) is the ordinary empirical survival function.
- If the largest observation is censored, \( \hat R \) never reaches 0; if it is a failure, \( \hat R \) drops to 0 there.
Memory Tricks
Don’t Confuse
| This | vs | That |
|---|---|---|
| Hazard \( h(t) = \frac{f(t)}{R(t)} \): a rate, can exceed 1 | density \( f(t) \): unconditional; \( F'(t) \) | |
| Type I: fixed time, random number of failures | Type II: fixed number of failures, random time | |
| TTT without replacement: \( \sum t_{(i)} + (n - r)t_{(r)} \) | TTT with replacement: \( n\,t_{(r)} \) | |
| \( \hat\theta = \frac{T}{r} \) (divide by failures) | \( \frac{T}{n} \) or \( \frac{\sum t_{(i)}}{r} \): both wrong | |
| IFR: \( h \) itself increases | IFRA: only \( \frac{H(t)}{t} \) increases | |
| Kaplan-Meier jumps at failures only | censored times change the risk set only | |
| Exponential: constant hazard, memoryless | geometric: its discrete twin |
Common Mistakes
Practice
All questions below are original practice questions written for this page. They are not from any CSIR NET paper. In Level 3, as in Part C, one or more options may be correct.
Level 1: Foundation
- Write \( R(t) \) in terms of the hazard \( h \).
- For which values of the Weibull shape \( \beta \) is the hazard decreasing?
- In which censoring scheme is the number of failures random: Type I or Type II?
Answers
- \( R(t) = \exp\left(-\int_0^t h(u)\,du\right) \).
- \( 0 < \beta < 1 \). (\( \beta = 1 \) is constant, \( \beta > 1 \) increasing.)
- Type I: the test stops at a fixed time, so the number of failures before it is random. In Type II the number \( r \) is fixed and the stopping time is random.
Level 2: Part B Style
Practice Q1. A unit has hazard \( h(t) = 2t \), \( t \ge 0 \). Its mean time to failure is
(a) \( \frac{1}{2} \) (b) \( \frac{\sqrt\pi}{2} \) (c) \( \sqrt\pi \) (d) 1
Answer: (b)
\( R(t) = e^{-t^2} \), so \( \text{MTTF} = \int_0^\infty e^{-t^2}\,dt = \frac{\sqrt\pi}{2} \approx 0.886 \). Option (c) is \( \int_{-\infty}^{\infty} e^{-t^2}\,dt \), the integral over the whole real line; (d) would need \( R(t) = e^{-t} \), a constant hazard.
Practice Q2. The life of a component is exponential with mean 1000 hours. Given that it has worked for 500 hours, the probability that it works for a further 1000 hours is
(a) \( e^{-1.5} \) (b) \( e^{-0.5} \) (c) \( e^{-1} \) (d) \( 1 - e^{-1} \)
Answer: (c)
Memorylessness: \( P(T > 1500 \mid T > 500) = \frac{e^{-1.5}}{e^{-0.5}} = e^{-1} \approx 0.368 \). Option (a) is the unconditional \( P(T > 1500) \).
Practice Q3. Ten items with exponential lifetimes are tested without replacement until the fourth failure, which occurs at 120 hours; the earlier failures are at 30, 55 and 80 hours. The maximum likelihood estimate of the mean life (hours) is
(a) 71.25 (b) 100.5 (c) 251.25 (d) 300
Answer: (c)
\( T = (30 + 55 + 80 + 120) + 6 \times 120 = 1005 \); \( \hat\theta = \frac{T}{r} = \frac{1005}{4} = 251.25 \). (a) ignores the six survivors; (b) divides by \( n \); (d) is the with-replacement value \( \frac{10 \times 120}{4} \).
Practice Q4. Which of the following lifetime distributions has a decreasing failure rate?
(a) Uniform on \( (0, 1) \) (b) Weibull with \( \beta = 2 \) (c) \( R(t) = (1 + t)^{-2} \), \( t \ge 0 \) (d) Gamma with shape 3
Answer: (c)
For (c), \( H(t) = 2\ln(1 + t) \), so \( h(t) = \frac{2}{1 + t} \), decreasing. (a) \( h = \frac{1}{1 - t} \) increases; (b) \( h \propto t \) increases; (d) gamma with shape \( > 1 \) is IFR.
Level 3: Part C Style
Practice Q5. A lifetime has reliability function \( R(t) = (1 + t)e^{-t} \), \( t \ge 0 \). Which statements are true?
(a) The hazard is \( h(t) = \frac{t}{1 + t} \), so the distribution is IFR. (b) The mean time to failure is 1. (c) The mean residual life at \( t = 1 \) is \( \frac{3}{2} \). (d) \( h(t) < 1 \) for every \( t \ge 0 \).
Answer: (a), (c) and (d)
- (a) True. \( f(t) = -R'(t) = -[e^{-t} - (1 + t)e^{-t}] = t e^{-t} \), so \( h = \frac{t e^{-t}}{(1 + t)e^{-t}} = \frac{t}{1 + t} \), increasing. (This is the gamma distribution with shape 2 and rate 1.)
- (b) False. \( \int_0^\infty (1 + t)e^{-t}\,dt = 1 + 1 = 2 \).
- (c) True. \( \int_t^\infty (1 + u)e^{-u}\,du = (2 + t)e^{-t} \), so \( m(t) = \frac{2 + t}{1 + t} \) and \( m(1) = \frac{3}{2} \). Note \( m(0) = 2 \) is the mean and \( m \) decreases to 1, as IFR implies.
- (d) True. \( \frac{t}{1 + t} < 1 \) for all \( t \ge 0 \); the hazard increases towards the limit 1 (the rate of the gamma) without reaching it.
Practice Q6. Eight units give the following times to failure (months), where + marks a censored observation:
2, 3+, 4, 4, 6+, 7, 9+, 10
Let \( \hat R \) be the Kaplan-Meier estimate. Which statements are true?
(a) \( \hat R(3) = \frac{7}{8} \). (b) \( \hat R(5) = \frac{7}{12} \). (c) \( \hat R(8) = \frac{1}{2} \). (d) \( \hat R(t) = 0 \) for \( t \ge 10 \).
Answer: (a), (b) and (d)
Failure times and risk sets: at 2, \( n = 8 \), \( d = 1 \); at 4, \( n = 6 \) (the units at 2 and 3+ have left), \( d = 2 \); at 7, \( n = 3 \) (6+ has left), \( d = 1 \); at 10, \( n = 1 \) (9+ has left), \( d = 1 \).
- (a) True. \( \hat R(3) = \hat R(2) = \frac{7}{8} \): the censoring at 3 does not move \( \hat R \).
- (b) True. \( \hat R(5) = \frac{7}{8} \cdot \frac{4}{6} = \frac{7}{12} \approx 0.583 \).
- (c) False. \( \hat R(8) = \frac{7}{12} \cdot \frac{2}{3} = \frac{7}{18} \approx 0.389 \).
- (d) True. At 10 the single unit at risk fails: factor \( 1 - \frac{1}{1} = 0 \). The largest observation is a failure, so \( \hat R \) reaches 0.
Practice Q7. Six items with exponential lifetimes (mean \( \theta \)) are tested without replacement until the third failure. The failures occur at 40, 90 and 150 hours. Let \( T \) be the total time on test and \( \hat\theta \) the MLE. Which statements are true?
(a) \( \hat\theta = \frac{730}{3} \approx 243.3 \) hours. (b) \( \frac{2T}{\theta} \) has a chi-square distribution with 6 degrees of freedom. (c) \( \text{Var}(\hat\theta) = \frac{\theta^2}{6} \). (d) Had the test been run with replacement and the same three failure times observed, the MLE would also be \( \frac{730}{3} \).
Answer: (a) and (b)
- (a) True. \( T = 40 + 90 + 150 + (6 - 3) \times 150 = 730 \); \( \hat\theta = \frac{730}{3} \).
- (b) True. Type II: \( \frac{2T}{\theta} \sim \chi^2_{2r} = \chi^2_6 \).
- (c) False. \( T \sim \text{Gamma}(3, \theta) \), so \( \text{Var}(\hat\theta) = \frac{3\theta^2}{9} = \frac{\theta^2}{3} \). The variance is governed by \( r = 3 \), not \( n = 6 \).
- (d) False. With replacement \( T = n\,t_{(r)} = 6 \times 150 = 900 \) and \( \hat\theta = 300 \).
For the record, the expected duration of this test is \( E[t_{(3)}] = \theta\left(\frac{1}{6} + \frac{1}{5} + \frac{1}{4}\right) = \frac{37}{60}\theta \approx 0.617\,\theta \).
Quick Revision Table
| Idea | Fact |
|---|---|
| Hazard | \( h = \frac{f}{R} = -\frac{d}{dt}\ln R \) |
| Reliability from hazard | \( R(t) = e^{-H(t)} \), \( H(t) = \int_0^t h \) |
| MTTF | \( \int_0^\infty R(t)\,dt \) |
| Mean residual life | \( m(t) = \frac{\int_t^\infty R}{R(t)} \); exponential: constant \( \theta \) |
| Weibull | \( R = e^{-(t/\eta)^\beta} \); \( \beta < 1 \) DFR, \( = 1 \) exponential, \( > 1 \) IFR; mean \( \eta\Gamma(1 + \frac{1}{\beta}) \) |
| Classes | IFR \( \Rightarrow \) IFRA \( \Rightarrow \) NBU \( \Rightarrow \) NBUE; IFR \( \Rightarrow \) DMRL \( \Rightarrow \) NBUE |
| Series of independent units | hazards add |
| Censored likelihood | \( \prod f(t_i)^{\delta_i} R(t_i)^{1 - \delta_i} \) |
| Exponential MLE | \( \hat\theta = \frac{T}{r} \) (TTT over failures) |
| Type II | \( \frac{2T}{\theta} \sim \chi^2_{2r} \); \( \hat\theta \) unbiased, variance \( \frac{\theta^2}{r} \) |
| TTT | without replacement \( \sum t_{(i)} + (n - r)t_{(r)} \); with replacement \( n t_{(r)} \) |
| Kaplan-Meier | \( \prod_{t_j \le t}(1 - \frac{d_j}{n_j}) \); jumps at failures only |
Key Takeaways
What’s Next
- Series and Parallel Systems: put these unit lifetimes together into systems, compute system reliability and MTTF, and see why a parallel pair of exponentials is not IFR.
- Markov Chains and Classification of States: the Poisson process behind tests with replacement is the simplest continuous-time Markov chain.
- Reliability and Operations Research: Quick Revision.