Residual Lifetime and Mean Residual Life

Introduction

A component that has operated for several years is not in the same informational position as a newly installed one: we already know it has survived those years. The relevant question is therefore conditional: given survival to time $t$, how long might it last from now? The distribution of that remaining time is the residual lifetime distribution.

Residual lifetime is useful in reliability, clinical follow-up, and other settings where survival to a landmark is known. Its mean, called the mean residual life, summarizes expected remaining time among units that have reached that landmark. The exponential and Weibull models show two important possibilities: remaining lifetime may be unchanged by survival so far, or it may depend strongly on how long survival has already lasted.

Conditional Distribution of Remaining Lifetime

Definition: Let $T$ be a nonnegative lifetime with survival function $S(t)=P(T>t)$. For any $t\geq 0$ such that $S(t)>0$, the residual lifetime at $t$, denoted $R_t$, is the conditional random quantity $T-t$ given $T>t$. Its conditional survival function is $$ P(R_t>x)=P(T>t+x\mid T>t),\qquad x\geq 0. $$

The conditioning event matters: this is not the unconditional distribution of $T-t$. It describes remaining time only among lifetimes that exceed $t$. By the definition of conditional probability,

P(R_t>x)=\frac{S(t+x)}{S(t)},\qquad x\geq 0.

Thus the conditional distribution function is

F_{R_t}(x)= \begin{cases} 0, & x<0,\\[4pt] 1-\dfrac{S(t+x)}{S(t)}, & x\geq 0. \end{cases}

For an absolutely continuous lifetime with density $f$, the residual lifetime also has a conditional density, given almost everywhere for $x>0$, by shifting the original density and dividing by the probability of surviving to $t$:

f_{R_t}(x)=\frac{f(t+x)}{S(t)},\qquad x>0.

Result: If $T$ is nonnegative and $S(t)>0$, its mean residual life is $$ m(t)=E[T-t\mid T>t] =\int_0^\infty \frac{S(t+x)}{S(t)}\,dx =\frac{\displaystyle\int_t^\infty S(u)\,du}{S(t)}. $$ The integral, and hence the mean, may be infinite.

To justify the mean formula, use the tail-integral identity for a nonnegative random variable $Y$: $E[Y]=\int_0^\infty P(Y>x)\,dx$, with both sides allowed to be infinite. Indeed, pointwise $Y=\int_0^\infty \mathbf{1}_{\{Y>x\}}\,dx$; Tonelli’s theorem permits the expectation and integral to be interchanged because the integrand is nonnegative. Apply this identity under the conditional probability given $T>t$, and substitute the conditional survival function just derived. Finally, set $u=t+x$ to obtain the last expression. This proves the result without requiring a finite mean in advance. $\square$

Exponential Lifetimes: Remaining Time Is Unchanged

Suppose $T$ is exponential with rate $\lambda>0$, so $S(t)=e^{-\lambda t}$ for $t\geq 0$. For every $t,x\geq 0$,

P(R_t>x) =\frac{e^{-\lambda(t+x)}}{e^{-\lambda t}} =e^{-\lambda x}.

This is the survival function of an exponential random variable with rate $\lambda$, regardless of $t$. Therefore the conditional residual lifetime has the same distribution as the original lifetime. Its mean is

m(t)=\int_0^\infty e^{-\lambda x}\,dx=\frac{1}{\lambda}.

This invariance is the memoryless property: for an exponential lifetime, knowing that the unit has survived to $t$ does not change the distribution of its remaining lifetime. It does not mean that elapsed time is literally erased; it means that the conditional distribution has this particular form.

Worked Example: Exponential Remaining Lifetime

A device has an exponential lifetime with rate $\lambda=0.2$ per month. Given that it has operated for 6 months, find the probability it operates for at least 3 additional months, the probability it fails within the next 3 months, and its mean remaining lifetime.

The probability of at least 3 more months is the conditional survival probability:

P(R_6>3)=\frac{S(9)}{S(6)} =\frac{e^{-0.2(9)}}{e^{-0.2(6)}} =e^{-0.6}\approx 0.5488.

The event of failure within the next 3 months is $R_6\leq 3$, so its probability is the complement:

P(R_6\leq 3)=1-e^{-0.6}\approx 1-0.5488=0.4512.

The mean remaining lifetime is $1/\lambda=1/0.2=5$ months. The same mean applies at any survival time $t$, not just at 6 months.

Weibull Lifetimes: Residual Time Depends on the Landmark

Use the Weibull parameterization with shape $k>0$ and scale $\eta>0$:

S(t)=\exp\left[-\left(\frac{t}{\eta}\right)^k\right], \qquad t\geq 0.

Substituting this survival function into the conditional survival formula gives

P(R_t>x) =\exp\left\{-\left[ \left(\frac{t+x}{\eta}\right)^k- \left(\frac{t}{\eta}\right)^k \right]\right\},\qquad x\geq 0.

In general, this expression depends on $t$, so the remaining-time distribution changes with the survival landmark. When $k=1$, the Weibull survival function is exponential, and the difference in brackets reduces to $x/\eta$, independent of $t$. For other shape values, the conditional distribution is generally not the original Weibull distribution with time simply reset to zero.

Weibull mean residual life: Define the upper incomplete gamma function by $$ \Gamma(a,z)=\int_z^\infty v^{a-1}e^{-v}\,dv,\qquad a>0,\ z\geq 0. $$ For a Weibull lifetime with shape $k$ and scale $\eta$, $$ m(t)=\frac{\eta}{k}\exp\left[\left(\frac{t}{\eta}\right)^k\right] \Gamma\left(\frac{1}{k},\left(\frac{t}{\eta}\right)^k\right). $$

To derive this expression, start with the mean residual life integral and substitute $z=(u/\eta)^k$. Then $u=\eta z^{1/k}$ and $du=(\eta/k)z^{1/k-1}dz$. Writing $z_t=(t/\eta)^k$, the numerator becomes

\int_t^\infty S(u)\,du =\frac{\eta}{k}\int_{z_t}^\infty z^{1/k-1}e^{-z}\,dz =\frac{\eta}{k}\Gamma\left(\frac{1}{k},z_t\right).

Dividing by $S(t)=e^{-z_t}$ proves the stated formula. The integral is finite for every finite $t$ when $k,\eta>0$, since the Weibull survival function has an integrable tail.

Worked Example: Weibull Residual Probability and Mean

Let $T$ have a Weibull distribution with shape $k=2$ and scale $\eta=10$ months. Given survival to $t=10$ months, find the probability of lasting at least 5 more months and calculate the mean residual life.

Here $t/\eta=1$ and $(t+5)/\eta=1.5$. The conditional survival probability is

P(R_{10}>5) =\exp[-(1.5^2-1^2)] =e^{-1.25} \approx 0.2865.

For the mean, $z_t=(10/10)^2=1$. The Weibull formula gives

m(10)=\frac{10}{2}e^1\Gamma\left(\frac12,1\right).

Since $\Gamma(1/2,1)=\sqrt{\pi}\,\operatorname{erfc}(1)\approx 0.2788$, this is

m(10)\approx 5(2.7183)(0.2788)\approx 3.79\text{ months}.

The conditional survival probability and the mean are both calculated among devices that have already reached 10 months. The mean remaining time is not the mean lifetime at installation.

Worked Example: Weibull Mean at Installation

For a Weibull lifetime with $k=2$ and $\eta=10$ months, calculate the initial mean lifetime and the probability of lasting more than 5 months from installation.

At $t=0$, the residual lifetime is just the original lifetime. In the mean formula, $z_t=0$, so

m(0)=\frac{10}{2}\Gamma\left(\frac12,0\right) =5\Gamma\left(\frac12\right) =5\sqrt{\pi} \approx 8.86\text{ months}.

The probability of lasting more than 5 months is

P(T>5)=\exp[-(5/10)^2] =e^{-0.25}\approx 0.7788.

This initial mean of about 8.86 months differs from the mean residual life of about 3.79 months at $t=10$. The difference reflects conditioning on survival to 10 months; the two means answer different questions.

Interpretation and Common Pitfall

The survival ratio $S(t+x)/S(t)$ is the central calculation: it compares survival through $t+x$ with survival through $t$, and therefore measures survival for an additional $x$ time units among those who reached $t$. Integrating this ratio over all $x\geq 0$ gives the expected remaining time. A long conditional tail can make the mean residual life large even when the probability of surviving any one fixed additional interval is modest.

Common mistake: treating $T-t$ as if it had the unconditional distribution of $T$, even after conditioning on $T>t$. That shortcut is correct for exponential lifetimes because their survival ratio reduces to $e^{-\lambda x}$, but it is not generally correct for Weibull lifetimes. For example, with $k=2,\eta=10,t=10$, the chance of 5 more months is $e^{-1.25}$, not the unconditional Weibull chance $e^{-(5/10)^2}=e^{-0.25}$. Always condition by dividing $S(t+x)$ by $S(t)$.

Key takeaway: For any nonnegative lifetime with $S(t)>0$, the residual survival function is $S(t+x)/S(t)$, and the mean residual life is its integral over $x\geq 0$. Exponential lifetimes retain the same residual distribution at every $t$; Weibull residual distributions generally change with $t$.

Check Your Understanding

Try these questions using the conditional survival and mean residual life formulas.

For a lifetime with survival function $S$, write the conditional probability of lasting more than $x$ additional time units given survival to $t$.
Explain why the mean residual life formula can give an infinite value, and identify the integral whose finiteness determines whether it is finite.
An exponential lifetime has rate $0.1$ per year. What is its mean residual life after surviving 4 years, and why does it not depend on 4?
For a Weibull lifetime, describe how you would calculate the probability of lasting at least $x$ more time units after surviving to $t$.
Why are the initial mean lifetime and the mean residual life at a positive time different quantities in general?