Survival analysis and reliability

Hazard rates, proportional hazard regression, life testing, mean time to failure

March 12, 2019 — March 7, 2022

point processes
time series
Figure 1

1 Estimating survival rates

Here’s the set-up: looking at a data set of individuals’ lifespans you would like to infer the distributions—Analysing when people die, or things break etc. The statistical problem of estimating how long people’s lives are is complicated somewhat by the particular structure of the data — loosely, “every person dies at most one time”, and there are certain characteristic difficulties that arise, such as right-censorship. (If you are looking at data from an experiment and not all your subjects have died yet, they presumably die later, but you don’t know when.)

Handily, the tools one invents to solve this kind of problem end up being useful to solve other problems, such as point process inference, and look not so far from the densities/intensities relations.

So let’s say you have a random variable \(X\) of positive support, according to which the lifetime of your people (components, machines, whatever) are distributed, which possesses a pdf \(f_X(t)\) and cdf \(F_X(T)\).

We define several useful functions:

The survival function (which is also the right tail CDF)
the hazard function
the cumulative hazard function
\(\Lambda(t) :=\int_0^t\lambda(s) \textrm{d} s.\)

Why? Because it happens to come out nicely if we do that, and these functions acquire intuitive interpretations once we squint at them a bit. The survival function is the probability of an individual surviving to time \(t\) etc. The hazard function will turn out to be the rate of deaths at time \(t\) given that one has not yet occurred.

Using the chain rule we can find the following useful relation:

\[S(t)=\exp[-\Lambda (t)]={\frac {f(t)}{\lambda (t)}}\]

The hazard function can be pretty much any non-negative function of non-negative support (or more generally, a Schwartz distribution, but let’s ignore that possibility for the moment.)

Figure 2

1.1 Life table method

Over intervals of time \([t,u]\) we define the cumulative hazard increment

\[ H(t,u) :=\int_t^u h (s) \textrm{d} s = H(u)-H(t) \]

and the survival increment

\[ \chi(t,u) :=\frac{\chi(u)}{\chi(t)} \]

The following relations are useful

\[ \chi(t)=\exp[-H (t)]={\frac {f(t)}{h (t)}}. \]


\[ \chi(t,u)=\frac{\exp[-H (u)]}{\exp[-H (t)]}=\exp[H (t)-H (u)]=\exp[-H (t,u)] \]

and so

\[-\log\chi(t,u)=H (t,u).\]

We estimate hazard via the life table method. Given a time interval \([t_{i}, t_{i+1})\) and survival counts \(N(t_{i})\) and \(N(t_{i+1})\) at, respectively, the beginning and end of that interval, (assuming no immigration) the life table estimate of a survival increment is

\[\hat{\chi}(t_i, t_{i+1}):= \frac{N(t_{i+1})}{N(t_{i})}\]

Plugging this in, we obtain cumulative hazard increment estimates

\[\begin{aligned} \hat{H} (t_i, t_{i+1})&=-\log \hat{\chi}(t_i, t_{i+1})\\ &=\log \frac{ N(t_{i}) }{ N(t_{i+1}) } \end{aligned}\]

From this we construct further point estimates of \(H\) at \(t\in[0, t_1, t_2,\dots]\) as

\[\hat{H} (t)=\sum_{t_i\leq t}\hat{H}(t_{i},t_{i+1})\] By introducing assumptions on the functional form, can estimate the entire hazard function. For example, we can take \(h (t)\) to be piecewise constant, so that

\[\begin{aligned} h (t)=\sum_i\mathbb{I}\{t_{i}<t<t_{i+1}\} h_i \end{aligned}\]

This corresponds to the assumption that \(H\) is piecewise linear and continuous; we are constructing a piecewise linear interpolant. Thus, for \(t\in(t_i,t_{i+1}],\) we such an interpolant \(\hat{H}\) for \(t\in[0,t_M]\) by a first order polynomial spline with knots \(0,t_1,t_2,\dots, t_M\) and values \(\hat{H}(0), \hat{H}(t_1), \hat{H}(t_2) \dots,\hat{H}(t_M).\)

Figure 3

1.2 Nelson-Aalen estimates

a.k.a. Empirical Cumulative Hazard Function estimator.

The original Aalen paper on this is notoriously beautiful because of clever construction of a life point process and associated martingale. Clear and worth reading. Spoiler: despite the elegant derivation, the actual estimator is something a high-school student could discover by guessing.


2 Other reliability stuff

Reliawiki has handy stuff, e.g. comprehensive docs on the Weibull law. It’s in support of some software package their are trying to sell, I think?

We can calculate an “effective age” if we want an intuitive risk measure (Brenner, Gefeller, and Greenland 1993).

3 tools

4 Score function versus hazard function

  • The score function and log-hazard rates are similar beasts. We can exploit that, e.g. in a Langevin dynamics algorithm? But would we gain anything useful from that?

5 Incoming

Social Desirability Bias: How Psych Can Salvage Econo-Cynicism

Social desirability bias is the tendency of respondents to answer questions in a manner that will be viewed favorably by others. It can take the form of over-reporting “good behavior” or under-reporting “bad,” or undesirable behavior. The tendency poses a serious problem with conducting research with self-reports, especially questionnaires. This bias interferes with the interpretation of average tendencies as well as individual differences.

6 References

Aalen, Odd O. 1978. Nonparametric Inference for a Family of Counting Processes.” The Annals of Statistics.
Aalen, Odd O., Borgan, and Gjessing. 2008. Survival and Event History Analysis: A Process Point of View. Statistics for Biology and Health.
Andersen, Per Kragh, Borgan, Gill, et al. 1997. Statistical models based on counting processes. Springer series in statistics.
Andersen, Per Kragh, and Keiding. 2014. Survival Analysis, Overview.” In Wiley StatsRef: Statistics Reference Online.
Andersen, Per K., and Vaeth. 2015. Survival Analysis.” In Wiley StatsRef: Statistics Reference Online.
Appendix 1: The Delta Method.” 2011. In Applied Survival Analysis.
Bagnoli, and Bergstrom. 1989. “Log-Concave Probability and Its Applications.”
Brenner, Gefeller, and Greenland. 1993. Risk and rate advancement periods as measures of exposure impact on the occurrence of chronic diseases.” Epidemiology (Cambridge, Mass.).
Cox, D. R. 1972. Regression Models and Life-Tables.” Journal of the Royal Statistical Society: Series B (Methodological).
Cox, D. R, and Oakes. 2018. Analysis of Survival Data.
Cutler, and Ederer. 1958. Maximum utilization of the life table method in analyzing survival.” Journal of Chronic Diseases.
Deddens, and Koch. 2014. Survival Analysis, Grouped Data in.” In Wiley StatsRef: Statistics Reference Online.
Efron. 1988. Logistic Regression, Survival Analysis, and the Kaplan-Meier Curve.” Journal of the American Statistical Association.
Fink, and Brown. 2006. Survival Analysis.” Gastroenterology & Hepatology.
Griffiths, and Ghahramani. 2011. The Indian Buffet Process: An Introduction and Review.” Journal of Machine Learning Research.
Hjort. 1990. Nonparametric Bayes Estimators Based on Beta Processes in Models for Life History Data.” The Annals of Statistics.
———. 1992. On Inference in Parametric Survival Data Models.” International Statistical Review / Revue Internationale de Statistique.
Hjort, West, and Leurgans. 1992. Semiparametric Estimation Of Parametric Hazard Rates.” In Survival Analysis: State of the Art. Nato Science 211.
Hosmer, and Lemeshow. 1999. Applied Survival Analysis: Regression Modeling of Time to Event Data. Wiley Series in Probability and Statistics.
Hosmer, Lemeshow, and May. 2008. Descriptive Methods for Survival Data.” In Applied Survival Analysis: Regression Modeling of Time-to-Event Data. Wiley Series in Probability and Statistics.
Klein. 2014. Survival Distributions and Their Characteristics.” In Wiley StatsRef: Statistics Reference Online.
Kleinbaum. 2010. Survival Analysis: A Self-Learning Text. Statistics for Biology and Health 1.0.
Laird, and Olivier. 1981. Covariance Analysis of Censored Survival Data Using Log-Linear Analysis Techniques.” Journal of the American Statistical Association.
Lu, Goldberg, and Fine. 2012. On the Robustness of the Adaptive Lasso to Model Misspecification.” Biometrika.
Nelson. 1969. Hazard Plotting for Incomplete Failure Data.” Journal of Quality Technology.
———. 2000. Theory and Applications of Hazard Plotting for Censored Failure Data.” Technometrics.
Omi, Ueda, and Aihara. 2020. Fully Neural Network Based Model for General Temporal Point Processes.” arXiv:1905.09690 [Cs, Stat].
Parametric Regression Models.” 2011. In Applied Survival Analysis.
Peng. 2021. Quantile Regression for Survival Data.” Annual Review of Statistics and Its Application.
Peterson. 1977. Expressing the Kaplan-Meier Estimator as a Function of Empirical Subsurvival Functions.” Journal of the American Statistical Association.
Pölsterl. 2020. Scikit-Survival: A Library for Time-to-Event Analysis Built on Top of Scikit-Learn.” Journal of Machine Learning Research.
Schoenberg. 2003. Multidimensional Residual Analysis of Point Process Models for Earthquake Occurrences.” Journal of the American Statistical Association.
Shaked, and Shanthikumar. 1988. On the First-Passage Times of Pure Jump Processes.” Journal of Applied Probability.
Sill. 1997. Monotonic Networks.” In Proceedings of the 10th International Conference on Neural Information Processing Systems. NIPS’97.
Simon, Friedman, Hastie, et al. 2011. Regularization Paths for Cox’s Proportional Hazards Model via Coordinate Descent.” Journal of Statistical Software.
Sy, and Taylor. 2000. Estimation in a Cox Proportional Hazards Cure Model.” Biometrics.
Taleb. 2020. On the Statistical Differences Between Binary Forecasts and Real-World Payoffs.” International Journal of Forecasting.
Tibshirani. 1997. The Lasso Method for Variable Selection in the Cox Model.” Statistics in Medicine.