AI Glossary / Deep learning / Intermediate
Proportional Hazards Model
Definition
A proportional hazards model describes how characteristics multiply the instantaneous rate of an event among cases that have not yet experienced it. For fixed characteristics, the hazard ratio remains constant over time. Cox regression is a widely used member of this model family.
What it means
The outcome is a time until something happens: a component fails, an account closes or a job finishes. Observations also record whether that event was actually seen. A machine still working when monitoring ends contributes information about its observed lifetime without supplying its eventual failure date.
Hazard is a conditional rate, not the probability of an event by a deadline. Proportionality concerns the ratio of rates between specified profiles; the underlying rates may rise or fall considerably with time.
This is a statistical survival-analysis concept. It does not require a deep neural network, although neural methods can extend the way characteristics enter a survival model.
How it works
A common formulation is h(t|x) = h0(t) × exp(βx). The baseline hazard h0(t) describes a reference profile, while the coefficient and characteristics determine its multiplier. With several characteristics, βx becomes a weighted sum.
Cox regression estimates coefficients using comparisons at observed event times through a partial likelihood. Its baseline hazard is not restricted to a specified parametric curve. Other proportional hazards models can impose a particular baseline distribution.
Analysts must define the event, time origin, follow-up and censoring process. They examine whether effects plausibly remain proportional. Residual plots and formal diagnostics can reveal changes over time; a non-significant test does not prove the assumption. Depending on the question, stratification or time-dependent effects may be more appropriate than forcing a constant ratio.
Three practical examples
- An equipment team studies time from installation to first failure, retaining units still operating at the study end as censored observations.
- A subscription analyst examines time to account closure while distinguishing closure from the end of available records.
- A computing team studies job completion times; here the event is a desirable completion, so a higher hazard need not mean a worse outcome.
Think of it this way
Think of two changing clocks whose speeds always maintain the same ratio. Both can speed up or slow down while one runs twice as fast as the other. The ratio can remain fixed even though neither speed is constant.
The analogy illustrates proportional rates only. A hazard model describes populations conditional on remaining event-free; it does not give every individual a deterministic countdown or say that every lifetime is divided by the hazard ratio.
Worked example
A hazard ratio is not a risk ratio
In an invented model, a binary characteristic has coefficient ln(2), approximately 0.693. Changing that characteristic from zero to one, with everything else fixed, multiplies the hazard by exp(0.693), approximately two.
Suppose the reference profile has survival probability 0.80 at a particular deadline. Under this proportional hazards model, the comparison profile has survival probability 0.80 squared, or 0.64. Their probabilities of experiencing the event by that deadline are therefore 0.20 and 0.36.
Doubling the hazard has not doubled the deadline probability to 0.40. The ratio of those probabilities is 1.8. These are calculations from stipulated model values, not empirical findings. An association between a characteristic and hazard also does not establish that changing the characteristic would cause the predicted difference.
Illustrative case study
Fictional Northfield Equipment compares failures across two operating environments. Its first dataset treats every unit withdrawn for inspection as simply lost to observation. Staff explain that unusual vibrations often trigger those withdrawals.
That process matters: observation may end precisely because a failure is becoming more likely. The team reviews withdrawal reasons and the assumptions needed for its analysis instead of treating all incomplete records as harmless omissions.
It also discovers that the difference between environments appears strongest early in service. Before publishing one summary hazard ratio, analysts inspect time-dependent diagnostics and consider a model that can represent the changing relationship. The case illustrates why clean timestamps alone do not establish a suitable survival model.
Common misconceptions and mistakes
Do not translate “twice the hazard” into “twice the probability” without a specified time horizon and survival calculation. Proportional hazards also does not mean a constant baseline hazard.
A censored record is neither an observed failure at the censoring date nor proof that the event never occurs. Standard analyses need appropriate assumptions about why follow-up ends, conditional on the information used.
Different event types can complicate interpretation. For example, retiring equipment permanently is not automatically equivalent to observing its mechanical failure. Decide how such competing events relate to the question before choosing the analysis.
Business use
Businesses can use survival models when timing and incomplete follow-up matter. A dashboard should identify the event, reference profile, observation window and uncertainty, and distinguish a relative hazard from an absolute deadline probability.
For operational decisions, examine calibration at relevant time horizons and performance on later cohorts. A model that ranks cases adequately may still give poor absolute probabilities. Record changes in maintenance, customer policy or monitoring because they can change both event behavior and the observation process.
People also ask
Is every proportional hazards model a Cox model?
No. Cox regression leaves the baseline hazard unspecified, whereas parametric models may specify its shape.
Does proportionality require flat hazards?
No. Hazards can change over time while their ratio remains constant for fixed profiles.
Can characteristics change over time?
Yes, suitable extensions can represent time-dependent covariates. Their interpretation and predicted survival require the relevant covariate history; a single fixed-profile comparison no longer describes every trajectory.
Check your understanding
A fitted hazard ratio is two and the reference deadline event probability is 20%. Can you report a comparison probability of 40%?
Answer: No. Under the simple proportional hazards setup, reference survival is 0.80, comparison survival is 0.80² = 0.64, and comparison event probability is 36%. The hazard ratio alone is not a deadline probability.
Related terms
- Survival analysis
- Survival function
- Censoring
- Calibration
- Cox regression
- hazard function
- competing risks
- time-dependent covariates
Learn next
Learn survival curves and right censoring first. Then compare hazard ratios with cumulative event probabilities and inspect proportionality diagnostics on a labeled teaching dataset. Finally examine how competing events and informative follow-up change the question the model can answer.
More in Deep learning
- Preprocessing - Preprocessing is the preparation and transformation of data before it is used by a model or another analysis step.
- Probabilistic Topic Model - A probabilistic topic model is a statistical model that represents a collection of documents through latent themes and probability distributions.
- Promoter based genetic algorithm - The promoter based genetic algorithm, or PBGA, is a neuroevolution method that uses promoter-like switches to control which encoded neural units are...
- Pulse-coupled networks - Pulse-coupled networks are networks of interacting units that influence one another through pulses.
- Quantum neural network - A quantum neural network is a learning model that uses quantum operations in a network-like or trainable computational structure.
- Radial Basis Network - A radial basis network is a neural network that represents inputs through functions of their distance from selected centers.
Sources
Sources: Cox, Regression Models and Life-Tables: https://rss.onlinelibrary.wiley.com/doi/pdf/10.1111/j.2517-6161.1972.tb00899.x R survival documentation, fitting Cox models: https://stat.ethz.ch/R-manual/R-devel/library/survival/html/coxph.html R survival documentation, proportionality diagnostics: https://stat.ethz.ch/R-manual/R-devel/library/survival/html/cox.zph.html R survival documentation, predicted curves: https://www.stat.ethz.ch/R-manual/R-devel/library/survival/html/survfit.coxph.html
Previous: Proportional hazards Next: Proportional improvement in prediction (PIP)