1 Definition and role in survival analysis

1.1 Hazard function basics

In survival analysis, the hazard function describes the instantaneous rate at which an event occurs at time \(t\), conditional on the event not having occurred before \(t\). Unlike a cumulative event probability, which accumulates over time, the hazard focuses on the moment-by-moment risk level. This makes it a natural target for modeling time-to-event data, particularly when risk changes as follow-up proceeds.

1.2 Baseline hazard versus total hazard

The baseline hazard isolates the event risk attributable to time alone, after fixing covariates at their reference values. The total hazard typically combines:

  • a baseline component that varies with time, and
  • a covariate-driven component that scales the baseline.

In proportional hazards models, the covariates act multiplicatively on the baseline hazard, so the baseline captures the “starting” risk trajectory that covariates then amplify or reduce.

1.3 Reference covariate values and interpretation

“Baseline” depends on how the model is parameterized. If covariate effects are expressed as relative risks, the baseline corresponds to setting covariates to 0 (for a log-linear specification) or to chosen reference levels (for categorical variables). Under either convention, interpreting the baseline requires clarity about what “zero effect” or “reference level” means in the particular dataset and modeling code. When covariates are centered or scaled, the implied baseline corresponds to those transformed values as well.

2 Mathematical formulation

2.1 Proportional hazards model setup

A common proportional hazards form is \[ h(t\mid X)=h_0(t)\exp(\beta^\top X), \] where \(h(t\mid X)\) is the hazard at time \(t\) given covariates \(X\), \(h_0(t)\) is the baseline hazard, and \(\beta\) are regression coefficients. The exponential term ensures the covariate effect is positive and yields a constant (in time) hazard ratio between individuals with different covariates.

2.2 Relationship to survival and cumulative hazard

2.2.1 Survival function via baseline hazard

The survival function \(S(t\mid X)=P(T>t\mid X)\) can be expressed through the cumulative hazard \(H(t\mid X)\): \[ S(t\mid X)=\exp\{-H(t\mid X)\}. \] Under proportional hazards, \[ H(t\mid X)=\int_0^t h_0(u)\exp(\beta^\top X)\,du = \exp(\beta^\top X)\,H_0(t), \] where \[ H_0(t)=\int_0^t h_0(u)\,du \] is the cumulative baseline hazard. Thus, the survival curve inherits the baseline time-shape through \(H_0(t)\), while covariates stretch or compress it via the multiplicative factor \(\exp(\beta^\top X)\).

2.2.2 Cumulative baseline hazard and estimation targets

Because the baseline hazard often appears inside integrals, estimation frequently targets \(H_0(t)\) or its increments rather than \(h_0(t)\) directly. Estimators may be reported as step functions in time, reflecting the discrete nature of observed event times in finite samples. For interpretability, cumulative quantities are often preferred since they are directly linked to survival through the exponential relationship.

2.3 Time-varying baseline hazard

In some models, the baseline hazard is allowed to vary more flexibly with time than a simple parametric form permits. Even within the proportional hazards framework, \(h_0(t)\) is not restricted to a constant: it may rise, fall, or oscillate depending on the underlying event mechanism. When time-varying effects of covariates are also introduced, the baseline remains the “reference-time” hazard component, while additional structure models how relative risks change across follow-up.

3 Estimation methods

3.1 Nonparametric estimation

3.1.1 Nelson–Aalen style cumulative hazard ideas

A classic approach in survival analysis is to estimate cumulative hazard nonparametrically using event counts and at-risk sets. The Nelson–Aalen estimator produces a stepwise estimate of \(H(t)\) by adding increments at observed event times proportional to the inverse of the number at risk. In the presence of covariates under proportional hazards, the analogous goal becomes estimating the cumulative baseline hazard \(H_0(t)\), typically using weighted information derived from the partial likelihood fit.

3.2 Semiparametric approaches

3.2.1 Breslow-type baseline hazard estimation

In proportional hazards models estimated via partial likelihood, the baseline hazard is commonly recovered through a Breslow-type estimator. Conceptually, after estimating \(\beta\), the method uses the fitted covariate relative risks to adjust contributions from each event time, producing an estimate of \(H_0(t)\) and, by extension, a baseline hazard increment at each observed time. This yields a baseline hazard estimate that is consistent with the proportional structure assumed by the model.

3.2.2 Handling tied event times

When multiple events occur at the same recorded time, the risk set contributions must be handled carefully. Common practical strategies approximate the contribution of tied event sets using methods such as Breslow’s approximation, Efron’s adjustment, or exact partial likelihood computations (depending on the software and model assumptions). These choices affect how the cumulative baseline hazard increments are allocated across ties, particularly in samples with many repeated event times.

3.3 Parametric baseline hazard models

3.3.1 Exponential baseline hazard

An exponential baseline hazard assumes \[ h_0(t)=\lambda, \] so the hazard is constant over time. This implies a survival function with a simple exponential decay in time, modified by covariates via the proportional factor \(\exp(\beta^\top X)\). The exponential model is mathematically convenient but may fit poorly when hazard rates change markedly with follow-up.

3.3.2 Weibull baseline hazard

The Weibull model allows monotone hazard trends and is often used when the hazard increases or decreases systematically with time: \[ h_0(t)=\alpha \gamma t^{\gamma-1}, \] with shape parameter \(\gamma\) and scale-related parameterization. Depending on \(\gamma\), the hazard can be increasing (\(\gamma>1\)) or decreasing (\(\gamma<1\)). Under proportional hazards, the covariate effects still scale the baseline hazard multiplicatively.

3.3.3 Gompertz and other common forms

The Gompertz hazard increases approximately exponentially with time and is used in settings where late follow-up experiences accelerating risk. More generally, parametric families may include log-logistic or log-normal formulations for survival time, which correspond to specific hazard shapes. The benefit of parametric baselines is that they reduce estimation variance and allow extrapolation beyond observed follow-up, but they can be biased if the chosen functional form does not match the true hazard trajectory.

3.4 Model selection and diagnostics for baseline choice

Choosing a baseline form typically involves balancing fit, interpretability, and stability. Diagnostics may include:

  • graphical checks of implied survival curves versus observed survival summaries,
  • residual and influence measures,
  • comparisons across candidate baseline families using information criteria or likelihood-based tests (within the confines of the fitted modeling framework).

Because baseline specification affects the time-shape of risk, selection procedures often emphasize whether the fitted model captures early versus late event dynamics rather than only overall calibration.

4.1 Cox model interpretation and reporting

In the Cox proportional hazards model, the baseline hazard \(h_0(t)\) is left unspecified during estimation of \(\beta\), focusing on the partial likelihood for relative effects. After fitting, analysts commonly report estimated baseline survival or cumulative baseline hazard to describe the time pattern of risk for a reference covariate profile. The main interpretive output of the Cox model remains hazard ratios, but baseline hazard estimates are useful for translating model predictions into absolute risk over time.

4.2 Accelerated failure time models (baseline interpretation parallels)

Accelerated failure time (AFT) models parameterize covariates by how they “stretch” or “shrink” survival time rather than scaling hazard multiplicatively. Although AFT models do not use a baseline hazard in the same proportional form, the baseline distribution of event times still plays an analogous role: it describes event timing when covariates are at reference values. The resulting baseline survival or density functions can be used to generate predicted event-time distributions for other covariate patterns.

4.3 Additive hazards models and contrasts

Additive hazards models represent hazard as \[ h(t\mid X)=h_0(t)+\beta^\top X, \] so covariate effects add to the baseline hazard rather than multiplying it. This changes interpretation: differences in hazard rates are modeled directly, and covariate impacts may interact with time implicitly through \(h_0(t)\). Baseline hazard remains the time-varying component at reference covariates, but the way it combines with covariates is fundamentally different from the proportional hazards setting.

4.4 Stratified baseline hazards

4.4.1 Stratification logic and implications

Stratification allows the baseline hazard to differ across categories of a stratifying variable while keeping covariate effects in other variables shared (or constrained) across strata. In practice, the model estimates separate baseline hazards \(h_{0k}(t)\) for each stratum \(k\), while the partial likelihood aggregates information without requiring proportionality across strata for the stratifying factor. This approach can improve fit when the proportional hazards assumption fails due to systematic baseline shifts across groups, though it may reduce efficiency when strata are small.

5 Practical interpretation and communication

5.1 Visualizing baseline hazard and cumulative hazard

Baseline hazards are often visualized as step functions or smoothed curves over time. Since hazard increments can be noisy, cumulative baseline hazard plots may be easier to interpret: the monotone growth of cumulative hazard provides a clearer view of how risk accumulates. When accompanied by uncertainty bands, these plots help convey estimation variability, especially near the tail where fewer subjects remain under observation.

5.2 Comparing baseline hazard shapes across subgroups

In stratified or subgroup modeling, analysts may compare how baseline hazard varies across groups by plotting estimated cumulative baseline hazards side by side. Care is needed: differences can reflect both genuine time dynamics and differences in sample composition, follow-up patterns, or censoring mechanisms. Presenting hazard ratios together with baseline shape summaries can clarify whether the main differences are in relative effects (multiplicative shifts) or in the underlying time trajectory of risk.

5.3 Relating baseline hazard to event-time dynamics

Baseline hazard helps translate abstract model parameters into time-relevant statements. For example, a rising baseline hazard indicates that instantaneous risk grows as time passes even for the reference covariate profile. Conversely, a declining baseline suggests a “fatigue” pattern where the remaining risk concentrates among a susceptible subset earlier in follow-up. When paired with covariate effects, baseline hazard supports predictions like expected time-to-event changes and absolute risk over a specified horizon.

6 Special considerations

6.1 Censoring assumptions and identifiability

Standard survival modeling relies on censoring being non-informative conditional on the covariates included in the model. If censoring depends on unmeasured factors related to event risk, baseline hazard estimates can become distorted because the observed risk sets no longer represent the true event-generating process. Under appropriate assumptions, baseline hazards are identifiable through the relationship between observed event times, at-risk sets, and the modeled hazard structure.

6.2 Left truncation and delayed entry

With left truncation (delayed entry), individuals enter the risk set only after their observation begins. This modifies the at-risk sets used for estimating baseline hazard or cumulative hazard increments. If handled correctly, the baseline hazard reflects instantaneous risk among those actually under observation at each time. Mis-specifying delayed entry can bias the estimated time-shape by improperly including individuals who were not truly at risk.

6.3 High-dimensional covariates and regularization (impact on baseline estimation)

When covariate dimension is large, regularization methods may be used to stabilize coefficient estimates. Because baseline hazard recovery depends on fitted relative risks from the covariate model, changes in coefficient shrinkage can indirectly affect baseline hazard estimates. Additionally, multicollinearity or separation-like behaviors in high dimensions can affect the weighting used in Breslow-type estimators, leading to unstable baseline increments. Practical workflows often include careful tuning, cross-validation, and sensitivity checks focused on how baseline curves behave under alternative regularization strengths.

6.4 Numerical stability and implementation details

Implementation involves computing risk scores \(\exp(\beta^\top X)\), aggregating weighted risk sets, and accumulating baseline hazard increments at event times. Numerical issues can arise when risk scores are extremely large or small, causing overflow, underflow, or loss of precision. Common remedies include scaling, using stable log-space calculations, and imposing sensible time grids for smoothing when needed. Correct handling of tied times, covariate transformations, and storage of intermediate cumulative hazards is also important for reproducible baseline hazard estimates.