Estimating Baseline Survival in a Cox Model

Why Estimate a Baseline Survival Curve?

A fitted Cox model summarizes how covariates multiply the hazard, but a hazard ratio alone does not give an estimated survival probability at a particular time. To make that prediction, we also need an estimate of the model’s baseline hazard. The Breslow estimator constructs this estimate from the event times and the fitted risk sets, without requiring a particular parametric shape for the baseline hazard.

The result is useful for describing survival for specified covariate values, while keeping clear what the model does and does not estimate. The baseline curve is tied to a covariate reference value; it is not automatically the survival curve of an average participant.

Definition: In the Cox proportional hazards model, subject $i$, with covariate vector $Z_i$, has hazard $$ \lambda_i(t\mid Z_i)=\lambda_0(t)\exp(\beta^\mathsf{T}Z_i). $$ Here $\lambda_0(t)$ is the baseline hazard, the hazard at the reference value $Z_i=0$, and $\beta$ is the vector of regression coefficients. The baseline cumulative hazard is $H_0(t)=\int_0^t \lambda_0(u)\,du$, and the corresponding baseline survival function is $S_0(t)=\exp\{-H_0(t)\}$.

For a fixed covariate vector $z$, the model implies $H(t\mid z)=H_0(t)\exp(\beta^\mathsf{T}z)$, and therefore $S(t\mid z)=\exp\{-H_0(t)\exp(\beta^\mathsf{T}z)\}$. These are conditional, covariate-specific quantities. When the reference value is changed—for example, by centering a continuous covariate—the numerical baseline curve changes, even though the fitted survival predictions for the same people need not.

The Breslow Estimator

Suppose the distinct observed event times are $t_{(1)},\ldots,t_{(K)}$. Let $d_k$ be the number of events at $t_{(k)}$, and let $R(t_{(k)})$ be the risk set just before that time. Using the fitted Cox coefficients $\hat\beta$, define the risk-set sum $S^{(0)}(\hat\beta,t_{(k)})=\sum_{j\in R(t_{(k)})}\exp(\hat\beta^\mathsf{T}Z_j)$. It is a sum of relative-risk weights, not a count of subjects unless all weights equal one.

Definition: The Breslow estimator of the cumulative baseline hazard is $$ \widehat H_0(t)=\sum_{k:t_{(k)}\leq t} \frac{d_k}{\sum_{j\in R(t_{(k)})}\exp(\hat\beta^\mathsf{T}Z_j)}. $$ The estimated increment at event time $t_{(k)}$ is $d_k/S^{(0)}(\hat\beta,t_{(k)})$. The corresponding estimated baseline survival is $\widehat S_0(t)=\exp\{-\widehat H_0(t)\}$.

This is a step-function estimate: it increases at observed event times and remains constant between them. With tied events, the Breslow formula puts all $d_k$ events over the same risk-set denominator. This is the standard Breslow treatment of ties; it does not remove tied-event subjects one at a time from the denominator. The fitted coefficients $\hat\beta$ are used in every risk-set sum, so this baseline estimate is calculated after fitting the Cox regression.

Why the Risk-Set Formula Has This Form

The denominator adjusts for the relative hazards of the subjects who could experience an event at that time. A subject with a larger fitted relative risk contributes more to the risk-set sum, so the estimated baseline increment needed to account for a given number of observed events is smaller.

One justification considers the contribution of a small baseline cumulative-hazard increment $\Delta H_{0k}$ at event time $t_{(k)}$, holding $\hat\beta$ fixed. The Breslow construction gives the increment a log-likelihood contribution of the form $d_k\log(\Delta H_{0k})-\Delta H_{0k}S^{(0)}(\hat\beta,t_{(k)})$, up to terms not depending on that increment. Differentiating with respect to $\Delta H_{0k}$ gives $d_k/\Delta H_{0k}-S^{(0)}(\hat\beta,t_{(k)})$. Setting this derivative to zero yields $\Delta H_{0k}=d_k/S^{(0)}(\hat\beta,t_{(k)})$; the second derivative is $-d_k/(\Delta H_{0k})^2<0$, so this stationary point maximizes the contribution. This argument also shows why the event count appears in the numerator and the weighted risk set in the denominator.

Worked Example: Calculating Cumulative Baseline Hazard

Consider hypothetical data from a fitted Cox model with a binary covariate $X\in\{0,1\}$ and $\hat\beta=\log(2)$. Thus, subjects with $X=1$ have fitted relative risk $2$, and those with $X=0$ have relative risk $1$. Suppose the event-time risk sets are summarized as follows.

Event time	At risk with $X=0$	At risk with $X=1$	Events $d_k$
2 months	6	2	1
5 months	3	2	2
8 months	1	1	1

At 2 months, the weighted risk-set sum is $6(1)+2(2)=10$, so the estimated baseline-hazard increment is $1/10=0.1000$. At 5 months, it is $3(1)+2(2)=7$, giving increment $2/7=0.2857$. At 8 months, it is $1(1)+1(2)=3$, giving increment $1/3=0.3333$.

By 8 months, the cumulative estimate is $0.1000+0.2857+0.3333=0.7190$, using unrounded fractions for the sum. For the reference covariate value $X=0$, the estimated survival is $\exp(-0.7190)\approx0.4872$. This is the fitted survival estimate at 8 months for $X=0$, not an estimate for the whole covariate-mixed sample.

From Baseline Hazard to Covariate-Specific Survival

After estimating the baseline cumulative hazard, substitute the covariate vector of interest and the fitted coefficient into the Cox survival formula. The estimated relative risk multiplies the cumulative baseline hazard; it does not multiply the survival probability itself.

Worked Example: Predicting Survival for a Covariate Value

Suppose a fitted model has one covariate $x$, coefficient $\hat\beta=0.4$ per unit, and estimated cumulative baseline hazard $\widehat H_0(t)=0.24$ at a specified follow-up time. Calculate survival for $x=0$ and $x=1$.

For $x=0$, the relative risk is $\exp(0.4\times0)=1$. Thus, $\widehat S(t\mid x=0)=\exp\{-0.24(1)\}=\exp(-0.24)\approx0.7866$.

For $x=1$, use the unrounded relative risk $\exp(0.4)$ in the calculation: $\widehat S(t\mid x=1)=\exp\{-0.24\exp(0.4)\}$. Since $\exp(0.4)\approx1.491825$, the cumulative hazard is $0.24\exp(0.4)\approx0.3580$, and the survival estimate is $\exp\{-0.24\exp(0.4)\}\approx0.6990$. The covariate is measured in units of $x$, so the coefficient describes the log-hazard change per one-unit increase.

The lower estimate for $x=1$ is consistent with its fitted relative risk being greater than one: under the model, the higher hazard corresponds to lower survival at the same time. These are plug-in estimates; they do not include uncertainty in $\hat\beta$ or $\widehat H_0(t)$.

Worked Example: Changing the Covariate Reference Value

Suppose a hypothetical fitted model has covariate $x$, coefficient $\hat\beta=\log(1.5)$ per unit, and estimated baseline cumulative hazard $0.30$ at 12 months. Calculate the fitted survival estimates at $x=0,1,$ and $2$.

At $x=0$, the relative risk is $1$, so $\widehat S(12\mid0)=\exp(-0.30)\approx0.7408$. At $x=1$, it is $1.5$, giving cumulative hazard $0.30(1.5)=0.4500$ and survival $\exp(-0.4500)\approx0.6376$. At $x=2$, the relative risk is $\exp\{2\log(1.5)\}=1.5^2=2.25$, so cumulative hazard is $0.30(2.25)=0.6750$ and survival is $\exp(-0.6750)\approx0.5092$.

If the covariate were recoded as $x^*=x-1$, the reference value $x^*=0$ would mean $x=1$. The baseline survival would then be the curve for $x=1$, not $x=0$. The prediction for any fixed original value of $x$ remains the same when the intercept-free Cox linear predictor is re-expressed consistently.

Assumptions, Checks, and Alternatives

The formula is useful only insofar as the fitted Cox model and its risk sets represent the data appropriately. In particular, the proportional hazards model assumes that covariate effects multiply the hazard by a factor that does not vary with time. The model also relies on the recorded covariates and risk sets being correctly specified, and on censoring being noninformative conditional on the modeled covariates for the purpose of estimating survival. Each person must be counted as at risk only while that person is under observation and has not yet had the event. The standard risk-set definition can incorporate delayed entry by including a person only after entry, when that design applies.

Practical checks include confirming that event counts and risk-set membership match the event and censoring records; checking that each denominator is positive; and verifying that the estimated cumulative hazard is nondecreasing and changes only at event times. A calculation that uses the number at risk in place of the weighted sum is wrong unless all fitted relative risks are one. Implausibly large increments may result from a very small weighted risk set, and unstable predictions can arise when a requested covariate profile is poorly represented in the data. Proportional-hazards diagnostics are a separate part of model assessment; a survival curve computed from the formula does not itself establish that the assumption holds.

Kaplan–Meier estimation is a nonparametric estimate of an unadjusted survival curve. It does not condition on covariates, whereas a Cox prediction uses the fitted covariate effects and a baseline curve to estimate survival at specified covariate values. With no covariates, or when all subjects have the same covariate value, the Cox baseline estimate is closely related to a Nelson–Aalen cumulative-hazard estimate, with survival obtained by exponentiating the negative estimated cumulative hazard. Parametric survival models instead specify a distributional form for the baseline hazard or survival. They may offer more efficient or extrapolatable estimates if that form is appropriate, but impose additional assumptions.

The Breslow baseline estimate is a step function and does not, by itself, provide a smooth hazard curve. Its survival estimates are conditional predictions, not causal effects, and extrapolation beyond the observed follow-up is not justified by the estimator alone. Standard errors or confidence intervals require accounting for uncertainty in both the regression coefficients and the baseline estimate; the plug-in calculations above are point estimates only.

Common Pitfalls

Calling $\widehat S_0(t)$ the sample-average survival. It is survival at the reference covariate value. Average survival across a population would require averaging covariate-specific predictions over an appropriate covariate distribution.
Using an unweighted count in the denominator. The Breslow denominator sums fitted relative risks over the risk set; using only the number at risk ignores the fitted covariate effects.
Mixing rounded and unrounded values. Carry full precision through the product and exponential, then round the reported result. For example, $0.24\exp(0.4)\approx0.3580$; multiplying $0.24$ by the rounded value $1.492$ instead gives $0.3581$.
Treating a survival prediction as certainty. A value such as $0.6990$ is a model-based estimate for a covariate profile and time point, not a guarantee for an individual or a probability that the model is correct.

Key takeaway: The Breslow estimator adds $d_k$ divided by the fitted relative-risk sum in the risk set at each event time. Exponentiating the negative cumulative estimate gives baseline survival, and multiplying the cumulative hazard by the fitted relative risk gives survival for a specified covariate profile.

Check Your Understanding

Try each problem, then check the concise solutions below.

At an event time, four subjects with relative risk $1$ and three subjects with relative risk $2$ are at risk. If two events occur, what is the Breslow baseline-hazard increment?
If $\widehat H_0(t)=0.50$ and $\hat\beta=\log(2)$ for a binary covariate, calculate the estimated survival at covariate value $1$.
A continuous covariate is recoded from $x$ to $x^*=x-1$. Which original covariate value does the new baseline survival represent?
Why is a Cox prediction at $Z=0$ generally not the Kaplan–Meier estimate for the study sample?

Solutions. (1) The weighted sum is $4(1)+3(2)=10$, so the increment is $2/10=0.2000$. (2) The relative risk is $2$, so survival is $\exp\{-0.50(2)\}=\exp(-1)\approx0.3679$. (3) Since $x^*=0$ exactly when $x=1$, the new baseline survival represents the original value $x=1$. (4) The Cox prediction at $Z=0$ is conditional on the reference covariate profile, while Kaplan–Meier estimates an unadjusted curve for the observed sample; the two need not coincide when covariates vary.