Hazard Difference Estimator in Survival Analysis

• Hazard Difference Estimator is a class of estimators quantifying absolute, additive hazard contrasts used in survival analysis and clinical trials.
• It employs methodologies like standardized average hazards, additive hazards regression (HDi), and competing risks models to ensure statistical robustness.
• The estimator enables direct treatment effect comparisons through average hazard differences, yielding interpretable metrics while requiring careful causal assumptions.

A hazard difference estimator denotes a class of estimators for contrasts on the additive hazard scale rather than the multiplicative hazard-ratio scale. In current survival-analysis usage, the term appears in at least three distinct forms: the difference in average hazard over a prespecified window $[0,\tau]$, $\mathrm{DAH}(\tau)=\eta_1(\tau)-\eta_0(\tau)$; the constant treatment coefficient $\theta$ in an additive hazards model $\haz(t;D,\mathbf Z)=\lambda_0(t)+D\theta+\beta^\top \mathbf Z$; and the cause-specific conditional treatment effect $\beta_j$ in competing risks models $h_j(t\mid A,Z)=\beta_j A+\lambda_j(t,Z)$ (Qian et al., 2024, Hou et al., 2019, Rava et al., 2021). These formulations share an absolute event-intensity interpretation, but they differ in target parameter, adjustment strategy, asymptotic theory, and causal status.

1. Terminological scope and neighboring concepts

In the average-hazard literature, the basic arm-specific quantity is the average hazard with survival weight,

$\eta(\tau)=\frac{1-S(\tau)}{\int_0^\tau S(u)\,du},$

interpreted as an average person-time incidence rate over $[0,\tau]$ under a hypothetical setting where censoring before $\tau$ is removed. The associated hazard difference is

$\mathrm{DAH}(\tau)=\eta_1(\tau)-\eta_0(\tau).$

In additive-hazards regression, by contrast, the hazard difference is encoded directly as the additive treatment coefficient. For a binary treatment $\mathrm{DAH}(\tau)=\eta_1(\tau)-\eta_0(\tau)$0, the model

$\mathrm{DAH}(\tau)=\eta_1(\tau)-\eta_0(\tau)$1

treats $\mathrm{DAH}(\tau)=\eta_1(\tau)-\eta_0(\tau)$2 as the additive effect of treatment on the event hazard. In competing risks, the analogous formulation is

$\mathrm{DAH}(\tau)=\eta_1(\tau)-\eta_0(\tau)$3

so that $\mathrm{DAH}(\tau)=\eta_1(\tau)-\eta_0(\tau)$4 is the cause-specific hazard difference (Qian et al., 2024, Hou et al., 2019, Rava et al., 2021).

Several adjacent literatures use hazards as inputs without defining a standalone hazard-difference estimator. "Transforming cumulative hazard estimates" develops a generic plug-in framework

$\mathrm{DAH}(\tau)=\eta_1(\tau)-\eta_0(\tau)$5

for obtaining survival-scale or mean-survival-scale parameters from cumulative hazard estimators, but it does not introduce a quantity explicitly called a direct estimator of $\mathrm{DAH}(\tau)=\eta_1(\tau)-\eta_0(\tau)$6 (Ryalen et al., 2017). Likewise, "Efficient and Debiased Learning of Average Hazard Under Non-Proportional Hazards" develops inference for the log average hazard ratio

$\mathrm{DAH}(\tau)=\eta_1(\tau)-\eta_0(\tau)$7

while explicitly noting that a difference $\mathrm{DAH}(\tau)=\eta_1(\tau)-\eta_0(\tau)$8 is not its primary target (Meng et al., 13 Feb 2026).

2. Stratified estimation through average hazard with survival weight

The most explicit recent construction of a hazard difference estimator on an absolute scale is the stratified difference in average hazard proposed for randomized time-to-event studies with stratification factors. For stratum $\mathrm{DAH}(\tau)=\eta_1(\tau)-\eta_0(\tau)$9 and treatment group $\theta$0, with Kaplan–Meier estimator $\theta$1, the stratum-specific average hazard estimator is

$\theta$2

The method then defines a standardized survival curve

$\theta$3

and a standardized average hazard

$\theta$4

Its estimator is

$\theta$5

The stratified treatment contrasts are then

$\theta$6

This construction is explicitly based on direct standardization rather than variance-weighted pooling of stratum-specific effects (Qian et al., 2024).

The standardization step gives the estimator a population meaning tied to the chosen stratum distribution $\theta$7. The method is presented as an alternative to conventional stratified pooling, which either assumes a common effect across strata or uses CMH-type or stratified-Cox-type weighting schemes with opaque population interpretation when stratum-specific effects differ. By standardizing group-specific survival first and contrasting standardized hazards second, the procedure does not require the treatment effect to be the same across strata, allows user-chosen weights that represent the intended target population, and yields both absolute and relative treatment summaries directly tied to that target population (Qian et al., 2024).

3. Assumptions, inference, and empirical behavior of DAH

The stratified DAH estimator is developed under independent censoring within each treatment-by-stratum cell,

$\theta$8

a finite number of strata $\theta$9, and positive asymptotic stratum proportions

$\haz(t;D,\mathbf Z)=\lambda_0(t)+D\theta+\beta^\top \mathbf Z$0

Unlike the stratified Cox model, the method does not assume proportional hazards within strata or a common hazard ratio across strata. Large-sample theory is formulated through

$\haz(t;D,\mathbf Z)=\lambda_0(t)+D\theta+\beta^\top \mathbf Z$1

which converge to mean-zero normal limits with variances estimated by plug-in procedures based on Kaplan–Meier and Nelson–Aalen quantities. This yields Wald-type confidence intervals for both $\haz(t;D,\mathbf Z)=\lambda_0(t)+D\theta+\beta^\top \mathbf Z$2 and the contrasts $\haz(t;D,\mathbf Z)=\lambda_0(t)+D\theta+\beta^\top \mathbf Z$3 and $\haz(t;D,\mathbf Z)=\lambda_0(t)+D\theta+\beta^\top \mathbf Z$4, as well as difference-based and log-ratio-based Wald tests for $\haz(t;D,\mathbf Z)=\lambda_0(t)+D\theta+\beta^\top \mathbf Z$5 (Qian et al., 2024).

The interpretation is explicitly absolute: $\haz(t;D,\mathbf Z)=\lambda_0(t)+D\theta+\beta^\top \mathbf Z$6 is a person-time event rate over $\haz(t;D,\mathbf Z)=\lambda_0(t)+D\theta+\beta^\top \mathbf Z$7, and $\haz(t;D,\mathbf Z)=\lambda_0(t)+D\theta+\beta^\top \mathbf Z$8 is the absolute reduction or increase in event intensity due to treatment. In the ARASENS prostate cancer example, using $\haz(t;D,\mathbf Z)=\lambda_0(t)+D\theta+\beta^\top \mathbf Z$9 months and weights based on stratum sizes, the adjusted average hazard was $\beta_j$0 for darolutamide and $\beta_j$1 for placebo, producing an adjusted DAH of $\beta_j$2 with 95% CI $\beta_j$3 and $\beta_j$4, and an adjusted RAH of $\beta_j$5 with 95% CI $\beta_j$6 and $\beta_j$7. The paper reports AH as “events per 100 person-months,” so DAH is directly interpretable as an absolute rate difference. Its simulation study, based on 3,000 simulations with random censoring, no random censoring, sample sizes $\beta_j$8 and $\beta_j$9 per arm, and truncation times $h_j(t\mid A,Z)=\beta_j A+\lambda_j(t,Z)$0 months, found negligible bias for AH, DAH, and $h_j(t\mid A,Z)=\beta_j A+\lambda_j(t,Z)$1, near-nominal 95% coverage, improvement when the risk set size at $h_j(t\mid A,Z)=\beta_j A+\lambda_j(t,Z)$2 was adequate, and stability under both censoring regimes (Qian et al., 2024).

4. Additive-hazards treatment effect estimation and the HDi estimator

A separate but closely related line of work treats the hazard difference as the additive treatment effect in a regression model. In the high-dimensional setting, "Estimating Treatment Effect under Additive Hazards Models with High-dimensional Covariates" studies right-censored data with binary treatment $h_j(t\mid A,Z)=\beta_j A+\lambda_j(t,Z)$3, covariates $h_j(t\mid A,Z)=\beta_j A+\lambda_j(t,Z)$4, observed time $h_j(t\mid A,Z)=\beta_j A+\lambda_j(t,Z)$5, and event indicator $h_j(t\mid A,Z)=\beta_j A+\lambda_j(t,Z)$6, under

$h_j(t\mid A,Z)=\beta_j A+\lambda_j(t,Z)$7

The paper first derives an orthogonal score for $h_j(t\mid A,Z)=\beta_j A+\lambda_j(t,Z)$8, then introduces the Hazards Difference (HDi) estimator as a closed-form special case with a balancing interpretation. Using propensity-based weights

$h_j(t\mid A,Z)=\beta_j A+\lambda_j(t,Z)$9

it defines weighted Breslow-type cumulative hazard estimators

$\eta(\tau)=\frac{1-S(\tau)}{\int_0^\tau S(u)\,du},$0

and the HDi estimator

$\eta(\tau)=\frac{1-S(\tau)}{\int_0^\tau S(u)\,du},$1

Under the additive hazards model, $\eta(\tau)=\frac{1-S(\tau)}{\int_0^\tau S(u)\,du},$2 estimates $\eta(\tau)=\frac{1-S(\tau)}{\int_0^\tau S(u)\,du},$3 while $\eta(\tau)=\frac{1-S(\tau)}{\int_0^\tau S(u)\,du},$4 estimates $\eta(\tau)=\frac{1-S(\tau)}{\int_0^\tau S(u)\,du},$5, so their difference targets the cumulative additive treatment contribution $\eta(\tau)=\frac{1-S(\tau)}{\int_0^\tau S(u)\,du},$6 (Hou et al., 2019).

The HDi estimator is developed to preserve asymptotically valid inference in high dimensions and to tolerate nuisance-model difficulty. The paper proves classical model double robustness—consistency when either the additive hazards outcome model is correct or the treatment model is correct—as well as a more novel sparsity double robustness, under which one of the nuisance models may be fully dense if the other is sufficiently well behaved and cross-fitting is used. In the SEER-Medicare prostate cancer application, radical prostatectomy versus conservative management produced treatment effects around $\eta(\tau)=\frac{1-S(\tau)}{\int_0^\tau S(u)\,du},$7 in hazard-rate units, interpreted as a lower overall mortality hazard after adjustment for high-dimensional confounding (Hou et al., 2019).

5. Competing risks and semiparametric hazard-difference estimation

For competing risks data, the hazard difference estimator is typically cause-specific. "Doubly Robust Estimation of the Hazard Difference for Competing Risks Data" considers observational studies with treatment $\eta(\tau)=\frac{1-S(\tau)}{\int_0^\tau S(u)\,du},$8, baseline covariates $\eta(\tau)=\frac{1-S(\tau)}{\int_0^\tau S(u)\,du},$9, event time $[0,\tau]$0, cause $[0,\tau]$1, censoring time $[0,\tau]$2, observed time $[0,\tau]$3, and counting processes

$[0,\tau]$4

The cause-specific hazard model is

$[0,\tau]$5

where $[0,\tau]$6 is left unspecified and $[0,\tau]$7 is the conditional treatment effect on the hazard scale for cause $[0,\tau]$8. The paper derives the efficient score for $[0,\tau]$9, then constructs two orthogonal estimating scores,

$\tau$0

both doubly robust with respect to the pair $\tau$1 and the competing-risks outcome models $\tau$2 (Rava et al., 2021).

The methodological emphasis is not only classical double robustness but rate double robustness. If nuisance estimators satisfy product-rate conditions such as

$\tau$3

root-$\tau$4 asymptotic normality of the hazard-difference estimator is retained even when nuisance functions are estimated by machine learning or other slower nonparametric methods. Simulation with 500 datasets of size 1000, true $\tau$5, treatment prevalence around 40–50%, and censoring 10–30% found that both score-based estimators were nearly unbiased with good coverage when either the propensity or the outcome model was correct; flexible propensity estimation via twang often performed well under misspecification; and random survival forests improved bias and coverage for censoring adjustment under dependent censoring. In the Honolulu Heart Program / Honolulu-Asia Aging Study application, mid-life drinking increased both the hazard of later-life cognitive impairment and the hazard of death without cognitive impairment (Rava et al., 2021).

6. Causal interpretation, selection, and identifiability

Absolute hazard contrasts are often presented as more interpretable than hazard ratios, but their causal status is contested. "Bias of the additive hazard model in the presence of causal effect heterogeneity" defines the causal hazard difference

$\tau$6

under the structural model

$\tau$7

It distinguishes this target from the observed hazard difference, which, under randomization and no confounding, equals a marginal causal hazard difference among those still at risk at time $\tau$8. When there is no effect heterogeneity on the hazard scale, the observed hazard difference equals the causal hazard difference. When effect heterogeneity is present, however, survivor selection on favorable values of $\tau$9 induces bias, and the observed hazard difference becomes

$\mathrm{DAH}(\tau)=\eta_1(\tau)-\eta_0(\tau).$0

rather than the full-population average $\mathrm{DAH}(\tau)=\eta_1(\tau)-\eta_0(\tau).$1. The paper concludes that hazard differences should not be used as causal estimands unless strong assumptions are made, particularly no confounding and no effect heterogeneity on the hazard scale (Post et al., 2022).

A different response is to redefine hazard causally. "Causal Inference in the Multiverse of Hazard" contrasts the conventional conditional hazard with an interventional counterfactual process hazard

$\mathrm{DAH}(\tau)=\eta_1(\tau)-\eta_0(\tau).$2

equivalently

$\mathrm{DAH}(\tau)=\eta_1(\tau)-\eta_0(\tau).$3

Here prior survival is not merely conditioned on; it is intervened upon. In that framework, a hazard difference is interpreted as

$\mathrm{DAH}(\tau)=\eta_1(\tau)-\eta_0(\tau).$4

a controlled direct causal contrast between risks in possible worlds where prior deaths have been intervened away. This suggests a different resolution of the selection problem than additive-hazards collapsibility: rather than treating the observed risk set as the target population, it constructs a hypothetical comparable population at each time point (Lai et al., 2024).

7. Related methodologies and methodological boundaries

Several neighboring developments shape the practical use of hazard difference estimators even when they do not define one directly. "Transforming cumulative hazard estimates" shows that hazard-based estimation can be used as an intermediate step and then consistently transformed to other scales through differential equations, with covariance estimation obtained by plug-in martingale recursions. Its motivation is that hazards can have built-in selection effects that prevent simple causal interpretations, even in randomized trials, and that survival-scale or mean-survival-scale summaries may therefore be preferable targets (Ryalen et al., 2017).

Within the average-hazard literature, "Efficient and Debiased Learning of Average Hazard Under Non-Proportional Hazards" establishes pathwise differentiability and an efficient influence function for $\mathrm{DAH}(\tau)=\eta_1(\tau)-\eta_0(\tau).$5, then constructs cross-fitted, doubly robust estimators under product-rate conditions. The paper explicitly targets the log ratio of average hazards rather than the difference, but it strengthens the inferential infrastructure around $\mathrm{DAH}(\tau)=\eta_1(\tau)-\eta_0(\tau).$6 and thereby around any DAH constructed from arm-specific average hazards (Meng et al., 13 Feb 2026). The validity of the Kaplan–Meier plug-in estimator for average hazard,

$\mathrm{DAH}(\tau)=\eta_1(\tau)-\eta_0(\tau).$7

including truncation times $\mathrm{DAH}(\tau)=\eta_1(\tau)-\eta_0(\tau).$8 that are not observed event times, is defended in "Comment on 'Average Hazard as Harmonic Mean' by Chiba," which reports negligible finite-sample bias over $\mathrm{DAH}(\tau)=\eta_1(\tau)-\eta_0(\tau).$9 even with $\mathrm{DAH}(\tau)=\eta_1(\tau)-\eta_0(\tau)$00 in an exponential simulation with hazard $\mathrm{DAH}(\tau)=\eta_1(\tau)-\eta_0(\tau)$01 (Uno et al., 21 Jul 2025).

Other hazard-estimation papers remain adjacent rather than central. "A new kernel estimator of hazard ratio and its asymptotic mean squared error" is concerned with direct kernel estimation of the hazard function $\mathrm{DAH}(\tau)=\eta_1(\tau)-\eta_0(\tau)$02, not a hazard difference (Moriyama et al., 2016). "An Asymptotic Linear Representation for the Breslow Estimator" provides a nearly $\mathrm{DAH}(\tau)=\eta_1(\tau)-\eta_0(\tau)$03-remainder linearization for the Cox-model baseline cumulative hazard estimator, a result that can support comparisons of cumulative hazards but does not itself define a hazard-difference estimand (Lopuhaa et al., 2015). "Large Deviations of the Estimated Cumulative Hazard Rate" shows that cumulative hazard estimation errors are asymmetric and tend toward overestimation in small samples and low-tail-probability regimes, a feature that can plausibly affect any comparison built from estimated cumulative hazards (Hohmann, 2019).

Taken together, these literatures indicate that a hazard difference estimator is not a single universally standardized object. It is instead a family of additive hazard-scale contrasts whose most developed forms are the standardized difference in average hazard for stratified randomized studies, the HDi estimator for high-dimensional additive hazards models, and doubly robust cause-specific estimators for competing risks. Their appeal lies in absolute-scale interpretability and freedom from some of the interpretive limitations of hazard ratios, while their limitations arise from dependence on target-population definition, censoring and nuisance modeling, and unresolved causal selection issues under heterogeneity.