Comparison

Cox Proportional Hazards vs Deep Recurrent Survival

Classical semi-parametric vs neural time-to-event models

The short answer

Cox proportional hazards is interpretable, fast to fit, and sufficient when the hazard ratio between groups is approximately constant over time. Deep recurrent survival models (DeepSurv, DRSA) relax the proportionality assumption and capture time-varying covariate effects, at the cost of interpretability and data requirements. Use Cox as the first model; escalate only when residual diagnostics reject proportionality.

Cox PH is the workhorse of survival analysis. It assumes the hazard ratio between any two covariate profiles is constant over time — the proportional hazards assumption — and estimates coefficients via partial likelihood. When the assumption holds, Cox produces interpretable hazard ratios that map cleanly to business decisions.

Deep recurrent survival models (DeepSurv, DRSA, Nnet-survival) replace the linear covariate combination with a neural network and relax the proportional hazards assumption. They can capture non-proportional effects (e.g., a covariate whose risk contribution changes over the customer lifecycle) and time-varying inputs, but require more data, more compute, and produce models whose individual parameters are not directly interpretable.

At a glance

DimensionCox Proportional HazardsDeep Recurrent Survival
Hazard assumptionConstant ratio (proportional)Non-proportional allowed
InterpretabilityHazard ratios per covariateSHAP / partial dependence only
Data requirementThousands of eventsTens of thousands or more
ComputeSeconds on laptopGPU for training
Handles time-varying covariatesExtended Cox onlyNatively
Concordance (typical)0.70–0.780.72–0.83

Use Cox Proportional Hazards when

  • First model; always start here
  • When you need coefficients you can explain to stakeholders
  • Smaller datasets where deep learning overfits
  • Regulated industries requiring model transparency

Use Deep Recurrent Survival when

  • Proven proportionality violations in Cox residuals
  • Subscription businesses with complex billing-cycle effects
  • Large datasets with time-varying covariates

Deeper reading

Related concepts