I am slightly confused about the way accelerated failure time models work in comparison to proportional hazard models.

  1. Is it a correct statement that while the PH model assumes covariates multiply the hazard function (hence proportional hazards), the covariates multiply the survival function in an AFT model?

  2. Is $S(t,x)=S_0(te^{x'\beta})$ the general expression for all AFT models? If $e^{x'\beta}=2$ (i.e. the acceleration factor), doesn't that mean the function is stretched relative to the baseline, rather than "accelerated"? If $e^{x'\beta}$ is the acceleration factor then shouldn't the function divide by this term instead of multiplying by it?

  3. I believe a Weibull model acts as both an AFT and PH model simultaneously. Does this mean that if $e^{x'\beta}=2$, the hazard for somebody with that set of covariates is multiplied by 2, AND the survival function is compressed by a factor of 2?

Any help is appreciated!


1 Answer 1

  1. For the PH model, the covariates appear in the exponential that multiplies the baseline hazard:

$$ h_\text{PH}(t|\mathbf x_i, \boldsymbol\theta,\boldsymbol\beta)=h_0(t|\boldsymbol\theta)\exp\left\{\mathbf x_i^\mathsf T\boldsymbol\beta\right\}.$$

In an AFT model, they multiply the time and the baseline hazard:

$$h_\text{AFT}(t|\mathbf x_i, \boldsymbol\theta,\boldsymbol\beta)=h_0\left(t\{\exp\left\{\mathbf x_i^\mathsf T\boldsymbol\beta\right\}|\boldsymbol\theta\right) \exp\left\{\mathbf x_i^\mathsf T\boldsymbol\beta\right\} .$$

  1. As you can see from the previous formulation, the exponential of the linear predictor affects both the time level and baseline hazard level. The AFT model can be reformulated as a log linear regression model. So, the covariates have an interpretation in terms of how the affect the survival time (see: Survival Analysis: Parametric Models).

  2. If the baseline hazard is Weibull (i.e. $\propto t^{\nu-1}$), then the AFT = PH, as the linear predictor affects simultaneously the baseline hazard and the time scale. However, this happens at different scales (see Survival Analysis: Parametric Models). To get the interpretation in terms of the AFT model, you need to re-parameterize the model and interpret in the corresponding scale. Note that $e^{x^\mathsf T\beta}$ affects both scales in the AFT model. You can find information about how to fit PH and AFT models in R at Example: Regression .


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.