Weibull parameterizations

Our course notes (and (Klein, Moeschberger, et al. 2003)) define the Weibull hazard as: \[\lambda(t) = \gamma \alpha t^{\alpha - 1}\] Base R defines the Weibull parameterization for rweibull(n, shape=\(\alpha\), scale=\(\sigma\)) as \[\lambda(t) = \frac{\alpha}{\sigma} \left(\frac{t}{\sigma}\right)^{\alpha - 1}\] The survival package parameterizes the Weibull, with intercept=\(\mu\), scale =\(\tau\), as \[\lambda(t) = \frac{1}{\tau e^{\mu/\tau}} t^{1/\tau - 1}\] Thus, we can see that the following identities hold: \[\begin{align*} \gamma & = \frac{1}{\sigma^\alpha} \implies \sigma = \frac{1}{\gamma^{1/\alpha}} \\ \gamma & = e^{-\mu/\tau} \implies \mu = -\tau \log(\gamma) \end{align*}\] This also implies that regression coefficients from survreg are interpreted differently from the typical interpretation from a proportional hazards model. The proportional hazards Weibull model is typically written \[\gamma e^{\boldsymbol{\beta}^T\mathbf{z}_i} \alpha t^{\alpha - 1}\] But survreg parameterizes the model as \[\frac{1}{\tau e^{(\mu + \boldsymbol{\theta}^T \mathbf{z}_i)/\tau}} t^{1/\tau - 1}\] This means that: \[\begin{align*} \boldsymbol{\beta} & = -\boldsymbol{\theta} / \tau \\ \gamma & = e^{-\mu/\tau} \end{align*}\] Thus, a positive coefficient in the parametric hazard which indicates that the variable increases hazard, all else being equal, will be negative in survreg’s coefficient results and vice versa.

References

Klein, John P, Melvin L Moeschberger, et al. 2003. Survival Analysis: Techniques for Censored and Truncated Data. Vol. 1230. Springer.