I have three questions concerning accelerated failure time models (AFT), one statistical, one regarding how to implement these models in R, and one related to finding out information about what R is doing. In short my questions are;

1) What is the relationship between the Gumbel and Weibull distributions?

2) How can I use (1) to simulate a AFT model using Gumbel errors and fit this model in R?

3) Where can I find formulae regarding exactly what distribution specification R is using when fitting a Weibull distribution, and exactly what model is being fitted?

I am having difficulties implementing 2), which may be due to my mis-understanding of 1), but which I can't seem to resolve due to 3). Question (3) is self-explanatory but (2) and (3) require more detail;

1) It seems a standard result that if $U\sim Gumbel(\alpha,\beta)$ then $V:=\exp(U)\sim Weibull(\lambda,\sigma)$ where $\alpha=\log(\sigma)$ and $\beta=1/\lambda$. However using the definition of the Gumbel and Weibull distributions commonly used (for example Wikipedia), when I do the derivation I can only get the transformation $V':=1/\exp(U)$ to give this result but where $\alpha=-\log(\sigma)=\log(1/\sigma)$. Thus can anyone confirm or not any knowledge of this relationship, or perhaps suggest where I have gone wrong (for brevity in the first instance I do not supply the detail)?

2) My approach is to use

$Y_{i}:=\log\left(\frac{1}{T_{i}}\right)=\beta_{0} + \beta_{1}x_{i} + e_{i},\hspace{20pt}i=1,...,N$,

as a data-generating mechanism for the logarithm of the time to event where $e_{i}\sim Gumbel(\alpha,\beta)$, where $i$ indexes subjects, $x_{i}$ is a scalar covariate, and the $e_{i}$ are all independent. I choose $\alpha=-\beta*c$ where $c$ is Euler's constant in order to ensure $E[e_{i}]=\alpha+c\beta=0$. This gives

$Y_{i}\sim Gumbel(\beta_{0} + \beta_{1}x_{i}+\alpha,\beta)$,

and using (1)

$T_{i}\sim Weibull(1/\beta,\exp[-(\beta_{0} + \beta_{1}x_{i}+\alpha)])$

The code at the end of this post is a minimal working example of this approach, where I censor subjects if $T_{i}$ is greater than the median of the $N$ theoretical medians of $\{T_{1},...,T_{N}\}$, and create an event if not. This gives $50-60\%$ of subjects being censored, the balance having events, and I interpret this to be right-censoring (say the end of a study).

I then use the survreg package in R to try to fit an AFT to $Y_{i}$ using the "dist=weibull" option. Using $\beta_{0}=-10$ and $\beta_{1}=0$ gives the following output

enter image description here

which gives the intercept being positive when it should be negative. Things get worse when using $\beta_{0}=-10$ and $\beta_{1}=2$ which gives the following output

enter image description here

which is obviously wrong. Thus I would like to know what model I am actually fitting when using the survreg package.

The code below is a minimal working example (apart from some code to produce plots which can be helpful).

# minimal working example
#params of the gumbel(alpha_gum,beta_gum) distribution so that E[X]=0
beta_gum = 1/5 #
alpha_gum = -(beta_gum*(-digamma(1)))

#calc the mean of the errors using Eulers constant as the negative of the diagamma function
mu_e = alpha_gum + (beta_gum*(-digamma(1)))#should be 0   

# regression parameters
intercept = -10;
beta1 =0;
#beta1 =2;

#number of subjects

# vector of uniform random numbers
U = runif(N)

#vector for gumbel distributed errors
e = matrix(,nrow=N,ncol=1)

# log of time to event, time to event, mean LTTE
logTTE = matrix(,nrow=N,ncol=1)
Xbeta_LTTE= matrix(,nrow=N,ncol=1)
TTE = matrix(,nrow=N,ncol=1)
TTE2 = matrix(,nrow=N,ncol=1)

#censoring variable
censor = matrix(,nrow=N,ncol=1)

#simulate covariate from a normal distribution
covariate1 = rnorm(N,6,4)

for (i in 1:N)
  # calculate the Gumbel RV from the inverse CDF of the Gumbel
  e[i,1] = alpha_gum + (-beta_gum*log(-log(U[i])))

  #generate the mean log TTE  
  Xbeta_LTTE[i,1] = intercept + (beta1*covariate1[i])

  #add the errors
  logTTE[i,1] = Xbeta_LTTE[i,1] + e[i,1]  

  #transform to raw time variable - this is a Weibull dist
  #TTE_i ~ Weibull[1/beta_gum , exp(-[logTTE_i+alpha_gum])
  TTE[i,1] = 1/exp(logTTE[i,1])      

#calc the median the TTE given TTE ~ Weibull[1/beta_gum , exp(-[X_i^t*beta+alpha_gum])
lambda_array = exp(-(Xbeta_LTTE + alpha_gum + (beta_gum*(-digamma(1)))))
kappa = 1/beta_gum
median_TTE_array = (lambda_array)*(log(2)^(1/kappa))
median_TTE = median(median_TTE_array)

# calculate the censoring variable
for (i in 1:N)
  #censoring: subjects with a TTE >median_TTE will be right-censored
  #i.e. study ends at T=median_TTE say
  if (TTE[i,1]>median_TTE)

#calculate the percentage of censored subjects and do a plot
pc_censored = sum(censor)/N

#fit AFT model
datframe_surv = data.frame(covariate1)

m.surv = Surv(TTE2,censor,type="right")
m.surv.fit = survreg(m.surv~covariate1,dist="weibull",scale=1)
sum = summary(m.surv.fit)

###################  plots ########################

#histogram of the errors - gumbel dist
h1 = hist(e, breaks=50, plot=FALSE) 

#histogram of the mean log TTE - gumbel dist
h2 = hist(logTTE, breaks=50, plot=FALSE) 

#histogram of the fixed means
h3 = hist(Xbeta_LTTE, breaks=50, plot=FALSE) 

#histogram of the TTE - weibul dist
h4 = hist(TTE, breaks=50, plot=FALSE) 

#calc the mean of the log TTE given logTTE ~ Gumbel(X_i^t*beta+alpha_gum,beta_gum)
median_logTTE_array = Xbeta_LTTE + alpha_gum - (beta_gum*(log(log(2))))
median_logTTE = median(median_logTTE_array)

#calc the means
ylim_h1 = c(min(h1$density),max(h1$density) )
xlim_h1 = c(mu_e,mu_e )

ylim_h2 = c(min(h3$density),max(h3$density) )
xlim_h2 = c(median_logTTE,median_logTTE )

ylim_h3 = c(min(h3$density),max(h3$density) )
xlim_h3 = c(mean(Xbeta_LTTE),mean(Xbeta_LTTE) )

ylim_h4 = c(min(h4$density),max(h4$density) )
xlim_h4 = c(median_TTE,median_TTE )


plot(h1$mids,h1$density,col='red',main="errors - gumbel dist",xlab="errors (log time)")

plot(h3$mids,h3$density,col='red',main="mean log TTE (X*beta) - fixed",xlab="mean log TTE (log time)")

plot(h2$mids,h2$density,col='red',main="log TTE - gumbel dist",xlab="log TTE (log time)")

plot(h4$mids,h4$density,col='red',main="TTE - Weibull dist",xlab="TTE (time)")

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.