4
$\begingroup$

Could you recommend an R package for estimating a (frequentist) multilevel Weibull regression model?

I need to model random intercepts, random slopes, as well as a cross-classified structure.

UPDATE: It seems like there is currently no "easy" solution for that. I decided to leave it for now by estimating a multilevel discrete hazard model with glmer and a multilevel Cox PH model with coxme (proposed by EddieMcGoldrick). With regards to the latter, I have still to figure out if implementing a cross-classified structure is possible.

$\endgroup$

2 Answers 2

2
$\begingroup$

You can add frailty terms in the survreg function, but that will not help you if you want random slopes.

If you don't need to use Weibull, you could use the coxme package for fitting a Cox regression model with random intercepts and slopes. Or a discrete time approach using a mixed effects logit with the lmer package.

$\endgroup$
1
  • $\begingroup$ Unfortunately, that is true and that is also the reason for my post. Random slopes are essential for testing my hypotheses. I have already estimated the model in glmer (assuming discrete data) and it works like a charm. However, the Weibull model would be really my preferred choice. I had a look at Pinheiro/Bates (2000) and wonder if it is possible to estimate the model with nlme? However, I haven't yet found an example which would proof me right. $\endgroup$
    – majom
    Aug 11, 2012 at 0:10
2
$\begingroup$

I am not aware of any R package that can do this. You could possibly coerce the lme4 or nlme packages to do it, if you wrote your own functions for that particular family. Another (related) possibility is to use AD Model Builder (ADMB). This is a flexible optimizer and forms the basis of the glmmADMB package in R (which although it implements quite a few families, does not implement the Weibull distribution).

Perhaps you can explain why this must use a frequentist approach? If you really want that, another more "canned" approach could possibly be done in SAS using proc nlmixed which allows random effects models where you specify your own (log) likelihood function. I work out an extensive example of this for random coefficient zero-inflated Poisson models here. That might not quite work as I have since updated the dataset that is based on (it was a simulated multilevel dataset).

$\endgroup$
2
  • $\begingroup$ Thanks for your answer. All models I have estimated so far for this project are frequentist models. Thus, I have a strong preference to stick with the frequentist approach and to avoid confusing reviewers anymore. I see that Bayes would make my life in this case probably easier. I'm rather surprised that I don't find too much information on such a model (and of course implementations). $\endgroup$
    – majom
    Aug 11, 2012 at 3:41
  • $\begingroup$ You can use the data cloning approach to implement frequentist statistical analysis by using MCMC sampling methods. For more details see this [paper] (rd.springer.com/article/10.1007/s11222-011-9254-z) $\endgroup$ Aug 11, 2012 at 9:19

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.