clustered survival model

[Tracy Clegg]

I am trying to fit a survival model to clustered data (animals within herds). I have tried using both parametric shared frailty models using streg and a mixed effect parametric model using mestreg. Can anyone tell me how to compare these models? Do I just use an AIC? Are there any diagnostics available for the mixed effect parametric model? If so can anyone point me towards a good reference material?

Thanks

[Steve Samuels]

mestreg has a built-in comparison to the model without a random effect. In Example 1 of the Manual entry it is the line: "LR test vs. Weibull model:".

[Tracy Clegg]

Sorry I have only just seen your reply. Thank you for responding. I want to compare my model to one with a frailty term rather than one without any random effect, the random effect within my model is significant (P<0.001), I'm just not sure whether a random effect model is better than a frailty model and how to check whether a random effect model is a good fit to the data ie how can I check the residuals and outliers etc? Thanks

[Steve Samuels]

A frailty model is also random effects model. mestreg has Gaussian l random effects, whereas for stcox: (Manual section on "Cox regression with shared frailty":

Across groups, the frailties are assumed to be gamma-distributed latent random effects that affect the hazard multiplicatively, or, equivalently, the logarithm of the frailty enters the linear predictor as a random offset

The main difference between the two models is that the Cox model is semiparametric (doesn't assume a parametric survival distribution) whereas mestreg fits several parametric distributions. Your main concern for mestreg is selection of the survival distribution. Note that mestreg claims to fit a residual, but I dont know how it does that for censored observations.

If herds are your only level of clustering, then you should use plain streg, not mestreg, because streg also fits frailty models, but has two choices of frailty distributions and a larger selection of parametric survival distributions.

The bottom line answer is that you should fit a random effects model; that's the way your data are structured and you already have evidence from mestreg that there is between-herd variation beyond that expected from the parametric model.

[Tracy Clegg]

Thanks Steve, I have been using streg but I got different results from mestreg with lower AIC so thats why I was wondering how to check the mestreg model, but maybe I should just stick to the streg ones. stcox with a frailty would not even run using my data it just gave an error message about the matsize! Thanks again for your help