Clustering in Cox proportional hazards model MLM vs. sandwich estimator This question is about a paper I am reviewing, so I cannot give a lot of detail, but I can say it involves patients clustered in hospitals and a Cox proportional hazards model.
My instinct for such data would be to use multilevel modeling, since hospitals may vary on ways that are not accounted for. The authors used a robust sandwich estimator with clustering for hospitals. 
Is this a legitimate choice? What are its drawbacks and advantages vs. a multilevel model?
 A: Neglecting the clustering is, I think, usual practice when analysing multicentre clinical trials based on a time-to-event outcome. That is, the standard Cox model is used. In that case, the treatment effect has a population-averaged (marginal) interpretation, the effect being averaged over all centres. It is known that the Cox model leads to a consistent estimate of the population hazard ratio (under some mild conditions, I guess), but the standard error is not because of the correlation between the survival times. Adjustment of the standard error, though, is possible by using the jackknife, leading to some kind of sandwich estimator. The method is available in R (cf. the cluster() function to be used within coxph()). Alternatively, multilevel modelling can also be used for such type of data, as you suggest. That is, the centre effect enters the Cox model as a random effect. In survival analysis, this is called a frailty model. In the frailty model, the treatment effect has a centre-specific (conditional) interpretation. The method is also available in R (cf. the frailty() function to be used within coxph(), or coxme()).  
Both approaches are correct, provided that the hazard ratio is well interpreted (population-averaged versus centre-specific). In general, the population-averaged effect is attenuated compared to the centre-
specific effect.  I would say that the choice of one method over another depends on the interpretation we want to give to the hazard ratio. The centre-specific interpretation is, according to me, particularly relevant in the context of clinical trials as it compares “like-for-like”.

References that discuss this issue include Glidden and Vittinghoff (2004), PhD thesis by Snavely (chapter 1), and Duchateau and Janssen (2008, Chapter 3).
