Autoscaling for [the priors on] association parameters is based on the range (if 2 values in x) or SD (if more than 2 values in x, where x is the association term i.e. the implicit "covariate" in the event submodel).
However, if an "etaslope" association term is used, the SD of the slopes may be really small in some cases, and therefore the scale for the prior may be large. This is seen for example when running a model with "etaslope" on the pbc*_subset data --> the large scale on the prior makes model estimation difficult (Stan takes much longer), and the estimates appear pretty unstable (however there is only 40 patients in that dataset). But, my guess is that this would only be an issue when there is very little information in the slopes that can be used to inform the log hazard ratio (in other words, it only makes model estimation difficult when "etaslope" is likely to be an inappropriate association structure to be using for the data at hand anyway). But, it's something to be weary of for now, and to potentially assess using a variety of different datasets.
An alternative is to always specify the prior explicitly when using an "etaslope" association structure, but users should be able to rely on the defaults, so the autoscaling needs to produce something sensible.
Autoscaling for [the priors on] association parameters is based on the range (if 2 values in x) or SD (if more than 2 values in x, where x is the association term i.e. the implicit "covariate" in the event submodel).
However, if an "etaslope" association term is used, the SD of the slopes may be really small in some cases, and therefore the scale for the prior may be large. This is seen for example when running a model with "etaslope" on the pbc*_subset data --> the large scale on the prior makes model estimation difficult (Stan takes much longer), and the estimates appear pretty unstable (however there is only 40 patients in that dataset). But, my guess is that this would only be an issue when there is very little information in the slopes that can be used to inform the log hazard ratio (in other words, it only makes model estimation difficult when "etaslope" is likely to be an inappropriate association structure to be using for the data at hand anyway). But, it's something to be weary of for now, and to potentially assess using a variety of different datasets.
An alternative is to always specify the prior explicitly when using an "etaslope" association structure, but users should be able to rely on the defaults, so the autoscaling needs to produce something sensible.