Closed fpavone closed 9 months ago
Hello Federico,
Thanks for your message, you really found one bug. However, the bug was not in graph_lme()
. It was in sample_spde()
. I have now fixed it. Can you confirm that it is working now? Please, install the most recent version of the package with:
remotes::install_github("davidbolin/metricgraph", ref = "devel")
Also, please note that the parameters you estimate when using "isoCov" do not need to be "close" to the ones you chose, since you sampled them from the Whittle-Matern model. This should be true, however, when you estimate using the "WM1" model.
I will close the issue. If you still have problem, please send another message and we can reopen the issue.
Best, Alexandre
The function
graph_lme(...)
seems to have a different behavior depending whether the observation locations (PtE
) are ordered (according to the edge indexing) or shuffled. In particular, when data are shuffled the model seems to be not correctly fitted, resulting in inconsistent estimates of parameterstau
andkappa
, for both theisoCov
and theWM1
models.I think this might be an unexpected behavior since I haven't found any explicit recommendation to order the data locations in the documentation.
An example: