Warning when attempting to fit a spatio-temporal model

[JoseCansec]

I have been trying to fit a spatio-temporal model using bru. I am observing which temporal autocorrelation best explains the data and I have tried weeks, days (they both work fine) but hours gives me two errors. Heres my code and the error:

ss_commercial2.3$Hour <- ss_commercial2.3$hour - min(ss_commercial2.3$hour) + 1 Hour <- ss_commercial2.3$Hour nhour <- length(unique(ss_commercial2.3$Hour))

---------------------------------------------------------

---- Spatio-temporal model by hour---

cmp.krill.cd.h = PRC_NASC ~ Intercept+ myspde(map = coordinates, group=Hour, ngroup=nhour, model = Hosts.spde, mesh = mesh.not.com, control.group=list(model="ar1"))

------------- Model fitting

bru.krill.cd.h = bru(cmp.krill.cd.h,
family = "gamma", data= ss_commercial2.3)

The warning I get is this one:

Warning in (function (formula, family = "gaussian", contrasts = NULL, data, : f(myspde, ...): Number of unused groups >= the number of groups used: 488 >= 1, this is usually a misspesification

Additionally, sometimes I get the following error:

Error in sparseMatrix(i = i, j = (1L + A.loc@j + n.spde * (group.i - 1L) + : all(dims >= dims.min) is not TRUE

Please help!

[finnlindgren]

Hi, Difficult to say without seeing the data, but I suspect the issue is in the contents of Hour. unique(Hour) should contain the integers 1,2,3,4,... etc

[JoseCansec]

The data consists consists of acoustic data (NASC) per cell and time lapse (hour, day, week). I have the time variables data with 1.... n as integers and the codes work when working with days and weeks either for presence absence or for conditional density. It works when fitting the model with hours as a random effect for the presence-absence of one data set but fails for conditional density. In the other data set, it fails on both presence-absence and conditional density. Could it be that the data set is too big and my computer doesn´t have the capacity?

[finnlindgren]

The errors are unrelated to the size of the data. What precisely did you mean by "sometimes", for the spareMatrix error? I also notice the family="gaussian", despite setting family="gamma" in the bru() call. Are you able to share the data and code via email? Or generate a small synthetic data set of the same structure that you can share, with the code that generates the warnings/errors? But first, make sure to let me know what inlabru version you're running, and if possible, also check if you get the warnings/errors when using the current devel version from github.

[JoseCansec]

I apologize for not having responded but I had to work on another project during this time. Regarding the spareMatrix error, it appears when defining my temporal variable as it is in my data frame without applying this:

ss_commercial2.3$Hour <- ss_commercial2.3$hour - min(ss_commercial2.3$hour) + 1

So the temporal structure is set as follows:

Hour <- ss_commercial2.3$hour nhour <- length(unique(ss_commercial2.3$hour))

I can share the code and data, just let me know what email to send it.

Best regards, Jose

[finnlindgren]

If you have hourly data, so that the model should be build on an hourly basis, then the Hour values should start at 1, and nhour should be max(ss_commercoal2.3$Hour) (note the uppercase Hour).

You can send the code that has the problem to finn.lindgren@gmail.com

[finnlindgren]

I couldn't tell from my email collection if we had solved this, but since the package code has changed considerably, it would need to be re-reported if it's still an issue with it.