LOO fails for data with only one observation

[paul-buerkner]

Reproducible example with the dev version of brms:

fit1 <- brm(count ~ log_Age_c, epilepsy)
LOO(fit1, newdata = epilepsy[1, ])
> Error in colSums2(x) : 
>   Argument 'dim' must be an integer vector of length two. 

I think there is a drop = FALSE argument missing somewhere when indexing observations. If possible it would be amazing if this could be fixed before releasing loo 2.0.

[jgabry]

Thanks Paul, I’ll look into it before releasing.

…

On Thu, Apr 5, 2018 at 8:35 AM Paul-Christian Bürkner < ***@***.***> wrote: Reproducible example with the dev version of brms: fit1 <- brm(count ~ log_Age_c, epilepsy) LOO(fit1, newdata = epilepsy[1, ])> Error in colSums2(x) : > Argument 'dim' must be an integer vector of length two. I think there is a drop = FALSE argument missing somewhere when indexing observations. If possible it would be amazing if this could be fixed before releasing loo 2.0. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#71>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AHb4QzzdDsfsvi4dRvssD60QOT9Ydwzsks5tlg97gaJpZM4TIXyT> .

[jgabry]

Should be fixed now. You had the correct diagnosis. (Just a heads up that the reported SEs are going to be NA in this case since they rely on a variance calculation.)

[paul-buerkner]

Thanks! I currently have this one-observation case in brms when I compute LOO for models with non-conditionally-independent observations such as for SAR models. I remember that Aki wrote something about how to split such dependent observations and produce a LOO SE regardless, but I can't remember where I have seen it.

[avehtari]

@paul-buerkner can you email me with details of that SAR model case? I can't remember the answer, but I guess I'll remember or figure it out when I see the details.

[paul-buerkner]

Sure, will do!