From 20a867acfcb20fd5d83011c80b2295bd3d4618e0 Mon Sep 17 00:00:00 2001 From: MikeDMorgan Date: Tue, 23 Jul 2024 13:48:34 +0100 Subject: [PATCH] more detail on solver recommendations --- vignettes/milo_glmm.Rmd | 3 +++ 1 file changed, 3 insertions(+) diff --git a/vignettes/milo_glmm.Rmd b/vignettes/milo_glmm.Rmd index 09dd29d..49fe360 100644 --- a/vignettes/milo_glmm.Rmd +++ b/vignettes/milo_glmm.Rmd @@ -156,6 +156,9 @@ solver, the constrained HE-NNLS re-frames the parameter estimation as a non-nega knock-on effect that the model solving might be close to singular - Milo will generate a warning for each nhood where this is the case. We therefore recommend you set `warning=FALSE` if running in an Rstudio notebook, or wrap the function call in `suppressWarnings()`. +As a guideline, we recommend that you use `solver="Fisher"` in `testNhoods` with a GLMM - if the system is well conditioned then the performance is superior to the HE/HE-NNLS +approaches and faster. + As we are estimating variances, we also want these to be unbiased, so we use restricted maximum likelihood (`REML=TRUE`). Note that NB-GLMMs are by construction slower than NB-GLMs as there are additional matrix inversion steps that don't scale very nicely. While we have made every effort to reduce the computational burden we are still limited by the bounds of matrix algebra!