Consider the multinomial GAM with mgcv in R

library(mgcv)

multinom = gam(list(sound ~ word + s(long, lat),
                          ~ word + s(long, lat),
                          ~ word + s(long, lat),
                          ~ word + s(long, lat)
                          ~ word + s(long, lat)),
                          data=pronunciation, family=multinom(K=5))

sound (numeric 0–5) and word are 6 categories each. I need word 1 as reference level but word 1 does not feature sound 3 (i.e., 0 occurances). The model only works with optimizer="efs" and gives unrealistic results for the linear predictor for the sound missing in the reference level.

Is there a way to overcome this issue by not changing the reference level to a word that features all sound categories?

Changing the reference level to a word that features all sound categories works.