Thread by @ericJpedersen

@bolkerb @ucfagls @noamross @millerdl you still specify it in the first term; y ~ s(time) + s(time, id, bs = "fs") will work. gamm4 doesn't actually differentiate between terms specified with smooths and random effects using (1|id) form; it transforms smooths into random effects under the hood

@bolkerb @ucfagls @noamross @millerdl ?mgcv::smooth.construct.fs.smooth.spec has some details/caveats about specifying this type of term in gamm or gamm4

@bolkerb @ucfagls @noamross @millerdl Also: when fitting models w/ both a main effect and a group level term (e.g. s(time) + s(time, id, bs="fs")) using the gam() function, it's really important to use method = "REML" or "ML"; using the default GCV criteria often leads to very flat main effects

@bolkerb @ucfagls @noamross @millerdl Final thought (otherwise I'll have to start tweet-storming this): if you don't want to use gamm4 due to harder down-stream issues (which I agree, can be tricky): using bam can also give a substantial speedup relative to gam w/ large #'s of groups