`gratia::assemble()` may be handling pattern matching incorrectly

[StefanoMezzini]

If I fit a model with offsets of log-transformed variables specified via stats::offset(), I get a warning for every smooth term that also contains log-transformed variables. I'm not sure that this is actually causing downstream issues, but I think there is potential for them.

The warnings do not occur if I specify the offset via the offset argument of gam() (or bam()).

library(mgcv)

d <- gamSim(n = 100)
d$logz <- runif(100)

# specifying offset via `offset()` produce a warning for each `s(log( ))`
m <- gam(y ~ offset(log(x0)) + s(log(x1)) + s(log(x2)) + log(x3) + s(logz),
         data = d)
gratia:::draw.gam(m, parametric = TRUE)
# Warning messages:
# 1: In grep(p.terms[!ind], m.terms[off]) :
#   argument 'pattern' has length > 1 and only the first element will be used
# 2: In grep(p.terms[!ind], m.terms[off]) :
#   argument 'pattern' has length > 1 and only the first element will be used
# 3: In grep(p.terms[!ind], m.terms[off]) :
#   argument 'pattern' has length > 1 and only the first element will be used
gratia:::assemble.gam(m, parametric = TRUE)
# ── Assembly of 4 plots ──
# 
# 1. s(log(x1))
# 2. s(log(x2))
# 3. s(logz)
# 4. log(x3)
# 
# Warning messages:
# 1: In grep(p.terms[!ind], m.terms[off]) :
#   argument 'pattern' has length > 1 and only the first element will be used
# 2: In grep(p.terms[!ind], m.terms[off]) :
#   argument 'pattern' has length > 1 and only the first element will be used
# 3: In grep(p.terms[!ind], m.terms[off]) :
#   argument 'pattern' has length > 1 and only the first element will be used

# specifying `offset` argument does not produce warnings
m <- gam(y ~ s(log(x1)) + s(log(x2)) + log(x3) + s(logz), data = d, offset = log(x0))
gratia:::draw.gam(m, parametric = TRUE)
gratia:::assemble.gam(m, parametric = TRUE)
# ── Assembly of 2 plots ──
# 
# 1. s(log(x1))
# 2. s(log(x2))

The plot produced in all four instances is the same:

[gavinsimpson]

Thanks @StefanoMezzini I'll take a look; I've been putting this off for a bit as NSE is a pain and Simon makes it harder than necessary to recreate the data as he resets the environments of the formula to avoid huge model objects with large data. It's why gratia isn't great at handling models with terms involving function calls.

Guess I need to bite that bullet.

[StefanoMezzini]

no problem! it hasn't been much of an issue for now since I can just use the offset argument, but I thought I would flag it in case you weren't aware of the bug. I appreciate all the work you've put into {gratia}, and I've been finding it quite useful.