Add branch for list of families

[ericward-noaa]

Per conversation with @James-Thorson-NOAA , it's worth adding a branch for cases when (1) multiple datasets are being passed in, and (2) we want to pass in a list of families that may vary by dataset.

We have this branch here for dealing with field observations of eDNA, and lab samples
#292

But I think the goal is to extend that to have multiple kinds of field observations (e.g. eDNA, line transect, acoustic surveys)

On the TMB side, family is passed in a vector so would be compatible

sdmTMB/src/sdmTMB.cpp

Line 218 in d9482b9

DATA_IVECTOR(family);

But lots of decisions to make on data handling , etc

[James-Thorson-NOAA]

Yeah, I still think the tinyVAST user-interface (passing a named list of families that match a specified column of data) is pretty straightforward, and it gets parsed here to a IVECTOR for family and link on the CPP side as you have in that referenced part of the sdmTMB.cpp code. But I'm also curious to see if y'all come up with a better option that I can then adapt to tinyVAST

[seananderson]

I had started something in this integrated branch. I got as far as turning most of the family and link parts into vectors/arrays. My plan was to adopt the tinyVAST interface. An alternative would be a list of families matching the rows of data. I was originally thinking that could be awkward as a user:

replicate(2, gaussian(), simplify = FALSE)
lapply(1:2, \(x) gaussian())
c(replicate(2, gaussian(), simplify = FALSE), replicate(3, binomial(), simplify = FALSE))

although this kind of approach wouldn't be too bad:

d <- data.frame(data_type = c(rep(1, 2), rep(2, 1)))
families <- list(gaussian(), binomial())
families[d$data_type]

#> [[1]]
#> 
#> Family: gaussian 
#> Link function: identity 
#> 
#> 
#> [[2]]
#> 
#> Family: gaussian 
#> Link function: identity 
#> 
#> 
#> [[3]]
#> 
#> Family: binomial 
#> Link function: logit

or

d <- data.frame(data_type = c(rep("catch", 2), rep("presence", 1)))
families <- list(catch = gaussian(), presence = binomial())
families[d$data_type]
#> $catch
#> 
#> Family: gaussian 
#> Link function: identity 
#> 
#> 
#> $catch
#> 
#> Family: gaussian 
#> Link function: identity 
#> 
#> 
#> $presence
#> 
#> Family: binomial 
#> Link function: logit

For reference, the tinyVAST interface is like this:

.family <- list(
  "Encounter" = binomial(link = "cloglog"),
  "Count" = poisson(link = "log"),
  "Biomass_KG" = tweedie(link = "log")
)
data$data_type <- c("Encounter", "Encounter", ... )
...
tinyVAST(
  ...
  family = .family,
  distribution_column = "data_type"
)

Which isn't that different from above. It means the user supplies the distribution column itself instead of a list of families that could be created with the distribution column.

At this point I don't see a major advantage either way and all else equal I think we should follow the tinyVAST format for consistency.

[seananderson]

Just a note that I have this working in the 'integrated' branch now using the same user-facing interface as in tinyVAST with a distribution_column and a named family list. I still need to expand the unit tests and fix the model summary:

• update print/summary to place nicely with integrated models
• expand unit tests including tests with offsets
• via existing unit tests: make sure existing functionality hasn't broken
• do some speed tests to make sure this hasn't regressed