Fitting glms on matricies instead of dataframes?

[lendle]

I'm looking for something like R's glm.fit function, which is called by glm after the design matrix and response vector are constructed and can be called directly. This would be useful if you already have a design matrix and response, and you don't need extra features that are provided by DataFrames. This would also avoid the overhead of having a DataFrame in memory in addition to your design matrix.

It looks like it would not be too hard to provide this. In glmfit.jl, fit takes a GlmMod. Other than the ModelFrame, the GlmMod object can be constructed from a response vector and design matrix instead of a Formula and DataFrame, and it looks like the ModelFrame is only used for column names in the coef table. (I'm exactly sure what the assign field is for in a ModelMatrix, important here?)

[dmbates]

Good idea. I patterned the division into glmfit and glm on the corresponding R code (but the implementation is considerably different from that in R, thankfully).

It is a matter of determining how much information needs to be retained at what types in a type hierarchy. You mentioned the assign field, which is important if you have used Formula/Data arguments and want to evaluate analysis of variance (lm) or analysis of deviance (glm) tables. (The assign field maps columns in the model matrix to terms in the formula.)

Coming from an R background I prefer the Formula/Data specification but to others using a pre-built model matrix is just fine.

I'll propose a series of types that could allow for more general approaches. I hope that @lindahua and @andreasnoackjensen will be able to comment on such types.

[simonster]

Perhaps if a model is fit with a design matrix, we can just create a ModelMatrix with the assign field populated as 1:size(X, 2)?

[simonster]

cc @johnmyleswhite, who probably saw this anyway but might have some thoughts since he's currently working on the Formula code. Also cross-referencing JuliaStats/Roadmap.jl#11.

[simonster]

I've implemented fitting GLMs on matrices in a somewhat naive way in simonster/GLM.jl@e517ede, although that code is built on top of #56, so I'll wait for the outcome on that before I create a PR for it.

[lendle]

Thank you!