RegressionFormulae

Extended StatsModels.jl @formula syntax for regression modeling.

Note that the functionality in this package is very new: please verify that the resulting schematized formulae and model coefficient (names) are what you were expecting, especially if you are combining multiple "advanced" formula features.

Supported syntax

Nesting syntax

a / b expands to a + fulldummy(a) & b.

Numeric constants are special cased so that / performs division, making it possible to e.g. convert time to speed in the formula:

julia> fit(MyModelType, @formula(time_in_milliseconds / 1000 ~ 1 + x), my_data)

Raising terms to a power

Generate all main effects and interactions up to the specified order. For instance, (a+b+c)^2 generates a + b + c + a&b + a&c + b&c, but not a&b&c.

NB: The presence of interaction terms within the base will result in redundant terms and is currently unsupported.

Approach

Extended syntax is supported at two levels. First, RegressionFormulae.jl defines apply_schema methods that capture calls within a @formula to the special syntax (^, /, etc.). Second, we define methods for the corresponding functions in Base (Base.:(^), Base.:(/), etc.) for arguments that are <:AbstractTerm which implement the special behavior, returning the appropriate terms. This allows the syntax to be used both within a @formula and for constructing terms programmatically at run-time.

If using apply_schema directly, please note that you need to pass an appropriate model type as context. Currently, the extensions here are defined for StatsAPI.RegressionModel and subtypes:

f = apply_schema(f, s, RegressionModel)