Modeling categorical data

To convert categorical data into a numerical representation suitable for modeling, StatsModels implements a variety of contrast coding systems. Each contrast coding system maps a categorical vector with $k$ levels onto $k-1$ linearly independent model matrix columns.

The following contrast coding systems are implemented:

How to specify contrast coding

The default contrast coding system is DummyCoding. To override this, use the contrasts argument when constructing a ModelFrame:

mf = ModelFrame(@formula(y ~ 1 + x), df, contrasts = Dict(:x => EffectsCoding()))

To change the contrast coding for one or more variables in place, use

StatsModels.setcontrasts!Function
setcontrasts!(mf::ModelFrame; kwargs...)
setcontrasts!(mf::ModelFrame, contrasts::Dict{Symbol})

Update the contrasts used for coding categorical variables in ModelFrame in place. This is accomplished by computing a new schema based on the provided contrasts and the ModelFrame's data, and applying it to the ModelFrame's FormulaTerm.

Note that only the ModelFrame itself is mutated: because AbstractTerms are immutable, any changes will produce a copy.

source

Interface

StatsModels.AbstractContrastsType

Interface to describe contrast coding systems for categorical variables.

Concrete subtypes of AbstractContrasts describe a particular way of converting a categorical data vector into numeric columns in a ModelMatrix. Each instantiation optionally includes the levels to generate columns for and the base level. If not specified these will be taken from the data when a ContrastsMatrix is generated (during ModelFrame construction).

Constructors

For C <: AbstractContrast:

C()                                     # levels are inferred later
C(levels = ::Vector{Any})               # levels checked against data later
C(base = ::Any)                         # specify base level
C(levels = ::Vector{Any}, base = ::Any) # specify levels and base

Arguments

  • levels: Optionally, the data levels can be specified here. This allows you to specify the order of the levels. If specified, the levels will be checked against the levels actually present in the data when the ContrastsMatrix is constructed. Any mismatch will result in an error, because levels missing in the data would lead to empty columns in the model matrix, and levels missing from the contrasts would lead to empty or undefined rows.
  • base: The base level may also be specified. The actual interpretation of this depends on the particular contrast type, but in general it can be thought of as a "reference" level. It defaults to the first level.

Contrast coding systems

  • DummyCoding - Code each non-base level as a 0-1 indicator column.
  • EffectsCoding - Code each non-base level as 1, and base as -1.
  • HelmertCoding - Code each non-base level as the difference from the mean of the lower levels
  • SeqDiffCoding - Code for differences between sequential levels of the variable.
  • HypothesisCoding - Manually specify contrasts via a hypothesis matrix, which gives the weighting for the average response for each level
  • StatsModels.ContrastsCoding - Manually specify contrasts matrix, which is directly copied into the model matrix.

The last two coding types, HypothesisCoding and StatsModels.ContrastsCoding, provide a way to manually specify a contrasts matrix. For a variable x with k levels, a contrasts matrix M is a k×k-1 matrix, that maps the k levels onto k-1 model matrix columns. Specifically, let X be the full-rank indicator matrix for x, where X[i,j] = 1 if x[i] == levels(x)[j], and 0 otherwise. Then the model matrix columns generated by the contrasts matrix M are Y = X * M.

The hypothesis matrix is the k-1×k matrix that gives the weighted combinations of group mean responses that are represented by regression coefficients for the generated contrasts. The contrasts matrix is the generalized pseudo-inverse (e.g. LinearAlgebra.pinv) of the hypothesis matrix. See HypothesisCoding or Schad et al. (2020) for more information.

Extending

The easiest way to specify custom contrasts is with HypothesisCoding or StatsModels.ContrastsCoding. But if you want to actually implement a custom contrast coding system, you can subtype AbstractContrasts. This requires a constructor, a contrasts_matrix method for constructing the actual contrasts matrix that maps from levels to ModelMatrix column values, and (optionally) a coefnames method:

mutable struct MyCoding <: AbstractContrasts
    ...
end

contrasts_matrix(C::MyCoding, baseind, n) = ...
coefnames(C::MyCoding, levels, baseind) = ...

References

Schad, D. J., Vasishth, S., Hohenstein, S., & Kliegl, R. (2020). How to capitalize on a priori contrasts in linear (mixed) models: A tutorial. Journal of Memory and Language, 110, 104038. https://doi.org/10.1016/j.jml.2019.104038

source
StatsModels.ContrastsMatrixType

An instantiation of a contrast coding system for particular levels

This type is used internally for generating model matrices based on categorical data, and most users will not need to deal with it directly. Conceptually, a ContrastsMatrix object stands for an instantiation of a contrast coding system for a particular set of categorical data levels.

If levels are specified in the AbstractContrasts, those will be used, and likewise for the base level (which defaults to the first level).

Constructors

ContrastsMatrix(contrasts::AbstractContrasts, levels::AbstractVector)
ContrastsMatrix(contrasts_matrix::ContrastsMatrix, levels::AbstractVector)

Arguments

  • contrasts::AbstractContrasts: The contrast coding system to use.
  • levels::AbstractVector: The levels to generate contrasts for.
  • contrasts_matrix::ContrastsMatrix: Constructing a ContrastsMatrix from another will check that the levels match. This is used, for example, in constructing a model matrix from a ModelFrame using different data.
source

Contrast coding systems

StatsModels.DummyCodingType
DummyCoding([base[, levels]])
DummyCoding(; base=nothing, levels=nothing)

Dummy coding generates one indicator column (1 or 0) for each non-base level.

If levels are omitted or nothing, they are determined from the data by calling the levels function on the data when constructing ContrastsMatrix. If base is omitted or nothing, the first level is used as the base.

Columns have non-zero mean and are collinear with an intercept column (and lower-order columns for interactions) but are orthogonal to each other. In a regression model, dummy coding leads to an intercept that is the mean of the dependent variable for base level.

Also known as "treatment coding" or "one-hot encoding".

Examples

julia> StatsModels.ContrastsMatrix(DummyCoding(), ["a", "b", "c", "d"]).matrix
4×3 Matrix{Float64}:
 0.0  0.0  0.0
 1.0  0.0  0.0
 0.0  1.0  0.0
 0.0  0.0  1.0
source
StatsModels.EffectsCodingType
EffectsCoding([base[, levels]])
EffectsCoding(; base=nothing, levels=nothing)

Effects coding generates columns that code each non-base level as the deviation from the base level. For each non-base level x of variable, a column is generated with 1 where variable .== x and -1 where variable .== base.

EffectsCoding is like DummyCoding, but using -1 for the base level instead of 0.

If levels are omitted or nothing, they are determined from the data by calling the levels function when constructing ContrastsMatrix. If base is omitted or nothing, the first level is used as the base.

When all levels are equally frequent, effects coding generates model matrix columns that are mean centered (have mean 0). For more than two levels the generated columns are not orthogonal. In a regression model with an effects-coded variable, the intercept corresponds to the grand mean.

Also known as "sum coding" or "simple coding". Note though that the default in R and SPSS is to use the last level as the base. Here we use the first level as the base, for consistency with other coding systems.

Examples

julia> StatsModels.ContrastsMatrix(EffectsCoding(), ["a", "b", "c", "d"]).matrix
4×3 Matrix{Float64}:
 -1.0  -1.0  -1.0
  1.0   0.0   0.0
  0.0   1.0   0.0
  0.0   0.0   1.0
source
StatsModels.HelmertCodingType
HelmertCoding([base[, levels]])
HelmertCoding(; base=nothing, levels=nothing)

Helmert coding codes each level as the difference from the average of the lower levels.

If levels are omitted or nothing, they are determined from the data by calling the levels function when constructing Contrastsmatrix. If base is omitted or nothing, the first level is used as the base. For each non-base level, Helmert coding generates a columns with -1 for each of n levels below, n for that level, and 0 above.

When all levels are equally frequent, Helmert coding generates columns that are mean-centered (mean 0) and orthogonal.

Examples

julia> StatsModels.ContrastsMatrix(HelmertCoding(), ["a", "b", "c", "d"]).matrix
4×3 Matrix{Float64}:
 -1.0  -1.0  -1.0
  1.0  -1.0  -1.0
  0.0   2.0  -1.0
  0.0   0.0   3.0
source
StatsModels.SeqDiffCodingType
SeqDiffCoding([base[, levels]])

Code each level in order to test "sequential difference" hypotheses, which compares each level to the level below it (starting with the second level). Specifically, the $n$th predictor tests the hypothesis that the difference between levels $n$ and $n+1$ is zero.

Differences are computed in order of levels. If levels are omitted or nothing, they are determined from the data by calling the levels function when constructing ContrastsMatrix. If base is omitted or nothing, the first level is used as the base.

Examples

julia> seqdiff = StatsModels.ContrastsMatrix(SeqDiffCoding(), ["a", "b", "c", "d"]).matrix
4×3 Matrix{Float64}:
 -0.75  -0.5  -0.25
  0.25  -0.5  -0.25
  0.25   0.5  -0.25
  0.25   0.5   0.75

The interpretation of sequential difference coding may be hard to see from the contrasts matrix itself. The corresponding hypothesis matrix shows a clearer picture. From the rows of the hypothesis matrix, we can see that these contrasts test the difference between the first and second levels, the second and third, and the third and fourth, respectively:

julia> StatsModels.hypothesis_matrix(seqdiff)
3×4 Matrix{Int64}:
 -1   1   0  0
  0  -1   1  0
  0   0  -1  1
source
StatsModels.HypothesisCodingType
HypothesisCoding(hypotheses::AbstractMatrix; levels=nothing, labels=nothing)

Specify how to code a categorical variable in terms of a hypothesis matrix. For a variable with $k$ levels, this should be a $k-1 imes k$ matrix. Each row of the matrix corresponds to a hypothesis about the mean outcomes under each of the $k$ levels of the predictor. The entries in the row give the weights assigned to each of these $k$ means, and the corresponding predictor in a regression model estimates the weighted sum of these cell means.

For instance, if we have a variable which has four levels A, B, C, and D, and we want to test the hypothesis that the difference between the average outcomes for levels A and B is different from zero, the corresponding row of the hypothesis matrix would be [-1, 1, 0, 0]. Likewise, to test whether the difference between B and C is different from zero, the hypothesis vector would be [0, -1, 1, 0]. To test each "successive difference" hypothesis, the full hypothesis matrix would be

julia> sdiff_hypothesis = [-1  1  0  0
                            0 -1  1  0
                            0  0 -1  1];

Contrasts are derived the hypothesis matrix by taking the pseudoinverse:

julia> using LinearAlgebra

julia> sdiff_contrasts = pinv(sdiff_hypothesis)
4×3 Matrix{Float64}:
 -0.75  -0.5  -0.25
  0.25  -0.5  -0.25
  0.25   0.5  -0.25
  0.25   0.5   0.75

The above matrix is what is produced by constructing a ContrastsMatrix from a HypothesisCoding instance:

julia> seqdiff_hyp = HypothesisCoding(sdiff_hypothesis;
                                      levels=["a", "b", "c", "d"],
                                      labels=["b-a", "c-b", "d-c"]);

julia> StatsModels.ContrastsMatrix(seqdiff_hyp, ["a", "b", "c", "d"]).matrix
4×3 Matrix{Float64}:
 -0.75  -0.5  -0.25
  0.25  -0.5  -0.25
  0.25   0.5  -0.25
  0.25   0.5   0.75

The interpretation of the such "sequential difference" contrasts are clear when expressed as a hypothesis matrix, but it is not obvious just from looking at the contrasts matrix. For this reason HypothesisCoding is preferred for specifying custom contrast coding schemes over ContrastsCoding.

Keyword arguments levels and labels give the names (in order) of the hypothesis matrix columns (corresponding to levels of the data) and rows (corresponding to the tested hypothesis). The labels also determine the names of the model matrix columns generated by these contrasts.

References

Schad, D. J., Vasishth, S., Hohenstein, S., & Kliegl, R. (2020). How to capitalize on a priori contrasts in linear (mixed) models: A tutorial. Journal of Memory and Language, 110, 104038. https://doi.org/10.1016/j.jml.2019.104038

source
StatsModels.hypothesis_matrixFunction
hypothesis_matrix(cmat::AbstractMatrix; intercept=needs_intercept(cm), tolerance=1e-5)
hypothesis_matrix(contrasts::AbstractContrasts, n; baseind=1, kwargs...)
hypothesis_matrix(cmat::ContrastsMatrix; kwargs...)

Compute the hypothesis matrix for a contrasts matrix using the generalized pseudo-inverse (LinearAlgebra.pinv). intercept determines whether a column of ones is included before taking the pseudoinverse, which is needed for contrasts where the columns are not orthogonal to the intercept (e.g., have non-zero mean). If tolerance != 0 (the default), the hypotheses are rounded to Ints if possible and Rationals if not, using the given tolerance. If tolerance == 0, then the hypothesis matrix is returned as-is.

The orientation of the hypothesis matrix is opposite that of the contrast matrix: each row of the contrasts matrix is a data level and each column is a predictor, whereas each row of the hypothesis matrix is the interpretation of a predictor with weights for each level given in the columns.

Note that this assumes a balanced design where there are the same number of observations in every cell. This is only important for non-orthogonal contrasts (including contrasts that are not orthogonal with the intercept).

Examples

julia> cmat = StatsModels.contrasts_matrix(DummyCoding(), 1, 4)
4×3 Matrix{Float64}:
 0.0  0.0  0.0
 1.0  0.0  0.0
 0.0  1.0  0.0
 0.0  0.0  1.0

julia> StatsModels.hypothesis_matrix(cmat)
4×4 Matrix{Int64}:
  1  0  0  0
 -1  1  0  0
 -1  0  1  0
 -1  0  0  1

For non-centered contrasts like DummyCoding, without including the intercept the hypothesis matrix is incorrect. So while intercept=true is the default for non-centered contrasts, you can see the (wrong) hypothesis matrix when ignoring it by forcing intercept=false:

julia> StatsModels.hypothesis_matrix(cmat, intercept=false)
3×4 Matrix{Int64}:
 0  1  0  0
 0  0  1  0
 0  0  0  1

The default behavior is to coerce to the nearest integer or rational value, with a tolerance of the tolerance kwarg (defaults to 1e-5). The raw pseudo-inverse matrix can be obtained as Float64 by setting tolerance=0:

julia> StatsModels.hypothesis_matrix(cmat, tolerance=0) # ugly
4×4 Matrix{Float64}:
  1.0  -2.23753e-16   6.91749e-18  -1.31485e-16
 -1.0   1.0          -2.42066e-16   9.93754e-17
 -1.0   4.94472e-17   1.0           9.93754e-17
 -1.0   1.04958e-16  -1.31044e-16   1.0

Finally, the hypothesis matrix for a constructed ContrastsMatrix (as stored by CategoricalTerms) can also be extracted:

julia> StatsModels.hypothesis_matrix(StatsModels.ContrastsMatrix(DummyCoding(), ["a", "b", "c", "d"]))
4×4 Matrix{Int64}:
  1  0  0  0
 -1  1  0  0
 -1  0  1  0
 -1  0  0  1
source

Special internal contrasts

StatsModels.FullDummyCodingType
FullDummyCoding()

Full-rank dummy coding generates one indicator (1 or 0) column for each level, including the base level. This is sometimes known as one-hot encoding.

Not exported but included here for the sake of completeness. Needed internally for some situations where a categorical variable with $k$ levels needs to be converted into $k$ model matrix columns instead of the standard $k-1$. This occurs when there are missing lower-order terms, as in discussed below in Categorical variables in Formulas.

Examples

julia> StatsModels.ContrastsMatrix(StatsModels.FullDummyCoding(), ["a", "b", "c", "d"]).matrix
4×4 Matrix{Float64}:
 1.0  0.0  0.0  0.0
 0.0  1.0  0.0  0.0
 0.0  0.0  1.0  0.0
 0.0  0.0  0.0  1.0
source
StatsModels.ContrastsCodingType
StatsModels.ContrastsCoding(mat::AbstractMatrix[, levels])
StatsModels.ContrastsCoding(mat::AbstractMatrix[; levels=nothing])

Coding by manual specification of contrasts matrix. For k levels, the contrasts must be a k by k-1 Matrix. The contrasts in this matrix will be copied directly into the model matrix; if you want to specify your contrasts as hypotheses (i.e., weights assigned to each level's cell mean), you should use HypothesisCoding instead.

source

Further details

Categorical variables in Formulas

Generating model matrices from multiple variables, some of which are categorical, requires special care. The reason for this is that rank-$k-1$ contrasts are appropriate for a categorical variable with $k$ levels when it aliases other terms, making it partially redundant. Using rank-$k$ for such a redundant variable will generally result in a rank-deficient model matrix and a model that can't be identified.

A categorical variable in a term aliases the term that remains when that variable is dropped. For example, with categorical a:

  • In a, the sole variable a aliases the intercept term 1.
  • In a&b, the variable a aliases the main effect term b, and vice versa.
  • In a&b&c, the variable a alises the interaction term b&c (regardless of whether b and c are categorical).

If a categorical variable aliases another term that is present elsewhere in the formula, we call that variable redundant. A variable is non-redundant when the term that it alises is not present elsewhere in the formula. For categorical a, b, and c:

  • In y ~ 1 + a, the a in the main effect of a aliases the intercept 1.
  • In y ~ 0 + a, a does not alias any other terms and is non-redundant.
  • In y ~ 1 + a + a&b:
    • The b in a&b is redundant because it aliases the main effect a: dropping b from a&b leaves a.
    • The a in a&b is non-redundant because it aliases b, which is not present anywhere else in the formula.

When constructing a ModelFrame from a Formula, each term is checked for non-redundant categorical variables. Any such non-redundant variables are "promoted" to full rank in that term by using FullDummyCoding instead of the contrasts used elsewhere for that variable.

One additional complexity is introduced by promoting non-redundant variables to full rank. For the purpose of determining redundancy, a full-rank dummy coded categorical variable implicitly introduces the term that it aliases into the formula. Thus, in y ~ 1 + a + a&b + b&c:

  • In a&b, a aliases the main effect b, which is not explicitly present in the formula. This makes it non-redundant and so its contrast coding is promoted to FullDummyCoding, which implicitly introduces the main effect of b.
  • Then, in b&c, the variable c is now redundant because it aliases the main effect of b, and so it keeps its original contrast coding system.