Modeling tabular data

Most statistical models require that data be represented as a Matrix-like collection of a single numeric type. Much of the data we want to model, however, is tabular data, where data is represented as a collection of fields with possibly heterogeneous types. One of the primary goals of StatsModels is to make it simpler to transform tabular data into matrix format suitable for statistical modeling.

At the moment, "tabular data" means a Tables.jl table, which will be materialized as a Tables.ColumnTable (a NamedTuple of column vectors). Work on first-class support for streaming/row-oriented tables is ongoing.

The @formula language

StatsModels implements the @formula domain-specific language for describing table-to-matrix transformations. This language is designed to be familiar to users of other statistical software, while also taking advantage of Julia's unique strengths to be fast and flexible.

A basic formula is composed of individual terms—symbols which refer to data columns, or literal numbers 0 or 1—combined by +, &, *, and (at the top level) ~.

Note

The @formula macro must be called with parentheses to ensure that the formula is parsed properly.

Here is an example of the @formula in action:

julia> using StatsModels, DataFrames

julia> using StableRNGs; rng = StableRNG(1);

julia> f = @formula(y ~ 1 + a + b + c + b&c)
FormulaTerm
Response:
  y(unknown)
Predictors:
  1
  a(unknown)
  b(unknown)
  c(unknown)
  b(unknown) & c(unknown)

julia> df = DataFrame(y = rand(rng, 9), a = 1:9, b = rand(rng, 9), c = repeat(["d","e","f"], 3))
9×4 DataFrame
 Row │ y          a      b         c
     │ Float64    Int64  Float64   String
─────┼────────────────────────────────────
   1 │ 0.585195       1  0.236782  d
   2 │ 0.0773379      2  0.943741  e
   3 │ 0.716628       3  0.445671  f
   4 │ 0.320357       4  0.763679  d
   5 │ 0.653093       5  0.145071  e
   6 │ 0.236639       6  0.021124  f
   7 │ 0.709684       7  0.152545  d
   8 │ 0.557787       8  0.617492  e
   9 │ 0.05079        9  0.481531  f

julia> f = apply_schema(f, schema(f, df))
FormulaTerm
Response:
  y(continuous)
Predictors:
  1
  a(continuous)
  b(continuous)
  c(DummyCoding:3→2)
  b(continuous) & c(DummyCoding:3→2)

julia> resp, pred = modelcols(f, df);

julia> pred
9×7 Matrix{Float64}:
 1.0  1.0  0.236782  0.0  0.0  0.0       0.0
 1.0  2.0  0.943741  1.0  0.0  0.943741  0.0
 1.0  3.0  0.445671  0.0  1.0  0.0       0.445671
 1.0  4.0  0.763679  0.0  0.0  0.0       0.0
 1.0  5.0  0.145071  1.0  0.0  0.145071  0.0
 1.0  6.0  0.021124  0.0  1.0  0.0       0.021124
 1.0  7.0  0.152545  0.0  0.0  0.0       0.0
 1.0  8.0  0.617492  1.0  0.0  0.617492  0.0
 1.0  9.0  0.481531  0.0  1.0  0.0       0.481531

julia> coefnames(f)
("y", ["(Intercept)", "a", "b", "c: e", "c: f", "b & c: e", "b & c: f"])

Let's break down the formula expression y ~ 1 + a + b + c + b&c:

At the top level is the formula separator ~, which separates the left-hand (or response) variable y from the right-hand size (or predictor) variables on the right 1 + a + b + c + b&c.

The left-hand side has one term y which means that the response variable is the column from the data named :y. The response can be accessed with the analogous response(f, df) function.

Note

To make a "one-sided" formula (with no response), put a 0 on the left-hand side, like @formula(0 ~ 1 + a + b).

The right hand side is made up of a number of different terms, separated by +: 1 + a + b + c + b&c. Each term corresponds to one or more columns in the generated model matrix:

  • The first term 1 generates a constant or "intercept" column full of 1.0s.
  • The next two terms a and b correspond to columns from the data table called :a, :b, which both hold numeric data (Float64 and Int respectively). By default, numerical columns are assumed to correspond to continuous terms, and are converted to Float64 and copied to the model matrix.
  • The term c corresponds to the :c column in the table, which is not numeric, so it has been contrast coded: there are three unique values or levels, and the default coding scheme (DummyCoding) generates an indicator variable for each level after the first (e.g., df[:c] .== "b" and df[:c] .== "a").
  • The last term b&c is an interaction term, and generates model matrix columns for each pair of columns generated by the b and c terms. Columns are combined with element-wise multiplication. Since b generates only a single column and c two, b&c generates two columns, equivalent to df[:b] .* (df[:c] .== "b") and df[:b] .* (df[:c] .== "c").

Because we often want to include both "main effects" (b and c) and interactions (b&c) of multiple variables, within a @formula the * operator denotes this "main effects and interactions" operation:

julia> @formula(y ~ 1 + a + b*c)
FormulaTerm
Response:
  y(unknown)
Predictors:
  1
  a(unknown)
  b(unknown)
  c(unknown)
  b(unknown) & c(unknown)

Also note that the interaction operators & and * are distributive with the term separator +:

julia> @formula(y ~ 1 + (a + b) & c)
FormulaTerm
Response:
  y(unknown)
Predictors:
  1
  a(unknown) & c(unknown)
  b(unknown) & c(unknown)

Julia functions in a @formula

Any calls to Julia functions that don't have special meaning (or are part of an extension provided by a modeling package) are treated like normal Julia code, and evaluated elementwise:

julia> modelmatrix(@formula(y ~ 1 + a + log(1+a)), df)
9×3 Matrix{Float64}:
 1.0  1.0  0.693147
 1.0  2.0  1.09861
 1.0  3.0  1.38629
 1.0  4.0  1.60944
 1.0  5.0  1.79176
 1.0  6.0  1.94591
 1.0  7.0  2.07944
 1.0  8.0  2.19722
 1.0  9.0  2.30259

Note that the expression 1 + a is treated differently as part of the formula than in the call to log, where it's interpreted as normal addition.

This even applies to custom functions. For instance, if for some reason you wanted to include a regressor based on a String column that encoded whether any character in a string was after 'e' in the alphabet, you could do

julia> gt_e(s) = any(c > 'e' for c in s)
gt_e (generic function with 1 method)

julia> modelmatrix(@formula(y ~ 1 + gt_e(c)), df)
9×2 Matrix{Float64}:
 1.0  0.0
 1.0  0.0
 1.0  1.0
 1.0  0.0
 1.0  0.0
 1.0  1.0
 1.0  0.0
 1.0  0.0
 1.0  1.0

Julia functions like this are evaluated elementwise when the numeric arrays are created for the response and model matrix. This makes it easy to fit models to transformed data lazily, without creating temporary columns in your table. For instance, to fit a linear regression to a log-transformed response:

julia> using GLM


julia> lm(@formula(log(y) ~ 1 + a + b), df)
StatsModels.TableRegressionModel{LinearModel{GLM.LmResp{Vector{Float64}}, GLM.DensePredChol{Float64, LinearAlgebra.CholeskyPivoted{Float64, Matrix{Float64}, Vector{Int64}}}}, Matrix{Float64}}

:(log(y)) ~ 1 + a + b

Coefficients:
──────────────────────────────────────────────────────────────────────────
                  Coef.  Std. Error      t  Pr(>|t|)  Lower 95%  Upper 95%
──────────────────────────────────────────────────────────────────────────
(Intercept)   0.0698025    0.928295   0.08    0.9425  -2.20165    2.34126
a            -0.105669     0.128107  -0.82    0.4410  -0.419136   0.207797
b            -1.63199      1.12678   -1.45    0.1977  -4.38911    1.12513
──────────────────────────────────────────────────────────────────────────

julia> df.log_y = log.(df.y);

julia> lm(@formula(log_y ~ 1 + a + b), df)            # equivalent
StatsModels.TableRegressionModel{LinearModel{GLM.LmResp{Vector{Float64}}, GLM.DensePredChol{Float64, LinearAlgebra.CholeskyPivoted{Float64, Matrix{Float64}, Vector{Int64}}}}, Matrix{Float64}}

log_y ~ 1 + a + b

Coefficients:
──────────────────────────────────────────────────────────────────────────
                  Coef.  Std. Error      t  Pr(>|t|)  Lower 95%  Upper 95%
──────────────────────────────────────────────────────────────────────────
(Intercept)   0.0698025    0.928295   0.08    0.9425  -2.20165    2.34126
a            -0.105669     0.128107  -0.82    0.4410  -0.419136   0.207797
b            -1.63199      1.12678   -1.45    0.1977  -4.38911    1.12513
──────────────────────────────────────────────────────────────────────────

The protect function can be used to block the normal formula-specific interpretation of +, *, and &:

julia> modelmatrix(@formula(y ~ 1 + b + protect(1+b)), df)
9×3 Matrix{Float64}:
 1.0  0.236782  1.23678
 1.0  0.943741  1.94374
 1.0  0.445671  1.44567
 1.0  0.763679  1.76368
 1.0  0.145071  1.14507
 1.0  0.021124  1.02112
 1.0  0.152545  1.15255
 1.0  0.617492  1.61749
 1.0  0.481531  1.48153

Constructing a formula programmatically

A formula can be constructed at runtime by creating Terms and combining them with the formula operators +, &, *, and ~:

julia> Term(:y) ~ ConstantTerm(1) + Term(:a) + Term(:a) & Term(:b)
FormulaTerm
Response:
  y(unknown)
Predictors:
  1
  a(unknown)
  a(unknown) & b(unknown)
  
julia> Term(:y) ~ ConstantTerm(1) + Term(:a) * Term(:b)
FormulaTerm
Response:
  y(unknown)
Predictors:
  1
  a(unknown)
  b(unknown)
  a(unknown) & b(unknown)

The term function constructs a term of the appropriate type from symbols or strings (Term) and numbers (ConstantTerm), which makes it easy to work with collections of mixed type:

julia> ts = term.((1, :a, "b"))
1
a(unknown)
b(unknown)

These can then be combined with standard reduction techniques:

julia> f1 = term(:y) ~ foldl(+, ts)
FormulaTerm
Response:
  y(unknown)
Predictors:
  1
  a(unknown)
  b(unknown)

julia> f2 = term(:y) ~ sum(ts)
FormulaTerm
Response:
  y(unknown)
Predictors:
  1
  a(unknown)
  b(unknown)

julia> f1 == f2 == @formula(y ~ 1 + a + b)
true

Constructing a FunctionTerm programmatically

It is also possible to create a FunctionTerm programmatically, matching the behavior of what happens when a call to a function like log is encountered inside the @formula macro, although it takes a bit of care to get right. In the future we may add more convenience methods to "lift" functions into the "term domain" but for now they must be constructed manually, like so:

julia> log_term(t::AbstractTerm) = FunctionTerm(log, [t], :(log($(t))))
log_term (generic function with 1 method)

julia> log_term(term(:y))
(y)->log(y)

julia> f = log_term(term(:y)) ~ sum(ts)
FormulaTerm
Response:
  (y)->log(y)
Predictors:
  1
  a(unknown)
  b(unknown)

julia> response(f, df)
9-element Vector{Float64}:
 -0.5358107653592508
 -2.5595706990153952
 -0.3331980664948834
 -1.1383191195688154
 -0.4260357285735626
 -1.4412188661761132
 -0.34293563140185523
 -0.5837776723176953
 -2.980055366491228
 
julia> lm(f, df)
StatsModels.TableRegressionModel{LinearModel{GLM.LmResp{Vector{Float64}}, GLM.DensePredChol{Float64, LinearAlgebra.CholeskyPivoted{Float64, Matrix{Float64}, Vector{Int64}}}}, Matrix{Float64}}

:(log(y)) ~ 1 + a + b

Coefficients:
──────────────────────────────────────────────────────────────────────────
                  Coef.  Std. Error      t  Pr(>|t|)  Lower 95%  Upper 95%
──────────────────────────────────────────────────────────────────────────
(Intercept)   0.0698025    0.928295   0.08    0.9425  -2.20165    2.34126
a            -0.105669     0.128107  -0.82    0.4410  -0.419136   0.207797
b            -1.63199      1.12678   -1.45    0.1977  -4.38911    1.12513
──────────────────────────────────────────────────────────────────────────

Compared with the example above, the result is the same.

Fitting a model from a formula

The main use of @formula is to streamline specifying and fitting statistical models based on tabular data. From the user's perspective, this is done by fit methods that take a FormulaTerm and a table instead of numeric matrices.

As an example, we'll simulate some data from a linear regression model with an interaction term, a continuous predictor, a categorical predictor, and the interaction of the two, and then fit a GLM.LinearModel to recover the simulated coefficients.

julia> using GLM, DataFrames, StatsModels

julia> using StableRNGs; rng = StableRNG(1);

julia> data = DataFrame(a = rand(rng, 100), b = repeat(["d", "e", "f", "g"], 25));

julia> X = StatsModels.modelmatrix(@formula(y ~ 1 + a*b).rhs, data);

julia> β_true = 1:8;

julia> ϵ = randn(rng, 100)*0.1;

julia> data.y = X*β_true .+ ϵ;

julia> mod = fit(LinearModel, @formula(y ~ 1 + a*b), data)
StatsModels.TableRegressionModel{LinearModel{GLM.LmResp{Vector{Float64}}, GLM.DensePredChol{Float64, LinearAlgebra.CholeskyPivoted{Float64, Matrix{Float64}, Vector{Int64}}}}, Matrix{Float64}}

y ~ 1 + a + b + a & b

Coefficients:
───────────────────────────────────────────────────────────────────────
               Coef.  Std. Error      t  Pr(>|t|)  Lower 95%  Upper 95%
───────────────────────────────────────────────────────────────────────
(Intercept)  1.01518   0.0400546  25.34    <1e-42   0.935626    1.09473
a            1.97476   0.0701427  28.15    <1e-46   1.83545     2.11407
b: e         3.01269   0.0571186  52.74    <1e-69   2.89925     3.12614
b: f         4.01918   0.065827   61.06    <1e-75   3.88844     4.14992
b: g         4.99176   0.0593715  84.08    <1e-88   4.87385     5.10968
a & b: e     5.98288   0.0954641  62.67    <1e-76   5.79328     6.17248
a & b: f     6.98622   0.107871   64.76    <1e-77   6.77197     7.20046
a & b: g     7.92541   0.109873   72.13    <1e-82   7.70719     8.14362
───────────────────────────────────────────────────────────────────────

Internally, this is accomplished in three steps:

  1. The expression passed to @formula is lowered to term constructors combined by ~, +, and &, which evaluate to create terms for the whole formula and any interaction terms.
  2. A schema is extracted from the data, which determines whether each variable is continuous or categorical and extracts the summary statistics of each variable (mean/variance/min/max or unique levels respectively). This schema is then applied to the formula with apply_schema(term, schema, ::Type{Model}), which returns a new formula with each placeholder Term replaced with a concrete ContinuousTerm or CategoricalTerm as appropriate. This is also the stage where any custom syntax is applied (see the section on extending the @formula language for more details).
  3. Numeric arrays are generated for the response and predictors from the full table using modelcols(term, data).

The ModelFrame and ModelMatrix types can still be used to do this transformation, but this is only to preserve some backwards compatibility. Package authors who would like to include support for fitting models from a @formula are strongly encouraged to directly use schema, apply_schema, and modelcols to handle the table-to-matrix transformations they need.