Fix nonpositive weights bug

[pearlzli]

This pull request fixes a bug in which passing in a weight vector with nonpositive elements would throw the following error:

julia> using DataFrames, Econometrics, Random;

julia> Random.seed!(123);

julia> x = randn(100); y = 2*x + randn(100); w = randn(100);

julia> data = DataFrame(x = x, y = y, w = w);

julia> fit(EconometricModel, @formula(y ~ x), data; wts = :w)
ERROR: MethodError: no method matching StatsBase.FrequencyWeights(::Vector{Union{Missing, Float64}})

Closest candidates are:
  StatsBase.FrequencyWeights(::AbstractVector{<:Real})
   @ StatsBase ~/.julia/packages/StatsBase/XgjIN/src/weights.jl:16
  StatsBase.FrequencyWeights(::var"#24#V", ::var"#22#S") where {var"#22#S"<:Real, var"#23#T"<:Real, var"#24#V"<:AbstractVector{var"#23#T"}}
   @ StatsBase ~/.julia/packages/StatsBase/XgjIN/src/weights.jl:13

Stacktrace:
 [1] decompose(f::StatsModels.FormulaTerm{StatsModels.Term, StatsModels.Term}, data::DataFrame, contrasts::Dict{Symbol, Union{StatsModels.AbstractContrasts, StatsModels.AbstractTerm}}, wts::Symbol, panel::Nothing, time::Nothing, estimator::Type{EconometricModel}, vce::Econometrics.VCE)
   @ Econometrics ~/.julia/dev/Econometrics/src/formula.jl:160
 [2] EconometricModel(estimator::Type{EconometricModel}, f::StatsModels.FormulaTerm{StatsModels.Term, StatsModels.Term}, data::DataFrame; contrasts::Dict{Symbol, Union{StatsModels.AbstractContrasts, StatsModels.AbstractTerm}}, wts::Symbol, panel::Nothing, time::Nothing, vce::Econometrics.VCE)
   @ Econometrics ~/.julia/dev/Econometrics/src/main.jl:64
 [3] fit(estimator::Type{EconometricModel}, f::StatsModels.FormulaTerm{StatsModels.Term, StatsModels.Term}, data::DataFrame; fit::Bool, kw::Base.Pairs{Symbol, Symbol, Tuple{Symbol}, NamedTuple{(:wts,), Tuple{Symbol}}})
   @ Econometrics ~/.julia/dev/Econometrics/src/main.jl:216
 [4] top-level scope
   @ REPL[13]:1

Before this change, while the following lines of decompose did filter out the rows of data with nonpositive weights (as well as missing values of the other variables in the formula)...

Econometrics.jl/src/formula.jl

Lines 76 to 80 in f8751d1

	data = materializer(data)(
	row
	for
	row ∈ rows(data) if all(pn -> !ismissing(getproperty(row, pn)), propertynames(row))
	)

... this wasn't being done for the wts vector itself, which was defined earlier using all rows of the dataset:

Econometrics.jl/src/formula.jl

Line 59 in f8751d1

wts = ifelse.(getproperty(data, wts) .≤ 0, missing, getproperty(data, wts))

This same unfiltered weight vector was later passed into FrequencyWeights(...) with missing values, resulting in the error:

Econometrics.jl/src/formula.jl

Lines 160 to 161 in f8751d1

	wts =
	isnothing(wts) ? FrequencyWeights(ones(length(y))) : FrequencyWeights(collect(wts))

[Nosferican]

Thanks! Could you rebase (it should pass the tests then) and then add a test? You can add it to a new test file issues.jl and use the example above.

[pearlzli]

Sorry for the delay! I've now rebased and added the test.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 90.92%. Comparing base (f8751d1) to head (446b217).
Report is 2 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master      #87      +/-   ##
==========================================
- Coverage   91.32%   90.92%   -0.41%     
==========================================
  Files           9        9              
  Lines         715      716       +1     
==========================================
- Hits          653      651       -2     
- Misses         62       65       +3     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[Nosferican]

Thanks!