Multiple fixed effects using FixedEffects.jl

Specific Domains Statistics
, , ,
1

Hi

I’m using FixedEffects.jl to residualize matrices with fixed effects. I might be missing something with the syntax, but I couldn’t get the same results using a simple regression approach. Below is the code from the readme file

using FixedEffects
# define fixed effects:
p1 = FixedEffect(repeat(1:5, inner = 2))
# combine fixed effects
p2 = FixedEffect(repeat(1:2, outer = 5), repeat(1:2, inner = 5))

x = rand(10)
x0 = copy(x)

# Regression approach
# fixed effect matrix
D = [p1.refs .== unique(p1.refs)' p2.refs .== unique(p2.refs)']
@time x1 = x0 - D*inv(D'*D)*D'*x0

#HDFE approach
@time solve_residuals!(x, [p1, p2])

# compare two methods
[x1 x]
2

It looks like your design matrix is wrong for the by-hand approach. It’s non-singular since you included all the indicator variables. The standard for the OLS dummy variable approach to fixed effects is to drop one level of the fixed effect. With the right D matrix, the answers are the same:

# ...same until here...
D = let dv_1 = p1.refs .== unique(p1.refs)[2:end]',
           dv_2 = p2.refs .== unique(p2.refs)[2:end]'
      [ones(Bool, 10) dv_1 dv_2]
    end

 @time x1 = let β = D \ x0
           x0 - D * β
       end

@time solve_residuals!(x, [p1, p2])

x ≈ x1
1 Like
3

Thanks!