Strange behavior on Mac?

[MaximilianJHuber]

Hi,

I faced a strange behavior on a Mac with FixedEffectModels v0.10.2, CategoricalArrays v0.7.1 and Julia 1.1.0:

A, B and ID are categorical Int64.

reg(simulated_est, @model(Y ~ A * B + fe(ID)))

===================================================================================
Number of obs:                     14350   Degrees of freedom:                    9
R2:                               -0.001   R2 Adjusted:                      -0.001
F Statistic:                     17.5056   p-value:                           0.000
R2 within:                         0.000   Iterations:                         2407
Converged:                          true
===================================================================================
                   Estimate   Std.Error    t value Pr(>|t|)   Lower 95%  Upper 95%
-----------------------------------------------------------------------------------
A: 0              0.011071  0.00139711    7.92422    0.000  0.00833252  0.0138096
A: 1            0.00848418  0.00640107    1.32543    0.185 -0.00406274  0.0210311
B: 1            -1.011e-5 0.000857439 -0.0117909    0.991  -0.0016908 0.00167058
A: 1 & B: 1       0.0003191  0.00893943  0.0356958    0.972  -0.0172033  0.0178415
===================================================================================



reg(data, @model(Y ~ A * B+ ID))

=====================================================================================
Number of obs:                      14350   Degrees of freedom:                     9
R2:                                 0.000   R2 Adjusted:                       -0.000
F Statistic:                    0.0439767   p-value:                            1.000
=====================================================================================
                    Estimate   Std.Error     t value Pr(>|t|)   Lower 95%  Upper 95%
-------------------------------------------------------------------------------------
A: 1            -0.00249048   0.0063197   -0.394082    0.694  -0.0148779 0.00989695
B: 1           -4.09534e-6 0.000857097 -0.00477815    0.996 -0.00168412 0.00167593
ID: 44        -2.68066e-6  0.00134907 -0.00198704    0.998 -0.00264704 0.00264167
ID: 49         1.11412e-5  0.00134896  0.00825912    0.993 -0.00263299 0.00265527
ID: 60          2.3912e-6  0.00134897  0.00177261    0.999 -0.00264177 0.00264656
ID: 76         1.74582e-6  0.00134896   0.0012942    0.999 -0.00264238 0.00264588
A: 1 & B: 1       2.2338e-5  0.00893653  0.00249962    0.998  -0.0174944  0.0175391
(Intercept)       0.0104727  0.00104674     10.0051    0.000  0.00842094  0.0125244
=====================================================================================

However, on a different Windows machine with the same version numbers, I get the latter results with either regression specification.

The high iteration count in the first specification hints at some co-linearity, right?

Any thoughts? Given the limited replicability of this behavior, just close this issue if not.

[matthieugomez]

Thanks. Could you give the code to construct the data frame to replicate this?

[MaximilianJHuber]

Please, copy it from my Dropbox and tell me when you did.

[matthieugomez]

I get the exact same coefficient for the two specifications.

Not sure what is happening with your computer. Iterations should be 1 because there is only one fixed effect (demeaning once is enough).

Btw, use @formula instead of @model.

julia> reg(simulated_est, @formula(L_ij ~ MDP*Post + MDP_id))
                                    Linear Model                                    
=====================================================================================
Number of obs:                      14350   Degrees of freedom:                     9
R2:                                 0.000   R2 Adjusted:                       -0.000
F Statistic:                    0.0439767   p-value:                            1.000
=====================================================================================
                     Estimate   Std.Error     t value Pr(>|t|)   Lower 95%  Upper 95%
-------------------------------------------------------------------------------------
MDP: 1            -0.00249048   0.0063197   -0.394082    0.694  -0.0148779 0.00989695
Post: 1           -4.09534e-6 0.000857097 -0.00477815    0.996 -0.00168412 0.00167593
MDP_id: 44        -2.68066e-6  0.00134907 -0.00198704    0.998 -0.00264704 0.00264167
MDP_id: 49         1.11412e-5  0.00134896  0.00825912    0.993 -0.00263299 0.00265527
MDP_id: 60          2.3912e-6  0.00134897  0.00177261    0.999 -0.00264177 0.00264656
MDP_id: 76         1.74582e-6  0.00134896   0.0012942    0.999 -0.00264238 0.00264588
MDP: 1 & Post: 1    2.2338e-5  0.00893653  0.00249962    0.998  -0.0174944  0.0175391
(Intercept)         0.0104727  0.00104674     10.0051    0.000  0.00842094  0.0125244
=====================================================================================


julia> reg(simulated_est, @formula(L_ij ~ MDP*Post + fe(MDP_id)))
                                 Fixed Effect Model                                 
=====================================================================================
Number of obs:                      14350   Degrees of freedom:                     9
R2:                                 0.000   R2 Adjusted:                       -0.000
F Statistic:                     0.102596   p-value:                            0.902
R2 within:                          0.000   Iterations:                             1
Converged:                           true   
=====================================================================================
                     Estimate   Std.Error     t value Pr(>|t|)   Lower 95%  Upper 95%
-------------------------------------------------------------------------------------
MDP: 1            -0.00249048   0.0063197   -0.394082    0.694  -0.0148779 0.00989695
Post: 1           -4.09534e-6 0.000857097 -0.00477815    0.996 -0.00168412 0.00167593
MDP: 1 & Post: 1    2.2338e-5  0.00893653  0.00249962    0.998  -0.0174944  0.0175391
=====================================================================================

[MaximilianJHuber]

Thanks! I will reopen, if I find out something more.

[MaximilianJHuber]

Just to make sure: you did run reg(simulated_est, @model(L_ij ~ MDP*Post + MDP_id)) too?
I am asking because the two results you posted are both from reg(simulated_est, @model(L_ij ~ MDP*Post + fe(MDP_id))).

[MaximilianJHuber]

Does not matter, reg(simulated_est, @model(L_ij ~ MDP*Post + MDP_id)) is not the problematic specification anyway.