Implement fstatistic

[OndrejSlamecka]

To be used by JuliaStats/StatsModels.jl#225

I am not sure how much sense does nobs::Float64 make but Float64 is the return type of nobs.

[OndrejSlamecka]

Thank you @palday for the suggestions, I implemented them, except un-deleting the two modelframe related functions for which I commented above. Is it desirable to also export fstatistic in GLM.jl?

[palday]

Thank you @palday for the suggestions, I implemented them, except un-deleting the two modelframe related functions for which I commented above. Is it desirable to also export fstatistic in GLM.jl?

I only focused on some Julia/style issues; I'll let @nalimilan or @kleinschmidt comment on API choices. 😄

[samusz]

Seems related to issue #400 , if I am not mistaken.

[nalimilan]

Sorry for not commenting earlier @OndrejSlamecka. How about making this a method of ftest? Since the p-value is printed, this is not just a function to get the F-statistic. That also simplifies the API.

[OndrejSlamecka]

Re "making this a method of ftest": so in

GLM.jl/src/ftest.jl

Line 87 in 504e518

function ftest(mods::LinearModel...; atol::Real=0.0)

turn, for example the following one into the one below? (To be clear, I just made up 42 and 0.42.)

julia> ftest(nullmodel.model, model.model)
F-test: 2 models fitted on 12 observations
─────────────────────────────────────────────────────────────────
     DOF  ΔDOF     SSR     ΔSSR       R²     ΔR²        F*  p(>F)
─────────────────────────────────────────────────────────────────
[1]    2        3.2292           -0.0000                         
[2]    3     1  0.1283  -3.1008   0.9603  0.9603  241.6234  <1e-7
─────────────────────────────────────────────────────────────────

julia> ftest(nullmodel.model, model.model)
F-test: 2 models fitted on 12 observations
─────────────────────────────────────────────────────────────────
     DOF  ΔDOF     SSR     ΔSSR       R²     ΔR²        F*  p(>F)
─────────────────────────────────────────────────────────────────
[1]    2        3.2292           -0.0000              42.0   0.42      
[2]    3     1  0.1283  -3.1008   0.9603  0.9603  241.6234  <1e-7
─────────────────────────────────────────────────────────────────

That'd be nice API-wise. A disadvantage is that it wouldn't show when viewing the model summary (JuliaStats/StatsModels.jl#225) which is something to consider.

If you want to go ahead with the "extending ftest" approach, I'll update the code and the comment (I reckon it's a matter of dropping/replacing the two NaNs here

GLM.jl/src/ftest.jl

Line 125 in 504e518

fstat = (NaN, (MSR1 ./ MSR2)...)

).

---

I also just rebased on latest master.

[nalimilan]

Adding this information to the first row of the table could be confusing, as it doesn't have the same meaning as in other rows. I was just thinking about using the ftest function with a single argument. The returned object could be a different type from FTestResult, similar to your FStatistic type (but with a name like SingleFTestResult or FTest1Result). Can you also put the new code in ftest.jl?

[codecov-commenter]

Codecov Report

Merging #424 (06ba845) into master (96f6eeb) will increase coverage by 0.44%.
The diff coverage is 100.00%.

❗ Current head 06ba845 differs from pull request most recent head 795a0df. Consider uploading reports for the commit 795a0df to get more accurate results

@@            Coverage Diff             @@
##           master     #424      +/-   ##
==========================================
+ Coverage   83.39%   83.84%   +0.44%     
==========================================
  Files           6        6              
  Lines         771      786      +15     
==========================================
+ Hits          643      659      +16     
+ Misses        128      127       -1     

Impacted Files	Coverage Δ
src/ftest.jl	100.00% <100.00%> (+1.47%)	⬆️
src/linpred.jl	76.57% <100.00%> (+0.21%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 96f6eeb...795a0df. Read the comment docs.

[OndrejSlamecka]

@nalimilan I've done that now. Please let me know what you think.

[nalimilan]

Thanks! This is something we could consider moving to StatsModels so that other packages don't have to reimplement it. Would probably be worth filing an issue there to discuss it, if you feel like it.