link to l

[yakir12]

No description provided.

[yakir12]

While this looks good, I am getting these errors from trying this:

julia> fm = fit!(GeneralizedLinearMixedModel(@formula(homebound ~ 1 + Run + (1+Run|Id)), df, Normal(), SqrtLink()))
ERROR: ArgumentError: invalid NLopt arguments: zero step size
Stacktrace:
 [1] chk(::NLopt.Opt, ::NLopt.Result) at /home/yakir/.julia/packages/NLopt/eqN9a/src/NLopt.jl:237
 [2] initial_step! at /home/yakir/.julia/packages/NLopt/eqN9a/src/NLopt.jl:340 [inlined]
 [3] NLopt.Opt(::OptSummary{Float64}) at /home/yakir/.julia/packages/MixedModels/eXZQr/src/optsummary.jl:85
 [4] #fit!#59(::Bool, ::Bool, ::Int64, ::typeof(fit!), ::GeneralizedLinearMixedModel{Float64}) at /home/yakir/.julia/packages/MixedModels/eXZQr/src/generalizedlinearmixedmodel.jl:182
 [5] fit!(::GeneralizedLinearMixedModel{Float64}) at /home/yakir/.julia/packages/MixedModels/eXZQr/src/generalizedlinearmixedmodel.jl:160
 [6] top-level scope at REPL[9]:1

Maybe specifying a non identity link function is simply not cosher?

[palday]

See #94, #95 .

[palday]

The fix for the NLopt error (#201) hasn't landed yet because the test case we have for that depends on the gamma functionality. If we had a good test case that didn't depend on that, I would split the PR and go ahead and get the boundary fit issue in.

[codecov-io]

Codecov Report

Merging #222 into master will not change coverage.
The diff coverage is 100%.

@@           Coverage Diff           @@
##           master     #222   +/-   ##
=======================================
  Coverage   92.89%   92.89%           
=======================================
  Files          18       18           
  Lines        1295     1295           
=======================================
  Hits         1203     1203           
  Misses         92       92

Impacted Files	Coverage Δ
src/generalizedlinearmixedmodel.jl	91.19% <100%> (ø)	⬆️

---

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 376bc71...38aa075. Read the comment docs.

[yakir12]

If we had a good test case that didn't depend on that, I would split the PR and go ahead and get the boundary fit issue in.

Can I help with that? I have data that isn't gamma related...

[palday]

Yes, please. Esp if it's a small dataset (both for repository and speed-of-tests reasons).

[yakir12]

I am unsure exactly what this would entail, but here is the CSV for the tiny data-set:
https://s3.eu-central-1.amazonaws.com/vision-group-file-sharing/Data%20backup%20and%20storage/Yakir/coffee%20beetles/beetles.csv
Not sure what else you need. I can give a short explanation, or the code to test it, let me know...

[palday]

The code to fit the model (so formula, family, link).

[yakir12]

A "simple" case:

using CSV, DataFrames, MixedModels
df = CSV.File("beetles.csv") |> DataFrame
df = df[!, Not(:outbound)]
fm = fit!(GeneralizedLinearMixedModel(@formula(homebound ~ 1 + Run + (1+Run|Id)), df, Normal(), LogLink()))

and more complicated:

df = CSV.File("beetles.csv") |> DataFrame
df = stack(df, variable_name = :Direction, value_name = :Speed)
fm = fit!(GeneralizedLinearMixedModel(@formula(Speed ~ 1 + Run * Direction + (1 + Run |Id)), df, Normal(), LogLink()))

[dmbates]

I think we should merge this change (it corrects an obvious error) and move the discussions to #94 and #95