SUMM: diagnostic tests, measures and plots

[josef-pkt]

(this is another topic that is missing consistent implementation across models.)
(I might add a topic_diagnostic label, but don't want to go through issues right now. Search shows many open isssues.)

This is mainly about checking underlying assumption and specification tests.
Postestimation in general is more about interpretation of model.

main tasks, design

• implementation of functions and methods across models (or check generalization and applicability)
• easy access for users (e.g. get_diagnostics, get_plots, check_sanity)
• documentation (suggest possible remedies? maybe not that)

main issues by models

• linear regression: ok
• tsa: PR by Chad for some models

- GLM/discrete: essentially nothing

issues by topic

...

[josef-pkt]

location for new code

• stats.diagnostic is currently for linear models mainly OLS, will become to large if we add discrete and other diagnostics
• stats._diagnostic_other not public, mostly foundation, CMT and similar, includes dispersion_poisson functions
• discrete._diagnostics_count not public, used in my notebook, includes plots, helper function (merge bins), gof for counts, zero-inflation in Poisson
• outlier_influence for OLS and GLM, includes reset_ramsey for OLS

not committed

• diagnostic_reset my draft version, but should be part of a more generic module
• draft version for functions and plots in notebooks, e.g. Hosmer-Lemeshow

current plans

• stats.diagnostic_gen : add current draft version for Hosmer-lemeshow and rest tests, link tests
  (I cannot come up with a better name, gen is general or generic, even though we put binary/multinomial also in it, i.e. for specific non-linear_model Models, as well as generic diagnostics like reset tests.)
• stats.diagnostic_tools It might be better for reusability to put binning and other helper functions into a separate module.

No idea yet where to put plots.

[josef-pkt]

non-public in analogy to discrete._diagnostics_count
discrete._diagnostics_multinomial as dumping ground for plots and functions for ordinal/multinomial, so I can collect them somewhere instead of misplacing them in my private notebooks.
i.e. it works, but interface and options are very likely to have substantial changes.

[josef-pkt]

semi-random links for GLM diagnostic plots in R

http://www.sthda.com/english/articles/36-classification-methods-essentials/148-logistic-regression-assumptions-and-diagnostics-in-r/#influential-values

http://biometry.github.io/APES/Stats/stats23-GeneralizedLinearModels-GLM.html

https://stat.ethz.ch/R-manual/R-patched/library/boot/html/glm.diag.plots.html
https://stat.ethz.ch/R-manual/R-patched/library/boot/html/glm.diag.html

https://www.rdocumentation.org/packages/LogisticDx/versions/0.2/topics/plot.glm

plot(glm) in R base is or was just using plot.lm, see answer in
https://stats.stackexchange.com/questions/121490/interpretation-of-plot-glm-model