ToDos: Implementation of AutoDML

[PhilippBach]

Some todos for the package implementation

• [x] in AutoDebiaased_backends implement interactions between covars
• [ ] allow for clustered standard errors
• [x] Robustness checks
  • [x] Different dictionaries for ATE and poly_order etc for APD
• [ ] Complete documentation for new (exported) functions
  • [ ] Elements in Output lists for DataATEAutoDML and DataAPDAutoDML
• [ ] Write unit tests
• [ ] Set up CRAN-only unit test
• [ ] Finalize open points regarding interface etc.
  • [ ] Naming: AutoDML or AutoDebiased?
  • [x] Handling intercept
  • [ ] Flexibility in terms of model matrix
  • [ ] Provided options and defaults
  • [ ] Wrappers for data types as well as entire estimation (?!)
• [ ] Issue release (github (🆕 ) and CRAN)
• [ ] Add Sylvia and Max as authors of package
• [x] Another thing is adding a wrapper for "first-differencing" so that panel data can be used, panel wrapper
• [ ] Edit data splitting code so that the warning "In split.default(sample(n, n, replace = FALSE), as.factor(1:L)) : data length is not a multiple of split variable " does not show up every time
• [x] Dobule Check data.table warning on reference/copy
• [ ] Notebooks /autodml_hdm.Rmd - add more to PD panel example

[klosins]

@PhilippBach thank you! I checked off first difference task bc that is already implemented. Question, for ATE do we still want to use older dictionary style de-bias or the new "M" style that we did for APD?

[PhilippBach]

HI @klosins ,

thanks! I'm sorry, I think I have forgotten to delete the old implementations of the ATE and APD , which I now did in https://github.com/MartinSpindler/hdm/commit/a31d23a2ff82f8062b1840b1f55117485cd49b03

So: We can use the same rlassoAutoDML function for any data backend, either the one for ATE or the one for APD

One thing I had in mind would be to have a data backend wrapper that is based on a dictionary as provided by the user. We could use this not only to offer more flexibility (then we cannot guarantee corectness, e.g. of derivatives), but also for unit tests

[klosins]

in commit d5c0291 I added the ability to add weight vector

[klosins]

In commit #f298824 I added the interactions between covariates themselves

[PhilippBach]

Hi @klosins ,

one thing that is not yet done is passing the unit tests for the case with weights, so I commented out https://github.com/MartinSpindler/hdm/blob/d123e757a938baaf815657797069e7fde2f91d30/tests/testthat/test_ATEAutoDML.R#L52 and https://github.com/MartinSpindler/hdm/blob/d123e757a938baaf815657797069e7fde2f91d30/tests/testthat/test_APDAutoDML.R#L41 . I think it could be that I didn't use proper weights ( I just sampled randomly from 0,1, here ) . Do you have an idea how to fix this?

I think this is the line of code

https://github.com/MartinSpindler/hdm/blob/440786e715b8d25bcfca361a6acc065a89a9a300/R/AutoDebiased_backends.R#L211

We can also chat about it

Update: I think this was resolved already and is not an open point anymore

[klosins]

@PhilippBach sorry for the delay - just seeing this now - Ill take a look at it, and how about we chat sometime next week ? Might be easier to talk

[PhilippBach]

Related issue #12

[klosins]

Fixed "Dobule Check data.table warning on #reference/copy" code - should be all good now

[PhilippBach]

Thanks! I'll have a look