- Save groupvar() in HDFE.estimate_dof()
- Docs
- Update help file
- Fix html help builder
- update website
- Optimization
- enable prune code for G=2 and G>2 (carefully)
- (?) allow weights with cg+sd, in their scalar adjustements (does hestenes has weights now?)
- Add LSMR as with Mathieu's implem.
- Improve the case with id#c.var and id##c.var . Currently it relies on a lot of calls to invsym. This is unnecessary, as we can i) do it once and save it, or ii) use something better than actually inverting the matrix, and saving the steps (see qrsolve or cholsolve)
- Memory usage:
- run
st_data
andpartial_out
in chunks, with thepool(#)
option compact
option: preserve, clear, and restore the dataset to save memory (slower but might be necessary for large datasets)- allow
nosample
option
- run
- Fix inference
- wild bootstrap (boottest) helps if firm and CEO numbers are growing at sqrt(N) rate
- Verdier's approach (or a feasible implementation of Cattaneo et al) help if firm numbers or CEO numbers grow proportional to N
- The first case can be seen as an expanding interconnected economy
- The second, as adding more independent or poorly connected countries to a dataset
- hdfe.ado
- Improve degrees-of-freedom with group3hdfe (or Mata implem of it)
savecache
andusecache
(orhold
) options?
If the dim of the 2nd FE is not that high, can I compute M and M:^2? Mata works up to 20k
We now have three fixes for rob/clus: CJN, V, WildB vce(wild) vce(m2) vce(overlap|subgraphs) (pick a good name)
Do this afterwards..!
http://web.stanford.edu/group/SOL/software/lsmr/
https://github.com/timtylin/lsmr-SLIM https://github.com/timtylin/lsmr-SLIM/blob/master/lsmr.m https://github.com/timtylin/lsmr-SLIM/blob/master/lsmrtest.m
https://github.com/PythonOptimizers/pykrylov/tree/master/pykrylov/lls https://github.com/PythonOptimizers/pykrylov/blob/master/pykrylov/lls/lsmr.py
http://web.stanford.edu/group/SOL/software/craig/ (we could use CRAIG for the alphas?)