bug in uniform kernel, uses and supersedes PR #1208 thanks to hmgaudecker
add unit tests against Stata see #1231
fix bug in confint calculation missing sqrt for K
fitted values, prediction on a grid agree with Stata
confint and se: problems because Stata uses a different way to calculate the standar errors
Also Stata returns NaNs in some cases, worst is cosine
Coverage remained the same when pulling aeddd6c on josef-pkt:kernels_bug_1208 into a3d7835 on statsmodels:master.
TravisCI is green, and I will be giving up here, enough detour. I have one example script to clean up.
It would be nice to have efficient kernel regression for reuse in other areas, there are recent papers on diagnostic tests and for example estimating non-parametric heteroscedasticity or non-parametrically smoothed heteroscedasticity robust covariances.
Code from last years GSOC is still in sandbox.
Fixed bug in uniform kernel (returned a scalar instead of array; made…
… KDEUnivariate fail).
BUG: fix kernels smoothconf standard error, rename Biweight.smoothcon…
…f see #1231
TST: kernels smoothconf against Stata, confint is only approx. see #1231
TST: adjust precision kernel.smoothconf
TST: keep disabled cosine2 of Stata, fitted matches, strange se
Coverage remained the same when pulling 3892ef9 on josef-pkt:kernels_bug_1208 into c0a62a0 on statsmodels:master.