lowess confidence intervals

772 views

Skip to first unread message

Joe Hamman

unread,

Aug 3, 2015, 1:29:32 AM8/3/15

to pystatsmodels

Is anyone aware of a non-bootstrap approach to calculating confidence intervals from the LOWESS smoother? The current statsmodels LOWESS model does not calculate confidence intervals and, in some cases, bootstrapping can be quite slow.

This question seems to have come up a fair number of times, for example:

- http://stackoverflow.com/questions/31104565/confidence-interval-for-lowess-in-python

- http://slendermeans.org/lowess-speed.html

xref: https://github.com/mwaskom/seaborn/pull/553

josef...@gmail.com

unread,

Aug 3, 2015, 3:45:45 AM8/3/15

to pystatsmodels

I've never seen it.

Stata doesn't have confidence intervals for lowess, but has it for lpoly, i.e. local polynomial regression. That's similar to our split between lowess and kernel regression.

A brief search

I don't see it in the documentation of R loess.

qqplot2 `stat_smooth` has it

loess references chapter 8 in Chambers statistical models in S, which has the formulas in section 8.4.1, based on a very brief skimming.

Hopefully someone has more or better information.

In any case, I guess we need to get optionally more information out of the cython code, loess for example calculates the trace of the hat matrix.