I was playing around with Census data tonight and wanted to see how a Lowess (bandwidth=.15) or polynomial (6-term) smoother compared. Both deal with the sampling variability of Census data by smoothing out a line as an approximation; the polynomial version takes all of the data into account while the Lowess version only incorporates nearby data (i.e. local vs. global smoothing).
Earlier this week I uploaded a working paper I wrote back in January that compared registration, turnout, and turnout of registered between 1996 and 2006. I used a polynomial smoother because that was readily available in Excel, but I was worried it might be biasing the edges. Hat tip to Avi Feller for suggesting the use of Lowess back then.
It looks like either is fine for registration or turnout, with Lowess being a little bit better at showing local changes as one would expect. I still need to look at turnout of registered though Adding in turnout of registered there is a stronger case for Lowess in that it better shows the 18 year old turnout bump among those who are registered.
Update: For those following along at home, here is Stata code you can use to try it out (data compliments of NBER):
drop if prcitshp == 5 | prtage gen voted = pes1 == 1
gen registered = voted == 1 | pes2 == 1
gen age = prtage if prtage > 0
collapse (mean) registered (mean) voted, by(age)
lowess registered age, bwidth(0.15) gen(registered_smooth) nograph
lowess voted age, bwidth(0.15) gen(voted_smooth) nograph
* Repeat if registered == 1
* Copy results into Excel and make a chart (see attached .xlsx file for Excel 2007/2008).