I’ve been checking polling sites like an addict recently, especially Pollster.com and FiveThirtyEight.com. Pollster sometimes employs 68% and 95% confidence intervals around their loess trend lines, like in this article about historic convention bounces:
Intuitively, it seems like the error might fluctuate over time depending on frequency of observations and the variance. The simpleboot package has a function for bootstrapping of loess fits, which will return the standard error from these fits.
Here’s the current graph from Pollster:
And here is a similar loess with +/- 1 and 2 standard error generated from bootstrapping fits.
Looking at the past 4 months shows McCain’s recent upturn in the polls.
It seems like the global constant for standard error is a decent assumption, at least from the graphs I generated from polling data.
Thanks to Brendan for the code for importing the Pollster data, from his post about polling loess at SocialScience++.
Here are my files for generating the graphs and the TSV data (convert .doc to .R or .tsv):