The week of 5/11-5/17 showed lower pageviews every day than had been typical. I don’t know why, but I didn’t post over the prior weekend, and only had two posts during the week. Posting frequency can sure complicate these statistics. The stat that intrigues me more is the peak the previous Thursday, 5/8. On 5/6 (probably in the evening) I posted a widely read and heavily commented post, Changes to Charting in Excel 2007. Perhaps this topic led to the pageview spike.

Thanks for your comments. That’s why I like this blog. I open my mouth and make a grand oversimplification, and someone not only calls me out, but gives me something interesting to read.

I know that moving averages are really more detailed than I blustered about. Monthly averages have to be tweaked to account for some months having more weekends and fewer weekdays, for example. December has two holidays (Christmas and New Years) and a light week between them.

I was going to show a couple charts here, but I decided it would make better sense to start another post for them. I don’t have enough data for a full blown seasonal analysis, not yet, but I’m intrigued.

This also will help my argument in Births by Day of the Year, a follow-up to a New York Times article. The premise of the article was that people scheduled deliveries in order to gain a tax deduction. My analysis indicated (to me, at least) that deliveries were scheduled for during the week all year round (why else are weekend births 1/3 lower than weekday births?), and for the days before major holidays, especially Thanksgiving, Christmas, and New Years, but also Independence Day, Labor Day, and Memorial Day. My premise is that the tax deduction was less of a driver than a doctor having to deliver on a weekend or holiday.

Without going too in-depth, it’s good to mention moving averages for those readers that may not be too familiar.

BTW – looks like PTS is recovering nicely from the ~5/18 decline.

http://www.census.gov/const/www/faq2.html#twelve

Clearly, the method I have described above is the simplest method of seasonal adjustment, and the US Census Bureau use more complex methods. From what I read from this page, this involves using a range either side of each data point (e.g. analagous to 1 week either side in your data) to determine the seasonal factors, so that they evolve over time, and would hence perform better at removing weekly variations.

Rick

Seasonal adjustment is not just something that can be done using moving averages. The way I remember it went something like this:

The seasonality factor is calculated by taking the average value for each (in your example) day of the week, then dividing this by the average of the entire dataset. This would give seven factors, presumably Saturday and Sunday would be 1.0 (above average)

Then each data point is ’seasonally-adjusted’ by dividing it by the relevant seasonality factor.

The resultant plot should reveal the outliers you mentioned in you previous post as deviations from a more-or-less straight line. Also I would expect you would pick up some change in the magnitude of weekday/weekend fluctuations over time.

Cheers,

Rick