More Power and Sample Size Discussion

Check out this post over at Climbing the Ladder, and my comment at the bottom of the post, to see more discussion on power, sample size, and alterantive methods for determining relationships between predictors and outcomes.

Just in case the comment gets removed, here it is:

Sadly, even the "fewer games" category is not statistically significant
when compared to the 50/50 chance most teams have of advancing.

Using this site (http://www.stat.ubc.ca/~rollin/stats/ssize/b1.html) and entering the following values:

p0 = 0.5

p1 = 0.672

1 sided Test

and default alpha and power levels

Yields a sample size of 51. This means this proportion of win percentage would have to be viewed over 51 matches to determine that fewer games was a good predictor vs. a coin flip.What may be more interesting is plotting win percentage vs. the actual number of fewer games. One could then run a Pearson correlation test to determine if correlation did exist, and if it did then run a linear regression to determine the strength. This could even be run for various metrics (win percentage, goal differential, etc.).

Sorry to be such a stickler, but I come from a viewpoint that most sports statistics actually aren't statistically significant. And it's a good thing they aren't - it's the random, 50/50 nature of any one sporting event that makes them interesting.