I think there is a fallacy in Roger's statistics, but I can't really tell because he doesn't provide enough details. First look at the spread.
1. 31.7% of measurements should fall beyond the standard deviation for a normal distribution. Look at the chart for the 5DII (= similar to 7D) vs 5DIII in the lensrental link. The spread on the y-axis for the 5DII is 650-750, that on the 5DIII is 720-770. That really is enough to notice if you are heavy cropping.
2. The fallacy appears to be this. He is calculating the standard deviation from the mean of the group, and the mean is not the best estimate of the best focussing spot. The best focus is at the very top of the chart! So, he just measuring a standard deviation from a value which is approximately a standard deviation from the mean.
Look at http://www.lensrentals.com/blog/2012/07/autofocus-reality-part-3a-canon-lenses
where there is an incredible difference for the series II lenses between the 5DII and 5DIII. Roger thinks they changed the feedback system between series II lenses and the rest, and the 5DIII but not the 5DII can handle it.