The data shown in the figure represent the average values (n=3-5 per pair of tires) of a subset of all such experiments I have performed, including only those where the temperature was between 15 and 25 deg C (ambient temperature for the experiments shown was 19.9 +/- 2.1 deg C). When mismatched pairs of tires were tested, I used the average value obtained during Al's roller tests, i.e., I assumed that my weight was equally distributed on the front and rear wheels of my TT bike. Finally, since I tested the Continental Ultra 2000 clincher tires using butyl tubes, whereas Al tested these (my) tires using latex tubes, I have adjusted the value he obtained upward by 0.00038, i.e., the average difference he has obtained in his roller tests when comparing butyl vs. latex tubes.
As can be seen in the figure, the roller and field test data agree quite closely, even though they have been performed by different individuals using different equipment and procedures. As might be expected, however, the Crr values I have obtained on an asphalt road are higher than what Al has measured using plastic rollers. Part of this difference, of course, is almost certainly due to differences in the two surfaces, and in fact when testing on aluminum rollers I have consistently obtained Crr values that are 18% lower than those found by Al, even when using identical procedures. It is also possible, however, that other factors contribute to the difference between the field-test and roller data, e.g., differences in the calibrations of our power meters (or scales), small biases in the values assumed for chain friction (field tests) or bearing friction (roller tests), etc.
The most important "take home" message, however, is the high correlation found between the roller and field test data (over a wide range of Crr values), which strongly supports the validity of the former as an approach for differentiating between the Crr of different tires. Indeed, the precision of roller testing is so much greater (i.e., by a factor of ~10x, in my experience) that it should be considered the method of choice for anyone who owns a power meter (and rollers).
(Note: I have previously posted the above plot to various web fora. If it differs from such prior versions, it is the result of more meticulously examining the data to spot errors, make certain that the brand, model, and width of tire that I tested was exactly the same as that tested by Al, etc.)