On 26 Jan 2003 08:39:29 -0800, [EMAIL PROTECTED] (Alan Acock) wrote: [ snip some] > URBAN COUNTIES > 1990 2000 diff > % no kids working 80% 82% 2% > % with kids work 42% 62% 20% > > The difference of the difference is 18%, so single mothers increased labor > force participation more than single women who were not mothers. > They have the same data for rural counties and lets say the difference of > difference for rural counties is 10%.
Let's say, It is awfully risky -- if not foolhardy, or plain ignorant -- to compare the *changes* in groups that are not comparable at time 1. In the first place, there is the SCALING. Statisticians, for instance, don't like to compare percentages when they occupy different ranges, because the scale is probably not 'additive.' - Note that if you take 80% in 1990 to 90% in 2000, the "diff" is 10%, which looks small compared to 20% (above). However, the Odds-Ratio is almost precisely the same, [80 changing to 90] versus [42 changing to 62]. The OR is not a sure-fire, guaranteed improvement, but we figure it is more robust than percentage-points. In the second place, it happens enough of the time to make it highly embarrassing to ignore it, that there will be outside conditions that make the comparisons look just plain dumb. So, you really have to be careful, even more than when your hap-hazard, grab-samples happen to look similar on the screening variables you thought to look at. -- Rich Ulrich, [EMAIL PROTECTED] http://www.pitt.edu/~wpilib/index.html . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
