You have to understand that this is borderline unscientific.
Reliable measurement of how one variable (car model) effects another variable (emissions) means controlling all other variables. "Real world" throws a lot more variables into the mix. How quickly does this driver accelerate? Is there an incline in the driving? City versus country? Do you follow speed limits or do what most drivers actually do and speed a little? What about weather? Oh my goodness, weather! Now, there is a lot more doubt about weather
I imagine the current system lets tests be conducted in far apart locations and still give very similar results. A single company can do what Top Gears does and have a single course and have they same instructor so race times of celebrities are consistent. Except for weather. However, if every car in the country has to go to (senior senator's state TBD) to get their test done in "real world" conditions, then it becomes much harder to test emissions. Or, rather they aren't repeatable. Plus, it just changes one set of certains for another set of certains that car makers can design their vehicle around.
There are a couple ways around this. Sample Size is one, though this involves having a representative sample of real world drivers and conditions. Accidentally pick to many drivers from the Southwest desert and the test is biased.
Another is simply to accept the fact that lab results are an ideal that won't be matched in the real world... however, a car that performs better in a lab condition will perform better in the "real world." Actually, this is sort of what we really do.