naturalistic observation only allows you to observe yourself and those immediately around you, as you can't easily watch people talk on cell phones and have accidents - it's not a large enough sample to be representative.
surveys run the risk of subject dishonesty, as self-reports almost always do, and again have the issue of generating a large enough sample to be statistically representative.
case study runs the risk of assuming that correlation equals causation (someone had an accident while talking on a cell phone, therefore the cell phone caused the accident... never you mind that the accident consisted of the guy hitting a moose while driving through Vermont in a snowstorm...)
experiment runs the risk of harming the subject (by causing someone a car accident!) and therefore is probably unethical.