When observational data is collected in artificially controlled laboratory conditions, that makes possible to isolate the action/treatment/behaviour we seek to observe from external conditions and to study its effects. In turn, it would be possible to infer casuality relations which are not contaminated by the effect of other variables that could be acting in the real (not controlled) world.
The problem with such experiments is how to generalize the results obtained outside the lab. The behaviour observed there may not happen at all in the real world, or differ in terms of the social conditions of each person's environment: gender, race, religious beliefs, social class, etc.
For example, imagine an experiment in which a woman has to punish one of the other people participating after some interactions. She punishes a man who tried to deceive her. Imagine this woman out of the lab, taking into account that she lives in a very religious patriarchal community. She would never dare to contradict a man in her real world (maybe she would if the social conditions were different). Hence, it could not be concluded that people always act in their better self-interest (punishing the meanest person in this case), but that sometimes self-interest is conditioned by social surrounding features that are present not in the lab.