Evidence on first language acquisition can be acquired in a controlled or naturalistic setting (and in other ways too, but this is a major distinction).
In a controlled setting all factors other than the one(s) you are interested in are held constant. Participants should have the same age, sex, socio-economic background or at least you include roughly equal numbers of both sexes etc. and check whether it makes a difference. Then you present the same stimuli in the same way and see what happens, or vary them in a specific manner between gropus. From an experimental and extremely unethical point of view, the ideal experiment would be taking children a birth, assigning them to different groups and presenting each group with input that differs in some way you are interested in. Needless to say, nobody will do exactly that.
In a naturalistic setting the researcher tries to influence the result as little as possible. You try to generalise from what you see happening 'in real life' - for example that 90 % of all children acquiring English in the UK from monolingual patterns have mastered irregular past tense verbs by age 3 (I'm making up the numbers here). One problem here is that the presence of a researcher might change the outcome.