Comment by slow_typist

The problem is in parts, how confirmatory statistics work, and how journals work. Most journals wouldn’t publish „we really tried very hard to get significance that x causes y but found nothing. Probably, and contrary to our prior beliefs, y is completely independent of x.“

Even if nobody would cheat and massage data, we would still have studies that do not replicate on new data. 95 % confidence means that one in twenty surveys finds an effect that is only noise. The reporting of failed hypothesis testing would really help to find these cases.

So pre-registration helps, and it would also help to establish the standard that everything needed to replicate must be published, if not in the article itself, then in an accompanying repository.

But in the brutal fight for promotion and resources, of course labs won’t share all their tricks and process knowledge. Same problem if there is an interest in using the results commercially. E.g. in EE often the method is described in general but crucial parts of the code or circuit design are held back.