The big problem with these "scientific studies" is they use a p-value of 5% as enough to declare a match.

It sounds like what they did is they gave pregnant women a survey, and then matched that with whether or not their kid has autism.

If the survey had 20 questions, and you use a p-value of 5%, then one thing is statistically guaranteed to match. I.e., if the women answered the survey randomly, 1 in 20 things would look like they're correlated to autism.

They didn't release the details, but I'm wondering if that's the mistake they made, false correlation observation due to too high p-value threshold.

To do it correctly, they would need 1000 pregnant women. Randomly split them in two groups. Give half of them tylenol, give half of them a placebo. See how many kids develop autism. That would be a proper experiment, rather than giving a survey and "discovering" a correlation after the fact.