You are handed a study reporting that a variable X is associated with an outcome Y, and asked whether X causes Y. Order the questions to ask, from the one that settles the most to the one that settles the least.

1. Did the researcher assign participants to levels of X at random?
2. If not, did the researcher at least control who received X and when?
3. If not, was X observed as it varied naturally, with the groups differing in other ways too?
4. Could a third variable plausibly drive both X and Y?
5. Could Y in fact be causing X, rather than the other way round?

The first question is first because a yes settles almost everything. Random assignment makes the groups equivalent on all other variables at once, which is why a true experiment does not need you to think of every possible confounder - it neutralises the ones you never considered.

Each subsequent step is a weaker version of the same protection, and when the answer to all three design questions is no, you are left doing by hand what randomisation would have done automatically: naming plausible third variables and arguing them away one at a time, and checking the direction of the arrow.

The reverse-causation question sits last not because it matters least but because it is the one people forget. Screen time is associated with poor sleep, and it is entirely plausible that not sleeping is what leaves someone scrolling at 2 a.m. A correlation is symmetric; the causal story you find natural is not.