This is a great point. People sometimes introduce trials as 'easy' because in the simplest cases they permit causal inference with very simple analyses. This is never true in observational settings afaik. But the trials-are-simple message can deceive…
1/
https://twitter.com/ADAlthousePhD/status/1248321233877258240

When you get into it, the analysis of trials is just never simple.
Did you restrict the randomisation procedure by centre? Then you'd better reflect this in the analysis. How? Will you use fixed or random intercepts? Something else?
2/

Did you measure baseline covariates? Then your analysis would be made more efficient by adjusting, but how best to do so?
Uh-oh, did someone say 'intercurrent events'? Yes, participants are real people and things happen to them! Now estimands get complicated.
3/

Hey, no-one said anything about missing data! Ok, what would the missing values have been if- wait, I don't know. So what am I supposed to do? Suppose people who were doing well stopped coming to- oh, wait, half of these people with missing data died
4/

What about checking modelling assumptions etc.?
Careful – in the RCT context we have to write our statistical analysis plan unambiguously before seeing the data (because if you decide on your analysis based on seeing the data, things go awry).
5/

*Unambiguously* is important because, these analyses being important, we double-code them: someone else will start with the same data as you, derive outcomes (if needed) and then follow the instructions in your statistical analysis plan.
6/

This is to make sure there aren't coding errors by the trial statistician. You know, little things like accidentally swapping the intervention labels, getting the outcome definition wrong, or taking '999' to be a real measurement rather that (stupid) code for 'missing value'
7/

It can be really hard to tell if you've given all the necessary details for people to follow your plan, especially if you've written it without seeing the data (top tip: simulate something then 'practice' on that). And hard to follow if you're the backup coder.
8/

Listen, I'm not saying that analysis of trials is harder than observational studies. I'm saying no analyses of real trials is easy. It's not a competition, so please don't fall for that lie.
Trial statos, what fun complexities of trial analysis did I miss?
9/end