I remember that when I was working as an undergrad research assistant in a physics lab, a PhD student and I had plotted some data, to see if the relationship be

I remember that when I was working as an undergrad research assistant in a physics lab, a PhD student and I had plotted some data, to see if the relationship between 2 variables was linear. We had cut off many of the data points to the left and right, keeping only a small region in the middle. When we showed it to our supervisor, he came back with "of course it looks linear if you only look at a small interval, that's just Taylor expansion." He then explained to us that it's incredibly scientifically sketchy to leave out perfectly good data from one's plots. Not fooling yourself is surprisingly hard when you're wishing for a specific outcome. On a related note to the supplementary data thing, I've recently seen a lot of people complaining online about not being able to reproduce the results of machine learning papers that don't have the code for their models included in the supplement. "Code, or it didn't happen." I guess each field has its own problems with replication of results.