I do a hybrid, where I posit an initial high-level mental model, then do a bottom-up analysis, looking at all of the details, looking for similarities and differences, and seeing what fits the initial model and what doesn’t. If I am lucky, I got it right, otherwise, I modify my initial model to fit the facts and details, then go back to the bottom up approach.
The benefit of this is I am less likely to overlook important things because of preconceived notions. The drawback is that I can easily get stuck in an infinite loop of looking at details, calibrating my mental model, looking at details, changing the model, looking at details, throwing the initial model out entirely, etc.
It once took me 6 months to choose which television to buy. Of course, when I finally figured it out, I got exactly what I wanted, and used it for 15 years before getting a new one.