The best way is to look at actual player behavior and mine the players for information then construct profiles based on that and let that inform you while adding forum/reddit/whatever feedback as secondary sources.
Shouldn't the very fact that they have a heat map give you a hint as to what they're thinking and their capabilities? Or the fact that they know what the romance rates are or how many times people have pet a damn thing?
This is where the real data is at, not a stupid survey, not these forums, not reddit.
Forum and other non game source feedback should be important I think. Some people play mostly in offline mode for various reasons (in my case it is due to late night internet issues), rather hard to analyze this data. Also, data analysis, while helpful, is probably only showing them a small amount of information. A high rate of death somewhere could mean hard content but also lag, the controls, stupidity, really anything. Petting the dog is an exploit to raise approval, it maybe shows how many people are using it, not how many would really pet the dog. Romances, well people are testing the game so there will be a high rate of this since people want to know what happens with everyone, not who they really like. Most people taking the "good" path doesn't show them how to fix the "evil" path.
Sure, just keep in mind that the data points I provided are what Larian has revealed to us they could have more and while your criticisms are valid keep in mind that the more data you have from player behavior the more accurate the aggregate profiles become. In each example you give if you have more data you can build a countervailing argument; in an area with high death you can reasonably ascertain that a percentage of the players died because of stupidity due to the actions they took leading up to the encounter starting from the beginning of the game.
And to address 'Most people taking the "good" path doesn't show them how to fix the "evil" path.' is not necessarily true because the game offers multiple scenarios/tests that allow for an 'evil' outcome and it's simply just mapping those, discarding which points were least touched upon and improving it then sending it out for another round of testing. You can pretty much create accurate profiles of what an evil player wants with enough data points by simply running them through as many scenarios as possible or if you already have enough data. Again, most social media platforms kind of do this already without ever receiving direct feedback from you in the form of a forum post or a reddit post etc.
But barring that then sure, I am willing to say that if you can't have rich data then go ahead and rake through all the forum stuff.
