Upgrade your account to a Community Supporter account and remove most of the site ads.

Reply to thread

Message: <blockquote data-quote="Jay Verkuilen" data-source="post: 7771248" data-attributes="member: 6873517">tl;dr: If all these data are being used for is to confuse and/or amuse some posters on EnWorld, no harm. If decisions are actually being made from them, for instance to guide future product development, I'm not sure that would be a good analysis, at least as presented. I'm not 100% sure what happened but I am an actual statistician IRL and know the kinds of mistakes that we make (having made many myself and seen more). Summarizing these data in a few charts would be incredibly difficult. In many ways it's like trying to compare the list of courses taken by college students in different majors across different years in school, trying to do it in only a few pages. The data may be intended to provide a broad overview, but I think the class and subclass data, at least as they got reduced down to a color ring chart (for which, <a href="https://www.businessinsider.com/pie-charts-are-the-worst-2013-6" target="_blank">ahem</a>), seem particularly confusing. That's almost always a sign of too much data reduction. The race data makes sense because it's an unchanging trait of the character chosen at creation. That's something that makes sense to break down as percentages. Notice that nobody's arguing about that! Class is a bit easier than subclass due to the fact that one could summarize it with some fairly simple rule, such as giving the percentages for single classed characters and then giving some kind of easy summary breakdown for multis, who often have a fairly obvious mix, such as Cleric 10/Fighter 2 or Cleric 1/Wizard 9; Cleric seems reasonable for the first character while Wizard seems reasonable for the second. Or, conditioning on being a multi, what's the breakdown? That might get messy due to the relatively large number of possible combinations but it's unambiguous as to what's being compared, particularly if broken down by tiers or by common level dips. I'm not sure how one would make a sensible analysis of the subclass data given that many characters won't have a subclass due to being ineligible to have one---this is like comparing students who are in a college where majors must be declared as freshmen to ones where majors are declared as juniors. My hunch is that conditional probabilities are being compared to unconditional probabilities but not across the proper margin, but I don't really know. Comparing within tier would make more sense (in my analogy this would be comparing freshmen to freshmen, sophomores to sophomores, etc.), but even that's tough. The 10% for Life Clerics, where Clerics make up 8% of the class breakdown is a sign that a conditional probability of some sort (percent of Life Clerics) is being compared to a probability based on a different denominator. All types of data have inherent biases separate from any calculation being made. Process or purchasing data can be very useful because they demonstrate "revealed preferences", that is the choices people actually make as opposed to what they say they want. However, they have a substantial potential biases and often shouldn't be assumed to be better than "stated preference data" or interview data gathered in a more structured way, which can ask about things like options that are not currently presented and thus not able to be chosen. As a very simple illustration, just looking in my own D&D Beyond account, I currently have five characters tagged to me. Of them, three (Minotaur Fighters of different subclasses) are trial builds for a friend of mine who's not particularly adept at building characters that have, nevertheless been sitting in my account for a few months, one is an NPC (half elf Valor Bard) that's maintained there for convenience. I'm actually only playing one... a 2nd level Variant Human Fighter. So what's my "revealed preference" here? Many other sites have run into this kind of problem in the past: Amazon, for instance, used to have serious issues with its recommendation engine if the same account was being used to make purchases for different people, meaning it was recommending to a chimera who didn't actually exist. Netflix had the same issue. At other times, Amazon would recommend a big ticket purchase right after people would make a big ticket purchase. Recommending someone buy a Panasonic 56" TV right after having bought one from Sharp seems... pointless. But this still happens. Right after I bought a new car a few years ago, I got a ton of ads for, you guessed it, a new car.</blockquote>

Verification