Essentialism!

This article by Albert Burneko doesn’t directly cite the developmental psychology literature on essentialism—indeed, it doesn’t cite any literature at all—but it’s consistent with a modern understanding of children’s thought. As Burneko says it:

Have you ever met a pre-K child? That is literally all they talk and think about. Sorting things into categories is their whole deal.

I kinda wish they’d stick to sports, though. Maybe some zillionaire could sue them out of existence?

18 thoughts on “Essentialism!

  1. This reminds me of folks who say things like, “I’d love to teach Bayes, but there’s no way that students will understand X” where X can be replaced by things like “priors” or “computation” or “likelihood” or “probability” or whatever you like. You could also replace “Bayes” with “using R” or “causal inference” or various other important but supposedly “difficult” topics. My point is that the people making those statements are usually not describing the students, but rather using a hypothetical student to talk about their own difficulties in understanding.

    This is especially sad because it leads to a kind of self-fulfilling prophecy. If the educator/parent never bothers to teach a topic, it makes it much easier for their own students/children to use the same excuse down the line, perpetuating the failure to teach and lending credence to the belief that the topic is just impossible to understand.

    • To be fair, this is what people have come to expect from statistics. No one has ever been able to explain how NHST makes sense, but everyone uses it so they assume it must be working somehow. The math is correct, but the premise is wrong (no difference between groups*).

      The root of it is an attempt to take methods appropriate for industrial quality control and apply it to scientific results. This simply doesn’t make sense because we have no idea what the costs of type I and type II errors may be. So then hypothesis testing was merged with significance testing leading us to what we have today.

      * with the exception of when that is actually predicted by a theory and all concievable effort has been made to control every relevant variable.

      • > this is what people have come to expect from statistics

        I agree, and I think part of that can attributed to the kind of “vicious cycle” I describe. But you point out another interesting factor, which is that the material we typically cover in intro stats courses (significance testing, type I/II errors) is no less “difficult” (possibly more so?) than alternative approaches. So at the same time that instructors eschew alternative approaches because they think students won’t understand them, instructors continue to demonstrably* fail to develop students’ understanding of the approaches they (we) do cover!

        * I say “demonstrably” because, as you point out, failures to properly apply NHST logic abound across science and industry despite the extensive training our undergrad and even Ph.D. students receive in those topics.

        • failures to properly apply NHST logic

          I’m not sure we are in agreement. I claim there simply is no way to “properly apply” NHST. That is why people have trouble understanding it: it doesnt make sense and is contrary to science. The only exception (mentioned above) is so different from the standard use that it deserves a seperate category.

          If you can explain the logic of checking whether theres a difference between cells/rats/people who got some treatment vs not then you would be the first to do. No one has ever done it, there are only pseudo-explanations involving at least one logical fallacy.

          And not fallacies in the heuristic, rule-of-thumb sense. Just plain false claims.

        • Like any approach, NHST can be properly applied to a situation that satisfies the requisite assumptions and in which the research question is one that can be addressed via its methods. Basically, such a question can be framed “would an observed data summary be so unlikely under model M that we should reject model M as a plausible way our data could have come about?” Using NHST to answer a question that does not have this form is a failure to properly apply NHST logic.

          I agree that most scientific questions do not have that form, though in a well-controlled experimental setting you could potentially use NHST to verify that your manipulation produced an effect that was detectable according to some criterion. This is analogous to quality control, and requires careful specification of the noise/null model M.

          Thus, other failures to properly apply NHST logic arise when that model is not specified correctly, or when other assumptions are not met, or when people consider rejection of a null model evidence for a specific alternative, or when people make other mistakes.

        • @gec

          Start by assuming everything is correlated with everything else. The NHST approach assumes correlations are rare, which is not supported by any evidence. This is the exact opposite of what has lead to successful research, which searches for “laws”.

          When people have checked they find that indeed with large enough sample size and precise enough measurements that pretty much all correlations are “significantly” different from zero.

          Eg, see:

          (1967). Theory-testing in psychology and physics: A methodological paradox. Philosophy of Science, 34, 103-115.
          https://meehl.umn.edu/sites/meehl.umn.edu/files/files/074theorytestingparadox.pdf

        • @Andrew

          Yep, reality can be “dense” but that does not mean your model should be. Just like a map doesn’t need to show every blade of grass to be useful.

          It is interesting to look at the progression of maps over time: https://en.wikipedia.org/wiki/Early_world_maps

          Not only does the area covered expand, but the maps also become more and more detailed. There is a limit to this though, the useful level of detail depends on the purpose of the map.

      • > …how NHST makes sense, but everyone uses it so they assume it must be working somehow. The math is correct, but the premise is wrong (no difference between groups*). […] The root of it is an attempt to take methods appropriate for industrial quality control and apply it to scientific results.

        This seems wrong, almost bizarrely so – and reversed from the truth.

        I struggle to imagine an “industrial quality control situation” where NHST is even conceptully ‘appropriate’ or even vaguely useful; indeed I’d bet heavily on you can’t come up any good ‘example’. Does this batch of steel have a different tensile strength than that batch? No NHST, indeed no sampling at all, is needed – the answer is NO. Does it have a material difference (as in, will the bridge I build collapse depending on my choice) – traditional NHST doesn’t even a “knob” to turn for that so the answer from NHST is ‘no help’ [you can add a knob, which would make you unpublishable and anyway but the result will remain perverse.]

        NHST for many scientific questions, though, definitely survives the ‘that’s not an obviously insane way to answer an interesting question’ test. Yes, only for about 10 seconds [after which the laughter deafens all] but it survives for a while.

      • > I do think there are conceptual issues that no one has figured out how to frame for many that want/need to learn about statistics

        I agree. I think part of the challenge comes from the mindset evinced by Schechter’s comment below: The idea that we have to get across “understanding” in a one big package. Some instructors feel overwhelmed at the idea of teaching Bayes because they think they need to spend a month on philosophy of probability, others feel overwhelmed using R in the classroom because they think they have to spend three months on algorithms and data structures. As a result, they think students won’t “understand” those topics and end up failing to prepare students to learn those things later.

        For example, I agree with Schechter that a child doesn’t have the context to understand everything that falls under the concepts of gender or sexuality. But I believe the main point of the labeling schemes we teach children (which also includes shapes, colors, numbers, etc.) is to develop language/schemata that can prepare them for a fuller understanding of those concepts later on. As children grow older and have more experiences, they can “hang” those experiences on those labels and, if the labels are well-structured, this can help children discover connections between ideas that might otherwise seem disparate. Eventually that collection of associated ideas might, at some point, be considered developed enough to call “understanding,” particularly if it prepares children to treat everyone they meet with respect and to empathize with different people’s life experiences.

        I think a mindset that focuses on preparing students to build understanding over time is more productive than one that focuses on trying to get everything across all at once. Unfortunately, most degree programs outside of stats/data science are structured such that students only ever take one course in statistics, so I can certainly understand why the “all-at-once” mindset remains dominant. But I think it is better for a student’s eventual understanding (of statistics, of philosophy of probability, of programming, etc.) to come out of that one semester with the ability to do some crude simulations in R than to perform the rituals of 20 different t tests.

        • > to come out of that one semester with the ability to do some crude simulations in R than to perform the rituals of 20 different t tests

          Preach! And even better if they learn Julia and a little more sophistication re more general simulation, like basic differential or difference equations, simulating dynamic processes, and such. T tests are like the least useful thing you could learn compared to something like “if a fish population grows by 10% per year in normal years but plummets to 20% of it’s previous value in flood years and 10% of it’s previous value in drought years, and floods happen on average with 3% probability and droughts happen with 5% probability but the year after a drought has 50% probability of also being a drought, what is the range of credible values for the fish population in 30 years if the current population is 5000 fish?

        • I pretty much agree with Daniel. For the long arc of ‘understanding over time’ it’s hard to beat a preparation of intro probability done right plus some general fluency in dynamics. The heart of ‘statistics’ is then thinking about how data arise, and how to connect them to useful quantities relevant to prediction and inference in various applied settings.
          The issue of course is we have many users and consumers of statistics who don’t have this background, and or even set of interests, and a system that isn’t going to require that of the majority any time soon.
          So we still have pre-packaged NHST ‘lexicographic rituals’ as the norm…

        • ” I think part of the challenge comes from the mindset evinced by Schechter’s comment below: The idea that we have to get across “understanding” in a one big package. ”

          Well, I don’t think that is the mindset that drives my comment. Nobody has yet proved, nor found a counterexample to, the Riemann hypothesis, but I still think we should teach kids multiplication. But trying to teach multiplication to a one year old is assuredly a waste of breath, even if the kid is one who will go on to win the Fields medal. Children’s ability to learn things grows over time, introducing things prematurely is not helpful, and, who knows, might actually be counterproductive. Developmental stages are a thing!

        • “The idea that we have to get across “understanding” in a one big package. ”

          Why is that true in statistics / social science but not true in physical sciences? I took five pure math courses for my undergrad then several more courses where applied math and chem was all of the course.

          I have plenty of ideas….

  2. > Have you ever met a pre-K child? That is literally all they talk and think about. Sorting things into categories is their whole deal.

    It’s not just children.

    The article is clever, but it’s worth noting that the author essemtializes “centrists” (w/o even providing any evidence in support).

  3. Sure, 6-year olds are first-rate at forming cognitive categories. But I’ve raised three children, and currently have a 6 year old grandchild, and I’m pretty sure none of them understood “love” in the adult, sexual, sense of the term at that point in their development.

    They probably do grasp that adults often form long-lasting pairs that live together, but I question whether they can recognize, let alone distinguish, the feelings between a couple that has sex and just housemates, or relatives or strangers who live together for a variety of reasons.

Leave a Reply

Your email address will not be published. Required fields are marked *