« previous post | next post »

The current xkcd:

Mouseover title: "You've got questions, we've got assumptions."

But it's unfair, in my opinion, to suggest that models never provide answers.


  1. Rob Grayson said,

    June 24, 2020 @ 7:41 am

    "But it's unfair, in my opinion, to suggest that models never provide answers."

    Does the cartoon suggest that? I don't think it does. To me, the phrasing "…is a powerful tool for…" means something like "…can very effectively be used for the purpose of…": in other words, it implies possibility, not necessity.

    [(myl) "Mouseover title: 'You've got questions, we've got assumptions.'"]

  2. ~flow said,

    June 24, 2020 @ 11:17 am

    Isn't linguistics pretty much based on models? I mean *are* there words, syllables, moras, phonemes in the way that sciences like physics or chemistry think of things that are real? To me it's more like these terms represent models without which a description of linguistic phenomena would be much harder or impossible. I assume there are moras as distinct from syllables in Japanese because that allows me to model timing and meter in utterances and also helps to explain specific features of kana orthography; that this orthography was shaped the way it was by pre-modern native speakers presumably without much theoretic background in linguistics and that it is overall so neatly patterned on what we today call moras in that language seems to support that model; on the other hand, the fact that I can model a plethora of mechanical setups—a bridge, a pendulum—with an analogue or digital computer does not imply bridges and pendulums work anything like capacitors or binary logic.

  3. mg said,

    June 24, 2020 @ 11:40 am

    I'm a statistician and I loved this xkcd! All models are based on assumptions – they have to be just for you to choose the model. In my training, we were taught to always be aware of the assumptions we're making and test them to the extent possible.

    And as a practical example, some of the worst garbage "research" we're seeing about COVID-19 is coming from modelers in mathematical but non-biomedical fields who are choosing to use models based on inappropriate assumptions (often wildly inappropriate).

  4. Daniel Barkalow said,

    June 24, 2020 @ 11:53 am

    I think the comic is more down on modeling studies than models. Models are great if you can show that their assumptions are true, or at least are a good approximation. I think the comic implies that something comes out as a modeling study precisely because the authors couldn't do that.

  5. Lai Ka Yau said,

    June 24, 2020 @ 12:05 pm

    @~flow: I agree that words and phonemes are not really 'real'. Words in particular escape a common crosslinguistic definition, and in many languages even a sensible language-particular definition is difficult to find. I suspect that what's real is not words the themselves, but transitional probabilities between forms and categories of forms are, as is our tendency to extend, generalise and categorise – and this is why in some languages, boundaries between forms cluster in a way that we can identify a 'word boundary' cluster. Similarly, I'm not aware of any compelling evidence for the reality of phonemes (e.g. Mitterer et al 2018 in perception), though clearly they are important tools for both synchronic description of phonological systems and historical reconstruction. I'm not sure about syllables and moras, because we have things like C-centre effects which show that the effects of syllables are not just phonological, but that would only be because we bring phonetics in – I'm not sure there's anything in linguistics we can call 'real' if it is outside the realm of phonetics.

    With that said, I'm still not sure I'd use the word 'model' to describe words, phonemes, etc. I've always found the use of the word 'model' in linguistics a bit strange, because to me a model is something that is able to make quantitative predictions based on assumptions. So it is not clear to me why a tree with a syllable node below which we find an onset and a rime node, for example, counts as a 'model', because it does not directly make any predictions about syllables, except a vague idea that the rime is a constituent in some way. Better examples of models, for me, would be task dynamics in AP or learning models in phonology, which have outputs that correspond more closely to experimental data, at least in principle. And of course things like regression models of all flavours are unambiguously models, as we can take the estimates of model parameters and their [sampling|posterior] distributions to make predictions about future data points as well as our uncertainty about them.

  6. mg said,

    June 24, 2020 @ 12:10 pm

    I agree with Rob Grayson. I didn't see the comic as implying that models never provide answers. Models always provide answers. Sometimes they are correct answers, sometimes they're wrong or even horribly wrong. As statistician George Box famously said, "all models are wrong, some are useful."

  7. KeithB said,

    June 24, 2020 @ 12:22 pm

    The mouseover text is my sig.

  8. Allan from Iowa said,

    June 24, 2020 @ 5:04 pm

    When you get down to the quantum level, even in sciences like physics you can ask whether photons and quarks and so on are real or just models.

  9. Robert said,

    June 24, 2020 @ 6:42 pm

  10. Andrew Usher said,

    June 25, 2020 @ 7:30 am

    Yes, models are only as good as the assumptions put in to them, and as far as I've seen, no model for COVID-19 has yet proven istelf better than educated guessing. Basically, a model is an extension of normal thinking about the problem, not a substitute for it as regrettably even many professional scientists seem to believe.

    One could also ask if that is even a sensible question …

    k_over_hbarc at

  11. Mark P said,

    June 25, 2020 @ 7:55 am

    I worked in a field that relied heavily on a very large, very complex, very old code that produced optical signatures for defense uses. Some users literally thought that simply running the program and getting an answer was the end to the process. Any answer, just as long as the program ran and they got an answer. No real analysis of the input or output, just as long as it ran.

  12. Philip Taylor said,

    June 25, 2020 @ 8:50 am

    I had a friend, a psychologist, who taught statistics to psychology students and postgrads. Of course, they used SPSS to carry out the statistical analyses but John never failed to stress that blindly accepting SPSS's answer was no better (or little better) than pulling a solution out of the hat. The were taught to think about the data, think about the problem, and think about the solution that SPSS offered. And then decide for themselves, "was this a sensible answer ?". With John's help, the College turned out some extremely able psychologists cum statisticans.

  13. Haamu said,

    June 26, 2020 @ 7:10 pm

    @Andrew Usher — There's nothing wrong with a little educated guessing, as long as you put the emphasis on "educated" rather than "guessing."

    One problem is that many people assume modeling fails as a pursuit if it doesn't deliver one correct model that accurately predicts the future. That framing — that our standard should be oracular certainty rather than learning and intelligent risk assessment — has been exploited by some in recent years for great mischief in public policy and other areas.

    If we stop insisting on One Right Answer, the comparison of conflicting models can be quite fruitful, sometimes highlighting good policy bets and sometimes pinpointing the salient gaps in our knowledge.

    With regard to COVID-19, you might be right to be unimpressed with any single model. But the aggregate picture of multiple models may be where the most value lies. See here (New England Journal of Medicine) for a view that is measured, but less dismissive than yours.

  14. Andrew Usher said,

    June 27, 2020 @ 7:56 am

    Well, I imagine that's at least the goal of modeling, even if it's not actually attained. On scientific questions there _is_ one right answer, even if we can't currently say with certainty. But – and this isn't specific to modeling – scientific answers don't necessarily equate to policy answers.

    How much we should be willing to sacrifice to reduce global warming, or to reduce the spread of COVID-19, are not questions that have scientific answers and I think it unfortunate that some people seem to treat them as such.

RSS feed for comments on this post