This is a phrase I use often in my work: “All models are wrong; some models are useful”. A meaningful portion of my job involves using statistical methods to create models. Some of those are the kind that one thinks of when you use the phrase “statistical model”, which is often some sort of regression analysis – you take a bunch of variables, apply some kind of mathematical rules to them, and then interpret the results. I’ve made vague reference to this before.
The other kind of model that it’s common for someone in my field to build is referred to properly as a ‘Health State Transition Model’, but colloquially called a ‘Markov Model’, or even more colloquially simply as ‘a model’. In these sorts of models, a matrix or array stands in as a possible state of health that someone can be in, and you describe mathematical rules that govern the likelihood that values held in one ‘state’ might migrate to another ‘state’.
Mine is certainly not the only field in which modeling is used. The most obvious example is stuff like engineering where computer models are used to simulate physical objects. Public health and epidemiology use disease transmission models. Psychology talks about ‘theories’, but those can also be thought of as a type of model that describes the relationship between different cognitive processes. Newtonian mechanics is a model; as are its various relativistic successors. Those of us who have learned chemistry probably learned about models like the nuclear model or the “blueberry muffin” model of the atom, both of which eventually gave way to quantum physics.
The key to models is this: all models are wrong. All of them. Every last one. However, some models, carefully designed, can help us test hypotheses about the world without having to somehow re-create a process in real life and then observe it directly. But the models are still wrong. They are, as a necessary consequence of their utility, reductive. They omit some data, they make assumptions, they do not explain every single observation, and they force some observations into states that they might not actually belong in the real world.
And so we constantly look to improve models. We strive to use the appropriate model to answer the appropriate question: the nuclear model is perfectly useful for answering questions about electron bonding and valence, but it’s less useful when we want to talk about the behaviour and movement of electrons. Newtonian mechanics is great if you want to predict what a baseball will do, but terrible if you want to predict what a quark will do. In the case where an old model fails to properly predict reality, we develop a more sophisticated model.
The important thing to take away from this is that there is no such possible thing as a ‘perfect’ model. A perfect model would be a full-scale replication of the entire universe with 100% perfect information about the relationship between every single quantum in existence. We’re not going to build that. So we use imperfect models, because that’s our only alternative.
The other important thing to take away is that Newtonian mechanics isn’t wrong, any more than special relativity is wrong. It is great for answering a certain kind of question, within a certain set of assumed parameters (i.e., we’re not talking about processes that happen on the submolecular level or speeds that approach the speed of light – if we need to build a bridge, Newton gets us there). Special relativity isn’t better than Newtonian mechanics when you want to build a bridge – in some cases it’s ponderous and unnecessary and overly complicated, even if it gives you the same answer.
Wisdom, therefore, is held in understanding the following: in what cases does the model we are using fail to apply? What are the limitations of the model? What new models might we need to develop in order to get the answers that lie outside the scope of the model we’re using? Under what circumstances would using Model A lead us to draw erroneous conclusions about the world that we wouldn’t make if we used Model B? What are the potential harms of using Model A instead of Model B?
Another important question to ask is the following: what information would we need to build a better model? How could we go about collecting this information? The ‘debunking’ of the “luminiferous ether” led us from a Newtonian model physics to a quantum model of physics. We needed a new model to explain the transmission of light through space – we collected data, we built a better model. The discoveries of Darwin led us away from Lamarckian models of evolution toward ones that were based on natural selection – discoveries made since then have improved on Darwin’s insights.
Related to that question is this one: how do we test the assumptions of our model? Knowing that our model is wrong, what kind of data can we find to expose the flaws in our model (since we know they’re there, because all models are wrong). How can we disconfirm our model, and what would we learn in the process of such falsification? This is essentially the scientific process in essence: build a model, and then find out if it works, then if it does, try to break it.
This is something that we should do with all models. It would be a terrible scientist indeed who insisted that the model we have to describe anything is perfect and immutable. We might know that Darwinian evolution is ‘less wrong’ than, for example, a model based on the assumption of supernatural spontaneous creation of all living creatures, but if we are wise, we also know that Darwin’s insights are not complete. Natural selection is one of many ways in which species differentiate. Hell, even the concept of “species” itself is just a model – useful at describing a lot of things, but counterintuitive and wrong when dealing with certain circumstances (e.g., a population that is gradually differentiating into two).
I put it to you that everything humans do is based on models. In the sciences we are accustomed to talking explicitly about models, but we do it other places as well. A nation, for example, is a model. It is not an immutable law of the universe that identity and territory are synonymous. Do I have more in common with a serial child rapist from Newfoundland than I do with a secular humanist viola player in Rwanda? Doubtful. But within the model we call “nation”, I agree to pay taxes for the health and safety of the serial rapist, while contributing nothing to the health and well-being of the Rwandan. Why? Because I agree to operate within the model we call “nation”, and the rapist is ‘in my nation’ whereas the violist is not.
Is “nation” a good model? In this hypothetical example, one might be strongly motivated to say it’s a terrible one. I deplore the actions of the rapist, I laud the actions of the violist. Within the universe of this particular model, “nation” is a pretty shit idea. And yet, in a universe in which federal tax revenues pay for cancer research (as well as humane prison conditions), maybe I am more willing to make the tradeoff – maybe “nation” is a useful model. Still ‘wrong’, but useful.
Of course, in this example I don’t have a lot of options about whether or not I accept the model of “nation”. I was born into a world and into a society that has accepted “nation” as a given. There are deleterious consequences to me (and, arguably, to others) if I decide that I simply don’t buy this whole “nation” thing. However, if I live in a just world, I can interrogate the model and say “I think we can have a better idea”. And then we get Kiva.
Some wish to completely abandon the model – whichever model we are talking about. We’ll stick with “nation” for now. Some people say “to hell with the nation of Canada, my nation is what I decide it is”. Others (whose perspective I am highly sympathetic to) say “we had a nation before Canada came along, and that’s the model of nation I identify with”. Still others say that the entire concept of “nation” is flawed and should be done away with.
The above radicals are all saying essentially the same thing: this model is wrong. The wrongness of the model is meaningful and the model needs to change. They are living in, or feel strongly about, the places where the model does not adequately describe the world. The model does not serve them.
This is distinct, I hasten to point out, from those who wish to keep the model but change the way the model behaves. People who want a better country don’t have a problem with the model of “country”, but rather they want to operate within the context of the existing model to make positive changes. They are not saying “the model is wrong”, they are saying “we can improve the model”. A subtle difference, but an important one to grok as we move forward.
The thing to recall, again, is that all models are wrong. The radicals and the non-radicals are both ‘correct’ in the sense that their beliefs follow logically from their prior beliefs about the model – they just have different prior beliefs. The nationalists think that the model should be kept, even if it isn’t perfect. The radical anti-nationalists think that a new model is needed. They are Newtonians vs. Einsteinians.
Conversations about society and social justice can be improved by recognizing that everything we’re talking about is a model, and that all models are wrong but some models are useful.
Let’s talk about a specific example: race. Race is a social construct, which is another way of saying it’s a model. One version of this model is to group people into 19th-century categories like “Negro” “Caucasian” “Mongoloid” and so forth. In an extremely broad-brush sense (that is, with a relatively limited set of prior beliefs and a metric assload of assumptions), that model is useful. We can talk about some things, like the migration of homo sapiens out of Africa and into the other continents of the world, using those categories.
A more precise (but still wrong) model of race includes South-Asian people as a distinct ‘race’ from East-Asian people, Indigenous Americans (both North and South – ‘America’ being, again, another model) as a distinct ‘race’, and then we end up with a five-part rather than a three-part model. It’s better, in that there are real differences between people from India and people from China that are obscured in “Asian”, and real differences between people from North America and those from South America.
An even more precise model of race is to recognize that race is more than biology and/or geography, and to include social consequence as a component of ‘race’. That’s how we get “black” people in the United States being a distinct group from Indigenous African people. It’s how we get our shifting definition of what “white” means over the generations. It’s how we get religion and history mixed up in biology and geography. Yes, our model is more complex, but if we want to talk about the effects that racial identity has in a contemporary context, we need that complexity because it gives us real information about the world that the “Negro/Caucasian/Mongoloid” model is woefully unequipped to impart.
And even within the most contemporary model of ‘race’, we see that it has flaws. What exactly, for example, do I have in common with Xavier and Jasmine? We all self-identify as ‘black’, but in what way are we actually, meaningfully, in the same group? Xavier and Jasmine are both Americans, whereas I am Canadian – our cultures and histories are distinct. Xavier and I are both black men, whereas Jasmine is a black woman – the consequences of our blackness and the stereotypes that drive how others treat us are distinct. Xavier is from the South, Jasmine lives in New York, I live in Vancouver – the consequences of our ‘race’ in these environments are all different from each other. I’ll let you ponder what it means that Xavier and Jasmine both have two ‘black’ parents, whereas I have only one, and mine is from the Caribbean rather than African-American.
Even the most contemporary model of race is wrong, depending on what question you’re asking. Because it’s a model, and all models are wrong. But, if you restrict the prior beliefs and assumptions of your model to a subset of all possible prior beliefs and assumptions – being racialized as black, dealing with the resulting stigma, aligning with a certain culture – then we are indeed members of the same ‘race’.
Wisdom, therefore, is held in understanding the following: what are the limitations and assumptions of the model of ‘race’ that I am using when I and these others call ourselves ‘black’? In what cases does the model hold? In what cases is the model going to give us erroneous information about the world (with respect to the three of us and our place in it). What are the potential harms of me saying that my blackness is the same as Xavier’s or Jasmine’s?
The question that follows is this: is the model of ‘race’ we are using so bad that we must throw out the entire model and come up with something new? Can we improve the existing model (or meld different models of things like ‘class’ and ‘nation’ and ‘culture’ and ‘history’) to give us a more complex, but ultimately more accurate understanding? Or, do we take the radical approach and say that the flaws in the model of “race” are so multifold that the entire model itself must be replaced?
Adding to this complexity is that there are multiple models of ‘race’. Some use a model of ‘race’ as “prejudice plus power”, incorporating class and political influence. Not everyone, not even all race scholars, accept that model in all cases (or indeed, in any case). There are some circumstances where that model breaks down: Barack Obama has ‘power’ in nearly every sense that a poor white woman doesn’t – how do we reconcile this with the idea that the “power plus prejudice” model leads us to the conclusion that anti-white racism at the hands of black people is logically impossible (and I recognize that this is an extreme and oversimplified conclusion to draw from that model).
Let’s look briefly at another example: gender. Gender as a binary is a really basic and pervasive model. And for (I think) the majority of the population, that model is fine. Men do X, women do Y, and that’s the model. But we know that the binary gender model leads us to make all kinds of harmful decisions. It’s often reductive to the point of absurdity, wherein it’s not “manly” to cry, and it’s not “feminine” to seek an authoritative role. These consequences are bad.
But even beyond that specific issue, there is a proportion of the population that does not comport with the idea of gender binaries at all. Rather, gender is a spectrum, radiating not necessarily between two points but between multiple possible forms of self-identity and self-expression. Still others operate within the binary, but violate the biological assumption in the “classical” gender binary and are ‘women’ and ‘men’ despite how the “classical” model would identify them based on chromosomes and hormones.
Mainstream cis feminism strongly critiques the idea that there are such things as “male roles” and “female roles”, or even the ideas of “masculinity” and “femininity”. Trans-inclusive feminism says that the very idea of gender is wrong and that a new model is needed. Other forms of gender skepticism arrive at different ideas. Sometimes they are in conflict with each other, sometimes they can be harmonized into a more complex model that nevertheless makes it possible for us to describe the world with more accuracy.
There can be no perfect model. All models are wrong; some models are useful. Some models are more useful than others. Some models can potentially be more harmful than others, despite being perfectly ‘useful’.
So the question that remains is this: what can we do to resolve the inherent tension in a world where all models are wrong? What can we do to reconcile multiple models, none of which is perfect, but which might lead us to different conclusions?
The answer may be unsatisfactory. Just like in the sciences, and just as described above, we need to be aware of the limitations of the different models we’re using. When does the prevailing sociological definition of ‘race’ fail to accurately describe the reality we’re discussing? When does the gender binary become obscure the truth or worse, lead us to engage in harmful behaviours and accept harmful beliefs?
In order to do this, we need to remember that models are used to answer questions, and the ‘appropriate’ model to use depends almost entirely on the question we are trying to answer. When discussing, for example, income inequality that falls along racial lines, is a model that acknowledges racial categories as prejudice and power more useful in reducing those inequalities in a just way than, say, a model that says that since race is a social construct we should look only at individuals? Does a model of gender that is binary make it more possible for us to deliver services to people with specific needs than one that embraces all possible gender identities? Is a more simplistic model ‘good enough’ for the issue we’re trying to address?
If we say ‘yes’, then our work is still not done. Remember: we have an obligation to understand the limitations of the model we’ve chosen. If we make sure that cis women have access to abortion, we have to realize that this strategy will not address the similar needs of trans men. If we make race-based admissions policies at schools in order to improve access and financial mobility, we have to realize that this strategy will not address the needs of white students who nevertheless might earn low incomes.
This is, to my eye, what social justice conversations are about. Someone proposes a model, and then others discuss the limitations, and we do what we can to accommodate the flaws of the model. We will not be able to address all things at all times, but we should at least recognize our failures rather than dusting our hands off and saying “this is good enough”. When deciding what kinds of models to build in the future, we can look to our knowledge of the failures of the previous model and incorporate what we’ve learned through the process of mindful self-correction.
Of course, the worst actors in these conversations are those who insist on their favoured model without giving thought to – or worse, outright denying the existence of – the flaws therein. The criticism that feminists of colour often level (and rightly so) against white feminists is that their model often doesn’t include race as a real and pervasive women’s issue. The criticism that trans-inclusive feminists level against cis feminists is that the definition of “woman” doesn’t include all of the ways in which that can potentially manifest. The criticism that male feminists (and, if we’re being as generous as humanly possible, MRAs) level against (some) female feminists is that the focus on the needs of women can leave the gendered issues facing boys and men without resources or advocates. And on, and on.
This doesn’t necessarily mean that “feminism” is wrong. It means that a single model of feminism cannot possibly address all situations, and that our conception of what that means must be flexible and inclusive. It doesn’t necessarily mean that “anti-racism” is wrong, it just means that focusing on the ‘black-white’ binary relationship neglects a lot of other people who have racial issues as well. It doesn’t necessarily mean that “nationhood” is wrong, it simply means that, like all models, we have to look at whose needs the model does not address, and what events and processes it fails to encapsulate.
All models are wrong; some models are useful. Not all models are equally wrong, but some models are too complex to use well. No model will address every possible situation, and models can and must learn from each other and adapt to observed reality.
There is another phrase that I am fond of using when discussing the use of models, and that phrase is ‘GIGO’, which stands for ‘Garbage In, Garbage Out’. If your model is based on assumptions about the world that are fundamentally untrue – if the structure of the elements of your model are not a good representation of reality – then you can’t help but get erroneous answers, no matter how well you do anything else. If you have perfect information, if somehow you managed to recreate every quantum in the universe, but you fail to properly represent the relationships between them, then your model will never give you the right answer. This is where discussions of privilege enter into the picture. If you are attempting to extrapolate the universe from a flawed piece of fairy cake, then your model will always give you false information. You need to inform your model with data, and to test the assumptions of your model using that data.
This is hard, and I recognize that. It’s a lot to keep in your mind at all times, and it’s impossible to have all possible perspectives on an issue at once, which is another reason that I am such a big proponent of diversity. Diverse teams allow us to incorporate, as a necessary consequence of our politics, multiple perspectives; multiple priors; multiple assumptions. We can build harmonious models that, while not perfect, are light years ahead of a model that is based on a single set of beliefs.
Whether we recognize it or not, this is a process we are constantly engaged in. Recognizing the limitations and flaws of our preferred model is a painful but necessary first step toward learning to adapt to the needs of others around us, and finding ways to not let our models get in the way of our goals.
Wow, that was a good and thorough discussion.
The anti-realist position that you take (i.e. that all models are wrong) is not as definite as you assert. There are philosophers who still defend realism (that is, that our models can approach actually describing the real world). At the very least, it’s worth noting that even if no model is right, some are more wrong than others (which you perhaps allude to in the garbage in/garbage out section).
At least, that’s what I’d say coming from a hard science like physics. In terms of sociology, I might be more agreeable to the “all models are wrong” argument, as human interactions are notoriously complicated and involve so much that no model can accurately describe all the relations. We therefore have to make simplifying assumptions to make any progress at understanding the world. But as you rightly point out, these assumptions must be critically examined.
Thanks for your comment, Ian.
I don’t think that physics models are shielded from my assertion about models at all. What we have found time and again is that our understanding of the universe shifts whenever we look beyond the nearest horizon. Our model improves as we test it under new assumptions. Unless you’re telling me that there is a model of physics that is capable of describing everything (in conflict with my understanding that such a thing is fundamentally impossible), then I think we are just talking about HOW GOOD a model can be.
As far as “approach(ing) actually describing the real world”, I definitely believe this to be the case. But I think that this happens asymptotically, and that there will always be something left unexplained or unexplainable.
There’s a difference between a theory and a model. The former is an explanatory framework in which nature is to be understood; the latter is an application of a theory (usually with some simplifications and approximations) to a specific scenario with the intent of explaining a specific thing. Dirac’s equation or classical electrodynamics (explicated by Maxwell’s laws) would be examples of theories. The solution of Dirac’s equation for a Coulomb potential would be a model of the hydrogen atom (and one that is imperfect, since it fails to account for the Lamb shift).
It’s one thing to say that all models are necessarily imperfect (I don’t think it’s right to claim they’re wrong), but another thing to say that a theory is necessarily wrong. Our current theories are imperfect, and we do have good reason to believe this (including both unexplained phenomena and incompatibility of general relativity and quantum field theory in their common domain of application), but there isn’t any reason to believe that there will never be a perfect theory. A theory, unlike a model, does not need to be a complete description of the actual universe—it need explicate the rules by which the universe operates. (Physics works under the assumption that the laws of nature are non-Humean.) A model based on the theory will always be imperfect, but the theory need not be.
Newton’s laws and quantum mechanics are theories rather than models, to be sure.
Also, I’m going to be a bit pedantic here: the luminiferous ether has nothing to do with quantum mechanics. It had to do with electrodynamics, and its rejection by Einstein was the basis for his investigations into relativity. (The symmetries of Newton’s laws and Maxwell’s laws are different, and Newton’s laws were assumed to be fundamental, so the rest frame of the ether was supposed to be the frame in which Maxwell’s laws were true. Einstein got rid of the ether and just said that the symmetries of Maxwell’s laws are the fundamental symmetries, and modified Newton’s laws accordingly.)