When the “similar” setting is even a little bigger than the minority setting, it is hard to achieve satisfiaction. For example, if the minority color is at 10% but satisfaction requires 15%, then you can’t find satisfaction. This tool shows that the absolute fastest way to be satisfied is to not care about the color of your neighbor. IE, the way to societal happiness is to be color blind.
EDIT: given that the model was developed in 1971, I think some of the below should be reconsidered. But I will leave it unaltered so that the thread discussion continues to make sense.
-----
I don't think it's a very insightful model, because I don't think people are segregating because they want some % of their neighbors to look like them. And even if they do, I think the tendency toward segregation is pretty trivially obvious and doesn't require a fancy agent based model to prove, although I'm sure it's fun to fiddle with the parameters and see what happens.
Hypotheses I'd look at first: perhaps they want to live near their relatives, who look like them, or they might only be able to afford a certain neighborhood, which when combined with race-class correlations reinforces segregation. Or there could be inertia from a past time when explicit preferences existed.
Explicit desire to live near a certain % of same skin color is really not the first hypothesis I'd reach for. And even if it is true, is it hard to see how segregation would result from this?
If the other factors are highly correlated with % of people in a group, then this simple segregation model still makes some sense.
Say group A has preferences for unmodeled factors that are strongly correlated with the % of people living in an area who are in group A. Think whole foods nearby, yoga studio, asian fusion restaurant. All the stuff in the SF marina neighborhood. As long as preferences are the same within the group, then these external factors can be represented by the % of people in group A.
This doesn't account for price and income though, which are obviously important.
I just think it's fairly obvious that if people prefer to live in places that would cause them to be segregated, then they will be segregated. Not sure what the model adds to this. Its parameters aren't a helpful guide to the underlying reality.
This model gives you numbers. It takes you from the very vague: "differences in between-group preferences can cause segregation" to something concrete. "X difference in between group preference for Y will lead to Z-level segregation." And X is surprisingly small for a given Z.
But there's no way to estimate the particular X in this model in the real world, and the notion that that X is the actual driver is just a bare possibility among others, not really argued for or evidenced.
If you made a model where groups have different preferences for "amenities" (nearby relatives, parks, yoga studios) that would at least represent a more plausible hypothesis. Although the parameters would still be impossible to estimate.
I get what you're saying though. Broadly speaking it might be theoretically interesting to set up various quantitative models for the XYZ chain.
It might be an outdated model of segregation? Maybe it explains the origins of segregated suburbs in the 1940s, but doesn't explain its persistence into the 2020s.
Segregation is not limited to places where people choose to live.
Large amounts of (self-)segregation are easy to observe in any mixed social setting, such as social networks (within them, across them), communities of whatever practice, scientific schools, etc.
I guess it depends on what you mean by "explain". Any small tendency is enough to cause segregation. Want to live near relatives. Want to live near grocery store that caries your ancestors cuisine's ingredients. etc...
I don't think the point of a model like this is so much "this is a full description of how the real world works," but "if you make this assumption, then here's how it works out." It's interesting when the result isn't what you'd naively expect.
A lot of people did find this particular model interesting. It might seem that if segregation is due to preferences, then when you see near-complete segregation it's because people preferred to almost completely avoid people who don't look like them. But actually, a relatively mild preference to just avoid being in the minority is sufficient to make neighborhoods almost completely segregated. That's true even if everyone actually preferred being in mixed neighborhoods (as long as they weren't in the minority).
Whether people actually do avoid being in the minority is a separate question. But if were get some data saying that's the case, then that would be sufficient to explain segregation.
The demo also shows a spectacular phase transition.
Make a mix like 10% / 90%, set tolerance to 50%. Even though some clusters form, a lot of unsatisfied dots keeps wandering around indefinitely, because there are not enough contiguous free areas.
Then change the tolerance to 51%, the smallest amount past 1/2. Very-very quickly clusters form and motion stops; the segregation is complete and stable.
Nicky Case's simulation is great and fun, but I think she reaches the wrong conclusion.
Her takeaway is this (in my words, not hers): "Diversity is good, therefore we should alter our behavior to increase diversity."
My takeaway is the opposite (again, my own words): "Segregated communities may arise in a decentralized way through the free choices of people optimizing for individual preferences for homogeneous communities. Engineering a system to promote diversity may cause harm. Such an engineered system must, by definition, pressure or force people to make a different choice than they freely would."
Hmm. Honestly, I think neither of these is the right takeaway?
My takeaway is "if people want some degree of diversity, and take individual actions based on what seem like reasonable heuristics, say, a mild desire for diversity but a stronger desire against minority status, then this can lead to systemic outcomes where there is less diversity than individuals want or expect."
I'm not sure where the "system engineering" part comes in, that doesn't seem to be part of the simulation, and at no point is anyone's "preference" being violated -- could you elaborate?
Nice interactives. The framing seems flawed since all preference is referred to as a "bias".
> Small individual bias can lead to large collective bias.
There are many reasons people might want to live near some other people who are similar to themselves. Some are purely practical, like language and food. It is not clear how much of the total preference is rooted in these sorts of practical realities as opposed to "bias" that would not exist in an ideal world. Referring to all preferences as biases strikes me as sloppy.
I'm fairly certain they're using a mathematical meaning for bias. Yes, all preference in this case is "bias" because it affects the system in a consistent way.
If you're an engineer, you're probably already familiar with this form of bias: a system that always exhibits a preference for some type of inputs, for example, a computer vision system that always categorizes green-tinged images as "cats", is exhibiting bias, not "preference" -- as in this simulation, there's nothing personal going on, it's literally just numbers in a matrix.
It's interesting how the two types in the simulation are said to be "happier to live in a diverse environment" but their only preference that leads to actual behavior is "when I'm in a 1/3 minority or less I'm unhappy and have to move".
Further down, when the shapes show their preference for neighborhood diversity in actual behavior, you get diverse results.
I’m not sure if this is related or not, but another explanation may simply be convenience. If you’re an immigrant, you might want to live in a place where you can easily buy ingredients to make your native cuisine, which implies a mild clustering force. It doesn’t mean you prefer to be around people of your own ethnicity per se. Probably more powerfully, you might want to be in a place that shares your values, irrespective of ethnicity; however, values probably correlate a bit with ethnicity, and thus you end up segregating.
With respect to wealth, you probably prefer to live in a place that maximizes value/money, which probably means living among other people in a similar socioeconomic context (economies of scale come into play when you have a lot of people with similar socioeconomic statuses in a small area).
So as always, we need to think critically rather than concluding “prejudice” at the first sign of disparity.
What does the model predict would happen if you put all the good schools in one location?
Does the model predict that people would need an armed guard to escort a child to school so she could be taught in an otherwise empty class because only one teacher was prepared to teach her and no parent would let their child be taught alongside her?
If not, then it's not really that effective at predicting the reality of the same era it was created in and so what's it actually good for except excusing a variety of terrible acts?
You might misunderstand what a model is supposed to do in such research papers.
Schelling's model elucidates a mechanism where small individual preferences (probably not even worth calling a 1/3rd preference a "bias") lead to segregated aggregate outcomes.
Models can isolate mechanisms. Other such examples might show stuff like
- separating good from bad candidates requires a condition on the visibility and cost of effort e.g. in education, but not usefulness (a degree need not be useful for the job to do its work)
- some scheme to build a public good (like a bridge) will never be cost neutral
- two people who would benefit from selling to each other may never find a way to contract because of two sided information asymmetries
- your choice of auction modality does not affect your revenue
etc.
Many of those things are "obvious", but before we understood the mechanism, we had all sorts of wild theories. This goes doubly for segregation. Now we know - no one needs to prefer segregation for it to exist.
Think model like "cube in vacuum", rather than model like "predicting who likes which Amazon Prime series".
Game theoretic models in particular do this "one weird move" very different from predictive modeling: They do not really tell us how reality would reach the equilibrium state. Like, if our model allows iterative best-replies, is that really what people would do in reality? Probably not, right?
And that's actually the point: Predicting human behavior based on some sort of "physical" model has failed in 100% of the cases. People aren't particles.
Instead, one should take these models in the following sense: If people interact, what can't be a stable outcome and why? If a stable outcome were to be reached, what would have to be true?
Flipping the logic in that way leads to these game theoretic models. One then learns about these mechanisms and understands reality better.
For example: Schelling has helped me understand the phenomenon of segregated exchange student groups at European universities. People come from all sorts of places in European and abroad, usually with the genuine desire to meet other cultures. So we can assume segregation is not what they want.
Yet, at the end of the semester, you sometimes end up with relatively cohesive groups (I remember Spanish and Latin Americans, Chinese groups) In other instances, there were no such groups at all. Schelling's work gives you some idea as to what might be going on.
Yes, the famously thorny problem of why people who come from the same different country and natively speak the same foreign language might hang out together in college.
Thank goodness the piercing logic of Game Theory managed to unravel this enigma for us.
And, in America, in 1971, you would obviously name this important scientific discovery the "segregation" model because thats a word you use when talking about social cliques in foreign students.
Might as well have called it the "Why did the Nazi Germany-era jews choose to live in the ghetto?" model if we were going to be that tone deaf about it.
Look, Game Theory is cool and nerdy, but like most parts of economics it's often used as a weapon. Let's not be niave.
What a weird take, but I do enjoy you make it, because it proves my point:
From your perspective, it is a simple mechanism. People speak the same language, so they must hang out together.
As I already explained, Game Theory can be good to check these naive inferences. Indeed, I have given some examples where results are counter-intuitive (but then obvious, once we understand the mechanism). Schelling is one: You'd think that if people prefer a diverse group of friends with only a little homogeneity, you'd never see a segregated society.
In the case of college groups, the hint was that sometimes we get these cohesive groups, but other times we don't. And yet, each year, students speak the same languages and arrive in comparable amounts.
If what you said was true, there'd be no such variation. Something else might be going on.
In any case, I want to be clear here that trying to understand "strategic interaction", the effects of complex inter-dependencies and their sometimes counter-intuitive results, is not a justification for morally or ethically unjustified behavior. Even if it can certainly be used in that way.
Indeed, in my frequent posts on mechanism design, you will see me fiercely criticizing our lax and relaxed treatment of how companies vaccuum in more and more data and develop better and better models to work around information asymmetries and appropriate consumer welfare.
I happily take an even more interventionist stance when it comes to issues of segregation, sustainability and inclusion.
Not sure if you were accusing me specifically, but better let this be said.
There is a free tool that implements Shelling's process. This work was a major contribution to his body of work for his Nobel prize. Website: http://ccl.northwestern.edu/netlogo/
38 comments
[ 2.1 ms ] story [ 93.8 ms ] thread-----
I don't think it's a very insightful model, because I don't think people are segregating because they want some % of their neighbors to look like them. And even if they do, I think the tendency toward segregation is pretty trivially obvious and doesn't require a fancy agent based model to prove, although I'm sure it's fun to fiddle with the parameters and see what happens.
Hypotheses I'd look at first: perhaps they want to live near their relatives, who look like them, or they might only be able to afford a certain neighborhood, which when combined with race-class correlations reinforces segregation. Or there could be inertia from a past time when explicit preferences existed.
Explicit desire to live near a certain % of same skin color is really not the first hypothesis I'd reach for. And even if it is true, is it hard to see how segregation would result from this?
Say group A has preferences for unmodeled factors that are strongly correlated with the % of people living in an area who are in group A. Think whole foods nearby, yoga studio, asian fusion restaurant. All the stuff in the SF marina neighborhood. As long as preferences are the same within the group, then these external factors can be represented by the % of people in group A.
This doesn't account for price and income though, which are obviously important.
If you made a model where groups have different preferences for "amenities" (nearby relatives, parks, yoga studios) that would at least represent a more plausible hypothesis. Although the parameters would still be impossible to estimate.
I get what you're saying though. Broadly speaking it might be theoretically interesting to set up various quantitative models for the XYZ chain.
Large amounts of (self-)segregation are easy to observe in any mixed social setting, such as social networks (within them, across them), communities of whatever practice, scientific schools, etc.
A lot of people did find this particular model interesting. It might seem that if segregation is due to preferences, then when you see near-complete segregation it's because people preferred to almost completely avoid people who don't look like them. But actually, a relatively mild preference to just avoid being in the minority is sufficient to make neighborhoods almost completely segregated. That's true even if everyone actually preferred being in mixed neighborhoods (as long as they weren't in the minority).
Whether people actually do avoid being in the minority is a separate question. But if were get some data saying that's the case, then that would be sufficient to explain segregation.
Make a mix like 10% / 90%, set tolerance to 50%. Even though some clusters form, a lot of unsatisfied dots keeps wandering around indefinitely, because there are not enough contiguous free areas.
Then change the tolerance to 51%, the smallest amount past 1/2. Very-very quickly clusters form and motion stops; the segregation is complete and stable.
Interactive, web-based, explained. Highly recommended.
Her takeaway is this (in my words, not hers): "Diversity is good, therefore we should alter our behavior to increase diversity."
My takeaway is the opposite (again, my own words): "Segregated communities may arise in a decentralized way through the free choices of people optimizing for individual preferences for homogeneous communities. Engineering a system to promote diversity may cause harm. Such an engineered system must, by definition, pressure or force people to make a different choice than they freely would."
My takeaway is "if people want some degree of diversity, and take individual actions based on what seem like reasonable heuristics, say, a mild desire for diversity but a stronger desire against minority status, then this can lead to systemic outcomes where there is less diversity than individuals want or expect."
I'm not sure where the "system engineering" part comes in, that doesn't seem to be part of the simulation, and at no point is anyone's "preference" being violated -- could you elaborate?
> Small individual bias can lead to large collective bias.
There are many reasons people might want to live near some other people who are similar to themselves. Some are purely practical, like language and food. It is not clear how much of the total preference is rooted in these sorts of practical realities as opposed to "bias" that would not exist in an ideal world. Referring to all preferences as biases strikes me as sloppy.
It would be sloppy if it were not deliberate.
I noticed nobody was accused of gentrification in this simulation.
The polygons are not people, they are trivial mathematical agents with explicit bias, they do not have preferences.
It is absolutely deliberate and not sloppy.
The author is clearly talking about people:
> So, fellow shapes, remember it's not about triangles vs squares, it's about deciding what we want the world to look like, and settling for no less.
That doesn’t change the fact that the polygons are not people.
Crash test dummies are also not people, despite usefully standing in for people in crash tests.
If you're an engineer, you're probably already familiar with this form of bias: a system that always exhibits a preference for some type of inputs, for example, a computer vision system that always categorizes green-tinged images as "cats", is exhibiting bias, not "preference" -- as in this simulation, there's nothing personal going on, it's literally just numbers in a matrix.
Further down, when the shapes show their preference for neighborhood diversity in actual behavior, you get diverse results.
With respect to wealth, you probably prefer to live in a place that maximizes value/money, which probably means living among other people in a similar socioeconomic context (economies of scale come into play when you have a lot of people with similar socioeconomic statuses in a small area).
So as always, we need to think critically rather than concluding “prejudice” at the first sign of disparity.
Does the model predict that people would need an armed guard to escort a child to school so she could be taught in an otherwise empty class because only one teacher was prepared to teach her and no parent would let their child be taught alongside her?
https://en.m.wikipedia.org/wiki/The_Problem_We_All_Live_With
If not, then it's not really that effective at predicting the reality of the same era it was created in and so what's it actually good for except excusing a variety of terrible acts?
Schelling's model elucidates a mechanism where small individual preferences (probably not even worth calling a 1/3rd preference a "bias") lead to segregated aggregate outcomes.
Models can isolate mechanisms. Other such examples might show stuff like - separating good from bad candidates requires a condition on the visibility and cost of effort e.g. in education, but not usefulness (a degree need not be useful for the job to do its work) - some scheme to build a public good (like a bridge) will never be cost neutral - two people who would benefit from selling to each other may never find a way to contract because of two sided information asymmetries - your choice of auction modality does not affect your revenue
etc. Many of those things are "obvious", but before we understood the mechanism, we had all sorts of wild theories. This goes doubly for segregation. Now we know - no one needs to prefer segregation for it to exist.
Think model like "cube in vacuum", rather than model like "predicting who likes which Amazon Prime series".
Game theoretic models in particular do this "one weird move" very different from predictive modeling: They do not really tell us how reality would reach the equilibrium state. Like, if our model allows iterative best-replies, is that really what people would do in reality? Probably not, right?
And that's actually the point: Predicting human behavior based on some sort of "physical" model has failed in 100% of the cases. People aren't particles.
Instead, one should take these models in the following sense: If people interact, what can't be a stable outcome and why? If a stable outcome were to be reached, what would have to be true?
Flipping the logic in that way leads to these game theoretic models. One then learns about these mechanisms and understands reality better.
For example: Schelling has helped me understand the phenomenon of segregated exchange student groups at European universities. People come from all sorts of places in European and abroad, usually with the genuine desire to meet other cultures. So we can assume segregation is not what they want. Yet, at the end of the semester, you sometimes end up with relatively cohesive groups (I remember Spanish and Latin Americans, Chinese groups) In other instances, there were no such groups at all. Schelling's work gives you some idea as to what might be going on.
Thank goodness the piercing logic of Game Theory managed to unravel this enigma for us.
And, in America, in 1971, you would obviously name this important scientific discovery the "segregation" model because thats a word you use when talking about social cliques in foreign students.
Might as well have called it the "Why did the Nazi Germany-era jews choose to live in the ghetto?" model if we were going to be that tone deaf about it.
Look, Game Theory is cool and nerdy, but like most parts of economics it's often used as a weapon. Let's not be niave.
From your perspective, it is a simple mechanism. People speak the same language, so they must hang out together. As I already explained, Game Theory can be good to check these naive inferences. Indeed, I have given some examples where results are counter-intuitive (but then obvious, once we understand the mechanism). Schelling is one: You'd think that if people prefer a diverse group of friends with only a little homogeneity, you'd never see a segregated society.
In the case of college groups, the hint was that sometimes we get these cohesive groups, but other times we don't. And yet, each year, students speak the same languages and arrive in comparable amounts. If what you said was true, there'd be no such variation. Something else might be going on.
In any case, I want to be clear here that trying to understand "strategic interaction", the effects of complex inter-dependencies and their sometimes counter-intuitive results, is not a justification for morally or ethically unjustified behavior. Even if it can certainly be used in that way.
Indeed, in my frequent posts on mechanism design, you will see me fiercely criticizing our lax and relaxed treatment of how companies vaccuum in more and more data and develop better and better models to work around information asymmetries and appropriate consumer welfare. I happily take an even more interventionist stance when it comes to issues of segregation, sustainability and inclusion.
Not sure if you were accusing me specifically, but better let this be said.