Sunteți pe pagina 1din 11

2.

1 Sorting and Peer Effects Introduction


This series of lectures discusses models that deal with the fact that people who form groups tend to look and act alike, two mechanisms will be discussed: 1. Sorting and 2. Peer effects

What is Sorting?
People choose people that have similar behavior. What is important is the choice in sorting there is an active choice that is the essential difference with peer-effects.

What are Peer Effects?


Acting or behaving like peers (people you hang out with), synchronizing behavior and beliefs. Not choosing people because they look and behave like you (Sorting) but rather: trying to look like the people you (want to) hang out with. Models discussed in this series of lectures Thomas Schelling (Schellings tipping model) Mark Granovetter (Granovetter Model) Standing Ovation Model Identification (the identification problem)

Note on the type of model used here: Not used: equation based model
(Example: exam score = 50 + 5 * hours) would be a linear model of someones grade Where 50 is some baseline and 5 times hours is the slope, this model says: if you spend more time preparing your grade will be higher.)

Used: Agent based model: Factors of importance are: 1. 2. 3. individuals behaviors (rules) outcome

The model seems simple but the answers it supplies are surprising.

Extra information
An agent-based model (ABM) is a class of computational models for simulating the actions and interactions of autonomous agents, with a view to assessing their effects on the system as a whole. Most agent-based models are composed of 1 numerous agents specified at various scales. 2. decision making heuristics. 3. learning rules or adaptive processes. 4. an interacton topology. 5. a non-agent environment.
Source: Wikipedia.

2.2 Schelling's Segregation Model


The chart of New York (Figure 1.) shows racial segregation, and also segregation by income. Schelling explored segregation with his model, he used an agent-based model consisting of 1. agents 2. rules they follow 3. outcome.

Figure 1.

Racial Segregation in New York: Red = Caucasian, Blue: African-American, Yellow: Latino, Green: Asian A simple example of Schelling's model: People (agents) living on a checkerboard environment, who have a threshold (a rule) that causes them to stay or move (outcome). Rich/Poor segregation example: Each cell can have: 1.No person (Blank) 2 Rich person (X) 3 Poor person (O) Xs = Subject of study, T = time. The Rule Xs follows is: if 44 % alike then stay else go (i.e. threshold = 44%). Figure 2. X O O

O Xs X

X O

Situation at T = 0, three of four neighbors are same, is above threshold of Xs so stay.

O O O

O Xs X

X O

Situation at T = 1, One person moved out and another moved in (opposite), is just under threshold so move.

Sample problem:
Suppose Sonja's current threshold is 35%, i.e. she needs at least 35% of her neighbors to be of the same type in order for her to remain at her location. If Sonja currently has four of eight neighbors of the same type, how many neighbors of the same type would have to relocate in order for Sonja to want to move? Assume that no new neighbors replace the old. A. 1 B. 2 C. 3 D. 4 Answer: This problem calls for a simple calculation. If one "like" neighbor relocated, she would still have 3 of 7 of the same type, which is 43%. But if two "like" neighbors relocated, Sonja would then have 2 of 6, which is only 33%. (B).

Netlogo Model (Schelling)


Explore the Netlogo model for the Schelling model. The insight you can gain from this model and the simulation is: Results at the macro-level cannot directly be explained by behavior on the micro-level, example: people can be very tolerant but the (macro) result can still be segregation.

In the simulation: when the micro-level tolerance would be low, then there would be a perpetual motion of people moving.

Tipping phenomenon of Shelling's model Some Examples of Tipping points Exodus Tip (People moving out cause others to move out) Figure 3. T = 0, rule for X is if 33% same then stay else move. O X 40% O O X 33% X 50% O At T = 0, its alright everybody is in balance, percentages given are percentages of same neighbors. For some reason the middle X is about to move. T=1 O X 25% O O X 33% O At T = 1, Middle X has moved causing Red X, and finally the green X (at T=3) to move out (tipping point) Figure 4. Genesis Tip (People moving in cause others to move out) T = 0, rule is if 33% same then stay else move. X 33% O O X 40% O X 33% T=1 O X 25% O O X 40% O X 33% Now O moving in causes a collapse first Red X moves out then the rest.

2.3 Measuring Segregation


In this lecture the index of dissimilarity is introduced, to measure segregation, recall the map of New York (Figure 1.) in this map patterns of segregation are made visible (segregated by race in Figure 1.) In lectures the following maps are shown study these carefully: Map of Chicago division: by race or income. Map of New York division: by race and by income.

Example: The index of dissimilarity


Place people on a grid, each cell (or block) has 10 people living on it. Blue represents rich. Yellow represents poor. Green represents mixed 50/50.

Figure 5. b = # blue living in block B = # blue total (150), 120 in big blue block + 30 in green block (1/2 * 10 * 6) y = # yellow living in block Y = # yellow total (90), 60 in yellow block and 30 in green block (1/2 * 10 * 6) Then b/B = the proportion of blue people in that block and y/Y = the proportion of yellow people in that block |(b/B) (y/Y)| = absolute difference of proportions, why this measure? Illustration: If there are 5 Blue and 3 Yellow living in a certain block, and the total of Blue is 150, total of Yellow = 90. Then the proportions would be 5/150 and 3/90 if you take the difference (absolute difference). Then youd get | (5/150) - (3/90) | = 0, this makes sense because it would mean the block is evenly populated regarding the proportions, this would be a perfectly mixed block. Take a look at Figure 5 above, the calculations below are the differences of the proportions Blue and Yellow in each block. |(10/150) (0/90)| = 1/15 in the Blue block |(0/150) (10/90)|= 1/9 in the yellow block |(5/150) (5/90)| = 1/45 in the green block Summing these numbers gives 72/45, what does that number mean? And how do we figure out how segregated the whole unit is? Whenever you construct a measure you can try to use extreme cases as a means to get information about the structure of the model and the meaning of the values. Illustration: extreme case of mixing (perfectly mixed). Figure 6 . All eight squares are green, perfect mixing. What does our calculation yield? Each block has 5 blue and 5 yellow, total people = 10 * 8 = 80, yellow 5 * 8 = 40 and also blue 5 * 8 = 40. For each block the proportion is 5/40 = 1/8 for yellow and blue. The difference of proportions for each block is |1/8-1/8|= 0.

Opposite case (extreme segregation).

Figure 7. Again total population is 80. Proportion in blue blocks is 10/40 in yellow blocks the proportion is the same 10/40. The difference of proportions is: |1/4 0| in each blue block so 1/ 4 in each block, for a total of 4 * = 1. In each yellow block the difference is |1/4 0| = 1/4 (again each block) , totaling 4 * = 1 if there is perfect segregation then you get a number of 2, if there is no segregation at all you get a number of 0. To get a number between 0 and 1 we can divide by 2. This number forms the index of dissimilarity.
That last part about dividing by 2 is important. Think of it this way: dividing by 2 is a tool to change the scale of Index of Dissimilarity so that our upper limit will be 1 rather than 2. This makes the Index more convenient, logical, and useful. Dividing by 2 should always be the last step in the calculation of Index of Dissimilarity, meaning there is never any reason to divide by any other number. Cool? Got it? Here's the thing: dividing by two doesn't have anything to do with how many groups we have or how many blocks we're looking at. Think of it as a formula - something we do at the end of any Index of Dissimilarity problem. If this still doesn't make sense, don't worry - there's a problem coming up where you can try it.

2.4 Peer Effects


"Sometimes the tail wags the dog." 1. Granovetter Model
First a point to make about predictability (no expert saw the fall of the Berlin wall, or the recent uprising in the middle-east) It is difficult to predict what is going to happen, this model will partly show why. Granovetter Model N individuals Each has a threshold Tj for person j, (Tj = threshold to join) Join if Tj (rule is to join if your threshold is exceeded)

So what causes a collective action to occur? We can use the Granovetter model to answer this question.

Wearing purple hats (example).


There are five people represented by a number that number is the threshold that those individuals have for wearing a purple hat. Example 1. Thresholds: 0, 1, 2, 2, 2 Once the first one starts wearing the hat (threshold 0) then the one with a threshold with 1 will wear a purple hat and finally the ones with threshold 2 will wear one. Example 2, Thresholds: 1, 1, 1, 2, 2 What happens here? Nothing happens, nobody will wear a hat, there is no starting point, no tail of threshold 0. Example 3, Thresholds: 0, 1, 2, 3, 4 Here the wearing of hats begins with individual 0 and cascades down to 4

The tail has wagged the dog. An extreme point has caused a cascade. 0, 1, 2, 2, 2 average = 1.4 1, 1, 1, 2, 2 average = 1.4 0 ,1, 2, 3, 4 average = 2 In the last example the average adversity to wear a purple hat is high that means that but through a cascading effect the "action" is spread anyway. This is the tail wagging the dog, a ripple can become a giant wave.

Question: Suppose that there are 10 people who have the following thresholds for joining a volunteer project: 2 will volunteer even if no one else does. 6 require 5 others to volunteer. And 2 will volunteer so long as anyone else does. How many people will volunteer total? A. 4 B. 2 C. 10 D. 6 Answer: (A) There will be 4 volunteers. First, the 2 who will volunteer regardless of others volunteer. Since these 2 volunteer, the 2 that will volunteer as long as at least 1 other person has volunteered will also volunteer. This takes us to 4 volunteers. However, the other 6 will not volunteer, because they will only volunteer if 5 others have volunteered, and only 4 others have volunteered. So we end up with 4 volunteers.

What can we learn from this? 1. If there are more people at the low end (of threshold) an action or belief can cascade through the population, so if you want to predict you need to know the distribution of thresholds (for example discontent). If there is more diversity then there is a greater chance of "cascading" or spreading.

2.

2.5 The Standing Ovation Model

The standing ovation model is an extension of the Granovetter model, it is an example of rule-based behavior. The standing ovation model has two important aspects 1. peer-effects and 2. information. (example of information is: the copying of the behavior of "better" informed individuals). The Framework: Threshold to stand: T (rule: IF people standing > threshold THEN stand) Quality: Q (a show has a quality between 0 and 100) Signal: S = Q + E (signal = quality + error) Initial Rule: IF S > T THEN stand Subsequent Rule: stand if more than X% stand The rule depends on two factors: 1. 2. Initial threshold for quality Threshold for how many people are currently standing (the subsequent rule).

A few claims can be made about the presumable outcomes (based on the rules). Claim 1: IF Higher Q (Quality) THEN more people will stand. Reasoning: Stand if Q + Error > T If Q is higher, more people will be likely to stand. Claim 2: IF lower T THEN more people stand. Reasoning: Stand if Q + Error > T If T is lower, more people will be likely to stand. Claim 3: IF lower X THEN more ovations. Reasoning: Stand If more than X % stand. If X increases the probability that more people stand increases. What would cause X (the threshold) to be Big or Small? X tells something about the individuals in the audience. Some have lower thresholds to stand than others. Explanation of the term Error used in the model: Recall: S = Q + E, Signal = Quality + Error (Diversity) E = Error or Diversity Diversity = interpretation of quality depending on different view. and Error is just noise in perception maybe someone has a "bad" day.

Examples:
Suppose there are 1000 people, T = 60 and Q = 50 50 < 60 as a consequence nobody stands up. The figure below shows how a bigger error can change the outcome. Signal = Quality + Error (Diversity) Here: S = 50 + error, and T = 60.
If the error is between -15 and +15 then S is between 35 and 65 people below 60 all sit. if the error is between -50 and +50 then 40% of people stand up, so if the error is bigger the chance of standing increases.

E in [-15, +15]
Mean
S= 35 50 60 65 0

E in [-50, + 50]
Mean
50 60 100

Stand

Sit

Claim 4: IF Q < T AND (IF more variation E) THEN more people stand. What would cause the Error (E) to be big? - Audience - diverse - unsophisticated - Performance - multidimensional - complex

Higher quality, lower threshold, larger peer effects, more variation, all contribute to higher probability of standing ovations. Without this model it is most likely you would not have found the 4th claim: that more variation leads to more standing ovations . Question: Imagine you are at a concert for which Q is greater than T. If X% is less than 50%, does increasing E increase the chance of a standing ovation? (Hint: Draw a plot, and assume signals are uniformly distributed between 0 and 100.) A. Yes B. No

Answer: B. No. Q>T means that more than half of the audience perceive the quality of the show to be above their threshold (more than 50% will stand up). If X% is less than 50%, there will always be a standing ovation (so long as Q>T), regardless of the variance.

2. Standing Ovation Model Advanced


Modeling often leaves a lot of factors out (abstraction), but some factors left out could be important. Lets investigate which factors could matter: first of all the location or seat in a theater could be an important factor. People in the front can see just part of the audience but almost the whole audience can see them (analogy: celebrity), people seated in the back can see a lot of people but are not seen by many (analogy: the academic). So people in the front have more influence than people in the back simply because more people see them. We know: High quality, lower threshold, larger peer effects and more variation can increase probability of standing ovations, we can now add: Claim 5: IF more celebrities stand THEN higher probability of standing ovation. Claim 6: IF Bigger Groups THEN more variation THEN more chance of standing ovation. Fertility (where else could we use this model?) We got some "clues" from this model that we might use to our advantage Some examples: Collective action: public uprising or charity, a celebrity could attract attention. Academic performance: suppose a school is underperforming, it could try to improve quality or use groups of students to stimulate each other, or even use celebrities. Urban renewal: Variation idea, if citizens get an amount of money to renew their house would this cause a cascade and improve the neighborhood? Fitness/Health: Peer effects by seeing others being healthy, or information: knowing more about health and seeing the importance.

2.6 The Identification Problem


How can we tell if we are dealing with sorting effects (Homophily) or peer effects. Two recommended books: The big sort, Bill Bishop Connected, Nicholas Christakis and James Fowler

An example from the Big Sort (illustrating sorting effects)

Examples from "Connected" (illustrating peer effects) Happiness distribution: in the chart you see a peer effect of happiness, people hanging out with happy people become happy.

Yellow Happy, blue not happy, green in between

Smoking: also a peer effect. Example from the book Connected

Other examples, where the kind of effect is not ascertained. Average number of hospice days per chronically ill patient

Medicare reimbursements per enrollee Is this because of sorting or a peer effect?

Why can't we always tell which it is sorting or peer effect? In Sorting: AABBAA

BBABBA People would move from places where there are more people like them. In Peer Effect people would start changing behavior so they become more like one another. AABBAA BBABBA The result in either case is: AAAAAA BBBBBB How can we tell whether the result was an effect of sorting or peer effects? To examine if we are dealing with sorting effects we seek for movements (choices people make), if there are peer effects people change behavior, that "micro-level" information is needed to tell if it is a sorting effect or a peer effect (dynamic data, data changing over time) are people moving or are they changing?

S-ar putea să vă placă și