Thursday, June 21, 2012

Social network analysis simulation - contagious depression


* Social network analysis simulation - contagious depression

* We would like to simulate a panel data set in which we have students who might be related to each other as friends.

* If one of the persons becomes depressed then that person will increase the likelihood of friends becoming depressed.



*****          *****
* Model Parameters *
*****          *****

clear

* Set the number of students to be simulated
local Num_students = 50

* Set the minimum probability that any two people are friends despite distance
local min_friend_prob = .01

* Set the max friend probability
local max_friend_prob = .75

* Set the distance coefficient. P(j&i being Friends)=max(`max_friend_prob'-`alpha'*distance(i,j),`min_friend_prob')
local alpha = 1

* Initial likelihood of facing depression P(depression)~base_likelihood + beta*#_friends_w_depression
local base_likelihood=.15
local beta = .1



*****       *****
* Generate Data *
*****       *****

set obs `Num_students'

gen stud_id = _n

* Lets imagine there is uniform three dimensional space x, y, and z (-1/2,1/2) in which students are connected to each other.

* This space is unobservable yet for the purposes of this simulation let's give it labels.

gen x = runiform()-.5
  label var x "Punk/Goth/Alternative (conformity scale)"

gen y = runiform()-.5
  label var y "Socio-economic status (wealth scale)"

gen z = runiform()-.5
  label var z "Attractiveness (physical appearance scale)"

* Now let's say the likelihood of any two people being friends is equal to (50% less the Euclidean distance from the other person with a minimum of 1%).

* To do this we will have a recursive loop that loops through all of the students and all of their potential "friendship" matches.

* First let's generate indicator variables to indicate if the student is friends with the other students
forv i=1(1)`Num_students' {
  gen fri_`i' = 0
    label var fri_`i' "This student is friends with student `i'"
}

* Count the number of friends
gen num_friends = 0

gen connection = 0
  label var connection "The connection number of the current friendship"
local num_connections = 0

* Now let's generate friendship levels
forv i=1(1)`Num_students' {
  * This recursively loops all of the students up to this point
  forv j=1(1)`i' {
    * A student cannot be friends with himself
if `i'!=`j' {
 * calculate the euclidean distance between student i and j
 local dist_ij = ((x[`i']-x[`j'])^2+(y[`i']-y[`j'])^2+(z[`i']-z[`j'])^2)^.5
 * Calculate the probability that i and j are friends
 local prob_ij = max(`max_friend_prob'-`alpha'*`dist_ij',`min_friend_prob')
 * Draw a bernoulli result for friendship
 local friends_ij = rbinomial(1,`prob_ij')
  di "`i'&`j' are " string(`dist_ij',"%9.2f") " apart ; friends prob=" string(`prob_ij',"%9.2f") " ; they are friends <`friends_ij'> 1=yes, 0=no"

 * This changes the variable indicators of how many friends students have.
 qui if `friends_ij' == 1 {
    local num_connections = `num_connections' + 1
    replace num_friends = num_friends+1 if stud_id==`i' | stud_id==`j'
replace fri_`i' = 1 if stud_id==`j'
replace fri_`j' = 1 if stud_id==`i'
 }
}
  }
}

di "--- `num_connections' connections (friendships) made ---"

tab num_friends

* My prediction is that the farther students get from the centers in any/all of the categories the less friends they will have.
reg num_friends x y z

* We cannot see this prediction hold true with this specification.

* That is because the effect is mirrored whether x y and z get larger than zero or less than zero.

* This simple OLS should fail.

* However, if we break the data into two observations positive and negative then I think we will come up with more interesting results.

foreach i in x y z {
  gen `i'_pos = max(0,`i')
    label var `i'_pos "Variable `i''s positive values"
  gen `i'_neg = min(0,`i')
    label var `i'_neg "Variable `i''s negative values"
}

* Now let's see what the results yeild
reg num_friends *pos *neg

* This is because in a uniform distribution (like almost all distributions) the closest any observation can get to the expected position of all other observations is the center of the distribution.


*****              *****
* Spread of depression *
*****              *****

* We would like to simulate the spread of depression through our network.

* First let us calculate the original level of depression:
* Initial likelihood of facing depression P(depression)~base_likelihood + beta*#_friends_w_depression

gen depression = rbinomial(1,.15)
  label var depression "Initial students depressed"

two (scatter y x if depression==0) (scatter y x if depression==1) , legend(label(1 "Healthy") label(2 "Depressed"))

gen depression2=0

* Next we will loop through every student to see if that student is not depressed if that student becomes depressed
forv i=1(1)`Num_students' {
  * If the student is depressed skip that student
  if depression[`i']==0 forv j=1(1)`Num_students' {
    * This checks if j is friends with i
 qui if fri_`j'[`i'] == 1 & depression[`j']==1  {
      local depression_spread=rbinomial(1,`beta')
 if `depression_spread' == 1 {
   replace depression2=1 if _n==`i'
noi di "Student `i' becomes depressed as a result of `j''s depression"
break
 }
}
  }
}

two (scatter y x if depression==0) (scatter y x if depression==1) ///
    (scatter y x if depression2==1) , legend(label(1 "Healthy") ///
label(2 "Depressed") label(3 "Newly depressed"))

*****                 *****
* Generate network graphs *
*****                 *****

* Can we plot one of those awesome social network graphs in Stata?
 gl two_list

* Person
forv i=1(1)`Num_students' {
  local x_pos = x[`i']
  local y_pos = y[`i']
  qui gen conct_x_`i' = `x_pos' if _n==1
  qui gen conct_y_`i' = `y_pos' if _n==1
  * This creates two variables for every person in the network.
  * It allows in effect a separate network for each person.

  local count_position = 1

  forv j=1(1)`i' {
    * This checks if j is friends with i
qui if fri_`j'[`i'] == 1  {
      replace conct_x_`i' = x[`j'] if _n==`count_position'+1
      replace conct_y_`i' = y[`j'] if _n==`count_position'+1
      replace conct_x_`i' = `x_pos' if _n==`count_position'+2
      replace conct_y_`i' = `y_pos' if _n==`count_position'+2
 * This draws a connection from the person i to the person j before returning to person i.
      local count_position = `count_position'+2
 * This moves the position up 2 spaces to make space for the next expansion of the set.
    }
  gl two_list ${two_list} (line conct_y_`i' conct_x_`i', mcolor(gs8))
  * This adds a entry in the list of graphs to be graphed for each person.
  }
}

* Warning this graph can tak
two ${two_list} (scatter y x if depression==0) (scatter y x if depression==1) ///
    (scatter y x if depression2==1) , legend(off)




* Note, this is not how network data is typically stored and created.

* A typical way of storing network data is as edgelists (see: http://en.wikipedia.org/wiki/Doubly_connected_edge_list).

* We can create an edge list from our data with the following code:

gen edge1 = .
gen edge2 = .

local count_position = 1

* This will loop through all of the students and add to the edge list one observation for every connection in each direction
forv i=1(1)`Num_students' {
  forv j=1(1)`i' {
    * This checks if j is friends with i
qui if fri_`j'[`i'] == 1  {
      replace edge1 = `i' if _n==`count_position'
      replace edge2 = `j' if _n==`count_position'
      local count_position = `count_position'+1
 noi di "edge1 edge2"
* This will makge sure we do not run out of observations in the data set to record our edges
 if `count_position'==_N set obs `=`count_position'+1'
}
  }
}


* netplot is a use written program that allows social network plots to be easily drawn from edgelist data.
* It works much faster than my network mapping code generated above.
cap netplot edge?
* If you do not have netplot installed the following code should install it.
if _rc!=0 {
  ssc install netplot
  netplot edge*
}

* This option allows network plots to be drawn as a circle.
cap netplot edge*, type(circle)

No comments:

Post a Comment