Exponential random graph (p*) models for social networks Workshop Harvard University February 2002 Philippa Pattison Garry Robins Department of Psychology.
Post on 19-Dec-2015
215 Views
Preview:
Transcript
Exponential random graph (p*) models for social networks
Workshop
Harvard University
February 2002
Philippa Pattison
Garry Robins
Department of Psychology
University of Melbourne
Australia
Plan for the workshop
Model construction: dependence graphs
Dyadic independence and Bernoulli graph models– Example: The network structure of a US law firm
Markov random graphs– Example: Mutual work ties among the lawyers
Estimation– Pseudo-likelihood estimation
– Monte Carlo maximum likelihood
Incorporating individual level variables– Directed dependence graphs
– Social selection models
– Example: Advice ties among the lawyers
Further steps
General framework for model construction
1. Regard each network tie as a random variable (often binary)
Xij = 1 if there is a network tie from person i to person j
= 0 if there is no tie
for i, j members of some set of actors N.
A directed network: Xij and Xji are distinct.
A non-directed network: Xij = Xji
X … matrix of all variables
x … matrix of observed ties (the network)
General framework for model construction
2. Formulate a hypothesis about interdependencies and construct a dependence graph
3. The Hammersley-Clifford theorem produces a general model : – each parameter corresponds to a configuration
in the network
4. Consider homogeneity constraints: should some parameters be equal?
5. Estimate model parameters
Dependence graphs
Spatial statistics (Besag, 1974)
Spread of a contagious disease in a field of plants.– Whether or not a plant has the disease in part
depends on whether nearby plants have the disease.
The probability of a plant having the disease is conditionally dependent on whether neighbouring plants have the disease.
What counts as neighbouring plants?
How to represent neighbours?
For example: a 2-dimensional lattice
1 2 3 4
65
9 10
7 8
11 12
More abstractly … a dependence graph!
Random variables
Zi = 1 if plant i has the disease
Zi = 0 if plant i does not have the disease
Z1 Z2Z3 Z4
Z6
Z5
Z9 Z10
Z7Z8
Z11Z12
The lattice then represents the contingencies among thevariables.
We can have neighbouring variables, just as we haveneighbouring plants.
And then ...
Besag (1974) used the Hammersley-Clifford theorem to derive a probabilistic description of the entire system
• in terms of the hypothesised local dependencies.
Cliques of the dependence graph.
• The sufficient statistics of the model are all the combinations of variables that are all neighbours of each other.
Cliques of the dependence graph
Z7, Z8, Z12 are all neighbours of each other.
Whether plant 7 has the disease depends not only on whether plant 8 or plant 12 has the disease, but also on whether both 8 and 12 have the disease simultaneously.
Z1 Z2Z3 Z4
Z6
Z5
Z9 Z10
Z7Z8
Z11Z12
For social networks
The dependence graph represents the contingencies among network variables Xij.
The cliques of the dependence graph represent local configurations in the network.
(a configuration is a small subgraph of the network)
There is one parameter in the model for each clique.
(ie one parameter for each configuration.)
The model then is an expression of the probability of the network expressed in terms of local configurations.
Example: Dyadic independence models
Hypothesis about a possible local process: • Whether person i considers person j a
friend is contingent on whether person j considers person i a friend.
Cliques of the dependence graph
{Xij}
{Xji}
{Xij , Xji}
Dependence graph:
Xij Xji
i j
i j
i j
The Hammersley-Clifford theorem
Pr(X = x) = p*(x) = (1/c) exp{all cliques AzA}
where:the summation is over all cliques A;
zA = xijA xij is the network statistic corresponding to the clique A;
A is the parameter corresponding to clique A;
c = X exp{AAzA(x)} is a normalising constant
(Besag, 1974)
For dyadic independence
Cliques: {Xij}, {Xji}, {Xij, Xij}
So: Pr(X = x) = (1/c) exp{all cliques AzA}
– with zA = A xij
becomes:
Pr(X = x) = (1/c) exp{i,j ij xij + i,j ij,ji xij xji }
Homogeneity: ij= ; ij,ji =
Pr(X = x) = (1/c) exp{ i,j xij + i,j xij xji }
= (1/c) exp{ L + M }where L is no. of ties, M is no. of mutual ties
HomogeneityThe model cannot be estimated (unidentifiable) unless
some parameters are equated:– ie, assume that certain effects are the same for all (or
at least large parts) of the network
– eg, a single tendency for mutuality across the entire network.
Homogeneity of isomorphic network configurations– parameters equated for the same types of
configuration ignoring the numbering on the nodes
– statistics become the counts of the configurations in the network
– parameter interpretation: tendency for the configuration to be present in the network, given the other configurations.
Isomorphic configurations within blocks
Bernoulli graphs: the simplest dependence structure
Dependence assumption: all ties are independent.
Dependence graph:
Xij Xkl
Cliques: {Xij} i j
Hammersley-CliffordPr(X = x) = (1/c) exp{i,j
ij xij }
Homogeneity: ij= Pr(X = x) = (1/c) exp i,j xij
Pr(X = x) = (1/c) exp ( L )where L is no. of ties.
Bernoulli block models
Suppose actors are either in block 1 or 2
Hammersley-Clifford: Pr(X = x) = (1/c) exp{i,j ij xij }
Block homogeneity:ij = 11 if i and j both in block 1
ij = 12 if i in block 1 and j in block 2, etc.
Pr(X = x) = (1/c) exp{11 L11+12 L12+21 L21+22 L22}
where L rs is the number of edges from block r to block s.
Example: The network structure of a US law firm(Lazega & Pattison, 1999)
Respondents: All 71 lawyers Blocks: 36 partners (block 1), 35 associates (block
2)
Coworker and Advice ties:
The general question: – How is collective participation organised?
– Small, flexible, and temporary work-teams must be able to form quickly and cooperate efficiently in order to react to complex and non-standard problems.
– Regularities in local patterns of exchange provide one possible solution.
Bernoulli and dyad-independent models for advice ties
a. Homogeneous Bernoulli:
Pr(X = x) = (1/c) exp ( L )
b. Bernoulli blockmodel:
Pr(X = x) = (1/c) exp{11 L11+12 L12+21 L21+22 L22}
c. Homogeneous Dyad-independent:
Pr(X = x) = (1/c) exp{ L + M }
d. Dyad-independent block
Pr(X = x) = (1/c) exp{r,s=1,2 rs Lrs+ M }
e. Dyad-independent block with block reciprocity parameters:
Pr(X = x) = (1/c) exp{r,s=1,2 (rs Lrs+ rs Mrs )}
Results:Using pseudolikelihood estimation (to come)
PL deviance is a measure of fitMAR = Mean absolute residual
PL deviance MAR Parameters
a. Bernoulli model 4677.8 .294 1
b. Bernoulli block 4341.1 .276 5
c. Homogeneous dyad-independent model 4391.1 .275
2
d. Dyad-independent block model 4072.6 .258
5
e. Dyad-indpt block with block reciprocity
4071.2 .257 8
Parameter estimates for model c
Advice ties: (P=Partner, A=Associate)
Parameter MPLE
11 -1.34 Density of ties in P
12 -3.51 Ties from P to A
21 -1.53 Ties from A to P
22 -1.98 Density of ties in A
1.52 Reciprocity
Markov Random Graphs
Dyadic independence is an unrealistic assumption– Both theoretically and empirically.
Markov Dependencies (Frank & Strauss, 1986)
– Assume that a tie from i to j is contingent only on other ties involving person i and person j.
• network ties assumed conditionally dependent if and only if they share a common actor
– There is an edge (Xij, Xkl) in the dependence graph if and only if {i,j} {k,l}
Markov dependence graph for a directed network
Cliques of size 1 or 2 :
{ Xij }
{ Xij , Xji }
{ Xij , Xik }
{ Xij , Xjk }
{ Xij , Xkj }
Xjk
Xij
Xik
Xji
Xkj
Xki
Markov dependence graph for a directed network
Cliques of size 3 :
1. Stars
{ Xij , Xik , Xil }
etc2. TriadsTransitive triads
{ Xij , Xjk , Xik }
3-Cycles{ Xij , Xjk , Xki }
etc
Xjk
Xij
Xik
Xji
Xkj
Xki
Parameters for a homogeneous Markov random directed graph model
Density (15)
Two-in-stars (14)
Two-mixed-stars (13)
Two-out-stars (12)
Reciprocity (11)
Three-Cycles (10)
Transitive triads (9)
8
7
plus higher order stars
NB: Bernoulli models- Density only; Dyad independent - Density plus reciprocity
For non-directed graphs, parameters are simpler
Density or edge ()
Two-stars (2)
Triangle ()
Three-stars (3)
and higher order stars…
Interpretation of Markov random graph models
Broadly, positive parameter indicates a high occurrence of the associated configuration.
But the effects are marginal to each other.– A positive parameter indicates a greater number
of the configuration than expected, given the observation of other configurations
– Interpret a higher order parameter in relation to its lower order constituents.
– e.g. interpret cyclic triads in relation to 2-mixed-stars (2-paths): Positive cyclic parameter suggests presence of more cycles than expected from the number of 2-paths.
Example: Markov random graph model for mutual work ties among the
lawyers
Model parameters -2LPL MAR
_________________________________________
1. Edge (density) 2119.0 .258
2. edge, 2-star, 3-star, triangle 1760.8 .213
_________________________________________
parameter configuration parameter estimate
– 2.785 (.369)
2 – 0.019 (.030)
3 0.002 (.002)
0.482 (.035)
Lazega AdviceWork.ppt
Pseudo-likelihood estimation:an approximate technique
Conditional form of the Hammersley-Clifford theorem:
where:
(1) The sum is over all cliques A that contain Xij;
(2) A is the parameter corresponding to clique A;
(3) dA(ij) is the change statistic
• the change in the value of the network statistic zA(x) when xij changes from 1 to 0
Note: this version of the theorem takes the form of a conditional logistic regression.
( )
( 1 " ")log ( )
( 0 " ")ij
ijA A
A Xij
P X restd ij
P X rest
Calculating the change statistic
Eg: Markov random graph model for a nondirected network.– Graph statistics: edges, 2-stars, 3-stars, triangles.
Set up standard logistic regression file:– “Cases” are each possible tie (i, j).
– Variables: Xij , edges, 2-stars, 3-stars, triangles.
Eg, to calculate 2-stars value for each (i, j):– count 2-stars in the network when Xij =1
– count 2-stars in the network when Xij =0
– take the difference.
Predict Xij =1in a standard logistic regression from the other variables
Some warnings
This looks like a logistic regression: It is not!!– The standard errors are at best approximate
(probably too small).
– Do not rely on the Wald statistic.
– You cannot assume the pseudo-likelihood deviance is asymptotically chi-squared
But the PL deviance still indicates how well the model fits the observed network. – If a variable does not contribute “substantially”
to the PL deviance, maybe exclude it?
– But maybe keep lower order variables in the model if higher order variables are important.
– Report PL deviance and mean absolute residual.
Monte Carlo Maximum Likelihood Estimation
Some recent developments in this area, eg Snijders 2002– Degenerate regions in the parameter space
– So far, studies only involve simple non-directed Markov random graph models
– Possible importance of model specification
Implications for Pseudolikelihood estimation– Do not treat PL as hypothesis testing; it is more
exploratory.
– More work is needed to determine types of models, and regions of parameter space, for which we can be confident about PL estimation.
– Nevertheless, PL is convenient, and for complex models is still our only available estimation technique
Incorporating individual level variables
Attribute variables:e.g. sex, rank, attitudes– Yi = 1 if actor i possesses the attribute (e.g.
actor i is female)
– Yi = 0 otherwise
Example: Social selection models (Robins, Elliott & Pattison, 2001).
Similarity or homophily hypothesis: There is a propensity for a tie to develop between actors with similar attributes.
Directed dependence graphs
One type of variable predictive of another– Two block chain graphs
Attribute variables Network variables
(possibly some non-directed dependencies among network variables)
Example: Dyadic independence Markov attribute models
Yi
Yj
Xij
Xji
Configurations for Bernoulli graph attribute
models:
Additional configurations for dyadic
independence models
Markov graph - Markov attribute models
Xij
Xji
Xki
Xik
Xjk
Xkj
Yi
Yj
Additional configurations
Markov graph - partial dependent attribute models
A possible tie between two people is contingent on a third person’s attribute if the third person is tied to at least one of the first two.
E.g. k’s attribute possibly shapes whether there is a tie from i to j
i j
k
Lawyers’ advice network with status attribute
Model - 2 LPL Parameters MAR
Bernoulli graph models
1.No attribute 4678 1 0.295
2.Markov attribute 4341 4 0.276
Dyadic independence models
3.No attribute 4391 2 0.275
4.Markov attribute 4071 7 0.257
Markov graph models
5.No attribute 2855 7 0.175
6.Markov attribute 2819 15 0.173
7.Partial dpt. attribute 2693 39 0.166
8.Reduced part. dpt. att. 2749 15 0.169
____________________________________________________
Configurations and estimates for reduced partial dependence model
-2.13 0.77
-2.06 2.44
1.60
-0.03
-0.15 -0.22
0.10 0.13
0.05 0.09
- 0.09
-0.14 0.29
Other applications of directed dependencies
• Social influence models– Robins, Pattison & Elliott (2001)
• Discrete time models– Robins & Pattison (2001)
• Tripartite graphs– Mische & Robins (2002)
Further stepsPattison & Robins (2002)
Partial dependence models:
observed ties create new neighbourhoods – longer range dependencies
– longer cycles (eg generalized exchange)
Further stepsPattison & Robins (2002)
Setting structures:A setting: a subset of possible ties
potential "site of social action" e.g., focus, Feld (1981)network-domain (White, 1995)
Limits the scope of the dependencies in the dependence graph:
Xjk
Xij
Xik
Xji
Xkj
Xki
top related