Frederickson-Technology and Music Performance in the Age of Mechanical Reproduction
Design of Randomized Experiments on Networks When Treatment …samsi... · 2014-08-13 · Jake...
Transcript of Design of Randomized Experiments on Networks When Treatment …samsi... · 2014-08-13 · Jake...
Background Estimating Causal E↵ects Testing Causal Hypotheses Summary/Next Steps
Design of Randomized Experiments on NetworksWhen Treatment Propagates: An Exploration
Bruce Desmarais
University of Massachusetts Amherst
Collaborators:Jake Bowers, Wayne Lee, Simi Wang, Mark Frederickson, Nahomi Ichino
Prepared for the 2014 Political Networks Conference, McGill University
Background Estimating Causal E↵ects Testing Causal Hypotheses Summary/Next Steps
How should experiments be designed to learn aboutinterference in networks?
Example: Ghana 2008 Voter Registration Fraud Experiment (Ichino)
868 Electoral RegistrationStations (vertices)connected by roads (edges)
Density of 2.2%
Graph Transitivity of70.3%
2 / 20
Background Estimating Causal E↵ects Testing Causal Hypotheses Summary/Next Steps
Notation for Causal Questions
Treatment assigned to i : Zi 2 {0, 1}
Vector of all treatment assignments: Z = {Z1, . . . ,Zn}
Sharp Null of No Treatment E↵ect: all potential outcomes are equal
Yi (Z) = Yi (Z0) = Yi (Zi = 1,Z�i ) = Yi (Zi = 0,Z0
�i ) 8Z 6= Z
0
Graph-conditioned exposure: Yit(Dit), Dit records exposure condition
6 / 20
Background Estimating Causal E↵ects Testing Causal Hypotheses Summary/Next Steps
Example of treatment propagation: The Ising Model
Initial treatment status drawn from Zi ⇠ Bernoulli(↵). “Infection”/“Exposure”
probability at each iteration is
pr(Ei = 1) =1
1 + exp( 2F (ki � 2mi ))
k is number of (directly adjacent) neighbors (0 ki � K )
m is number of already infected neighbors (0 mi � M)
F is “temperature” or “propensity to be infected”
This example: Only two time periods (t 2 {0, 1}). The Ising model controlsactual infection after an experimenter assigns Zi at t = 0. Record infection afterthe first period.
7 / 20
Background Estimating Causal E↵ects Testing Causal Hypotheses Summary/Next Steps
Example of exposure types on the Ghana nodesOne draw from the Ising propagations with pr(Zi = 1) ⇠ Bernoulli(.15) fort 2 {0, 1} and Temperature= 10.
d1 ⌘ Di (Zi = 1, 0 mit � M)
d(0,0) ⌘ Di (Zi = 0,mit = 0)
d(0,1) ⌘ Di (Zi = 0,mit � 1)
d01
d01
d01
d1
d1
d01
d01
d01d01
d01
d01d1
d01
d01d01d01
d01
d01
d01
d00
d00
d00
d00
d01
d00
d00
d01
d00
d00
d01d1
d01
d1
d00 d00
d01d01
d1
d00
d1
d01 d01
d1d01
d1
d01
d01
d01d00
d01d1
d1
d01 d01
d01
d01
d01
d01d01
d01d01
8 / 20
Background Estimating Causal E↵ects Testing Causal Hypotheses Summary/Next Steps
ATE: Aronow and Samii Estimator
Definitions:
Di indexes the exposure condition of observation i
⇡i (dk) is the probability i ends up in condition dkWe consider
d(0,1), untreated with at least one treated neighbor
d(0,0), untreated with no treated neighbor
Estimand:
µ̂(dk) =1
n
nX
i=1
I(Di = dk)Yi (dk)
⇡i (dk)
⌧̂(d(0,1), d(0,0)) = µ̂(d(0,1))� µ̂(d(0,0))
Testing:
Following Horvitz and Thompson (1952) and CLT:Var(⌧̂(d(0,1), d(0,0))) 1/N2
�Var(µ̂(d(0,1))) + Var(µ̂(d(0,0)))
�
9 / 20
Background Estimating Causal E↵ects Testing Causal Hypotheses Summary/Next Steps
Simulations to assess power and guide design
Baseline (pre-treatment) response
Y (0, 0) ⇠ U(0, 1)
Treatment allowed to propagate for one period and true (constant,multiplicative) model of causal e↵ects computed:
Y (1, 0) = Y (0, 1) = �Y (0, 0)
Consider parametersTreatment assignment probability ↵ 2 {0.05, 0.10, . . . , 0.50}‘Temperature’ T 2 {0, 10, . . . , 100}� 2 {0.26, 0.63}
Simulate 1,000 propagations at each combination of T and ↵.
For each realized propagation, assess power to reject H0 : ⌧ = 0 atsignificance level .05.
10 / 20
Background Estimating Causal E↵ects Testing Causal Hypotheses Summary/Next Steps
Power as a Function of N Indirectly Treated
� = 0.26 � = 0.63
100 200 300 400
0.0
0.2
0.4
0.6
0.8
Prop
ortio
n p=
=0.0
5
0.05
0.05
0.05
0.05
0.050.050.050.050.050.05
0.1
0.1
0.10.1
0.10.10.10.10.10.1
0.15
0.15
0.150.150.150.150.15
0.150.150.15
0.2
0.2
0.20.20.20.20.20.20.20.2
0.25
0.250.250.250.250.250.250.250.250.25
0.3
0.30.30.30.30.30.30.30.30.3
0.350.350.350.350.350.350.350.350.350.35
0.40.40.40.40.40.40.40.40.40.40.450.450.450.450.450.450.450.450.450.450.50.50.50.50.50.50.50.50.50.5
N Exposed to Treatment by Propagation in 1 Period100 200 300 400
0.00
0.05
0.10
0.15
Prop
ortio
n p=
=0.0
5
0.05
0.05
0.05
0.050.05
0.05
0.05
0.050.05
0.05
0.1 0.10.1
0.10.10.10.1
0.10.10.1
0.15 0.15 0.150.150.150.150.150.150.150.150.2 0.2 0.20.20.20.20.20.20.20.20.25 0.250.250.250.250.250.250.250.250.250.3 0.30.30.30.30.30.30.30.30.30.350.350.350.350.350.350.350.350.350.350.40.40.40.40.40.40.40.40.40.40.450.450.450.450.450.450.450.450.450.450.50.50.50.50.50.50.50.50.50.5
N Exposed to Treatment by Propagation in 1 Period
11 / 20
Background Estimating Causal E↵ects Testing Causal Hypotheses Summary/Next Steps
Power With Additive E↵ects Model
Y (0, 0) = 1 + U(0, 1)
Y (0, 1) = �+ U(0, 1)
� = 0.26 � = 0.63
100 200 300 400
0.00
0.05
0.10
0.15
0.20
0.25
0.30
Prop
ortio
n p=
=0.0
5
0.05
0.05
0.05
0.05
0.05
0.050.05
0.050.050.05
0.10.1
0.1
0.10.10.10.10.1
0.10.1
0.15 0.150.150.150.150.150.150.150.150.15
0.2 0.2 0.20.20.20.20.20.20.20.20.25 0.250.250.250.250.250.250.250.250.250.3 0.30.30.30.30.30.30.30.30.30.350.350.350.350.350.350.350.350.350.350.40.40.40.40.40.40.40.40.40.40.450.450.450.450.450.450.450.450.450.450.50.50.50.50.50.50.50.50.50.5
N Exposed to Treatment by Propagation in 1 Period100 200 300 400
0.00
0.01
0.02
0.03
0.04
0.05
0.06
0.07
Prop
ortio
n p=
=0.0
50.05
0.05
0.05
0.050.050.05
0.05
0.050.05
0.05
0.1 0.10.1
0.10.10.10.1
0.10.10.1
0.15 0.15 0.150.150.150.150.150.150.150.150.2 0.2 0.20.20.20.20.20.20.20.20.25 0.250.250.250.250.250.250.250.250.250.3 0.30.30.30.30.30.30.30.30.30.350.350.350.350.350.350.350.350.350.350.40.40.40.40.40.40.40.40.40.40.450.450.450.450.450.450.450.450.450.450.50.50.50.50.50.50.50.50.50.5
N Exposed to Treatment by Propagation in 1 Period
12 / 20
Background Estimating Causal E↵ects Testing Causal Hypotheses Summary/Next Steps
Power With Certain Propagation
Y (d00) = 1 + U(0, 1)
Y (d01) = �+ U(0, 1)
� = 0.26 � = 0.63
100 200 300 400
0.0
0.1
0.2
0.3
0.4
0.5
0.6
Prop
ortio
n p=
=0.0
5
0.05 0.050.05
0.050.05
0.050.050.050.050.05
0.1 0.10.1 0.10.1
0.10.10.1
0.1
0.1
0.15 0.15 0.150.150.150.150.150.150.150.15
0.2 0.2 0.20.20.20.20.20.20.20.20.25 0.250.250.250.250.250.250.250.250.250.3 0.30.30.30.30.30.30.30.30.30.350.350.350.350.350.350.350.350.350.350.40.40.40.40.40.40.40.40.40.40.450.450.450.450.450.450.450.450.450.450.50.50.50.50.50.50.50.50.50.5
N Exposed to Treatment by Propagation in 1 Period100 200 300 400
0.00
0.02
0.04
0.06
0.08
0.10
0.12
Prop
ortio
n p=
=0.0
5
0.05
0.050.05 0.050.05
0.05
0.05
0.05
0.05
0.05
0.1
0.10.1
0.10.10.10.10.10.10.1
0.15 0.15 0.150.150.150.150.150.150.150.150.2 0.2 0.20.20.20.20.20.20.20.20.25 0.250.250.250.250.250.250.250.250.250.3 0.30.30.30.30.30.30.30.30.30.350.350.350.350.350.350.350.350.350.350.40.40.40.40.40.40.40.40.40.40.450.450.450.450.450.450.450.450.450.450.50.50.50.50.50.50.50.50.50.5
N Exposed to Treatment by Propagation in 1 Period
13 / 20
Correlating Treatment with Degree
●
●
●●●●●●
●●
●●●
●●●
●
●●●●●
● ●
●
−0.10 −0.05 0.00 0.05 0.10
0.45
0.50
0.55
0.60
0.65
0.70
0.75
Prop
ortio
n Si
g. (p
=0.0
5)
Degree−Treatment Corr.
α = 0.05 ●
●
●●
●
●●●●●
●
●
●
●
●●
●
●
●●
●
●
● ●
●
−0.10 −0.05 0.00 0.05 0.10
0.35
0.40
0.45
0.50
0.55
0.60
0.65
Prop
ortio
n Si
g. (p
=0.0
5)
Degree−Treatment Corr.
α = 0.1 ●
● ●●●
●●●●
●●●
●
●●●
●●●●
●
●
●
●
●
−0.10 −0.05 0.00 0.05 0.10
0.2
0.3
0.4
0.5
Prop
ortio
n Si
g. (p
=0.0
5)
Degree−Treatment Corr.
α = 0.15
●
●
●
●
●
●●
●
●
●
●●
●
●
●●
●●●●●
●
●
●
●
−0.10 −0.05 0.00 0.05 0.10
0.100.150.200.250.300.350.400.45
Prop
ortio
n Si
g. (p
=0.0
5)
Degree−Treatment Corr.
α = 0.2 ●
●
●
●●
●
●
●
●
●
●
●●●
●
●●
●
●
●
●
●
●
●
●
−0.10 −0.05 0.00 0.05 0.10
0.05
0.10
0.15
0.20
0.25
Prop
ortio
n Si
g. (p
=0.0
5)
Degree−Treatment Corr.
α = 0.25 ●
●
●
●
●
●●●
●●
●●
●
●●
●
●
●●●●
●
●●
●
−0.10 −0.05 0.00 0.05 0.10
0.05
0.10
0.15
0.20
0.25
Prop
ortio
n Si
g. (p
=0.0
5)
Degree−Treatment Corr.
α = 0.3
●
●
●
●●
●
●
●
●
●
●●
●●
●
●●●
●●
●●
● ●
●
−0.10 −0.05 0.00 0.05 0.10
0.02
0.04
0.06
0.08
0.10
0.12
0.14
Prop
ortio
n Si
g. (p
=0.0
5)
Degree−Treatment Corr.
α = 0.35 ●
● ●●●●●●
●
●●●
●●
●
●
●
●
●
●
●
●
●
●
●
−0.10 −0.05 0.00 0.05 0.10
0.00
0.02
0.04
0.06
0.08
Prop
ortio
n Si
g. (p
=0.0
5)
Degree−Treatment Corr.
α = 0.4 ●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
−0.10 −0.05 0.00 0.05 0.10
0.01
0.02
0.03
0.04
0.05
Prop
ortio
n Si
g. (p
=0.0
5)
Degree−Treatment Corr.
α = 0.45
Background Estimating Causal E↵ects Testing Causal Hypotheses Summary/Next Steps
Tests of a Sharp Null: Distributional Di↵erences
Bowers, Jake, Mark M. Fredrickson and Costas Panagopoulos. 2013.“Reasoning about Interference Between Units: A General Framework.” PoliticalAnalysis 21(1):97–124.
Details: Anderson-Darling k-sample test statistic (Scholz and Stephens, 1987);randomization distributions via RItools (Fredrickson, Bowers, Hansen 2014).
15 / 20
Background Estimating Causal E↵ects Testing Causal Hypotheses Summary/Next Steps
Power as a Function of N Indirectly Treated
100 200 300 400
0.5
0.6
0.7
0.8
0.9
1.0
Power to Reject the Sharp Null of No Effects
N Exposed to Treatment by Propagation in 1 Period
Prop
ortio
n p=
0.05
0.05
0.1
0.15
0.2
0.25 0.30.350.40.450.5
0.050.10.150.20.250.30.350.40.450.5
100 200 300 400
0.5
0.6
0.7
0.8
0.9
1.0
Power to Reject the Sharp Null of No Difference D01 vs D00
N Exposed to Treatment by Propagation in 1 Period
Prop
ortio
n p=
0.05
0.05
0.1
0.15
0.2 0.25 0.3 0.350.40.450.5
0.050.10.150.20.250.30.350.40.450.5
16 / 20
Background Estimating Causal E↵ects Testing Causal Hypotheses Summary/Next Steps
Power With Additive E↵ects Model
100 200 300 400
0.70
0.75
0.80
0.85
0.90
0.95
1.00
Power to Reject the Sharp Null of No Effects
N Exposed to Treatment by Propagation in 1 Period
Prop
ortio
n p
<= 0
.05
0.05
0.1
0.15
0.2 0.25 0.30.350.40.450.5
0.050.10.150.20.250.30.350.40.450.5
100 200 300 400
0.70
0.75
0.80
0.85
0.90
0.95
1.00
Power to Reject the Sharp Null of No Difference D01 vs D00
N Exposed to Treatment by Propagation in 1 Period
Prop
ortio
n p
<= 0
.05
0.05 0.1
0.150.2 0.25
0.30.350.40.45
0.5
17 / 20
Background Estimating Causal E↵ects Testing Causal Hypotheses Summary/Next Steps
Power With Certain Propagation
100 200 300 400
0.70
0.75
0.80
0.85
0.90
0.95
1.00
Power to Reject the Sharp Null of No Effects
N Exposed to Treatment by Propagation in 1 Period
Prop
ortio
n p
<= 0
.05
0.05
0.1 0.15 0.2 0.25 0.30.350.40.450.5
0.050.10.150.20.250.30.350.40.450.5
100 200 300 400
0.70
0.75
0.80
0.85
0.90
0.95
1.00
Power to Reject the Sharp Null of No Difference D01 vs D00
N Exposed to Treatment by Propagation in 1 Period
Prop
ortio
n p
<= 0
.05
0.05 0.1 0.15 0.2 0.25 0.30.350.40.450.5
18 / 20
Background Estimating Causal E↵ects Testing Causal Hypotheses Summary/Next Steps
Summary
Under propagation, power optimizing proportion assigned to treatment maybe much less than 0.5.
Higher order parameters of treatment assignment distribution candramatically a↵ect statistical power
Trade-o↵ between power to detect network-moderated andnon-network-moderated e↵ects (see also Bowers, Fredrickson,Panagopoulos 2013).
Test statistics matter: even if interest focuses on ATE, the ATE mayprovide little power when the true model is not a simple distributional shift.
Limitation on power driven by the need for strong (i.e., isolated) controls.
19 / 20
Background Estimating Causal E↵ects Testing Causal Hypotheses Summary/Next Steps
Next Steps
More measures of design adequacy: RMSE, Type I error, Power.
Make pr(Zi = 1) depend on topology if not also covariates.degree-correlated assignment
community-wise assignment schemes
Consider higher-order spreading
Re-parameterize Ising for substantive interpretation.
Provide more general simulation system that accommodates di↵erent:models of propagation
models of e↵ects
statistical/causal inferential focus
20 / 20