07350015%2E2013%2E792263
Journal of Business & Economic Statistics
ISSN: 0735-0015 (Print) 1537-2707 (Online) Journal homepage: http://www.tandfonline.com/loi/ubes20
Comment
Bruce Sacerdote
To cite this article: Bruce Sacerdote (2013) Comment, Journal of Business & Economic
Statistics, 31:3, 275-275, DOI: 10.1080/07350015.2013.792263
To link to this article: http://dx.doi.org/10.1080/07350015.2013.792263
Published online: 22 Jul 2013.
Submit your article to this journal
Article views: 164
View related articles
Full Terms & Conditions of access and use can be found at
http://www.tandfonline.com/action/journalInformation?journalCode=ubes20
Download by: [Universitas Maritim Raja Ali Haji]
Date: 11 January 2016, At: 22:11
Sacerdote: Comment
275
One reason for inference on structural parameters may be
“science.” Researchers may want to characterize reality, as an
end in itself. A different reason is to predict treatment response
when a regime change (an uber treatment) alters part of a structural model in a known way, leaving other parts invariant. Then
response functions change in a way that can be predicted with
knowledge of the structural model but not otherwise.
Consider the linear-in-means response function
yj (t J ) =
α
γβ1 + β2
+ β1 tj +
E(t|x) + uj
1−γ
1−γ
Downloaded by [Universitas Maritim Raja Ali Haji] at 22:11 11 January 2016
+ ϕ0 + ϕ1 tj + ϕ2 E(t|xj ) + uj .
(6)
A regime change might alter some of the structural parameters
(α, β 1 , β 2 , γ ) in a known way while leaving others unchanged.
Then knowledge of (α, β 1 , β 2 , γ ) enables prediction of treatment
response but knowledge of (ϕ 0 , ϕ 1 , ϕ 2 ) does not.
REFERENCES
Kiefer, N., and Goldberger, A. (1989), “The ET Interview: Arthur S.
Goldberger,” Econometric Theory, 5, 133–160. [274]
Manski, C. (1993), “Identification of Endogenous Social Effects: The Reflection
Problem,” Review of Economic Studies, 60, 531–542. [273,274]
—— (2013), “Identification of Treatment Response with Social Interactions,”
The Econometric Journal, 16, S1–S23. [274]
Comment
Bruce SACERDOTE
Department of Economics, Dartmouth College, Hanover, NH 03755 and NBER (Bruce.Sacerdote@dartmouth.edu)
This is an innovative article and a nice addition to the literature on the estimation of endogenous and exogenous peer
effects. There are two main contributions. First, the authors
suggest that we can incorporate the endogeneity of peer choice
in a parsimonious way. The authors introduce an unobserved
individual specific parameter ξ i . For any two individuals i and
j, the distance between ξ i and ξ j affects the probability that i
and j form a link. Then this ξ i is introduced directly and linearly
(in Equation (6.1)) as a determinant of i’s outcome Yi . This is a
very clever approach and has the potential to greatly reduce the
complexity of an otherwise intractable problem.
The second advance of the article is to show that all the
model’s parameters (including the ξ i ’s) can in principle be estimated in a Bayesian framework using Monte Carlo methods.
This answers the obvious question of how we might estimate
the individual specific unobserved regressor.
The authors proceed to estimate their model using Ad Health
data and calculating exogenous and endogenous peer effects on
own Grade Point Average. The estimates seem quite plausible.
For example, an individual’s own past grades predict current
grades with a coefficient of 0.73. Peers’ past grades predict
own current grades with a coefficient of 0.11. Such estimates
are in the same ballpark as existing articles that have random
assignment to classrooms.
Interestingly the introduction of endogenous network formation (through the vector of etas) does not have a meaningful
impact on the estimated peer effects. Compare, for example,
the results in Tables 5 and 6, where the former table assumes
that peer choice is exogenous. My own experimentation with
the model found much the same result. Using data from a military academy with randomly assigned squadrons, I found that
accounting for an individual specific ξ i (which affects friend
choice) affects the outcome, but does not affect the estimated
peer effects. (My coauthors Scott Carrell and James West who
have access to the data were kind enough to run the code for
me.)
Besides being a fan of the article, I have two general comments. First, not all readers will accept the simple parameterization of friendship choice as having solved the peer selection
problem. I suspect that the authors’ formulation works particularly well in their example and in my example because in both
cases we have strong controls for own ability and peer (background) ability. Staiger and Kane have convinced me that in test
score value added models, having prior test scores does a great
deal to compensate for the selection of students into classrooms
and schools.
Second, economics researchers have become progressively
less interested in the linear-in-means model. Models beyond the
linear-in-means models allow the possibility for Pareto improving reallocations of students, such as tracking students (grouping
them into classrooms) by ability. Hoxby and Weingarth (2005)
and my own work with Imberman and Kugler finds that nonlinear models fit the data much better. I suspect that with a
minimum of tinkering the authors’ model could be extended
to a more flexible (nonlinear) formulation. Part of the beauty
of the Markov chain Monte Carlo method being used is that a
wide variety of models can be estimated even in cases where we
cannot conduct maximum likelihood estimation.
Overall I found this to be a thoughtful article and a worthwhile
contribution.
REFERENCE
Hoxby, C. M., and Weingarth, G. (2005), “Taking Race Out of the Equation:
School Reassignment and the Structure of Peer Effects,” Working Paper,
Harvard University. [275]
© 2013 American Statistical Association
Journal of Business & Economic Statistics
July 2013, Vol. 31, No. 3
DOI: 10.1080/07350015.2013.792263
ISSN: 0735-0015 (Print) 1537-2707 (Online) Journal homepage: http://www.tandfonline.com/loi/ubes20
Comment
Bruce Sacerdote
To cite this article: Bruce Sacerdote (2013) Comment, Journal of Business & Economic
Statistics, 31:3, 275-275, DOI: 10.1080/07350015.2013.792263
To link to this article: http://dx.doi.org/10.1080/07350015.2013.792263
Published online: 22 Jul 2013.
Submit your article to this journal
Article views: 164
View related articles
Full Terms & Conditions of access and use can be found at
http://www.tandfonline.com/action/journalInformation?journalCode=ubes20
Download by: [Universitas Maritim Raja Ali Haji]
Date: 11 January 2016, At: 22:11
Sacerdote: Comment
275
One reason for inference on structural parameters may be
“science.” Researchers may want to characterize reality, as an
end in itself. A different reason is to predict treatment response
when a regime change (an uber treatment) alters part of a structural model in a known way, leaving other parts invariant. Then
response functions change in a way that can be predicted with
knowledge of the structural model but not otherwise.
Consider the linear-in-means response function
yj (t J ) =
α
γβ1 + β2
+ β1 tj +
E(t|x) + uj
1−γ
1−γ
Downloaded by [Universitas Maritim Raja Ali Haji] at 22:11 11 January 2016
+ ϕ0 + ϕ1 tj + ϕ2 E(t|xj ) + uj .
(6)
A regime change might alter some of the structural parameters
(α, β 1 , β 2 , γ ) in a known way while leaving others unchanged.
Then knowledge of (α, β 1 , β 2 , γ ) enables prediction of treatment
response but knowledge of (ϕ 0 , ϕ 1 , ϕ 2 ) does not.
REFERENCES
Kiefer, N., and Goldberger, A. (1989), “The ET Interview: Arthur S.
Goldberger,” Econometric Theory, 5, 133–160. [274]
Manski, C. (1993), “Identification of Endogenous Social Effects: The Reflection
Problem,” Review of Economic Studies, 60, 531–542. [273,274]
—— (2013), “Identification of Treatment Response with Social Interactions,”
The Econometric Journal, 16, S1–S23. [274]
Comment
Bruce SACERDOTE
Department of Economics, Dartmouth College, Hanover, NH 03755 and NBER (Bruce.Sacerdote@dartmouth.edu)
This is an innovative article and a nice addition to the literature on the estimation of endogenous and exogenous peer
effects. There are two main contributions. First, the authors
suggest that we can incorporate the endogeneity of peer choice
in a parsimonious way. The authors introduce an unobserved
individual specific parameter ξ i . For any two individuals i and
j, the distance between ξ i and ξ j affects the probability that i
and j form a link. Then this ξ i is introduced directly and linearly
(in Equation (6.1)) as a determinant of i’s outcome Yi . This is a
very clever approach and has the potential to greatly reduce the
complexity of an otherwise intractable problem.
The second advance of the article is to show that all the
model’s parameters (including the ξ i ’s) can in principle be estimated in a Bayesian framework using Monte Carlo methods.
This answers the obvious question of how we might estimate
the individual specific unobserved regressor.
The authors proceed to estimate their model using Ad Health
data and calculating exogenous and endogenous peer effects on
own Grade Point Average. The estimates seem quite plausible.
For example, an individual’s own past grades predict current
grades with a coefficient of 0.73. Peers’ past grades predict
own current grades with a coefficient of 0.11. Such estimates
are in the same ballpark as existing articles that have random
assignment to classrooms.
Interestingly the introduction of endogenous network formation (through the vector of etas) does not have a meaningful
impact on the estimated peer effects. Compare, for example,
the results in Tables 5 and 6, where the former table assumes
that peer choice is exogenous. My own experimentation with
the model found much the same result. Using data from a military academy with randomly assigned squadrons, I found that
accounting for an individual specific ξ i (which affects friend
choice) affects the outcome, but does not affect the estimated
peer effects. (My coauthors Scott Carrell and James West who
have access to the data were kind enough to run the code for
me.)
Besides being a fan of the article, I have two general comments. First, not all readers will accept the simple parameterization of friendship choice as having solved the peer selection
problem. I suspect that the authors’ formulation works particularly well in their example and in my example because in both
cases we have strong controls for own ability and peer (background) ability. Staiger and Kane have convinced me that in test
score value added models, having prior test scores does a great
deal to compensate for the selection of students into classrooms
and schools.
Second, economics researchers have become progressively
less interested in the linear-in-means model. Models beyond the
linear-in-means models allow the possibility for Pareto improving reallocations of students, such as tracking students (grouping
them into classrooms) by ability. Hoxby and Weingarth (2005)
and my own work with Imberman and Kugler finds that nonlinear models fit the data much better. I suspect that with a
minimum of tinkering the authors’ model could be extended
to a more flexible (nonlinear) formulation. Part of the beauty
of the Markov chain Monte Carlo method being used is that a
wide variety of models can be estimated even in cases where we
cannot conduct maximum likelihood estimation.
Overall I found this to be a thoughtful article and a worthwhile
contribution.
REFERENCE
Hoxby, C. M., and Weingarth, G. (2005), “Taking Race Out of the Equation:
School Reassignment and the Structure of Peer Effects,” Working Paper,
Harvard University. [275]
© 2013 American Statistical Association
Journal of Business & Economic Statistics
July 2013, Vol. 31, No. 3
DOI: 10.1080/07350015.2013.792263