2. Why
do
we
care?
(alterna3ve
being
“laisser-‐faire”)
• Equity
– In
presence
of
scarcity
– In
urban
contexts
where
students
mobility
is
high
and
fashion/herding
can
create
conges3on
• Private
versus
social
preferences
over
school
composi3on
– Externali3es:
social
cohesion,
academic
diversity,
.
3. Top-‐down
or
BoQom-‐up?
• Premise:
We
want
to
take
parents
/
students’
preferences
as
much
as
possible
into
account
– Some
combina3on
of
top-‐down
and
boQom-‐up
– Different
countries
/
districts
locate
themselves
differently
on
this
scale
Need
to
be
able
to
handle
preference
informa3on
together
with
(poli3cal
/
school)
priori3es
4. Three
criteria
for
candidate
procedures
• Efficiency
– A
procedure
is
efficient
is
there
does
not
exist
another
alloca3on
of
students
to
schools
such
that
every
student
is
beQer
off
and
at
least
one
is
strictly
beQer
off
• No
jus1fied
envy
– There
is
no
student
that
has
a
place
in
a
school,
whereas
another
one
who
actually
has
priority
over
that
student
at
that
school,
and
prefers
that
school
to
the
school
he’s
assigned
to,
does
not
have
one.
• Strategic
simplicity
– It
should
be
in
the
interest
of
parents
to
reveal
their
true
preferences
instead
of
manipula3ng
them
– Equity
and
efficiency
considera3ons
5. School
choice
mechanisms
• Inputs:
– Reports
by
students
over
schools
(rank
order
list,
ROL)
– Quotas
and
student
priori3es
at
each
school
– School
capaci3es
• No
procedure
sa3sfy
all
three
criteria
when
priori3es
are
not
strict
at
all
schools
– Top
trading
Cycles
and
Deferred
Acceptance
best
in
class
6. Student-‐proposing
deferred
acceptance
algorithm
(Gale-‐Shapley)
• Students
submit
their
ROLs
and
schools
their
priori3es
over
students
(use
of
a
3e-‐breaker
if
necessary)
• Step
1:
Each
student
proposes
to
her
first
choice.
Each
school
tenta3vely
assigns
its
seats
to
its
proposers
one
at
a
3me
following
their
priority
order.
Any
remaining
proposer
is
rejected.
• …
• Step
k:
Each
student
who
was
rejected
in
the
previous
step
proposes
to
her
next
choice.
Each
school
considers
the
students
it
had
tenta3vely
accepted
in
the
previous
period
together
with
the
new
proposers
and
accepts
tenta3vely
those
with
the
highest
priori3es.
It
rejects
other.
• The
algorithm
terminates
when
no
more
requests
are
rejected.
7. Example
– 4
kids,
2
schools
with
2
seats
each
– Student
preferences:
– Student
a:
1
2
Student-‐proposing
DAA,
first
round:
– Student
b:
1
2
Students
apply
to
their
first
choice
– Student
c:
1
2
school.
School
1
rejects
student
c
Round
2:
Student
c
applies
to
school
– Student
d:
2
1
2
and
is
accepted
– Priori3es
over
students:
– School
1:
a
d
b
c
– School
2:
b
a
c
d
7
8. Comments
• Centraliza3on
is
necessary
to
make
this
run
smoothly
(takes
a
few
minutes
to
run
on
a
computer)
• Poli3cal
objec3ves
are
translated
into
priori3es
and
quotas
• Interface
for
parents
to
input
preferences
9. Ac3ve
field
of
policy
• Many
school
districts
are
revamping
their
school
choice
procedures
– Drivers:
technology
and
pressure
to
introduce
choice
• Not
a
“one-‐size-‐fits-‐all”
solu3on
– Tailoring
to
policy
objec3ves
needed
– Parents’
aspira3ons
and
poli3cal
acceptability
10. Ac3ve
field
of
research
• Proper3es
of
procedures
• Applica3ons
and
access
to
data
open
an
opportunity
to
answer
new
ques3ons
– Long
term
effects
of
school
choice
regula3on
on
school
composi3on
and
student
outcomes?
– Preference
forma3on?
• “Matching
in
Prac3ce”
network
gathers
informa3on
on
procedures
and
outcomes
across
Europe
12. Data
• Preschool
popula3on
in
Dutch-‐speaking
preschools
in
Brussels
as
of
1
October
2008
(10,867
kids,
150
schools,
entering
class
4079)
• Kid
characteris3cs:
age,
loca3on,
na3onality,
GOK
status,
socioeconomic
class
of
neighborhood,
whether
Dutch
is
spoken
at
home,
school
aQended
• School
characteris3cs:
loca3on,
network,
confessional
orienta3on,
establishments,
pedagogy
13. Legal
constraints
on
the
mechanisms
Current
procedure:
– Siblings
have
priori3es
over
other
kids
– 30%
quota
for
GOK
students
– 45%
quota
for
Dutch
na3ve
speakers
– Priori3es
and
quotas
implemented
through
early
registra3on
periods
– First
come,
first
served
as
a
3e-‐breaker
– Decentralized
New
GOK
decree
allows
them
to
experiment
with
distance
as
a
different
3e-‐breaker
14. Analysis
of
the
current
situa3on
–
heterogeneity
across
schools
Percentage
of
GOK
students
and
na1ve
speakers
across
schools
0.8
0.7
%
GOK
students
0.6
%
Dutch
@
home
0.5
0.4
0.3
0.2
0.1
0
10%
lowest
2
3
4
5
6
7
8
9
10%
highest
15. Analysis
of
the
current
situa3on
–
distance
to
school
Brussels
kids
going
to
preschool
in
Brussels
-‐
closest
school
0.35
0.3
whole
sample
0.25
low
socio
high
socio
gok
0.2
Dutch
@
home
0.15
0.1
0.05
0
closest
2
to
3
4
to
5
6
to
10
11
to
15
16
to
20
21
to
30
above
30
1141
incoming
students,
958
outgoing
students,
16. Genera3ng
a
counterfactual
policy
experiment
• LOP
Brussels
is
considering
to
replace
its
3me
priority
with
a
distance-‐based
3e
breaker
.
• How
will
kids
be
impacted?
How
will
schools
be
impacted?
• Main
challenge
:
We
do
not
observe
preferences
over
schools
17. Calibra3ng
preferences
Working
assump3ons:
– Current
procedure
can
be
approximated
by
a
student-‐
proposing
DAA
with
socioeconomic
status,
then
distance
as
a
3e-‐breaker
– Brussels-‐based
students
have
preferences
over
Brussels
schools
that
depend
on
their
socioeconomic
status
(top
30%,
GOK,
other)
uis
=
α1k
distanceis
+
α2k
qualitys
+
(1-‐
α1k
-‐
α2k
)εis
They
also
have
an
outside
op3on
(random
u3lity)
and
place
the
school
where
they
have
a
sibling
first
– Out-‐of-‐Brussels
students
have
preferences
that
take
the
form
uis
=
δ
qualitys
+
(1-‐
δ)εis
18. Calibra3ng
preferences
(con3nued)
Calibrate
these
preferences
so
that
predicted
outcome
(distribu3on
of
ranks
of
assigned
school)
close
to
actual
outcome
α1high
=
0.55
Weight
on
ε
set
to
0.05
α1GOK
=
0.70
α1rest
=
0.58
δ
=
0.75
21. Counterfactual
1:
Impact
on
school
popula3on
Propor1on
of
Dutch
na1ve
speakers
-‐
before
and
aQer
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
10%
2
3
4
5
6
7
8
9
10
%
lowest
highest
simulated
"before"
"ater"
actual
22. Impact
on
school
popula3on
(cont’d)
Propor1on
of
GOK
students
before
and
aQer,
per
decile
of
schools
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
10%
2
3
4
5
6
7
8
9
10
%
lowest
highest
simulated
"before"
"ater"
actual
23. Counterfactual
1:
Likely
long
term
residen3al
effects
• Mean
median
distance
to
school
goes
from
1.45
km
to
0.9
km
• Mean
max
distance
to
school
goes
from
11.17
km
to
10.54
km
– max
distance
goes
down
in
41
schools
out
of
147
– Min
max
distance
goes
from
0.94
km
to
0.45
km