METHODOLOGY
FOR
WASTEWISE
MULTI­
STAGE
STATISTICAL
ANALYSIS
The
method
used
to
estimate
the
WasteWise
impact
accounts
for
two
data
facts:
(
1)
not
all
WasteWise
partners
report
waste
prevention
and
recycling
data
and
(
2)
we
have
no
information
on
waste
prevention
and
recycling
among
nonpartners.
To
account
for
these
data
problems,
EPA
developed
a
method
that
relies
on
statistical
modeling
to
account
for
the
missing
data.
One
important
fact,
however,
is
that
although
we
do
not
have
waste
prevention
and
recycling
data
for
non­
reporting
partners
and
non­
partners,
we
do
have
access
to
other
information
on
these
companies
(
e.
g.,
financial
information,
size).
Use
of
this
other
information
helps
us
to
control
for
the
fact
that
we
are
missing
outcome
(
i.
e.,
waste
prevention
and
recycling)
data.
In
statistical
terms,
we
have
censored
data.
Thus,
although
we
are
missing
outcome
data
for
some
companies,
we
know
which
companies
are
missing
data
and
we
have
other
information
on
these
companies.

The
method
can
best
be
described
as
a
multi­
stage
statistical
analysis
that
uses
statistical
modeling
to
develop
control
variables
to
overcome
the
lack
of
outcome
data.
In
the
first
part
of
the
analysis,
we
look
at
what
influences
firms=
decisions
to
join
WasteWise.
From
this
analysis
we
create
a
variable
that
reflects
the
probability
that
a
firm
is
a
WasteWise
partner.
1
In
the
second
part
of
the
analysis,
we
measure
how
a
number
of
factors,
including
the
probability
of
being
a
WasteWise
partner,
affect
waste
prevention
and
recycling.
The
relationship
between
the
probability
of
being
a
WasteWise
partner
and
the
outcome
variables
(
waste
prevention
and
recycling)
will
determine
the
program
effect
of
WasteWise.

Using
the
probability
of
being
a
WasteWise
partner
adjusts
for
the
lack
of
outcome
data
for
nonpartners.
To
understand
this,
consider
the
situation
where
outcome
data
is
are
available
for
both
partners
and
nonpartners.
We
will
refer
to
this
as
the
full
data
scenario.
In
this
situation,
we
can
construct
a
binary
variable
equal
to
one
if
the
firm
was
a
partner
and
zero
if
not.
The
statistical
analysis
would
then
compare
outcomes
among
partners
(
binary
variable
equal
to
one)
to
nonpartners
(
binary
variable
equal
to
zero)
while
controlling
for
other
relevant
factors
(
size,
financial
status).
The
difference
between
outcomes
for
partners
and
nonpartners
after
controlling
for
other
relevant
factors
is
the
program
effect
of
WasteWise.
This
type
of
analysis
is
not
possible
because
we
do
not
have
outcome
data
for
nonpartners.

1
In
some
sense,
this
variable
is
an
artificial
construction
since
we
know
which
firms
are
and
which
are
not
partners.
We
need
to
construct
this
variable,
however,
to
deal
with
the
missing
outcome
data.
To
develop
an
estimate
of
the
program
effect
in
the
absence
of
outcome
data
for
nonpartners,
we
have
developed
a
statistical
approach
method
that
combines
some
well­
accepted
methods.
In
terms
of
the
discussion
above,
we
transform
the
binary
variable
into
one
that
is
continuous
between
zero
and
one:
i
(
i.
e.,
the
probability
of
being
a
partner).
In
the
full
data
scenario,
outcome
data
areis
available
for
both
partners
and
nonpartners.
In
our
situation,
however,
outcome
data
are
is
only
available
for
cases
where
the
binary
variable
equals
one,
making
direct
comparison
impossible
because
there
is
no
variation
in
the
program
participation
measure.
To
get
around
this,
we
use
the
probability
of
being
a
partner
to
measure
the
degree
to
which
a
firm
is
part
of
the
program.
The
probability
of
being
a
WasteWise
partner
will
exhibit
variation
among
partners.
Firms
that
have
a
higher
probability
of
being
a
partner
should
also
perform
more
waste
prevention
and
recycling.
A
detailed
technical
justification
of
this
method
can
be
found
in
the
full
project
report
(
currently
under
peer
review).

The
second
part
of
the
analysis
also
adjusts
for
two
other
problems:
(
1)
some
partners
will
have
A
zero@
for
the
outcome
variable,
but
in
actuality,
probably
are
conducting
waste
reduction
initiatives
but
do
not
report
and
(
2)
there
will
be
some
the
presence
offirm­
specific
effects.
The
occurrence
of
A
zero@
values
for
waste
prevention
and
recycling
amounts
reflects
censoring
of
the
outcome
variable.
That
is,
we
expect
that
actual
waste
prevention
and
recycling
amounts
are
positive
for
those
companies,
but
since
the
companies
did
not
report
a
value,
we
record
their
value
as
A
zero.@
This
type
of
problem
is
not
uncommon
in
statistical
research,
and
a
variety
of
methods
have
been
developed
to
account
for
the
missing
data
in
a
regression
model
framework.
We
use
a
A
Tobit@
model
specification
to
handle
this
issue,
which
.
This
is
discussed
in
more
detail
in
the
full
report.

The
presence
of
A
firm­
specific
effects2@
reflects
the
nature
of
the
data.
Specifically,
we
have
a
number
of
partners
measured
over
time.
This
scenario
is
commonly
referred
to
as
panel
or
longitudinal
data.
In
such
a
framework,
unmeasured
differences
(
i.
e.,
the
firm­
specific
effects)
between
firms
can
lead
to
biased
estimates
for
the
variables
included
in
the
analysis.
A
number
of
methods
have
been
developed
to
account
for
the
presence
of
these
firm­
specific
effects.
We
use
the
random
effects
procedure
to
account
for
these
firm­
specific
effects,
which
.
This
is
discussed
in
more
detail
in
the
full
report.

Our
method
also
allows
us
to
estimate
a
waste
prevention
and
recycling
amount
for
nonreporting
partners.
Thus,
once
the
analysis
has
defined
the
relationship
between
outcomes
and
the
control
factors,
we
can
use
the
estimated
statistical
model
to
fill
in
the
missing
values
for
nonreporting
partners.
We
discuss
this
in
more
detail
in
the
full
report.

Finally,
to
meet
EPA=
s
goal
of
identifying
the
impact
of
WasteWise
on
our
global
climate,
the
method
must
also
provide
material­
specific
estimates
of
the
amount
of
waste
prevention
and
recycling
that
are
attributable
to
the
WasteWise
program.
While
Tthe
statistical
analysis,
however,
provides
a
program
effect
for
categories
of
materials,
it
does
not
provide
a
material­
specific
estimate
needed
for
use
in
EPA=
s
WARM
model3.
There
are
two
separate
questions
issues
that
need
resolution
in
this
area:
(
1)
wWhat
will
be
the
program
effect
for
specific
materials
if
we
only
have
category­
level
estimates
of
the
program
effect?
and
(
2)
hHow
do
we
estimate
the
amount
of
waste
prevention
and
recycling
of
specific
materials
for
non­
reporting
partners
if
we
only
have
category­
level
estimates
of
these
amounts
for
non­
reporting
partners?.
The
first
can
be
resolved
with
a
reasonable
assumption.:
T
that
is,
we
assume
that
the
category­
level
estimates
of
the
program
effect
can
be
applied
to
the
materials
within
each
category.

2
By
A
firm
specific
effects,@
we
mean
factors
that
we
cannot
measure
that
might
affect
the
data.
Examples
of
such
factors
include
corporate
culture
and
how
firms
measure
waste
prevention.

3
For
more
information
on
EPA=
s
WARM
model,
see
<
http://
yosemite.
epa.
gov/
oar/
globalwarming.
nsf/
content/
ActionsWasteWARM.
html>
The
second
issue
requires
some
elaboration
before
describing
a
solution.
The
statistical
models
that
we
estimate
provide
a
category­
level
estimate
of
waste
prevention
and
recycling
for
non­
reporting
partners.
For
example,
the
statistical
model
for
plastics
waste
prevention
will
provide
an
estimate
of
the
amount
of
plastics
waste
prevention
doneachieved
by
non­
reporting
partners.
The
statistical
model
will
not,
however,
provide
material­
specific
estimates
of
waste
prevention
and
recycling.
This
is
not
an
issue
for
reporting
partners
since
they
have
reported
material­
specific
amounts.
To
determine
the
material­
specific
amounts
for
non­
reporting
partners,
we
use
the
distribution
of
materials
in
each
category
from
the
reported
partners,
stratified
by
industry
sector.
We
describe
this
approach
in
more
detail
in
the
full
report.
