Note
From:
Cody
Rice,
OIAA/
EAD/
ASB
Office
of
Environmental
Information
DATE:
September
24,
2003
TO:
Judith
Kendall,
OEI/
OIAA/
TRIPD
RE:
Terms
of
Clearance
for
TRI
ICR
Renewal
This
memo
provides
additional
information
on
revised
burden
hour
estimates
for
Toxics
Release
Inventory
(
TRI)
reporting
as
requested
by
the
Office
of
Management
and
Budget
(
OMB)
in
the
Terms
of
Clearance
dated
March
10,
2003
for
the
Information
Collection
Requests
(
ICRs)
for
the
TRI
Form
R
(
OMB
No.
2070­
0093)
and
Form
A
(
OMB
No.
2070­
0143).

In
the
Terms
of
Clearance,
OMB
stated
that
"
If
EPA
continues
to
believe
that
further
adjustments
[
to
burden
hour
estimates]
are
appropriate,
they
should
provide
additional
documentation
during
the
next
ICR
review
cycle.
This
documentation
should
specifically
address
the
issue
of
whether
the
revised
estimates
account
for
all
categories
of
burden,
including
time
for
data
tracking
and
assembly;
creation,
operation
and
maintenance
of
data
tracking
systems;
training;
and
compliance
determinations."

This
memo
addresses
the
origin
and
derivation
of
TRI
burden
hour
estimates
used
in
previous
ICRs.
This
memo
also
documents
and
further
describes
the
data
that
are
available
to
revise
estimates
of
the
burden
hours
associated
with
TRI
reporting.

Existing
Burden
Hour
Estimates
The
existing
Form
R
burden
hour
estimates
(
i.
e.,
those
in
the
most
recent
ICR
renewal
approved
by
OMB)
trace
their
roots
back
to
the
final
rule
implementing
TRI
reporting
in
the
late
1980s.
These
estimates
were
based
primarily
on
the
knowledge
and
informed
judgement
of
federal
personnel
who
analyzed
the
reporting
form
and
its
associated
requirements.
These
types
of
estimates
are
sometimes
referred
to
as
"
engineering"
estimates,
because
they
reflect
expert
judgement
rather
than
burden
hour
data
from
responding
facilities.
This
method
is
often
employed
to
estimate
reporting
burden
prior
to
actual
reporting.
Much
of
the
TRI
analysis
reflected
engineering
estimates
from
an
analysis
of
a
previous
reporting
rule
known
as
the
Comprehensive
Assessment
Information
Rule.

In
the
Regulatory
Impact
Analysis
of
the
final
rule
implementing
TRI
in
1988,
the
burden
for
rule
familiarization
and
compliance
determination
in
the
first
reporting
year
was
estimated
to
be
34.5
hours
for
facilities
with
50
or
more
employees
and
12
hours
for
facilities
with
less
than
50
employees.
Each
subsequent
year's
compliance
determination
was
assumed
to
require
only
one­
fourth
as
much
time,
as
facility
staff
became
familiar
with
the
form
and
its
data
requirements.
The
burden
of
completing
a
single
Form
R
was
estimated
to
be
33.2
hours
in
the
first
reporting
­
2­
year.
Based
on
additional
expert
judgement,
it
was
assumed
that
subsequent
year
report
completion
burden
would
be
68
percent
of
the
first­
year
time
requirements,
or
22.6
hours.
Recordkeeping
and
mailing
burden
was
assumed
to
be
an
additional
4
hours
per
form
in
both
first
and
subsequent
reporting
years.
The
estimates
for
report
completion
were
validated
with
a
pretest
of
the
proposed
Form
R
among
28
facilities
(
25
large
and
3
small)
in
the
manufacturing
industries
who
were
requested
to
estimate
the
time
required
to
provide
the
requested
information
for
one
chemical
at
their
facility.
The
pretest
average
time
for
completing
the
proposed
Form
R
was
29.7
hours
per
chemical
(
EPA
1988).

In
1993,
EPA
revised
the
burden
hour
estimates
for
TRI
reporting
based
on
several
years
of
reporting
experience
and
new
engineering
estimates
of
the
burden
associated
with
data
elements
added
to
the
From
R
due
to
the
Pollution
Prevention
Act
of
1990.
After
6
years
of
reporting
experience,
EPA
determined
that
facilities
would
tend
to
make
compliance
determinations
by
checking
for
changes
in
reporting
requirements
and
at
their
facilities.
As
a
result,
the
estimate
for
subsequent
year
compliance
determination
was
lowered
to
3
hours
per
facility.
At
this
time,
EPA
stated
that
"
previous
reporting
experience
and
annual
utilization
of
procedures
developed
by
respondents
to
the
program
should
tend
to
keep
the
time
required
for
compliance
determination
to
a
minimum."
For
report
completion,
EPA
adopted
a
revised
estimate
of
47
hours
per
Form
R
in
subsequent
reporting
years
based
on
an
engineering
assessment
for
the
additional
pollution
prevention
data
requirements
(
EPA
1993).

For
reporting
year
1995,
EPA
added
new
chemicals
to
the
list
of
reportable
substances,
and
created
a
certification
statement
(
Form
A)
as
a
burden
reduction
measure.
As
a
result
of
the
expansion
of
the
chemical
list,
EPA
raised
the
estimate
for
subsequent
year
compliance
determination
from
3
hours
to
4
hours.
Adopting
the
assumption
from
the
original
TRI
RIA
that
subsequent
year
compliance
determination
takes
one­
fourth
as
long
as
first
year
compliance
determination,
EPA
back­
calculated
the
revised
first
year
compliance
determination
burden
at
12
hours
(
EPA
1994b).
In
the
Chemical
Expansion
and
Alternate
Threshold
Rules
and
subsequent
ICRs,
EPA
continued
to
use
an
estimate
of
47
hours
per
Form
R
for
report
completion
in
subsequent
reporting
years
and
5
hours
per
Form
R
for
recordkeeping
and
mailing.
Based
on
the
unit
time
estimates
for
the
data
elements
on
the
Form
R
that
are
used
to
determine
eligibility
for
the
Form
A,
the
burden
of
the
Form
A
was
estimated
at
30.2
hours
for
calculations,
3
hours
for
recordkeeping/
mailing,
and
1.4
hours
for
form
completion
(
EPA
1994a).
These
unit
burden
estimates
were
used
to
generate
the
total
burden
hour
estimate
that
OMB
approved
in
the
last
TRI
ICR
renewal
in
March
2003.
­
3­
Respondent
Data
Addressing
TRI
Reporting
Burden
As
described
above,
the
existing
TRI
reporting
burden
estimates
primarily
reflect
a
series
of
engineering
estimates
developed
prior
to
actual
reporting.
Based
on
feedback
from
TRI
reporters,
it
appears
that
burden
hours
are
actually
less
than
previously
estimated.
A
number
of
factors
may
be
contributing
to
the
lower
realized
reporting
burden:

$
Computerization
and
automation
of
data
gathering,
calculations,
report
completion,
recordkeeping,
and
submission.

$
Increased
accessibility
of
information
to
facility
staff
from
EPA
guidance,
trade
associations,
and
the
internet.

$
Implementation
of
other
state
and
federal
reporting
requirements
that
serve
as
precursors
to
TRI
reporting
and
can
be
used
to
fulfill
TRI
reporting
requirements.

$
Previous
burden
hour
estimates
assumed
that
facilities
would
enter
data
in
all
sections
of
the
form,
although
this
is
not
the
case
for
most
Form
Rs.

A
review
of
burden
hour
data
collected
from
reporting
facilities
indicates
that
the
existing
burden
hour
estimates
substantially
overestimate
actual
reporting
burden
for
most
reporting
facilities.
The
existing
burden
estimates
for
subsequent
year
compliance
determination,
Form
R
calculations
and
form
completion,
and
recordkeeping/
mailing
are
above
the
95th
percentile
of
per
form
burden
reported
by
actual
TRI
respondents
(
EPA
2002).

For
the
ICR
renewal,
EPA
developed
a
revised
estimate
of
14.5
hours
for
Form
R
calculations/
report
completion
in
subsequent
reporting
years.
EPA
did
not
change
any
of
the
existing
estimates
for
first
year
reporting
burdens,
including
those
for
calculations/
report
completion.
EPA
also
left
burden
estimates
unchanged
for
subsequent
year
compliance
determination
(
4
hours
per
facility)
and
recordkeeping/
submission
(
5
hours
per
Form
R).
For
the
Form
A,
EPA
also
revised
the
estimate
for
subsequent
year
calculations/
certification
burden
to
9.3
hours
based
on
the
previous
estimate
from
the
Alternate
Threshold
RIA
that
calculations
for
a
Form
A
take
approximately
64
percent
of
the
time
of
calculations
for
the
Form
R.
This
estimate
was
validated
by
contacting
nine
facilities
that
filed
Form
As
in
reporting
year
2000.
The
average
of
facility­
level
burden
hours
per
chemical
certification
was
reported
at
11.2
to
15.5
hours.
EPA's
estimate
of
13.7
total
hours
(
including
3
hours
for
recordkeeping/
submission
and
1.4
hours
for
form
completion)
for
a
facility
certifying
one
chemical
on
a
Form
A
falls
within
this
range.
EPA
did
not
change
any
of
the
existing
estimates
for
first
year
reporting
burdens
associated
with
Form
A,
nor
did
EPA
change
estimates
of
burden
for
subsequent
year
recordkeeping/
submission
(
3
hours)
and
form
completion
(
1.4
hours)
for
Form
A.
EPA's
burden
hour
estimates
are
summarized
in
the
following
table.
Further
details
are
available
in
the
current
ICR
supporting
statements
and
a
background
memo
that
was
prepared
for
the
ICR
renewal
process
(
EPA
2002).
­
4­
TRI
Burden
Hour
Estimates
Activity
Hours
per
year
Comments
Existing
Estimate
Change
Revised
Estimate
First
Year
of
Reporting
Facility
Rule
Familiarization
34.5
0
34.5
No
change
in
baseline
estimates
of
first
year
reporting
burden
due
to
lack
of
data.

Estimates
based
on
expert
judgement,
and
made
prior
to
actual
reporting.

Estimates
date
to
beginning
of
TRI
program.
Compliance
Determination
16
0
16
Form
R
Calculations/
Form
Completion
69
0
69
Recordkeeping/
Submission
5
0
5
Form
A
Calculations/
Certification
44.5
0
44.5
Recordkeeping/
Mailing
3
0
3
Form
Completion
2.1
0
2.1
Subsequent
Years
of
Reporting
Facility
Compliance
Determination
4
0
4
As
above,
no
change.

Form
R
Calculations/
Form
Completion
47.1
­
32.6
14.5
Revised
based
on
data
from
180
reporting
facilities.

Recordkeeping/
Submission
5
0
5
As
above,
no
change.

Form
A
Calculations/
Certification
30.2
­
20.9
9.3
Revised
based
on
assumptions
about
relative
burden
of
Form
R
vs.
A.
Validated
by
contacting
9
respondents.

Recordkeeping/
Submission
3
0
3
As
above,
no
change.
Form
Completion
1.4
0
1.4
Note:
Additional
burden
reduction
of
15%
applied
to
forms
filed
with
TRI­
ME
software
based
on
responses
from
software
users.
­
5­
In
developing
the
revised
estimate
of
subsequent
year
burden
hours
for
Form
R
calculations/
report
completion,
EPA
relied
on
data
from
180
facilities
on
actual
burden
incurred
due
to
TRI
reporting.
These
data
were
available
from
the
following
sources:

°
1994
and
1995
Toxic
Release
Inventory:
Data
Quality
Report
°
1996
Toxic
Release
Inventory:
Data
Quality
Report
°
1999
Research
Triangle
Institute
(
RTI)
Informal
and
Formal
Surveys
of
TRI
Burden
The
specific
interest
expressed
by
OMB
in
the
Terms
of
Clearance
is
the
extent
to
which
the
available
data,
and
by
extension
the
revised
burden
hour
estimates,
account
for
all
categories
of
burden,
including
time
for
data
tracking
and
assembly;
creation,
operation
and
maintenance
of
data
tracking
systems;
training;
and
compliance
determinations.
This
interest
can
be
addressed
by
examining
the
context
of
each
data
collection
and
questions
that
were
asked.

Data
Quality
Reports
The
Data
Quality
Reports
for
reporting
years
1994­
1996
were
part
of
an
EPA
program
of
site
surveys
to
assess
the
quality
of
TRI
data
and
to
identify
areas
where
improved
guidance
would
be
useful
for
improving
the
accuracy
of
future
reported
data.
Facilities
were
selected
to
obtain
a
random
sample
of
facilities
from
key
industries
that
permitted
results
to
be
scaled
up
to
the
entire
industry
group.
The
survey
was
conducted
by
the
engineering
staff
of
an
EPA
contractor.
By
design,
the
identities
of
specific
facilities
were
never
revealed
to
EPA.
The
EPA
contractor
conducted
telephone
interviews
followed
by
site
visits
to
review
the
methodology
and
data
used
by
facilities
to
make
the
threshold
calculations
and
release
and
transfer
estimates
(
EPA
1998a,
1998b).

Most
of
each
survey
focused
on
how
and
where
facilities
obtained
data
on
the
use
and
waste
management
of
TRI
chemicals
in
their
operations,
and
how
they
used
these
data
to
complete
threshold
determinations
and
release
calculations.
As
a
result,
it
is
likely
that
the
respondents
were
particularly
aware
of
all
the
activities
related
to
reporting
that
resulted
in
the
expenditure
of
burden
hours.
Within
the
context
of
this
reporting
audit,
the
burden­
specific
question
was
framed
broadly.
Facilities
were
prompted
to
include
time
for
familiarization
with
the
regulation
and
reporting
requirements,
as
well
as
activities
to
assemble
data,
make
and
review
estimates,
and
document
work.
The
burden­
specific
questions,
which
are
reproduced
in
the
following
box,
asked
for
the
total
time
to
comply
with
the
TRI
reporting
requirements
of
EPCRA
section
313.
­
6­
.
Data
Quality
Reports:
Burden­
specific
questions
(
RY94)
What
is
your
estimate
of
the
time
needed
to
fulfill
the
reporting
requirements
of
Section
313
for
1994?
Please
include
familiarization
with
the
regulation
and
reporting
instructions,
completion
and
internal
review
of
the
reporting
forms,
and
documentation
of
all
information
in
your
reports.

(
RY95)
What
is
your
estimate
of
the
time
needed
to
fulfill
the
reporting
requirements
of
Section
313
for
1995?
Please
include
familiarization
with
the
regulation
and
reporting
instructions,
completion
and
internal
review
of
the
reporting
forms,
and
documentation
of
all
information
in
your
reports.

(
RY96)
What
is
your
estimate
of
the
time
needed
to
fulfill
the
reporting
requirements
of
Section
313
for
1996?
Please
include
familiarization
with
the
regulation
and
reporting
instructions,
completion
and
internal
review
of
the
reporting
forms,
and
documentation
of
all
information
in
your
reports.
(
This
is
the
total
time
for
all
Form
Rs.)

RTI
Surveys
The
RTI
surveys
were
small
scoping
activities
with
the
primary
intent
of
identifying
factors
influencing
variability
in
burden
hours
at
reporting
facilities.
Although
there
was
a
script
of
questions
for
interactions
with
the
facilities,
the
conversations
with
facilities
were
fairly
openended
Prior
to
asking
for
a
burden
hour
estimate,
the
respondents
were
asked
questions
about
the
typical
activities
involved
in
complying
with
reporting
requirements,
how
many
and
what
type
of
staff
were
involved
in
reporting,
what
information
sources
were
available,
and
which
methods
of
estimation
were
used
(
RTI
1999a,
1999b).

Based
on
the
results
of
the
first
(
informal)
survey,
the
burden­
specific
question
for
the
second
(
formal)
survey
was
modified
slightly
to
elicit
additional
information
on
the
specific
activities
comprising
the
burden
hours.
This
included
activities
that
contributed
to
the
facilities'
ability
to
complete
the
reporting
form,
but
which
were
the
result
of
other
regulatory
authorities
or
routine
operating
procedures.
The
questions,
which
are
reproduced
in
the
box
below,
asked
for
the
average
time
to
complete
a
single
Form
R
(
or
the
time
to
complete
all
the
Form
Rs
at
a
facility):
­
7­
RTI
Surveys:
Burden­
specific
questions
(
Informal)
How
long
does
it
take
on
average
to
fill
one
Form
R?
(
If
the
time
per
form
can
not
be
estimated,
then
how
long
does
it
take
to
do
all
the
forms?)

(
Formal)
On
average,
how
long
does
it
take
to
complete
one
Form
R?
(
If
the
time
per
form
can
not
be
estimated,
then
how
long
does
it
take
to
fill
all
the
forms)?

a)
Please
list
the
activities
that
you
included
in
deriving
this
estimate?

b)
What
percentage
of
this
estimate
is
related
to
other
ongoing
activities
(
e.
g.,
collecting
data
required
for
compliance
with
NPDES
permits
etc)?

Additional
Data
and
Analysis
During
the
public
comment
period
for
the
last
TRI
ICR
renewal,
the
American
Petroleum
Institute
(
API)
submitted
the
results
of
a
burden
study
covering
TRI
reporting
year
2001
activities
for
99
facilities
in
the
petroleum
refining
and
petroleum
terminal
and
bulk
station
industries
(
API
2002).
API
subsequently
provided
EPA
with
the
survey
form
and
the
data
for
individual
facilities.
The
API
survey
questions
addressed
facility­
and
form­
specific
burden
categories
separately.
The
burden­
specific
questions
from
the
API
data
collection
are
reproduced
in
the
following
box:

American
Petroleum
Institute
Data
Collection:
Burden­
specific
questions
Number
of
hours
spent
on
rule
familiarization,
including
reviewing
FR
notices,
instructions,
EPA
guidance,
and
so
forth.

Number
of
hours
spent
making
compliance
determinations,
including
determining
whether
reporting
thresholds
are
met.

Total
number
of
hours
spent
per
Form
R
(
Include
release
calculations,
completing
form,
mailing
and
recordkeeping,
etc.
Do
not
include
rule
familiarization
and
compliance
determination.)

Although
EPA
reviewed
the
API
data,
these
data
were
not
used
in
the
revised
burden
hour
estimate
for
the
March
2002
ICR
renewal
because
of
concerns
about
overweighting
observations
from
the
petroleum
refining
and
petroleum
terminal
and
bulk
station
industries.
API's
results
are
also
confounded
somewhat
by
the
first
year
of
reporting
on
lead
and
lead
compounds
at
lower
thresholds,
with
associated
higher
first­
year
reporting
burdens.
­
8­
Nevertheless,
the
API
results
for
total
reporting
burden
were
below
or
near
the
EPA
revised
estimates
when
similar
numbers
of
reports
were
assumed.

EPA
subsequently
conducted
a
statistical
analysis
using
the
API
data,
along
with
the
original
data
from
the
Data
Quality
Reports
and
the
RTI
surveys
(
see
Appendix
A).
The
analysis
used
a
prediction­
based
approach
to
estimate
the
burden
associated
with
TRI
reporting.
First,
linear
regressions
for
each
industry
were
estimated
using
the
available
data.
Then,
the
parameter
estimates
were
applied
to
the
census
data
(
i.
e.,
the
data
on
forms
per
facility
for
each
industry)
to
derive
estimates
of
total
reporting
burden,
the
average
time
per
Form
R,
the
standard
errors
and
confidence
intervals.
Overall,
the
average
reporting
burden
per
Form
R
was
found
to
be
11.7
hours
plus
or
minus
1.8
hours
(
Abt
2003).
This
is
actually
lower
than
the
revised
estimate
used
in
the
ICR
renewal
of
19.5
hours
per
Form
R
(
composed
of
14.5
for
calculation/
report
completion
and
5
hours
for
recordkeeping/
submission)
plus
4
hours
per
facility
for
compliance
determination.
It
may
be
appropriate
to
think
of
the
difference
in
estimates
of
total
hours
per
form
as
potentially
addressing
various
industry
comments
about
additional
time
spent
on
training,
guidance
review,
and
other
activities
that
are
not
individually
estimated
as
part
of
EPA's
revised
burden
hour
estimate
for
subsequent
year
reporting.

Conclusion
Although
EPA's
revised
estimate
of
19.5
hours
per
Form
R
plus
4
hours
per
facility
for
compliance
determination
is
more
of
an
aggregated
estimate
than
an
estimate
that
is
built­
up
from
numerous
discrete
activities,
there
is
little
danger
that
the
total
reporting
burden
is
underestimated.
EPA's
revised
estimate
is
based
on
responses
that
reveal
actual
facility
burden
hours,
and
it
is
substantially
higher
than
the
average
of
these
responses.

Furthermore,
it
should
not
be
assumed
that
the
only
direction
of
possible
bias
in
responses
is
downward
because
of
incomplete
specification
of
compliance
activities.
Based
on
previous
experience
interviewing
facilities
about
compliance
burden,
facilities
sometimes
include
burden
that
is
incurred
in
complying
with
other
regulations
or
in
standard
operating
practices
if
data
from
those
activities
are
ultimately
used
for
TRI
reporting.
Although
it
is
appropriate
to
attribute
time
spent
arranging
data
and
making
estimates
to
complete
Form
R
and
A
data
elements
to
TRI
compliance,
it
is
not
appropriate
to
attribute
time
spent
complying
with
other
regulations
to
TRI.
Also,
there
is
some
possibility
of
strategic
bias
in
industry
responses
to
survey
questions
about
the
reporting
burden
of
TRI
reporting
if
the
respondents
believe
that
the
responses
may
have
some
bearing
on
reporting
requirements.

Although
the
burden­
specific
questions
varied
somewhat
from
data
source
to
data
source,
facilities
were
encouraged
in
all
cases
to
think
comprehensively
about
the
overall
burden
of
TRI
reporting.
It
seems
reasonable
to
conclude
that
the
available
burden
data
are
appropriate
and
adequate
for
the
purpose
of
revising
unit
reporting
burden
estimates,
especially
in
light
of
the
validation
provided
by
more
recent
burden
data
independently
gathered
by
API.
The
sampled
facilities
reflect
the
experience
of
a
broad
range
of
industries
reporting
to
TRI,
and
the
data
­
9­
consistently
show
that
existing
estimates
of
reporting
burden
used
by
EPA
are
not
an
accurate
representation
of
current
reporting
burden.
The
revised
burden
estimates
in
the
latest
ICR
renewal
represent
an
improvement
over
the
previous
estimates
from
the
1980s,
and
their
use
would
provide
a
more
accurate
representation
of
the
burden
of
the
TRI
program
on
reporting
facilities.
The
revised
estimates
would
also
provide
a
more
accurate
baseline
for
evaluation
of
potential
investments
in
future
TRI
burden
reduction
initiatives.
­
10­
REFERENCES
Abt
Associates.
.
Memo
from
Bill
Rhodes
and
Susan
Day
to
Cody
Rice
(
USEPA/
OEI)
re:
Estimation
of
TRI
Reporting
Burden,
September
23,
2003.

American
Petroleum
Institute.
Comments
on
Information
Collection
Request
(
ICR)
for
Toxic
Chemical
Release
Reporting.
EPA
ICR
No.
1363.12,
OMB
No.
2070­
0093,
September
6,
2002.

Research
Triangle
Institute/
Center
for
Economics
Research,
Memo
from
Smita
Brunnermeier,
et
al
to
Joe
Callahan
(
USEPA/
OPPT),
Subject:
Informal
Survey
Results,
May
7,
1999a.

Research
Triangle
Institute/
Center
for
Economics
Research,
Memo
from
Smita
Brunnermeier,
et
al
to
Joe
Callahan
(
USEPA/
OPPT),
Subject:
Formal
Survey
Results,
June
15,
1999b.

U.
S.
Environmental
Protection
Agency.
Estimates
of
Burden
Hours
for
Economic
Analyses
of
the
Toxics
Release
Inventory
Program.
June
10,
2002.

U.
S.
Environmental
Protection
Agency.
Regulatory
Impact
Analysis
in
Support
of
Final
Rulemaking
under
Section
313
of
Title
III
of
the
Superfund
Amendments
and
Reauthorization
Act
of
1996.
Prepared
for
Office
of
Toxic
Substances
by
ICF.
Inc.
EPA
Contract
No.
68­
02­
4240.
Task
Order
No.
3­
3.
February
1988.

U.
S.
Environmental
Protection
Agency.
Regulatory
Impact
Analysis
of
the
EPCRA
Section
313
Alternate
Threshold
Final
Rule,
November
18,
1994a.

U.
S.
Environmental
Protection
Agency.
Regulatory
Impact
Analysis
of
the
Final
Rule
to
Add
Various
Chemicals
and
Chemical
Categories
to
the
EPCRA
Section
313
List
of
Toxic
Chemicals,
November
18,
1994b.

U.
S.
Environmental
Protection
Agency.
Review
and
Update
of
Burden
and
Cost
Estimates
for
EPA's
Toxics
Release
Inventory
Program,
August
1993.

U.
S.
Environmental
Protection
Agency.
1994
and
1995
Toxic
Release
Inventory:
Data
Quality
Report,
EPA
745­
R­
98­
002,
March
1998a.

U.
S.
Environmental
Protection
Agency.
1996
Toxic
Release
Inventory:
Data
Quality
Report,
EPA
745­
R­
98­
016,
December
1998b.
APPENDIX
A:

Memo
on
Estimation
of
TRI
Reporting
Burden
1
MEMORANDUM
TO:
Cody
Rice,
US
EPA
FROM:
Bill
Rhodes,
Susan
Day
DATE:
September
23,
2003
RE:
Estimation
of
TRI
Reporting
Burden
The
intent
of
this
analysis
is
to
estimate
the
average
reporting
burden
associated
with
filing
Form
Rs
and
to
provide
a
confidence
interval
for
this
burden.
Two
estimates
are
generated:
average
burden
per
facility
and
average
burden
per
Form
R.

The
Data
Data
reflecting
the
time
spent
filing
Form
Rs
comes
from
three
sources.
They
are
included
in
Appendix
A.
Each
source
reports:

°
The
total
time
required
to
complete
all
forms
filed
by
a
facility.
Time
is
either
reported
as
a
range
with
a
lower
and
upper
limit
or
as
a
point
estimate.
°
The
Standard
Industrial
Classification
(
SIC)
code.
Two
digit
SIC
codes
were
used
in
this
analysis
with
the
exception
of
SIC
5171.
°
The
number
of
Form
Rs
completed.
This
estimate
was
occasionally
reported
as
a
range
with
a
lower
and
upper
limit.
When
reported
as
a
range,
the
range
was
converted
to
a
point
estimate
equal
to
the
midpoint
of
that
range.

The
first
data
set
comes
from
the
1994,
1995,
and
1996
Data
Quality
Reports.
It
was
a
random
sample
of
facilities
in
key
industries,
defined
as
those
that
are
large
contributors
to
total
releases
(
162
observations
­
one
observation
was
excluded
as
an
outlier).
The
second
data
set
was
the
American
Petroleum
Institute
(
API)
survey,
which
was
limited
to
API
members
(
99
observations
in
SIC
codes
29
and
5171).
The
third
data
set
was
a
small
survey
of
facilities
(
18
observations
­
one
observation
was
excluded
because
there
was
only
one
facility
in
SIC
code
34)
by
RTI;
these
facilities
were
selected
to
explore
the
relationship
between
industry,
number
of
reports
and
other
factors
potentially
related
to
reporting
burden.
These
three
data
sets,
after
exclusions,
are
2
collectively
referred
to
as
the
calibration
data.
They
are
presented
in
Appendix
A.

The
TRI
data
for
RY2001
were
also
used
in
the
analysis.
These
data
represent
a
census
of
TRI
facilities
in
RY2001.
These
data
are
referred
to
as
the
prediction
data.
The
prediction
data
provided
information
on:

°
SIC
code
°
Number
of
forms
completed
The
calibration
data
came
from
a
period
that
predated
the
prediction
data.
It
is
assumed
that
the
time
spent
per
form
has
not
changed
materially
between
the
time
when
the
calibration
data
were
collected
and
the
time
when
the
prediction
data
were
collected.
It
is
also
assumed
that
the
calibration
data
reflects
subsequent
year
filings.

Estimation
Estimation
uses
a
prediction­
based
approach
described
by
Valliant,
Dorfman
and
Royall
(
2000).
The
authors
argue
that,
when
a
sampling
procedure
is
biased,
a
prediction­
based
approach
is
required
to
overcome
the
bias.
Moreover,
even
when
a
sampling
procedure
is
unbiased,
a
prediction­
based
approach
can
provide
estimates
with
lower
mean­
squared
error
than
are
provided
by
traditional
survey
estimators.
An
overview
of
the
prediction­
based
approach
follows:

It
is
assumed
that
the
relationship
between
burden
and
the
number
of
Form
Rs
filed
can
be
represented
as
a
linear
function:

B
Fe
ij
i
i
ij
ij
=
+
+
 
 
0
1
[
1]
where:

B
ij
The
burden
for
the
jth
facility
in
the
ith
industry.
F
ij
The
number
of
Form
Rs
completed
by
the
jth
facility
in
the
ith
industry.

The
represents
fixed
parameters
whose
values
may
vary
across
the
industries.
The
e
are
 
random
errors
terms.
They
are
normally
and
independently
distributed
but
groupwise
heteroscedastic
where
the
industry
defines
the
group.
In
practice,
model
[
1]
is
estimated
by
estimating
separate
regressions
that
are
specific
to
each
industry.

As
noted
earlier,
the
burden
estimates
were
frequently
reported
as
ranges,
so
model
[
1]
could
not
be
estimated
directly.
Instead,
estimates
were
based
on
a
likelihood
function.
When
burden
was
reported
as
a
range,
the
likelihood
was
written:
3
L
Z
Z
ij
ij
H
ij
L
=
 
 
 
(
)
(
)

where:

 (...)
is
the
standard
normal
central
distribution
function.

Z
B
F
ij
H
ij
U
i
i
ij
i
=
 
 
 
 
 
0
1
Z
B
F
ij
L
ij
L
i
i
ij
i
=
 
 
 
 
 
0
1
and
the
upper
limit
reported
by
the
jth
respondent
in
the
ith
industry.
 
ij
U
the
lower
limit
reported
by
the
jth
respondent
in
the
ith
industry.
 
ij
L
However,
when
the
burden
was
reported
as
a
point
estimate,
the
likelihood
was
written
as
the
density
function:

L
e
ij
i
B
F
ij
i
i
ij
i
=
 
 
 






1
2
2
1
2
0
1
2
2
  
 
 
 
Parameters
were
estimated
using
maximum
likelihood.
The
estimated
parameter
covariance
matrix
is
denoted
V.

To
derive
the
average
reporting
burden
for
an
industry,
the
above
regressions
were
estimated
using
the
reporting
burden
data
(
calibration
sample).
The
parameter
estimates
were
then
applied
to
predict
the
reporting
burden
for
the
census
data
(
prediction
data).
Thus
the
average
reporting
burden
for
a
member
of
industry
i
was
estimated
as:




B
F
J
i
i
j
J
i
ij
i
i
=
+
=

 
 
0
1
1
[
2]
where
is
the
number
of
facilities
that
are
in
the
prediction
data
for
industry
i.
The
sampling
J
i
4
variance
for
the
mean
burden
in
industry
i
was
estimated
as:

 
 
 
 
B
i
i
i
i
i
i
i
i
XV
X
J
2
2
=
 
 
+


[
3]

where:

X
i
is
a
matrix
with
ones
in
the
first
column
and
in
the
second
F
j
J
ij
i
(
)
=
1

column.
is
a
conformable
column
vector
with
every
element
equal
to
.
 
i
1
J
i
For
a
derivation,
see
Valliant,
Dorfman
and
Royall
(
2000),
page
29.

Another
way
to
express
the
estimation
is:

[
]
(
)
(
)
(
)
(
)
 
=

=








 
 
  
 
 
 
X
F
V
VAR
COV
COV
VAR
i
i
i
i
i
i
i
i
1
0
01
0
1
1
,

,

so
the
first
part
of
equation
[
3]
can
be
rewritten
as:

(
)
(
)
 
=
+
+
 
 
 
 
 
 
XVX
VAR
VAR
F
COV
F
i
i
i
i
i
i
i
0
1
2
01
1
2
(
)
,

Thus
the
first
component
of
depends
on
the
precisions
with
which
the
parameters
can
be
 
Bi
2
 
estimated.
Precision
will
increase
with
the
size
of
the
calibration
sample.
It
is
not
affected
by
the
size
of
the
prediction
sample.

Another
way
to
consider
the
first
component
of
is
that
it
is
the
sampling
variance
for
the
 
Bi
2
expected
value
of
conditional
on
.
The
actual
mean
will
vary
from
this
conditional
mean
B
i
F
i
because
of
the
randomness
in
reporting
burden
from
facility
to
facility
given
a
constant
.
The
F
ij
difference
will
be
smaller
as
the
size
of
the
census
data
gets
larger
 
hence
the
second
component
of
the
variance
term.
Note
that
the
size
of
this
second
term
is
not
affected
by
the
size
of
the
calibration
sample.
5
Results
Results
from
the
regression
analysis
are
summarized
in
appendix
B
table
B­
1.
For
each
industry
classification,
the
table
reports
regression
parameter
estimates
and
t­
scores.
The
t­
scores
are
the
parameter
estimates
divided
by
the
estimated
standard
errors.

Estimation
of
model
[
1]
was
attempted
for
ten
SIC
categories
where
data
were
available.
In
three
cases,
however,
that
was
not
possible.
First,
for
SIC
5171,
either
a
corporate
parent
or
API
apportioned
total
reporting
burden
across
several
facilities.
As
a
result,
most
facilities
in
the
data
seemed
to
have
the
same
reporting
burden
and
the
same
number
of
Form
Rs.
There
was
so
little
variation
in
the
number
of
Form
Rs,
in
fact,
that
the
regression
indicated
that
was

 
1i
negative,
an
implausible
result.
Consequently,
for
SIC
5171,
was
forced
to
equal
zero.
 
0
i
Second,
for
SIC
29,
the
calibration
sample
did
not
represent
the
prediction
sample.
For
reasons
discussed
subsequently,
the
regression
provided
more
plausible
predictions
when
the
constant
was
constrained
to
zero.
Finally,
for
SIC
35,
there
were
only
two
observations.
Estimates
for
the
prediction
sample
were
based
on
the
mean
for
these
two
observations.

Residual
plots
confirmed
that
the
regressions
provided
a
reasonable
specification
of
the
relationship
between
reporting
burden
and
Form
Rs
filed.
Furthermore,
with
the
exception
of
SIC
5171,
the
sample
appeared
to
be
balanced,
so
that
model
misspecification
should
have
little
effect
on
the
resulting
estimates.
On
this
point,
see
Valliant,
Dorfman
and
Royall
(
2000),
page:
49­
61.

Table
1
summarizes
the
data
used
in
the
model.
The
second
and
third
columns
present
the
sample
sizes
for
both
the
calibration
sample
and
the
prediction
census.
The
prediction
census
shows
the
number
of
facilities
that
filed
at
least
one
Form
R.
The
last
two
columns
show
the
average
number
of
Form
Rs
in
the
calibration
sample
and
the
prediction
census.

Table
1:
Sample
Size
and
Average
Number
of
Form
Rs
SIC
Code
Sample
Size
Average
Form
Rs
Calibration
Prediction
Calibration
Prediction
5171
75
499
7.6
7.0
28
52
3,330
6.2
5.1
33
29
1,946
3.2
3.6
29
26
511
26.0
8.1
30
24
1,828
3.3
2.2
25
23
326
2.5
2.0
37
20
1,325
3.7
3.4
36
14
1,193
2.9
2.6
26
12
525
6.3
6.2
35
2
1,086
1.5
2.5
Total
277
12,569
7.2
3.9
6
With
one
exception,
the
sample
is
balanced,
meaning
that
the
mean
for
the
sample
roughly
equals
the
mean
for
the
census.
The
exception
is
SIC
29,
whose
data
came
from
the
API
Survey.
(
SIC
5171,
which
came
from
the
same
survey,
did
not
suffer
from
the
same
problem.)
In
fact,
while
most
of
the
observations
from
the
calibration
sample
for
SIC
29
came
from
facilities
that
filed
a
large
number
of
Form
Rs,
most
of
the
observations
for
the
census
came
from
facilities
that
filed
fewer
Form
Rs.

Table
2
presents
average
reporting
burden
by
SIC
code.
The
observed
average
reporting
burden
by
SIC
code
is
shown
in
column
two.
This
observed
average
was
computed
using
the
midpoint
of
the
range
when
the
reporting
burden
was
reported
as
a
range.
Column
three
presents
the
estimated
average
reporting
burden
by
SIC
code.
The
standard
error
for
that
estimated
average
reporting
burden
is
shown
in
column
four.
The
weighted
average
reporting
burden
across
all
ten
SIC
codes
is
presented
in
the
last
row
of
the
table.
Weights
are
based
on
the
number
of
facilities
within
each
SIC
classification.

Table
2:
Average
Reporting
Burden
by
SIC
Code
SIC
Calibration
Sample
Prediction
Census
Observed
Average
Time
per
Facility
Estimated
Average
Time
per
Facility
Standard
Error
5171
48.5
43.40
2.30
28
69.4
58.53
13.12
33
35.0
37.36
5.17
29
598.9
200.00
27.10
30
29.2
25.57
3.70
25
21.7
18.63
1.86
37
61.1
57.30
14.80
36
32.4
25.82
4.82
26
43.0
41.71
4.88
35
33.0
33.00
2.12
Total
47.60
4.09
As
shown
in
Table
2,
on
average,
a
facility
spends
nearly
48
hours
filling
out
all
of
its
Form
Rs.
An
approximate
95
percent
confidence
interval,
based
on
two
standard
deviations,
is
about
44
to
52
hours
per
facility.
This
estimate
represents
a
weighted
average
where
burden
is
weighted
by
the
number
of
facilities
within
an
SIC
code.
With
the
exception
of
SIC
29
and
5171,
the
average
for
the
prediction
sample
is
within
two
standard
deviations
of
the
average
for
the
calibration
sample.
This
makes
sense
because
the
samples
are
reasonably
balanced.

For
SIC
29,
the
average
for
the
prediction
sample
is
only
about
one­
third
the
size
of
the
average
for
the
calibration
sample.
Furthermore,
the
predicted
average
for
SIC
29
is
much
higher
than
the
predicted
average
for
all
other
industries.
It
is
unclear
whether
this
reflects
reality
or
just
the
7
uncertainty
about
estimating
the
average
for
SIC
29.

Note
that
the
standard
error
for
SIC
5171
is
probably
biased
downward
because
of
the
way
reporting
burden
estimates
were
collected.
Specifically,
total
reporting
burden
at
the
corporate
level
was
apportioned
to
facilities
by
the
corporate
parent
or
by
API
for
several
facilities.
Thus,
in
the
data
set,
many
facilities
have
the
same
apparent
reporting
burden
and
the
same
apparent
number
of
Form
Rs.
One
implication
is
that
while
the
predicted
mean
for
industry
5171
appears
to
differ
significantly
from
the
calibration
mean,
this
is
probably
because
the
sampling
variance
is
underestimated.

Table
3
presents
the
estimated
average
reporting
burden
per
Form
R,
computed
by
dividing
the
average
reporting
burden
for
the
industry
(
from
Table
2)
by
the
average
number
of
forms
per
SIC.
The
confidence
interval
shown
is
based
on
two
standard
deviations
(
approximately
a
95
percent
confidence
interval).
In
the
last
row,
the
weighted
average
of
the
reporting
burden
across
all
SIC
codes
is
presented.
As
before,
the
weights
are
based
on
the
number
of
facilities
within
each
SIC.

Table
3:
Average
Reporting
Burden
Per
Form
R
SIC
Prediction
Census
Average
Time
per
Form
R
Standard
Error
95
Percent
Confidence
Interval
5171
6.24
0.33
5.58
6.90
28
11.43
2.56
6.31
16.55
33
10.24
1.42
7.40
13.07
29
24.62
3.34
17.95
31.30
30
11.72
1.70
8.32
15.12
25
9.11
0.91
7.29
10.93
37
16.74
4.32
8.10
25.38
36
10.02
1.87
6.28
13.76
26
6.77
0.79
5.19
8.35
35
13.40
0.86
11.68
15.12
Total
11.65
0.89
9.87
13.43
Overall,
it
appears
that
the
average
reporting
burden
is
about
11.7
hours
per
Form
R.
If
SIC
29
is
excluded
from
these
calculations,
the
overall
average
would
fall
to
about
11.1
and
the
standard
error
would
not
change
by
much.
Whatever
the
error
when
estimating
the
reporting
burden
for
SIC
29,
there
is
not
much
affect
on
the
overall
reporting
burden
per
Form
R.
8
Additional
Comments
These
estimates
are
limited
to
the
10
SIC
codes
for
which
there
was
data.
To
extend
these
estimates
to
other
SIC
codes
would
require
an
assumption
that
the
ten
industries
observed
in
the
calibration
sample
comprise
a
random
sample
of
all
industries
that
report
to
TRI.
Were
that
the
case,
then
random
error
models
would
be
appropriate,
and
it
would
be
possible
to
extend
the
estimates
to
industries
not
included
in
the
sample.
For
an
explanation,
see
McCulloch
and
Searle
(
2001).
However,
making
such
an
assumption
may
not
be
appropriate
given
that
selection
of
the
above
sample
was
purposeful.
This
extension
was
not
pursued.

Note
also
that
the
covariance
matrix
from
maximum
likelihood
estimation
has
an
asymptotic
justification.
A
sample
of
two
(
SIC
=
35)
probably
does
not
comprise
a
sample
of
sufficient
size
that
an
asymptotic
justification
would
apply.
The
same
might
be
said
of
other
samples.
Thus,
estimated
variances
should
be
treated
as
approximations.
Nevertheless,
as
shown
in
Appendix
B,
an
OLS
model
provides
variance
estimates
that
are
close
to
those
for
the
maximum
likelihood
estimates.
This
is
encouraging
because
the
OLS
estimates
are
unbiased
provided
the
error
term
is
normal
and
identically
distributed.

References
McCulloch,
C.
and
Searle,
S.
Generalized,
Linear,
and
Mixed
Models.
John
Wiley
&
Sons,
2001.

Valliant,
R.,
Dorfman,
A.
and
Royall,
R.
Finite
Population
Sampling
and
Inference.
John
Wiley
&
Sons,
2000.
9
Appendix
A
Calibration
Data
Set
SIC
FORM_
MINFORM_
MAX
HOUR_
MIN
HOUR_
MAXSOURCE
25
5
5
21
50DQ94
25
2
2
0
20DQ94
25
3
3
21
50DQ94
25
4
4
21
50DQ94
25
5
5
21
50DQ94
25
2
2
0
20DQ94
25
1
1
0
20DQ94
25
1
1
50
100DQ94
25
9
9
50
100DQ94
25
3
3
21
50DQ94
25
2
2
0
20DQ94
25
1
1
21
50DQ94
25
7
7
0
20DQ94
25
1
1
21
50DQ94
25
1
1
21
50DQ94
25
3
3
0
20DQ94
25
1
1
21
50DQ94
25
5
5
21
50DQ94
25
6
6
0
20DQ94
25
2
2
21
50DQ94
25
1
1
0
20DQ94
25
8
8
21
50DQ94
25
5
5
21
50DQ94
25
1
1
0
20DQ94
26
1
1
0
8DQ95
26
8
8
21
40DQ95
26
3
3
41
100DQ95
26
3
3
21
40DQ95
26
4
4
21
40DQ95
26
2
2
0
8DQ95
26
5
5
9
20DQ95
26
7
7
21
40DQ95
26
6
6
41
100DQ95
26
14
14
0
100DQ95
26
9
11
72
88RTI­
1
26
13
13
100
100RTI­
2
28
1
1
21
50DQ94
28
3
3
0
20DQ94
28
7
7
0
20DQ94
28
5
5
21
50DQ94
28
1
1
0
20DQ94
28
3
3
0
20DQ94
SIC
FORM_
MINFORM_
MAX
HOUR_
MIN
HOUR_
MAXSOURCE
10
28
9
9
50
100DQ94
28
4
4
0
20DQ94
28
4
4
21
50DQ94
28
9
9
50
100DQ94
28
5
5
0
20DQ94
28
7
7
100
200DQ94
28
14
14
50
100DQ94
28
2
2
21
50DQ94
28
4
4
0
20DQ94
28
5
5
0
20DQ94
28
3
3
0
20DQ94
28
8
8
21
50DQ94
28
2
2
0
20DQ94
28
2
2
21
50DQ94
28
5
5
21
50DQ94
28
4
4
21
50DQ94
28
8
8
100
200DQ94
28
1
1
0
20DQ94
28
9
9
0
20DQ94
28
1
1
0
20DQ94
28
6
6
21
50DQ94
28
2
2
0
20DQ94
28
6
6
0
20DQ94
28
1
1
21
50DQ94
28
4
4
21
50DQ94
28
2
2
0
20DQ94
28
4
4
0
20DQ94
28
1
1
0
20DQ94
28
3
3
0
20DQ94
28
2
2
0
20DQ94
28
5
5
41
100DQ95
28
11
11
41
100DQ95
28
19
19
0
100DQ95
28
5
5
41
100DQ95
28
15
15
41
100DQ95
28
1
1
0
8DQ95
28
2
2
21
40DQ95
28
2
2
0
8DQ95
28
13
13
41
100DQ95
28
13
13
41
100DQ95
28
29
29
400
400RTI­
1
28
9
11
160
160RTI­
1
28
4
5
320
480RTI­
1
28
5
5
380
380RTI­
2
28
10
10
480
480RTI­
2
28
20
20
160
160RTI­
2
29
26
26
248
424RTI­
1
29
23
23
640
640RTI­
2
SIC
FORM_
MINFORM_
MAX
HOUR_
MIN
HOUR_
MAXSOURCE
11
29
36
36
1143
1143API
29
25
25
304
304API
29
21
21
580
580API
29
39
39
1152
1152API
29
18
18
188
188API
29
18
18
654
654API
29
26
26
488
488API
29
12
12
173
173API
29
43
43
1057
1057API
29
30
30
588
588API
29
24
24
948
948API
29
18
18
188
188API
29
18
18
188
188API
29
12
12
173
173API
29
31
31
143
143API
29
35
35
1695
1695API
29
37
37
319
319API
29
12
12
173
173API
29
45
45
1333
1333API
29
25
25
174
174API
29
25
25
620
620API
29
36
36
1884
1884API
29
24
24
242
242API
29
18
18
188
188API
30
1
1
0
20DQ94
30
1
1
0
20DQ94
30
2
2
0
20DQ94
30
3
3
21
50DQ94
30
6
6
21
50DQ94
30
3
3
0
20DQ94
30
1
1
0
20DQ94
30
1
1
0
20DQ94
30
15
15
50
100DQ94
30
1
1
0
20DQ94
30
2
2
0
20DQ94
30
2
2
21
50DQ94
30
1
1
0
20DQ94
30
3
3
0
20DQ94
30
3
3
21
50DQ94
30
2
2
0
20DQ94
30
1
1
21
50DQ94
30
2
2
21
50DQ94
30
1
1
0
20DQ94
30
2
2
0
20DQ94
30
1
1
21
50DQ94
30
3
3
21
50DQ94
30
1
1
0
20DQ94
SIC
FORM_
MINFORM_
MAX
HOUR_
MIN
HOUR_
MAXSOURCE
12
33
1
1
41
100DQ96
33
4
4
21
40DQ96
33
5
5
9
20DQ96
33
1
1
0
8DQ96
33
9
9
21
40DQ96
33
4
4
41
100DQ96
33
2
2
9
20DQ96
33
5
5
120
120DQ96
33
2
2
21
40DQ96
33
1
1
0
8DQ96
33
4
4
21
40DQ96
33
1
1
41
100DQ96
33
1
1
9
20DQ96
33
3
3
21
40DQ96
33
10
10
41
100DQ96
33
4
4
9
20DQ96
33
1
1
21
40DQ96
33
1
1
0
8DQ96
33
1
1
0
8DQ96
33
1
1
0
8DQ96
33
3
3
0
8DQ96
33
1
1
0
8DQ96
33
2
2
9
20DQ96
33
2
2
21
40DQ96
33
1
1
41
100DQ96
33
3
3
21
40DQ96
33
2
2
6
6RTI­
1
33
4
4
32
32RTI­
1
33
13
13
160
160RTI­
2
34
1
1
8
8RTI­
2
35
2
2
30
30RTI­
1
35
1
1
36
36RTI­
2
36
1
1
0
8DQ96
36
7
7
21
40DQ96
36
1
1
9
20DQ96
36
6
6
41
100DQ96
36
2
2
9
20DQ96
36
4
4
9
20DQ96
36
1
1
0
8DQ96
36
1
1
21
40DQ96
36
4
4
41
100DQ96
36
5
5
9
20DQ96
36
4
4
21
40DQ96
36
1
1
9
20DQ96
36
3
3
41
100DQ96
36
1
1
41
100DQ96
37
2
2
0
8DQ96
SIC
FORM_
MINFORM_
MAX
HOUR_
MIN
HOUR_
MAXSOURCE
13
37
2
2
0
8DQ96
37
2
2
0
8DQ96
37
4
4
21
40DQ96
37
4
4
9
20DQ96
37
3
3
41
100DQ96
37
2
2
41
100DQ96
37
1
1
0
8DQ96
37
3
3
41
100DQ96
37
3
3
160
160DQ96
37
1
1
0
8DQ96
37
11
11
41
100DQ96
37
14
14
200
200DQ96
37
1
1
9
20DQ96
37
4
4
41
100DQ96
37
2
2
21
40DQ96
37
3
3
9
20DQ96
37
1
1
9
20DQ96
37
6
6
320
320RTI­
1
37
4
6
40
60RTI­
2
5171
8
8
40
40API
5171
7
7
66
66API
5171
8
8
40
40API
5171
8
8
40
40API
5171
8
8
40
40API
5171
7
7
66
66API
5171
8
8
40
40API
5171
8
8
40
40API
5171
7
7
66
66API
5171
7
7
66
66API
5171
7
7
66
66API
5171
8
8
25
25API
5171
8
8
25
25API
5171
7
7
66
66API
5171
8
8
25
25API
5171
8
8
25
25API
5171
7
7
66
66API
5171
8
8
25
25API
5171
7
7
66
66API
5171
7
7
66
66API
5171
7
7
66
66API
5171
8
8
60
60API
5171
7
7
66
66API
5171
7
7
66
66API
5171
8
8
40
40API
5171
8
8
60
60API
5171
7
7
66
66API
5171
8
8
60
60API
SIC
FORM_
MINFORM_
MAX
HOUR_
MIN
HOUR_
MAXSOURCE
14
5171
7
7
66
66API
5171
8
8
40
40API
5171
8
8
40
40API
5171
7
7
66
66API
5171
8
8
40
40API
5171
7
7
17
17API
5171
7
7
66
66API
5171
7
7
17
17API
5171
9
9
22
22API
5171
9
9
22
22API
5171
7
7
66
66API
5171
9
9
22
22API
5171
8
8
59
59API
5171
7
7
66
66API
5171
8
8
59
59API
5171
8
8
59
59API
5171
7
7
66
66API
5171
8
8
59
59API
5171
8
8
59
59API
5171
7
7
66
66API
5171
9
9
22
22API
5171
8
8
46
46API
5171
7
7
66
66API
5171
8
8
46
46API
5171
7
7
66
66API
5171
7
7
17
17API
5171
7
7
17
17API
5171
7
7
66
66API
5171
7
7
17
17API
5171
8
8
46
46API
5171
7
7
66
66API
5171
8
8
60
60API
5171
7
7
66
66API
5171
8
8
40
40API
5171
8
8
25
25API
5171
7
7
66
66API
5171
8
8
25
25API
5171
8
8
40
40API
5171
7
7
66
66API
5171
8
8
40
40API
5171
8
8
25
25API
5171
8
8
40
40API
5171
7
7
66
66API
5171
8
8
40
40API
5171
8
8
40
40API
5171
7
7
66
66API
5171
8
8
40
40API
15
Appendix
B
Regression
Results
Table
B­
1
presents
the
regression
results
based
on
the
calibration
data.

CONST
denotes
that
a
constant
entered
the
regression.
FORMS
denotes
that
the
number
of
forms
that
entered
the
regression.
R­
SQUARE
R2
corrected
for
degrees
of
freedom.
NA
indicates
that
no
R2
was
reported
when
the
model
lacked
a
constant.
STANDARD
ERROR
estimated
standard
error
for
the
regression
residuals.

Results
are
first
reported
for
the
OLS
estimates,
for
which
measured
burden
was
set
equal
to
the
midpoint
of
the
reported
range.
The
T­
score
is
the
parameter
estimate
divided
by
its
standard
error.
The
parameter
estimate
and
T­
score
have
similar
interpretations
for
the
maximum
likelihood
estimation,
for
which
the
T­
score
has
an
asymptotic
justification.

Table
B­
1:
Regression
Results
OLS
Estimation
Maximum
Likelihood
Estimation
Industry
Parameter
T­
score
Parameter
T­
Score
5171
FORMS
6.24
20.41
6.24
20.55
STANDARD
ERROR
20.20
20.06
6.12
R­
SQUARE
NA
28
CONST
12.18
0.61
10.20
0.53
FORMS
9.25
3.82
9.45
4.09
STANDARD
ERROR
95.93
91.49
4.98
R­
SQUARE
0.21
33
CONST
10.62
1.29
7.03
0.94
FORMS
7.68
4.01
8.31
4.72
STANDARD
ERROR
30.10
26.43
3.47
R­
SQUARE
0.35
29
FORMS
24.61
9.12
24.62
9.30
STANDARD
ERROR
380.37
372.90
3.60
R­
SQUARE
NA
30
CONST
23.15
3.59
22.08
3.97
OLS
Estimation
Maximum
Likelihood
Estimation
Industry
Parameter
T­
score
Parameter
T­
Score
16
FORMS
1.85
1.17
1.60
1.16
STANDARD
ERROR
18.54
14.25
2.87
R­
SQUARE
0.02
25
CONST
10.86
3.46
10.00
2.85
FORMS
4.30
5.27
3.73
2.18
STANDARD
ERROR
11.35
6.46
1.47
R­
SQUARE
0.55
37
CONST
10.55
0.46
9.79
0.44
FORMS
13.65
2.91
13.89
3.09
STANDARD
ERROR
68.30
64.87
3.14
R­
SQUARE
0.28
36
CONST
21.11
1.72
17.16
2.11
FORMS
3.86
1.12
3.36
1.43
STANDARD
ERROR
26.11
16.55
2.03
R­
SQUARE
0.02
26
CONST
13.16
0.98
1.85
0.21
FORMS
4.70
2.63
6.47
5.19
STANDARD
ERROR
25.11
15.53
1.94
R­
SQUARE
0.35
The
maximum
likelihood
parameter
estimates
correspond
to
the
parameters
from
equation
 
[
1]
in
the
main
text.
For
reasons
explained
in
the
main
text,
the
regression
specification
excluded
a
constant
(
CONST)
for
SIC
codes
5171
and
29.
The
t­
score
is
the
ratio
of
the
parameter
estimate
and
its
estimated
standard
error.
The
table
also
reports
the
estimated
residual
standard
error
for
the
regression.

For
example,
for
SIC
code
28,
the
constant
was
10.20
with
a
t­
score
of
0.53.
The
incremental
burden
per
form
was
9.45
with
a
t­
score
of
4.09.
Depending
on
the
criterion
used,
a
t­
score
in
excess
of
1.96
might
be
judged
as
being
statistically
significant.
The
parameter
associated
with
the
FORMS
variable
is
not
significant
for
SIC
codes
30
and
36
according
to
this
criterion.
Nevertheless,
a
variable
does
not
have
to
have
a
significant
parameter
to
be
a
useful
predictor,
so
the
FORMS
variable
is
always
used
when
making
predictions.
17
An
ordinary
least
squares
regression
was
estimated
as
a
check.
Parameter
estimates
are
similar
between
the
two
models.
The
OLS
model
provides
a
straightforward
estimate
of
explained
variance
(
R2).
Explained
variance
is
very
low
in
two
regressions;
software
does
not
compute
an
R2
when
the
model
specification
lacks
a
constant.
