MEMORANDUM
TO:
Docket
No.
OAR­
2002­
0058
FROM:
Jim
Eddinger,
U.
S.
Environmental
Protection
Agency,
OAQPS
(
C439­
01)

DATE:
January
2004
SUBJECT:
Statistical
Analysis
of
Mercury
Test
Data
Variability
in
Response
to
Public
Comments
on
Determination
of
the
MACT
Floor
for
Mercury
Emissions
BACKGROUND
In
the
October
2002
memorandum
"
MACT
Floor
Analysis
for
the
Industrial,
Commercial,
and
Institutional
Boilers
and
Process
Heaters
National
Emission
Standards
for
Hazardous
Air
Pollutants",
we
stated
that,
for
existing
sources,
the
calculation
of
numerical
MACT
floor
emission
limits
was
a
two­
step
analysis.
The
first
step
involved
calculating
a
numerical
average
of
an
appropriate
subset
of
the
emission
test
data
from
units
using
the
same
technology,
or
technologies,
as
the
units
in
the
top
12
percent.
The
second
step
of
the
process
involved
generating
and
applying
an
appropriate
variability
factor
to
account
for
unavoidable
variations
in
emissions
due
primarily
to
uncontrollable
differences
in
fuel
characteristics
and
ordinary
operational
variability.

For
the
proposed
rule,
the
variability
in
the
measured
emissions
was
calculated
for
each
unit
with
multiple
emission
tests
by
dividing
the
highest
three­
run
test
result
by
the
lowest
threerun
test
result.
The
overall
variability
in
the
measured
emissions
from
these
units
was
calculated
by
averaging
all
the
individual
unit
variability
factors.
This
overall
variability
factor
was
multiplied
by
the
overall
average
measured
emissions
performance
level
to
derive
a
MACT
floor
emission
limit
representative
of
the
average
emission
limitation
achieved
by
the
top
12
percent
of
units.

PUBLIC
COMMENTS
During
the
public
comment
period
for
the
proposed
rule,
we
received
many
comments
requesting
that
EPA
revise
the
MACT
floor
methodology
for
mercury
emission
limits.
The
commenters
contended
that:
1.
the
variability
factor
was
calculated
inappropriately,
2.
EPA
should
account
for
variability
in
fuel
composition
in
the
MACT
floor
analysis,
3.
the
MACT
floor
analysis
should
incorporate
variability
in
fuel
being
burned,
the
units
operating
condition,
and
sampling
and
analytical
errors,
4.
EPA's
calculation
of
variability
was
statistically
unsound,
and
5.
EPA
should
estimate
statistically
the
variance
in
the
distribution
of
control
technology
performance
rather
than
calculate
a
variability
factor.

RESPONSE
TO
COMMENTS
2
In
the
proposed
rule,
we
requested
commenters
to
provide
additional
emissions
information.
However,
only
one
source
provided
any
additional
mercury
emissions
data.
This
information
(
test
results
from
three
additional
coal­
fired
industrial
boilers
equipped
with
fabric
filters)
was
used
to
revise
the
mercury
emission
limit
for
existing
large
solid
fuel
units.
We
also
reviewed
the
mercury
emission
database
used
to
develop
the
MACT
floor
emission
limit
for
existing
sources.
After
review,
we
determined
that
a
revision
to
the
variability
factor
was
appropriate.
Incorporating
the
additional
data
and
the
revised
variability
factor
(
determined
using
the
same
approach
as
for
the
proposal)
into
the
MACT
floor
analysis
resulted
in
a
calculated
mercury
MACT
floor
emission
limit
of
9
lb/
trillionBtu
(
compared
to
7
lb/
trillion
Btu
in
the
proposal).
A
detailed
discussion
of
the
revised
MACT
floor
analysis
conducted
is
provided
in
the
memorandum
"
Revised
MACT
Floor
Analysis
for
the
Industrial,
Commercial,
and
Institutional
Boilers
and
Process
Heaters
National
Emission
Standards
for
Hazardous
Air
Pollutants
Based
on
Public
Comments"
in
the
docket.

Variability
of
the
emissions
data
were
incorporated
into
the
proposed
emission
limits.
The
approach
used
to
determine
the
variability
of
emissions
information
indirectly
incorporated
variability
in
fuel,
operating
conditions,
and
sampling
and
analytical
conditions
because
these
parameters
vary
from
emission
tests
conducted
from
one
unit
to
another,
within
one
unit,
and
over
time.
For
the
final
rule,
as
a
comparison
to
the
variability
analysis
conducted
for
the
proposed
rule,
we
did
conduct
a
statistical
analysis
of
the
data
to
identify
the
97.5th
percent
confidence
interval.
This
analysis
provided
similar
results
to
the
variability
analysis
conducted
in
the
proposed
rule.
Consequently,
we
decided
not
to
change
the
variability
methodology.
A
detailed
discussion
of
the
statistical
analysis
conducted
is
provided
below.

STATISTICAL
ANALYSIS
To
address
the
concern
on
whether
the
MACT
floor
approach
used
for
the
proposed
rule
reasonably
ensures
that
the
mercury
emission
limit
selected
as
the
MACT
floor
adequately
represents
the
average
level
of
control
actually
achieved
by
units
in
the
top
12
percent,
considering
ordinary
operational
variability,
we
conducted
a
statistical
analysis
similar
to
the
statistical
analyses
used
in
the
Integrated
Iron
and
Steel
Plants
MACT
and
the
proposed
Utility
MACT.
That
is,
we
estimated
the
mercury
emissions
limitation
achieved
for
a
source
at
the
97.5th
percentile
using
the
one­
sided
z­
statistics
test
(
i.
e.,
the
emission
limitation
which
the
emission
point
is
estimated
to
be
able
to
achieve
97.5
percent
of
the
time).
The
median
(
average)
of
the
97.5th
percentiles
of
the
top­
performing
sources
would
be
considered
as
the
MACT
floor.

PROCEDURE
In
the
determination
of
the
MACT
floor
for
mercury
emissions
from
existing
large
solid
fuel­
fired
boilers,
we
used
the
results
of
the
mercury
emissions
tests
for
9
industrial
boilers
equipped
with
fabric
filters.
Multiple
stack
tests
were
conducted
on
5
of
these
units.
The
mercury
emissions
(
lb
Hg/
trillion
Btu)
for
each
test
run
of
the
23
stack
tests
are
presented
in
Table
1,
together
with
the
unit
average
(
mean)
emissions
of
mercury
and
the
standard
deviation
of
3
the
individual
test
runs.

The
statistical
approach
used
was
the
one­
sided
z­
statistics
test
using
the
equation:

confidence
limit
=
average
+
z
*
standard
deviation,

where
the
value
of
z
is
a
function
of
the
degrees
of
freedom
and
obtained
from
the
statistical
table
listing
t
distribution
critical
values.
The
number
of
degrees
of
freedom
for
sample
size
n
is
simply
n­
1
for
a
one­
sample
mean
problem.
The
z
values
used
in
determining
the
97.5
percentile
confidence
limit
are:

Degrees
of
Freedom
z
value
2
4.303
3
3.182
5
2.571
11
2.201
26
2.056
63
2.000
DATA
The
data
consisted
of
stack
emission
measurements
(
lb
Hg/
TBtu)
from
9
units.
In
general,
three
replicate
runs
were
performed
at
each
unit.
For
4
of
the
units,
only
results
from
2
test
runs
were
available;
hence,
there
were
64
runs
overall.
The
averages
and
the
observed
within­
unit
variances
(
standard
deviations)
for
these
units
are
listed
in
Table
1.
For
test
runs
where
the
mercury
emissions
were
below
the
detection
limit,
the
detection
limit
is
reported.
4
TABLE
1
MERCURY
DATA
FOR
LARGE
SOLID
FUEL
BOILERS
EQUIPPED
WITH
FABRIC
FILTERS
Test
ID
Facility
Name
Run
1
Result
(
lb/
TBtu)
Run
2
Result
(
lb/
TBtu)
Run
3
Result
(
lb/
TBtu
Facility
Average
(
lb/
TBtu)
Standard
deviation
97.5th
Confidence
limit
(
lb/
TBtu)

E27.001
Delano
Energy
1.28
0.29DL
0.29DL
E27.008
Delano
Energy
0.43DL
0.43DL
0.43DL
E833.005
Delano
Energy
0.0077DL
0.0075DL
0.0074DL
E833.006
Delano
Energy
0.0076DL
0.0076DL
0.0075DL
0.266
0.370
1.08
E15.002
GWF
Power
0.529DL
0.515DL
0.533DL
E20.004
GWF
Power
0.508DL
0.782DL
0.223DL
0.515
0.177
0.97
E697.007
James
River
5.81
6.31
E697.011
James
River
7.89
11.7
7.928
2.667
16.41
E738.002
Kimberly­
Clark
1.22
1.03
0.992
E739.001
Kimbrtly­
Clark
0.075
0.686
0.0813
0.681
0.497
1.96
E1.006
National
Cogeneration
9.1
3.6
2.76
5.153
3.443
19.97
E794.001
Wheelabrator
1.38DL
2.7
1.46DL
E795.005
Wheelabrator
1.41DL
1.58DL
1.54DL
E795.013
Wheelabrator
1.45DL
1.27DL
1.39DL
E795.021
Wheelabrator
1.21DL
1.33DL
1.23DL
E797.010
Wheelabrator
1.64DL
1.69DL
7.45
E797.019
Wheelabrator
2.44DL
2.89DL
2.87DL
E798a.
010
Wheelabrator
0.504
1.4DL
6.49
E798a.
024
Wheelabrator
7.9
3.47
1.74
E798b.
006
Wheelabrator
1.4DL
1.4DL
1.4DL
2.320
1.907
6.24
MSU1
Michigan
State
4.361
5.587
4.974
0.867
15.99
MSU2
Michigan
State
5.488
4.910
5.199
0.409
10.39
MSU4
Michigan
State
0.203
0.248
0.225
0.032
0.63
Test
ID
Facility
Name
Run
1
Result
(
lb/
TBtu)
Run
2
Result
(
lb/
TBtu)
Run
3
Result
(
lb/
TBtu
Facility
Average
(
lb/
TBtu)
Standard
deviation
97.5th
Confidence
limit
(
lb/
TBtu)

5
Average
of
top
12%
3.03
8.18
In
Table
1,
this
average
mercury
emission
rate
is
3.0
lb/
TBtu.
This
average
number
has
two
elements
of
uncertainty
associated
with
the
value
of
the
number.
One
uncertainty
is
the
actual
value
of
the
long­
term
average
of
the
best
units,
since
there
are
only
a
few
tests
available
that
represent
the
best
12
percent
of
the
units
and
there
is
measurement
uncertainty
associated
with
each
test
set
of
3
measurements.
The
second
uncertainty
is
the
variability
of
emissions
for
these
best
units
under
the
worst
foreseeable
circumstances
under
normal
operating
conditions.
This
second
uncertainty
includes
operational
variability.
There
is
no
direct
measurement
of
operational
variability,
although
there
was
some
operational
variability
included
in
the
measurement
uncertainty
associated
with
each
test
set
of
three
measurements.

RESULTS
The
mercury
emissions
(
lb/
trillion
Btu)
for
each
of
the
9
units
representing
the
best
performing
12
percent
are
listed
in
Table
1,
together
with
their
average
(
mean)
mercury
emissions.
The
97.5th
percentile
confidence
limits
for
each
of
the
9
units
are
also
listed
in
Table
1.
It
is
assumed
that
the
confidence
limits
indicate
the
long­
term
performance
of
the
unit
due
to
unit
performance
variability
being
added
to
the
unit's
average
through
conventional
statistical
techniques.

The
average
of
these
confidence
limits
(
indicating
long­
term
performance)
for
the
bestperforming
12
percent
units
is
8.2
lb/
trillion
Btu.
If
the
statistical
approach
was
used,
this
would
be
the
MACT
floor
limit
since
this
value
is
the
average
of
the
units
representing
the
best
performing
12
percent.
This
level
is
slightly
lower
that
the
MACT
floor
limit
(
9.0
lb/
trillion
Btu)
determined
by
using
the
approach
used
for
the
proposed
rule.

If
we
would
conduct
the
same
statistical
analysis
on
the
data
set
as
a
whole
(
64
test
runs),
the
resulting
average
is
2.2
lb/
trillion
Btu
with
a
standard
deviation
of
2.58.
The
97.5th
percentile
confidence
limit
would
be
7.4
lb/
trillion
Btu.

CONCLUSION
We
continue
to
believe
that
by
considering
the
variability
of
emissions
information,
we
have
indirectly
incorporated
variability
in
fuel,
operating
conditions,
and
sampling
and
analytical
conditions
because
these
parameters
vary
from
emission
tests
conducted
from
one
unit
to
another,
over
time,
and
even
within
each
test
set
of
three
measurements
at
a
single
unit.
The
most
6
elementary
measure
of
variation
is
range.
Range
is
defined
as
the
difference
between
the
largest
and
smallest
values.
This
is
the
variability
methodology
used
in
the
proposed
rule.
That
is,
for
each
unit
with
multiple
3­
runs
emissions
tests
conducted
over
time,
the
variability
for
that
unit
was
calculated
by
dividing
the
highest
three­
run
test
result
by
the
lowest
three­
run
test
result.
The
overall
variability
was
calculated
by
averaging
all
the
individual
unit
variability
factors.
This
overall
average
emission
level
was
multiplied
by
the
overall
averaged
variability
factor
to
derive
a
MACT
floor
limit
representative
of
the
average
emission
limitation
achieved
by
the
top
12
percent
of
units.
We
believe
that
this
approach
adequately
accounts
for
inherent
fuel
supply
variability.

The
more
robust
statistical
analysis
(
t­
test)
of
the
mercury
emissions
data
used
in
the
MACT
floor
analysis
to
identify
the
97.5th
percent
confidence
limit
provided
similar
results
to
the
variability
analysis
conducted
in
the
proposed
rule.
However,
since
three
units
had
most
or
all
of
their
test
results
at
the
detection
limit,
we
have
concerns
about
the
effect
of
using
these
detection
limit
values
on
the
statistical
analysis.
Consequently,
we
concluded
not
to
change
the
variability
methodology.
