DOCKET SECTION
RECEM&YT-~
BEFORE THE
POSTAL RATE COMMlSSlON
POSTAL RATE AND FEE CHANGES,
1997
i
DEE30
3 34 PpI ‘97
::tSTLi
2,’ : :: ,: :,iv;‘,,-!I!
OFiitt: (11’j ..;( $y.,,$ ,>,iy
DOCKET NO. R97-1
DIRECT TESTIMONY OF
KEVIN NEELS
ON BEHALF OF
UNITED PARCEL SERVICE
TABLE
QUALIFICATIONS
DEFECTS
......................................
OF MY CONCLUSIONS
IN BRADLEY’S
A.
B.
APPROACH
2
...................................
. .
3
. ..
.
Bradley Fails to Use Appropriate Measures
ofVolume
. . . . . . .._....................................
.
..
2.
“Total Piece Handlings” Is Not a Suitable Proxy For
Volume . . . . . .._.................................
2.
3.
8
8
Hours Are Not a Suitable Proxy For Cost
Flaws In Bradley’s Implementation
.
of Cost and
1.
1.
C.
1
.................................................
NATURE OF MY ASSIGNMENT
SUMMARY
OF CONTENTS
. .
.
.
..
..
..
8
12
.
14
There Are Serious Shortcomings in the Piece Handling
Data Used in Bradley’s Econometric Analysis . . . .
.
15
Bradley’s Data “Scrubbing” Procedures
Substantively Altered His Results
.
23
Have
.
.
.
Bradley’s Results Imply Implausible Patterns of
Technological Change . . . .
. . ...
.
.
Bradley’s Analysis Sheds Little Light on the Long-Rufn Volume
Variability of Costs
.. .. . .. ..
. .. ..
.
34
39
BRADLEY’S CROSS-SECTIONAL
MODEL PROVIDES SUPERIOR
RESULTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
BRADLEY HAS NOT DEMONSTRATED
THAT VOLUME VARIABILITY
IS LESS THAN 100 PERCENT
. .. ... . . .. . . ... . .
A RELIABLE
A.
ECONOMETRIC
ESTIMATE
OF VOLUME VARIABILITY?
Bradley’s Cross-Sectional Model Provides An Appropriate
Starting Point . . . . . . . . . . . . . . . . . . . . . . . . .
-i-
.
. 44
..
44
.
45
B.
Some Modifications of Bradley’s Cross-Sectional Model Are
Appropriate
................. ..... .. .. ..
.
C.
Data ‘Scrubbing”
D.
Implementation of These Changes Still Leaves a Model That
Fails to Consider Either Actual Costs or Actual Volumes
.
CONCLUSION
APPENDIXA
.
.
.
.
.. .. ..
..
. .
. . . 45
. . . . . 46
47
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A-l
-ii-
J ST OF TABLES
Table 1 -
Table 2 --
Volume Variability Estimates Derived from
Modified Version of Bradley’s Cross-Sectional
Model
.
7
.
...
17
Number of Instances in Which Piece Handlings
Are Reported at a MODS Facility for a Direct
Activity for Only a Single Accounting Period .
Number of Instances in Which Gaps in Piece Handlilngs
Reported at a Facility Appear for MODS Direct Activiities
18
Table 4 --
Data Eliminated
25
Table 5 --
Effects of Discarding Usable Observations on Bradley’s
Estimates of the Volume Variability of Mail
Processing Labor Costs
. . ...... . ..
.
. 32
Volume Variability Estimates from
Bradley’s “Between” Model . . .
. 41
Table 3 --
Table 6 --
Due to Data “Scrubbing”
.
.
. .....
... .
Table A-l -- Results Obtained When Bradley’s Errors-in-Variables
Methodology Is Applied to All MODS Direct Activities
.
A-5
..
21
Criteria ., . .
. 28
ImIST OF FIGURES
Figure 1
Automatic vs. Manual Volume Variabilities
Figure 2
Manual Parcel Sorting Site #242 Threshold
Figure 3
Effects of Time Trend Variables in Bradley’s
Manual Letter Sorting Model . . . . . . . . . . .
Figure 4
.
.
Effects of Time Trend Variables in Bradley’s
Manual Parcel Sorting Model
. . . . .
.. .
-111.
.
.
.
..
.
.
.
36
. 37
BEFORE THE
POSTAL RATE COMMISSION
POSTAL RATE AND FEE CHANGES,
1997
:
DOCKET NO. R97-1
DIRECT TESTIMONY OF
KEVIN NEELS
ON BEHALF OF
UNITED PARCEL SERVICE
1
LIFICATIONS
My name is Kevin Neels.
2
I am a director at the management
3
economic consulting
4
economic
5
my work has involved the use of econometric
6
relating to product costing in a wide range of areas.
firm of Putnam, Hayes ,& Bartlett, Inc. I have provided
research and consulting
services for more than twenty years.
Much of
analysis and has addressed
I have analyzed pricing behavior in the context of allegations
7
and
8
antitrust violations
and conducted
9
services in areas ranging from municipal services to air traffic control.
issues
of
numerous studies on the pricing1 of government
10
connection
with this and other work, I have frequently
11
costs for purposes of determining
12
the costs associated
13
product firm, or the effects of incremental
conducted
the total cost of providing
In
investigations
of
a product or service,
with a specific product or service in the context of a multichanges in volume on total profit.
I have
1
developed
econometric
analyses for purposes of estimating
2
litigation; of providing
3
sales of changes in product design, pricing, or promotion;
4
structure of costs.
5
econometric
clients with forecasting
capability;
damages
in civil
of estimatilng the effects on
and of analyzing
the
I have also been called upon to provide critical evaluations
analyses developed
of
by others.
Prior to joining Putnam, Hayes 8 Bartlett, Inc. I held a number of
6
7
responsible
positions in economic research and consulting.
8
president
9
River Associates,
in the Transportation
I was previously
Program at the economic consulting
Inc., and senior economist at Abt Associates,
10
firm. I have also served as a senior research associate
11
Division of the Rand Corporation
12
Program of the Urban Institute.
13
University.
and as an associate
vice
firm of Charles
a policy research
in the Systems Sciences
in the Transportation
Studies
I hold a Ph.D. and a B.A., both from Cornell
NATURF OF MY ASSIGNfyljXt
14
In previous proceedings,
15
the Commission
labor costs are 100 percent volume variable.
has decided that mail
16
processing
17
Service witness Michael D. Bradley (USPS-T-14)
18
purports to demonstrate
19
changes in mail volume.
20
a series of cost equations.
21
in specific activities, facilities, and accounting
22
pieces of mail are handled in those activities, facilities, and acwunlting
that mail processing
In this proceeding,
has introduced
al study that
labor costs do not vary fully with
Bradley’s study is based on the econometric
These cost equations
-2-
Postal
estimation
relate the number of labor hours
periods to the number of times
periods.
of
1
For activities at Management
Operating
Data System (MODS) facilities,
2
relies on data drawn from the Postal Service’s MODS database.
3
Centers (BMCs), he relies upon data from the Productivity
4
System (PIRS) database,
5
Service for BMC activities.
another operational
database
I have been asked to review the approach
6
Bradley
F:or Bulk Mail
Information
maintained
Reporting
by the Postal
used by Bradley and to
7
determine whether it provides accurate and reliable estimates of the volume
8
variability
9
also been asked to review Bradley’s implementation
of mail processing
costs that are suitable for use in rate setting.
of the results that he presents.
of his approach
I have
and to assess
10
the soundness
Finally, I have been asked to offer
11
an opinion in the light of all of the available evidence about how mlail processing
12
costs should be treated in determining
postal rates.
SUMMARYOF My CONCl USIONS
13
Bradley’s econometric
14
the extent to tiich
analysis does not provide a reliable basis for
15
determining
16
Bradley’s econometric
17
determining
18
estimate the extent to which compensation
19
some affirmative evidence that compensation
20
One can advance reasonable
21
compensation
22
Bradley presents no evidence that would permit a determination
whether
mail processing
equations
labor costs vary with volume.
look not at cost, but,at labor hours.
Before
labor hours are a suitable proxy for cost, one must either
per hour varies with volume, or provide
per hour is independent
of volume.
arguments to support either the contention
per hour increases with volume or that it decreases
-3-
that
with volume.
of which is the
1
case.
In the absence of such evidence, the use of labor hours as a proxy for cost
2
could result in either an underestimate
3
or an overestimate
Also, Bradley’s econometric
equations
a measure of mail processing
of volume variability.
look not at volumes, but rather
4
at piece handlings,
5
changes
8
and other factors.
7
for volume, one must first estimate the extent to which piece handlings vary with
8
volume, or provide some affirmative evidence that the two are directly proportional.
9
Once again, Bradley presents no evidence on this question.
in volume, but also to changes in routing, sorting technology,
Before determining
whether piece handlings
10
such evidence, the use of piece handlings
11
either an underestimate
or an overestimate
13
problems with the implementation
14
questions
15
include:
64
about the robustness,
piece handlings
18
investigations;
(b)
22
issues, there are a number of
approach that raise serious
and relevance
of his results.
These
has been questioned
Bias in volume variability
in internal Postal Service
estimates due to errors in the measurement
of mail volumes;
20
21
In the absence of
Reliance on a dataset whose ability to provide accurate estimates of
17
19
are! a suitable proxy
of volume variability.
of Bradley’!
reliability,
error rates,
as a proxy for volumes could result in
Quite apart from these fundamental
12
16
steps that is sensitive not only to
(c)
Adoption
of a set of highly subjective data ‘scrubbing”
that result in the elimination
procedures
from the analysis of enormous quantities
1
of otherwise
2
analysis;
3
usable data and that significantly
alter the results of the
Use of a complex set of time trend variables that prodluces erratic and
(4
counter-intuitive
4
estimates of the effects of technologiical
5
Also, and perhaps most significant,
Bradley’s equations
6
estimates of the long-run response
of costs to changes in volume.
change.
cannot provide accurate
These factors lead me to conclude that the new study introduced
7
8
Bradley does not justt
9
determination
rejection of the Commission’s
that mail processing
well-established
labor costs are fully volume variiable.
10
sense indicates that labor costs should be fully variable.
11
unadorned
12
provide convincing
13
recommend
14
processing
Common
Simple, straight-forward
plots of the raw data tend to confirm this view. Bradley has failed to
evidence to contradict
that the Commission
the Commission’s
stand by its traditional
traditional
position.
I
position and treat mail
labor costs as 100 percent volume variable.
If the Commission
15
by
does elect to adopt some version of Bradley’s
16
econometric
17
starting point, This analysis is better able to provide estimates of long-run volume
18
variability,
19
of volume.
20
analysis, I recommend
adoption of his cross-sectional
and it is less subject to downward
l recommend
analysis as a
bias from errors in the measurement
a number of changes to the cross-sectional
21
Bradley presents in his workpapers.
22
in his recommended
23
should therefore
model that
The time trend variables that Bradley includes
model have little meaning in a cross-sectional1 context and
be dropped.
Also, the cross-sectional
analysis needs to be
1
extended to all of the activities that Bradley considers.
2
results are based upon a drastically
3
“scrubbing”
4
portion of the available data. I have rerun Bradley’s cross-sectional
5
dataset that uses all of the data. The volume variability
estimates
6
this improved cross-sectional
in Table 1.
process.
Bradley’s
reduced dataset that emerges from his
This process results in the loss of an unacceptably
analysis are summarized
Table 1 does not include volume variability
7
Furthermore,
8
and remote encoding
9
for these two activities is very limited.
large
analysis on a
resulting from
estimates for the registry
activities because the amount of data available for analysis
I do not believe that it is adequate
10
time to support an econometric
analysis of the volume variability
11
labor costs.
12
reaffirm its earlier finding that mail processing
13
variable.
Hence, for these two activities I recommend
-6-
at this
of mail processing
that the C:ommission
labor costs are 100 percent volume
1
Table 1
2
3
Volume Variability Estimates Derived from
Modified Version of Bradley’s Cross-Sectional Model
4
5
8
7
8
9
10
11
12
13
14
15
18
17
18
19
20
21
22
23
24
25
26
27
28
Activity
L”
k
32
132%
121%
121%
116%
125%
131%
BCS Sorting
npR Sorting
LS\ A Sorting
r-FSM Sorting
etter Sorting
11t-lat Sorting
7
..
110%
1 l^^’
lU”70
Sorting
‘P”
Ii”.”Kw.
8%
9i--ll-. J9%
I
n narn
U.U40L
0.1254
0.0901
-
0.0490
MODS Allied
Opening - Pref Mai1
,-..
-z
1 Opening - BUIK Business Mail
him
I
13A%
..,-
I
128%
131%
146%
I
.-.
BMC Direct
I Sack
- - -. . Snrtina
- -. _. il Machine
-- -
119%
Primary Parcel Sorting I Machine
Secondary Parcel Skorting Machine
IIr D-rrc,l
I a*YII Dnr
I ”A
+!?&
,naninn
y.,, 111
.y ”I Init
Nnn
,.-, , Mm-hinahln
l..lwll.llll.Outsides
159%
89%
84%
82%
78%
BMC Allied
29
30
31
Results Base,d Up:,/
tpmmym
Ellll
Es,
-.. ,,mple
MODS Din set
Platform
Floor Labor
45%
120%
I
* Source: WP I.
-l-
t
Cl l-If574
1
DEPPCTS
Bradley Fails to Use Appropriate
A.
2
3
IN BRADfJ=Y’S
APPROACH
Measures
of Cost
of costs must relate a
Any empirical study of the volume variability
4
5
suitable measure of cost to a suitable measure of volume.
6
threshold
7
of cost, nor a true measure of volume.
6
1.
requirement.
This is a fundamental
In Bradley’s study, however, there is neither a true measure
Hours Are Not a Suitable Proxy For Cost
At the most basic level, ‘cosr
9
is a quantity that is defined in monetary
10
units. However, no such measure appears anywhere
in Bradley’s datasets or
11
results.
into the various activities he
12
examines.
Cf course, labor hours and labor costs are related.
13
expressed
by average compensation
14
costs are equal to the product of average compensation
15
labor hours.
Instead, he focuses on labor hours ‘clocked”
cost per labor hour. By definition,
But average compensation
16
That relationship
is
total labor
per labor hour times total
per hour is influenced
by a variety of
At a given facility and a given point in time, there will be a
17
different factors.
16
schedule that spectiies hourly wage rates for different crafts, different levels of
19
seniority, different types of employees
20
and for different types of time (k,
21
schedule of wages may be higher or lower, depending
upon labor market
22
conditions,
and other factors.
(h,
permanent,
casual, temporary,
etc.),
straight time, overtime, holiday time, etc.). This
inflation, collective bargaining
agreements,
All else
1
equal, the higher these wage rates, the higher the average compensation
2
will be.
Average compensation
3
per hour will also be influenoed,
4
the mix of labor hours at a facility.
5
types of time (such as overtime),
6
more highly paid categories
7
hour even if the wage schedule
6
wage rates that are paid, the more powerfully
9
be influenced
per hour
however, by
A shift in the mix of hours toward more costly
higher paid crafts, more senior employees,
of employees
will raise average compensation
remains unchanged.
or
per
The greater the range of
average compensation
per hour will
by changes in the mix of types of hours worked.
Bradley ignores the effects that changes in the mix of types of hours
10
11
have on average compensation
12
asserts that:
per hour and therefore on total labor costs.
13
14
For mail processing labor cost, the variations
processing hours are the variations in cost.’
15
This statement
is not correct.
He
in mail
For example, it is obvisous that if wage
16
rates increase, costs will also increase even in the absence of a change in labor
17
hours.
16
increase if the mix of hours shifts in the direction of more highly paid types of time.
19
Overtime is perhaps the best illustration
20
hours are the same as cost; using hours as a proxy for costs will nlot capture the
1.
In addition, the cost associated
with a given number of labNor hours will also
of the fallacy in Bradley’s assertion that
Direct Testimony of Michael D. Bradley on Behalf of the United States Postal
Service, USPS-T-14, p. 12 (emphasis added).
-9-
1
total impact of increased volume on costs when the increase in vollume leads to a
2
need for higher paid overtime hours.
If either of these two factors -- the schedule of wage rates or the mix
3
4
of types of hours -- changes systematically
5
the overall variability
6
changes in average compensation
7
whether or the extent to which such volume-related
8
them into his estimates of volume variability.
of mail processing
with volume, this change will influence
costs.
Because Bradley’s analysis ignores
per hour, it is not capable of determining
effects exist or of factoring
While one might argue that the schedule of wage rates is determined
9
10
largely by general labor market conditions
rather than by mail volume, the same
11
cannot be said for the mix of types of time. There are a number of reasons for
12
believing that the mix of hours at a facility might vary systematically
13
High-volume
14
cost temporary
15
the involvement
of higher-cost
16
mail processing
schedules
17
maintenance
18
use of overtime and greater amounts of overtime pay.
periods could be characterized
or casual workers.
by the more extensive use of lower-
Conversely,
high-volume
senior or supervisory
periodls could require
personnel
and maintain service standards.
of service standards
with volume.
during high-volume
in order to meet
It is also possible that
periods could involve greater
We have no idea what the net effect is of these different factors.
19
20
only way to determine whether average compensation
21
systematically
22
between actual labor costs and volume.
The
per hour varies
with volume and, if so, by how much, is to examine the relationship
Bradley has not done this.
-lO-
Bradley argues that his use of hours rather than cost to measure cost
1
2
variation with volume is a virtue rather than a limitation of his study.
3
characterizes
4
he argues, ‘hours are directly comparable
5
them for inflation.” ’
hours as “a ‘real’ variable that inflation does not influence.”
These statements
6
He
through time, and I do not have to adjust
are not entirely correct.
7
focusing on hours Bradley has eliminated
a
shifts in the overall wage schedule
9
resulting measure of hours is comparable
Hence,
While it is true that by
changes in costs that are associated
with
rather than with volume, it is not true that the
across sites or across tirne, a
10
precondition
for the use of labor hours as a proxy for costs. The hours of
11
supervisory
personnel
12
unskilled casual workers.
13
hours of the same individual
14
differences
15
in the overall schedule of wages, these other differences
16
dealt with them.
and skilled craftsmen are not the same as the hours of
Nor is it even true that the straight time and overtime
are comparable,
since there are real resource cost
between the two types of time. Even after removing the effects of shifts
remain.
IBradley has not
The need to account for changes in overall wage levels cannot serve
17
18
as a justification
19
true that focusing on compensation
20
the effects of inflation and changes in wage levels, such adjustments
21
difficult to make.
2.
for failure to account for variations
USPS-T-14,
in the mix of hours.
While it is
costs would have necessitated1 adjustments
p. 13.
-ll-
are not
for
1
2.
“Total Piece Handlings”
2
Is Not a Suitable Proxy For Volume
It is also obvious that an econometric
study of the valriability of mail
3
processing
costs with changes in volume should involve an analysis of changes
4
the volume of mail delivered.
5
short. With the sole exception of registered
6
are devoid of any measure of the volume of mail actually delivered.
7
Bradley bases his conclusions
8
conceptually
9
easily lead to erroneous
On this count, too, Bradley’s analysizs comes up
mail, Bradley’s datasets and results
on an analysis of piece handlings.
distinct from volume and, therefore,
conclusions
Instead,
This measure is
using it as a prory for volume can
regarding the volume variability
On its route toward final delivery, a particular
10
in
of costs.
item of mail will
11
generate an additional
piece handling every time it passes though a processing
12
step. Thus, the more complex an item’s routing, the more piece handlings
13
require.
14
than one that must travel through twc SCFs, or one that goes from one SCF
15
through a BMC to another SCF, or one that must travel from an SCF through two
16
BMCs to another SCF. An item that requires both a primary and a secondary
17
will experience
18
presorted dropshipped
19
in a corner mailbox with other unsorted items.
it will
An item that passes through a single SCF will undergo less processing
sort
more piece handlings than one that requires only a single sort. A
item will require less processing
than one that is deposited
20
Even if Bradley is correct in his assertion that ‘[t]he primary driver of
21
costs in any activity is the number of pieces sorted in that activity.“3 he cannot draw
3.
USPS-T-14,
p. 13.
-13
1
conclusions
about the volume variability
of costs from an analysis of piece
2
handlings without first considering
3
number of times a piece of mail is handled tends to decline with volume, Bradley’s
4
analysis will overstate the volume variability
5
of mail is handled tends to increase with volume, Bradley’s analysis will understate
6
the volume variability
the volume variability
of piece handlings.
of costs; if the number of times a piece
of costs.4
A variety of different factors could alter the relationship
7
and volumes.
handlings
9
changes in the use of presort options by business mailers; (2) changes in the
proportion
of intra-BMC
11
handling procedures
That relationship
between piece
8
10
If the
could be affected, for example, by: (1)
or intra-SCF movements;
or (3) modifications
in mail
that result in the addition or deletion of processing
12
Changes
13
could mask significant
14
this might happen.
15
error sorting rates and could thereby result in mail having to be resorted afler being
16
delivered to the wrong processing
17
incorrect destination,
18
a second time through the sequence
19
the increase in error rates clearly leads to an increase in cost, Bradley’s
4.
in the relationship
diseconomies
between piece handlings
steps.
and volume
of scale. There are a variety of ways in which
For example, increases
center.
in volume could lead tlo increases
in
When such missorted mail arrives at the
it would then have to reenter the processing
of sorting and processing
stream and pass
steps.
Even though
analysis
In cross examination on his testimony regarding purchased transportation
costs, Bradley conceded that failure of a proxy measure of volume to directly
track a true measure of volume can bias estimates of volume variability: “I
did not study the relationship between cubic-foot miles and volume. And
certainly I think to the extent that that second analysis had ,a lower variability
then my capacity variabilities would be too high.” Tr. 713823-24.
-13-
1
would incorrectly
interpret this as an increase in productivity,
2
the original sort to handle pieces in less time.
It is possible for the number of times an item is handled to increase
3
4
systematically
5
become economical
8
individual processing
with volume, even over the long term. At higher volumes it may
to incorporate
greater specialization
steps into the overall processing
Bradley has provided no information
7
8
piece handlings
9
determine what his piece handling variability
10
since it would allow
variability
and volume.
of mail processing
on the relationshlip
between
the Commission
cannot
estimates imply for the volume
costs.
Apart from the fundamental
12
also more
sequence.
Without such information
B. Flaws In Bradlev’s
11
and therefore
Imnlementirtina
deficiencies
above, there are a number of significant
in conceptual
approach
13
discussed
problems in the implementation
14
of this approach
15
deficiencies
18
to which the data have been subjected, the treatment of technologilcal
17
the focus on the short-run response of hours to changes in piece handlings.
18
Correcting
19
sensitivity of the results to changes in sample selection,
20
method of estimation
21
enough and stable enough to provide a sound foundation
that call into serious question the reliability of its results.
have to do with the data sources relied upon, the ‘scrulbbing”
these problems generally
raises questions
These
process
change, and
results in higher estimates of -variability.
model specification,
The
and
about whether these results are reliable
-14-
for ratemaking.
1
2
1.
There Are Serious Shortcomings
Bradley’s Econometric Analysis
3
in the Piece Handling Data Used in
The data quality requirements
for a sound econometric
4
those for simpler forms of analysis.
5
the latter context can give rise to biased results when the same data are used to
a
develop regression
7
sources carefully before using them in sophisticated
a
of Bradley’s data source for Total Piece Handlings
9
problems.
equations.
Measurement
study exceed
For this reason, it is essential to scrutinize
The MODS piece handlings
10
errors that might cancel out in
econometric
studies.
indicates potentially
data
Scrutiny
serious
data that Bradley relies upon for major
11
portions of his analysis have been the target of considerable
12
review of measurement
13
found large variances
14
system and actual piece counts. 5 These variances were attributed; to a variety of
15
different causes, including inadequate
ia
and out-of-tolerance
17
substantial.
5.
a.
systems conducted
criticism.
by the U.S. Postal Inspection
between the piece handling figures contained
scales.
A recent
Service
in the MODS
conversion factors,6 improper data input,
The magnitudes
of these variances
c:ould be
In one instance, the count projected by the MODS system for 57 trays
. .
National Coormon
Audit.‘. Mall VolumeMeasurement
RepQdrna
p
United States Postal Inspection Service, December 1998, LR-HIn many situations, the MODS system records number of trays, sacks, or
pounds of mail and then uses conversion factors to convert these estimates
to piece counts. ti
LR-H-220, p. 8.
-15
1
was 29,837 pieces, while the actual piece count was 17,842 pieces -- an error of 88
2
percent.’
In one of his interrogatory
3
responses,
Bradley claimed that “several of
4
the reports findings are irrelevant for my analysis because much of the data set
5
used in my analysis is not based upon FHPs [First Handling
a
end-of-run
7
the piece handling variable used by Bradley; the MODS Manual’
a
Section 212.2 that Total Piece Handlings
9
Subsequent
data and machine counts.“*
Handling
Pieces.
Pieces;], but rather on
However, First Handling Pieces is a part of
states clearly in
is the sum of First Handling
Even if the MODS counts of downstream
Pieces and
handlings
10
are totally free of the measurement
11
Pieces, all of the problems surrounding
12
are still passed forward into Bradley’s analysis.
13
raised regarding the accuracy of the MODS piece handling data naturally lead one
14
to question whether the same problems infect the PIRS piece handling data upon
15
which Bradley based his analysis of BMC mail processing
In fact, examination
ia
problems that infect estimates of First Handling
the measurement
of First Handling
The questions
that, have been
costs.
of Bradley’s datasets reveals many problems.
17
There are, for example, hundreds
ia
handlings for a specific activity for only a single period out of the nine years
19
covered by Bradley’s dataset.
20
operational
of instances in which a site reports piece
See Table 2, below.
practices that would so frequently
7.
LR-H-220,
a.
Tr. 1115389.
9.
LR-H-147.
Pieces
p. 8.
-16-
It is difficult to imagine actual
bring an activity to life for only a
period.
Data entry errors, such as recording
1
single accounting
2
under the wrong activity or with the wrong facility identifier, would s;eem to provide
3
a more plausible explanation.
Table 2
4
5
a
Number of Instances in Which Piece Handlings;
Are Reported at a MODS Facility for a Direct
Activity for Only a Single Accounting
7
a
9
10
11
12
13
14
15
18
17
ia
19
20
21
22
23
24
25
piece handlings
I
Activity
Period
1 Sing la Period Observations
MODS Direct
59
BCS Sorting
OCR Sortin?
LSM Sorting I
FSM Sorting I
^ ting
Manual Letter SOI
.”
Manual Flat Sortir,U
Manual Parcel Sorb ng
Cnrtinn
Manual Priority Mail I vyI
.,. ,LI
SPBS - Priority Mail Sorting
SPBS - Non Priority Mail
Sorting I
m:ancellation and Mail Prep
Fatal
I
An
3
82
43
71
88
87
Al
7,
I
I
.-
12
23
I
--I
I
83
549
I
zl
Source: Wp II.
There are also numerous reporting gaps in Bradley’s datasets.
Oflen an activity will ‘disappear”
See
at a site for a single accounting
28
Table 3, below.
27
period or for a number of accounting
28
possible, of course, for an activity to be temporarily
29
number of instances
30
the data simply not to make their way into the MODS system.
periods, only to reappear
shut down.
at al later date.
It is
However, the large
in which this occurs suggests that it may also be common for
-17-
Table 3
Number of Instances in Which Gaps in Piece Handlings
Reported at a Facility Appear for MODS Direct Activities
25
Source: WP II.
28
Econometric
27
well-established
28
variable causes downward
29
clearly in a recent text:
30
31
studies are especially
econometric
sensitive to data errors.
principle that measurement
bias in coefficient
estimates.
error in a;n independent
This resullt is stated
As long as 02, [the variance of the measurement error in
the independent variable] is positive, b [its estimated
-18-
It is a
1
2
3
coefficient] is inconsistent, with a persistent bias toward
zero. . . . The effect of biasing the coefficient toward
zero is called attenuation.”
4
Bradley acknowledges
5
may be subject to measurement
the possibility
that his data on piece handlings
error:
When using operating data, there is always a concern
that the data might contain measurement error. If the
measurement error is in the dependent variable, hours,
it will simply be part of the specified error term in the
econometric regressions.
If the measurement error is in
the right-hand-side variables, however, traditional leastsquares methods will not accurately account for it. This
is called the “errors-in-variables”
problem.”
a
7
a
9
10
11
12
13
14
Although
Bradley correctly cites the existence of the errors-in-variables
15
fails to describe the full nature of the attenuation
ia
Bradley’s analysis, the attenuation
17
in the piece handlings
18
variability
effect.
In the specific context of
effect means that if there is measurement
variable, Bradley’s analysis will understate
of mail processing
problem, he
error
the true volume
labor costs.
The pattern of results reported by Bradley suggests that his results
19
20
may have been powerfully
influenced
21
handlings.
22
variability
23
type of mail. In the automated
by errors in the measurement
There are a number of instances
of piece
in which Bradley reports volume
estimates for both the manual and the automated
sorting of the same
sorting activities, the data on piece handlings
10.
Greene, William H.. Econometric
1997, p. 437.
11.
USPS-T-14,
Analysis. Third Edition, Prentice Hall,
p. 80.
-19-
come
from counters on the machinery.
In contrast, for manual activities, piece counts are
derived indirectly by applying conversion
volume, or some other proxy.
measurement
One would expect, therefore,
error and stronger attenuation
Bradley’s results confirm this expectation.
variability
factors to measures of weight, cubic
estimates for the mechanized
to find greater
effects in the manual .activities.
Figure 1 shows Bradley’s volume
and manual sorting of letters, flats, and
Priority Mail. In every case the volume variability
manual activities, sometimes by a substantial
-20.
estimates are lovver for the
margin.
1
2
3
Figure 1
Automatic vs. Manual Volume Variabilities
AutomatItVI. ManualVolumeVul~bllltlsc
-21-
In his direct testimony
1
Bradley presents the results of an analysis
2
that, he claims, quantifies
the effects of measurement
error in his piece handlings
3
variable.
4
analysis, but he does use it to support an argument that ‘measurernent
5
manual letter and flat piece handling volumes is not a critical problem for the
a
estimation
7
this analysis that call into question its ability to support these claims.
8
claims to have found upward bias in his estimate of the volume variability
9
manual letter sorting activity rather than the downward
He does not derive his final estimates of volume variability from this
of cost elasticities
for those activities.“”
error in
However, there are problems in
Bradley
of the
bias that Greene states is
10
the result of measurement
error. As shown in Appendix A to my te:stimony, the
11
formulas that Bradley himself presents in his direct testimony show clearly that
12
upward bias is a mathematically
13
is therefore a sign of serious and fundamental
impossible
result.
flaws in his analysis.
It is clear that the effect of measurement
14
Bradley’s finding of upward bias
error in the piecie handlings
15
variable is to cause Bradley to understate
ia
processing
17
this bias remains unknown.
18
claim that measurement
19
this analysis and the claim based upon it should be ignored.
12.
labor costs.
USPS-T-14,
the true volume variability
of mail
Despite Bradley’s attempt to quantify it, the magnitude
of
The analysis that Bradley puts forward to support his
error is not a critical problem is clearly unreliable.
pp. 83-84.
-22-
Both
1
2. Bradley’s Data ‘Scrubbing”
Procedures
Bradley’s volume variability
2
Have Substantively
A/tared His Results
estimates are derived from a dataset that is
3
the end product of an extensive editing process in which enormous
4
data are eliminated from his analysis based solely on subjective criteria.
5
volume variability
6
different from those derived from the initial dataset, calling into question the
7
reliability of Bradley’s estimates of volume variability.
9
10
The
estimates derived from this reduced dataset are substantially
In his direct testimony
8
amounts of
Bradley describes the elaborate,
process through which he discarded
multi-step
what he regarded as questionable
data.
These steps included:
(a)
11
Elimination
of observations
corresponding
to periods in which the site is
just starting the activity, and in which volumes are thus still ramping UP;‘~
12
13
@I
Elimination
observations
14
of all sites having less than 39 consecutive
usable
in the activity under examination;‘4
13.
This description is based upon statements contained in Bradley’s direct
testimony, USPS-T-14, p. 30, lines 22-25. Since the filing of that testimony,
his rationale for this particular step in the ‘scrubbing’ process has changed
somewhat. Examination of the computer programs used to do the
“scrubbing’ had indicated that this step in the process had eliminated not
just observations corresponding to the first periods in which- an activity‘was
present at a facility, but also long runs of observations in the middle of the
reporting periods for some established sites. When asked about this under
cross-examination,
Bradley indicated that his “general intent was to
eliminate from the data any period in which the level of activity fell below this
minimum, normal operating activity.” Tr. 1115571.
14.
For the allied activities he required 26 consecutive
1115475.
-23-
observations.
Tr.
If a site has more than one run of 39 or more consecutive
usable
2
observations
of all runs for
3
that site but the most recent;15
04
1
4
(‘4
Elimination
in the activity under examination,
of observations
corresponding
elimination
to high or low productivities
in
the activity under examination;
5
6
(e)
If the elimination
of high or low productivity
7
having less than 39 consecutive
8
previous steps in the process, elimination
9
and
0
10
If the elimination
observations
observations
of high or low productivity
results in a site
that have survived all
of the site from the analysis;
observations
11
having more than one run of 39 consecutive
12
survived all previous steps in the process, elimination
13
site but the most recent.
14
This process eliminates
results in a site
observationis that have
of all runs for that
an enormous amount of otherwise
usable data.
15
Details for the 23 activities Bradley ‘scrubs” are presented
16
of the 23 activities, he discards over ten percent of the data. In seven cases he
17
discards over 20 percent of the data. In two cases he discards ovlsr 30 percent of
18
the data, and in one case -- SPBS Priority -- he throws away a staaggering 49
19
percent of the potentially
20
50,000 observations.
15.
in TablIe 4, below.
In 21
usable data. Across all of his models he discards over
For the allied activities he required 26 consecutive
Ill5475
-24-
observations.
Tr.
1
Table 4
Data Eliminated Due to Data “Scrubbing”
2
Usable
ObSeNPtiOnS
3
Activity
sorting
SPBS - Priority Mail
Sorting
SPBS - Non Priority Mail
Sorting
Cancellation and Mail
Prep
Opening - Pref Mail
Opening - Bulk Business
Mail
Pouching
Platform
Plalform
FloorLabor
3.903
1,906
6,775
3353
26,260
6,470
I
I
19,634
17,660
-
38
Sources:
39
[I] LR-H-148-7, Table H148-1
[2] TR. 1115447
[3] Tr. 1l/5448
40
41
42
43
Observations
Discarded
PI = [41- 151
11 = 15lJ41
-25-
Obsarvaths
Remainina
Percent
Discarded
The extreme lengths to which Bradley takes his ‘scrubbing”
1
questions.
process raise
2
a number of significant
How can one be sure that the clata that are
3
‘scrubbed”
4
with volume? How justifiable
5
which observations
6
Bradley’s results been affected by his decisions to discard large plsrtions of the
7
data? And, of course, if the datasets Bradley is using are so dirty as to require such
8
extreme clean-up measures,
9
from them?
away contain errors and not important information
about how costs vary
are the subjective criteria Bradley used in deciding
to discard and which to retain? How have the character
how reliable can the conclusions
be that are derived
How can one be sure that the data that are “scrubbed”
10
of
a:way contain
11
errors and not important information
about how costs vary with volume?
One
12
cannot.
13
independent
14
‘scrubs”
15
the dataset.
16
attributed
17
the site was quite normal, but that the data were recorded
18
cases, the elimination
19
possible that the data were in fact recorded correctly but look unusual even though
20
they are normal for that site. In the latter case the elimination
21
harder to justii.
22
apply to all facilities, so they should reflect the variability
23
both “usual’ and ‘unusual.’
Bradley cites no external evidence that could be used to provide
verification
of the accuracy or inaccuracy
eliminate observations
that look ‘unusual”
of any of his data.
His
relative to other observations
in
The fact that they may look unusual to a particular observer can be
to either of two causes.
It is possible that what was in fact going on at
of observations
incorrectly.
might be appropriate.
In such
Homsver, it is also
of observations
is
The rates that will be set as a result of these prosceedings will
-26-
of costs at all facilities,
It is very possible that such “unusual”
information
about the true relationship
experienced
an enormous
observations
contain the most
between cost and volume.
A site that has
increase in volume may well be unusual, but it may also
provide the clearest possible picture of how processing
costs vary with volume.
It
is for this very reason that Dennis Cook, an authority on the analysis of outliers,
has warned against discarding
disproportionate
an observation
simply because it exerts a
influence over the estimated regression
results:
8
9
10
. . influential cases are not necessarily undesirable.
Often, in fact, they can provide more important information
than most other cases.16
11
Bradley’s decision to eliminate observations
12
handlings
13
Examination
14
extended periods of time. See, for example, Figure 2, below.
15
between piece handlings
16
relationship
17
systematically
18
that his results are applicable
19
observations
16.
also raises questions
involving low levels of piece
about the representativeness
of his results.
of the data reveals sites that exhibit low levels of piece handlings
appropriate
The relationship
and hours at such sites is part of the normal overall
between piece handlings
eliminated
over
and hours.
If these observations
have been
from Bradley’s analysis, we have no reason to believe
to their circumstances.
To give such sites and
weight, one needs to include their data in the analysis.
R. Dennis Cook and Sanford Weisberg, Residuals
Reqression, Chapman and Hall (1982) p. 104.
-27-
and Influence in
6601
8805
8808
8813
SD04
8908
6812
BOO3
BOOT
90-l
1
D-IO2
9106
9110
920
-I
9205
8209
9213
8304
9308
B312
9403
9407
94-9
9602
9606
8510
860
-606
Be-lo
1
Bradley’s decision to include only sites that have 39 periods of
1
2
continuous
data appears to be especially
3
Bradley’s models or analysis that mandates such a requirement.
4
include piece handlings for both the current accounting
5
accounting
6
of his fixed effects model, it is necessary
7
two consecutive
8
estimation
9
complete dat.a be available for three consecutive
period.
There is nothing about
IBradley’s models
period and the preceding
Hence, in order for a data point to be included in the estimation
accounting
periods.
only that complete data be available for
In order for a data point to be included in the
of his fixed effects model with serial correlation,
10
programs or calculations
11
periods.
The requirement
12
arbitrary.
inherently
it is necessary
accounting
perio’ds.
requires data for 39 consecutive
None of his
accounting
that a site have complete data for 39 consecutive
13
accounting
periods accounts for the largest portion of the observations
14
discarded
15
a rigorous defense of his decision to discard observations
16
Discussing
as a result of Bradley’s ‘scrubbing”
his continuity
only that
requirement
procedures.
that are
Bradley fails to present
with sulch abandon.
in his direct testimony,
Bradlley simply states:
17
18
19
20
21
The time dimension is an important part of the nature of
panel data and if possible, it is preferable to have
continuous data. Continuous data facilitate the estimation
of accurate seasonal effects, secular non-volume trends,
and serial correlation corrections.”
22
Nothing in this statement refers to his specific decision to require three
23
years of consecutive
17.
USPS-T-14,
data, and the assertions
p. 31.
-29-
which it does conta’in are not entirely
1
correct.
Estimation of accurate seasonal effects requires simply that the dataset
2
contain adequate
3
periods, a requirement
4
consecutively.
5
only that the dataset provide adequate
6
overall sample period.
numbers of usable observations
in each of the dlifferent seasonal
that is easily met. It is not necessary
Similarly, estimation of accurate secular non-volume
Again, it is not necessary
8
consecutive
observations
9
one period to prediction
that they occur consecutively.
corrections
does require
in order to make it possible to relate prediction
errors in the prior period.
are needed to contribute
observations
11
correlation
12
model with serial correlation
correction,
13
required.
requires 39 consecutive
To contribute
Neither estimation
errors in
However, only 1~ consecutive
10
coefficient.
trends requires
coverage of all of the dates within the
Estimation of accurate serial correlation
7
that they occur
to the estimation
to the estimation
of the Bakagi-Li
serial
of the final fixed effects
three consecutive
observations
are
observations.
Bradley’s decision to discard huge volumes of data has had a substantial
14
15
effect on his results.
To illustrate this point, I reran his ewnometriic
analyses on
16
the full dataset using all of the observations
17
available.
18
compares the estimated volume variabilities
19
results differ sharply from those presented
20
Using the full dataset produces volume variabilities
21
reported by Bradley.
22
volume variability for the MODS OCR sorting activity from 79 percent to 83 percent.
23
The estimated variability for MODS LSM sorting increases from 9’1 percent to
for which complete dalta were
The results of this exercise are shown in Table 5 belo\nr, which also
to those reported by 13radley. The
by Bradley in his direct testimony.
that are often higher than those
For example, using the full dataset raises thle estimated
-30.
1
98 percent.
For MODS bar code sorting, variability
2
108 percent, indicating the presence of disewnomies
3
Dramatic changes occur for most activities.
-31-
increases from1 95 percent to
of scale in this activity.
1
Table 5
2
3
Effects of Discarding Usable Observations on Bradley’s
Estimates of the Volume Variability of Mail Processing Labor Costs
4
5
6
7
8
9
IO
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
Bradley’s
“Scrubbed” Data [1]
Activity
F
Observations
[2]
Estimated Volume Variability of Mail Processing Labor Costs
MODS Direct
95%
I
70”’
102%
84%
90%
32%
42%
73%
36%
53%
79%
108%
81%
76%
03’,”
I”,
.I,
I1.1”“,
1.1
I” %
3
3acK sorring rdacnine
Primary Parcel Sorting Machine
Seconda
Parcel .W+inn w,,+inm
Irregular Parcel Poe ,t
1 unit“
Non MaClWIable Outsides
Source:
34
35
[l] USPS-T-14,
[2] WP IV
:i -70
a7
G%
72%
67%
I
BMC Allied
53%
I
Platform
Floor Labor
33
I
60%
p. 9
-32-
77%
62%
92%
61%
31%
81%
One should always be suspicious
1
alter the conclusions
of decisions
to discard data when
2
those decisions
of the analysis in substantively
important
3
ways.
4
reproducibility
5
possible for someone to independently
6
conclusions.
7
computer programs to produce the same set of computer outputs.
6
able to follow each of the steps of the analysis and verify that each of the specific
9
actions taken is an appropriate
One of the key elements of the scientific method is its emphasis
of results.
Independent
on the
If Bradley has done his work correctly, it should be
replication
replicate his analysis and arrive at the same
is more than simply rerunning
Bradley’s
It means being
response to the problems encountered.
10
has chosen to require 39 consecutive
11
analysis of MODS direct activities.
12
consecutive
13
such specific judgment calls regarding which data points to include and which to
14
discard do not pass this test. From an economic and policy perspective,
15
to discard data whose implications
are this significant
16
objective empirical and conceptual
justification
16.
observations
observations
Bradley
in order to include a site in his
Yet, Bradley himself has indicated that 26
may be sufficient. ”
Tr. 1115450.
-33-
Results that depen,d strongly on
decisions
require greater and more
than Bradley has provided.
1
3. Bradley’s Results Imply Implausible
Patterns of Technological
Bradley has included in his cost equations
2
Change
a number of time trend
3
variables
intended to account for the effects on productivity
of changes in mail
4
sorting technology.
5
his models is rather complex.
6
two. Both of these trend variables are mean-centered,
7
variables.
Both are entered into the equations
8
Interaction
terms between the time trends and the other independent
9
the model are also included.
The manner in which Bradley has introduced
First, he introduces
The specification
time trends into
not one time trend variable,
along with his other
in both linear and squared forms.
variables
that Bradley uses for the MODS
10
direct activities includes eight estimated coefficients
11
trend variables.‘g
for terms involving the time
12
In his direct testimony Bradley explains their presence
13
14
15
16
17
18
Thus, in my equations, the time trend’s coefficient measures
the rate of growth (or decline) in hours not attributable to
increases (or decreases) in piece-handlings.
A trend
approach is particularly well suited for looking at mail
processing labor costs because changes in technology
generate smooth changes in mail processing productivity.20
19
This statement accurately
trend variables to capture the effects of technological
21
driven by the gradual accumulation
22
and by their diffusion into practical application.
This specification
20.
USPS-T-14.
is presented
as follows:
describes the rationale for the use of time
20
19.
of knowledge
change.
Such change is
and improvement
in technique
The use of time trends for the
as equation (2) on page 36 of USPS-T-14.
-34-
but
in
1
econometric
measurement
2
the expectation
3
manifested
of the effects of technological
change is motivated by
(shared by Bradley) that the effects of such change would be
in a steady and gradual improvement
in productivity.
In Bradley’s analysis we have even more reason to expect the effects of
4
5
technological
change to be manifested
in smooth and steady changes
6
productivity.
Bradley’s activity-based
7
isolation.
8
we will never observe in his results the discontinuous
9
result from changes such as the move from manual to automated
analysis examines each sorting technology
Because he examines manual sorting and automated
10
we should see the effects of gradual refinements
11
improvement
12
reasons to expect the effects of technological
13
in a steady and gradual improvement
14
trend results can best be described
16
and steady change.
17
their effects are hard to summarize.
18
jumps in prolductivity that can
sorting.
Instead.
and
We thus have strong
change in Bradley’s analysis to result
in productivity.
However, the pattern of technological
change implied by Bradley’s time
as one of lively variation rather than smooth
His models contain so many terms involving time trends that
Figures 3 and 4 show results for the manual letter sortin!g and manual
19
parcel sorting activities.
20
covered by Bradley’s analysis.
21
trend variables
In each case, the horizontal
axis shows the time period
The vertical axis shows the products of the time
and their estimated coefficients.
-35-
in
sorting separately,
of the equipment
in workers’ familiarity with its operation.
15
in
Figure 3
Effects of Time Trend Variables in Bradley’s
Manual Letter Sorting Model
-.._
Fiscal Year Accounting Period
4
5O"rcKWP"I.
-36-
Figure 4
Effects of Time Trend Variables in Bradley‘s
Manual Parcel Sorting Model
4
60UnB:WPVI.
-3l-
The patterns revealed in these two manual activities are complex, and
1
2
strikingly different.
3
declining
4
and discontinuous
5
rise for a time, but eventually
6
manual parcel sorting we see a period of rising productivity,
7
falling productivity,
8
of falling productivity.
9
In manual letter sorting we see an initial period of sharply
productivity.
Then, suddenly,
jump in productivity.
in the beginning
From that point on productivity
it levels off and then begins to decline.
followed by a period of rising productivity,
The results presented
followed
continues
patterns that appear in Bradley’s time trend results.
11
technological
12
positive direction.
13
gaining in efficiency, while others show losses.
14
and declines and rapid alterations
15
productivity.
Again, in
by a period of
followed
by a period
progress is neither gradual nor smooth.
In Bradley’s world,
Nor does it always move in a
At the same points in time one finds some activities rapidly
The magnitudes
One also finds discontinuous
between increases and decreases
of these shifts are not trivial. The vertic:al axes in
Figures 3 and 4 are measured
18
the activity.
19
efficiency point and the highest efficiency point amounts to approximately
20
percent of labor hours.
21
approximately
23
jumps
in
17
22
in terms of percentages
of average labor hours in
In the case of manual parcel sorting, the gap between the lowest
Ii’
In the case of manual letter sorting it amounts to
13 percent.
The pattern of Bradley’s time trend results causes me to, question his
interpretation
to
in Figures 3 and 4 are indicative of the general
10
16
of 1993 there is a large
of their meaning.
Although
he has alluded to a “major restructuring”
-38.
that took place in FY 1993, he has not explained
change.
Moreover,
the nature or significance
he has provided little substantive
explanation
of this
of what changes
could account for such erratic patterns within the periods prior to and following the
restructuring.
I am hard pressed to envision a pattern of technologiical
would produce such variations
time trend coefficients
9
and hence I do not believe that his
are really picking up the effects of technological
C. Bradley’s
Lono-Run
7
8
in productivity,
change that
progress.
Analysis Sheds Little Light on the
Volume Variatv
of Costs
The fixed effects models that Bradley relies upon for his variability
10
estimates do not appear to be capable of providing
11
run variability
12
labor hours in a four-week
13
that same period and in the previous period.
14
a single accounting
15
changes that take place over longer periods of time. Their short-run view of labor
16
cost variability
17
of mail processing
reliable estimates of the long-
labor costs. Those models relate mail processing
accounting
period to the number of piece handlings
Because these models look back only
period, they are not capable of detecting or accounting
calls into question their relevance
The extent to which mail processing
in
for
to this proceeding.
labor costs vary with volume will
18
depend upon the time horizon over which volumes and costs chanmge. It is possible
19
that productivity
20
Workers might increase the pace of work, take fewer or shorter breaks, or adopt
21
other strategies for dealing with the added workload.
might increase in response to a temporary
-39-
surge in volume.
In his responses
to
1
interrogatories,
2
not be sustainable,
3
eventually
be necessary
4
workload.
Thus, after an initial surge it is likely that productivity
5
something
closer to its original level. Over an even longer period of time,
6
increases
in volume could facilitate greater reliance on specialization
7
and thereby lead to higher productivities.
8
volume variability
9
long a time is allowed for costs to respond to changes in volume.
10
Bradley concedes
this point.2’
Such increases
in productivity
may
however, and if the increase in volume persists it may
to hire additional
of mail processing
In past proceedings,
workers to handle the inlcreased
would decline to
or automation
Thus, the estimate that one gets for the
labor costs may differ, depending
the Commission
has relied upon evidence of the
11
long-run volume variability
of costs in its findings regarding
12
“Long-run,”
13
periods longer than a year.=
14
Bradley’s fixed effects models falls well short of this threshold.
in this context, has been interpreted
upon how
the attribution
of costs.
as changes that olccur over
The eight week adjustment period provided for in
BRADLEY’S CROSS-SECTIONAL
MODEL
PROVIDES SUPFRIOR RFSULTS
15
16
In his workpapers,
17
18
the ‘between”
19
“between”
Bradley presents regression
model for each of the mail processing
models are essentially
cross-sectional
activities he examines.
These
versions of the s;pecification he
21.
Tr. 1115512.
22.
. .
National AssocraWn of Greeting Card Publj&~
Service, 462 U.S. 810,815-16
(1983).
40-
results for what he calls
v. United St ates Postal
1
uses in his fixed effects analysis.
2
these cross-sectional
The volume variability
models are presented
estimates produced
by
in Table 6.
Table 6
Volume Variability Estimates from
Bradley’s “Between” Model
4
5
7
8
9
10
11
12
13
14
15
18
17
18
19
20
21
22
23
24
Volume~~ariability
Activity
6
MODS Direct
ppc
@orting
Manual Letter Sorting
Manual Flat Sorting
Cancellation
il Sorting
and Mail Prep
I
MODS Allied
Opening - Pref Mail
Openint 3 - Bulk Business Mail
Pour :hinQ
Platform
Source: USPS-T-14
Wp I and II
The results of Bradley’s cross-sectional
25
26
superior basis for estimating
27
They are more likely to show long-run effects, and they are less sensitive to data
28
quality problems and to judgments
29
30
Adoption
short-term
increases
the volume variability
analysis appear to provide a
labor costs.
about how to ‘scrub” the data.
of a cross-sectional
and decreases
of mail processing
approach
de-emphasizes
‘the effects of
in volume, focusing the analysis instead on the
-41-
]
1
long-run effects of changes
2
emphasizes
3
of mail they process and that have had the chance to adjust fully to those
4
systematic
5
levels, degree of automation,
6
factors affecting the volume variability
7
the volumes typically processed
A cross-sectional
approach
the contrast between facilities that differ systematically
differences.
indeed, one would expect decisions
layout of the processing
Under cross-examination
8
9
in mail volumes.
transportation
of processing
in the volume
regarding
staffing
flows, and other significant
costs to be closely related to
at a facility.
on his testimony
costs, Bradley readily conceded
regarding
purchased
the ability of cross-sectional
10
analysis to yield simpler and more direct estimates of the volume variability
11
costs:
of
And one of the advantages of a cross-sectional analysis is it
allows me to estimate how quickly cost rises or falls with
increases or decreases in volume without the necessity of
tracing the size of total accrued cost through time.=
12
13
14
15
16
Later he conceded
that cross-sectional
analysis is not just an appr,opriate way to
17
address these issues, but also perhaps the more common approach:
However, another approach which is widely used, perhaps
even more so for the type of analysis I’m doing, is to, for a
cross-section of facilities or in this case contracts, collect
the information on costs and cubic-foot miles and use the
variation between the small contracts, that is small cubic-
18
19
20
21
22
23.
Tr. 713809.
-42-
foot miles, and the large cubic-foot miles, to measure how it
is costs vary with cubic-foot miles.24
1
2
3
He concluded
4
Y . I believe it’s an entirely valid methodology
5
section of data and to analyze
6
with a strong affirmation
of the relevance
of cross-sectional
to collect a cross-sectional
transportation
testimony.
over and above those cited
7
in his purchased
8
attenuation
9
constructs
the observations
10
averaging
across time periods the values for the dependent
11
variables
12
the inputs to his cross-sectional
13
instances of over-estimation
14
cancel out, with the result that net measurement
15
will become proportionately
due to errors-in-variables
Cross-section
results are less subject to
bias than his fixed effects model results.
that enter into his cross-sectional
with each facility.
He
analysis by
and independent
The averages created in thiis way become
regression
model.
and under-estimation
In this averaging
of piece handliqs
process,
will tend to
error in the independent
variables
less significant.
The volume variabilities
16
cross-
. how it is costs increase with cub:ic-foot miles.“25
Bradley’s between model offers advantages
associated
analysis:
implied by the cross-sectional
models are often
17
higher than those reported by Bradley and are generally very close to 100 percent
18
(or greater than 100 percent, implying diseconomies
19
between the cross-sectional
20
the fact that the cross-sectional
21
variabilities
results and the fixed effects results
Tr. 713813-l 4.
25.
Tr. 713814.
can be
attributed
results are closer to the long-run volume
and are less subject to attenuation
24.
of sale). The differences
-43-
effects caused by measurement
to
error in the piece-handlings
variables.
results provide a more appropriate
For these reasons, the cross-sectional
basis for the attribution
of mail Iprocessing labor
costs.
4
5
BRADLEY HAS NOT DEMONSTRATED
THAT VOLtJME
VARIABII D-Y IS LESS THAN 100 PFRCFNT
Bradley’s analysis is not sufficiently
6
7
Commission’s
established
determination
8
percent volume variable.
9
address the right question,
reliable to justii
that mail processing
rejection of the
labolr costs are 100
Taken at their plain face value, these analyses
do not
nor do they use the right variables.
Even if one were to set aside concerns about the appropriateness
10
of
11
Bradley’s measures of cost and volume, ample reasons for skepticism
12
Many aspects of his data, his sample selection criteria, his model, and his results
13
seem questionable.
14
in approach,
15
different results.
16
estimates understate
His econometric
results do not appear to be stable.
or even changes in data “scrubbing”
Changes
criteria, lead to substantially
Moreover, there is good reason to believe that his econometric
the true variabilities.
A RELIABLE ECONOMETRIC
OF VOLUMF VARlARlt
17
18
The points discussed
19
remain.
20
with Bradley’s econometric
21
to draw any conclusions
22
costs.
Nonetheless,
ESTIMATE
D-Y?
above indicate the presence of selrious problems
results.
regarding
it is necessary
The Commission
should not use those results
the volume variability
of mail processing
at times to make policy decisions
-44
labor
even in the
absence of definitive empirical information.
question:
If the Commission
variability
that is achievable
This fact logically raises an important
is seeking the best econometric
estimlate of volume
given the available data and resources,
to which set of
results should they turn? This section of my testimony attempts to answer that
question.
A. Bradley’s Cross-Sectional
An Aoorooriate
[email protected]
6
7
As I have stated, the most accurate and reliable estimates of the volume
8
9
Model Provides
Point
variability
of mail processing
labor costs to be found in Bradley’s results are those
10
provided by the cross-sectional
analysis embodied
11
reasons for this belief are set forth above, but are worth repeating
12
important, the cross-sectional
13
estimates of long-run volume variability.
14
part of the cross-sectional
15
deleterious
16
piece handlings
approach
is inherently
here. Most
across time that is
error problems in the MODS and PIRS
data.
If the Commission
19
My
process mitigates some of the
B. Some Modifications
of Bradley’s
Model Are Aooropriate
17
18
models.
better able to provide
Also, the averaging
model development
effects of the measurement
in his “between’
20
processing
21
results would be in order.
22
interaction
Cross-Sectional
-
were to rely upon an econometric
labor cost variability,
two modifications
I recommend
terms from the model.
ana!lysis of mail
to Bradley’s “between”
dropping the time trend variables
In Bradley’s fixed effects analysis, these
-45-
model
and
variables
give rise to some curious coefficient
whether they are in fact capturing
the effects of technological
sectional context, they have little meaning.
Bradley’s “between’
approach
estimates that raise questions
change.
about
In the cross-
Also, it is clearly necessary
to extend
to all of his activities.
5
C. Data “Scrubbina”
6
If the Commission
decides to rely upon an econometric
approach
to the
it will have to decide how much ‘scrubbing”
of the
7
estimation
8
data is appropriate.
9
a clear and objective basis for deciding which data to use and whic:h data to
of volume variability,
In the absence of any external validity checks, it is hard to find
10
discard.
For this reason, as described
11
with all of the “scrubbing’
is to dispense
and run the analyses on the full set of dalta.
Bradley’s ‘threshold”
12
above, the best approach
scrub seemqdesigned
systematically
to eliminate
13
certain types of facilities or time periods from his analysis.
14
and time periods do exist and their effect on costs must be dealt with in the rate
15
setting process, I see no justification
for eliminating
16
eliminating
that the results will not be representative
17
cost-volume
relationship
18
recommend
dropping the threshold
19
recommend
against adoption of Bradley’s ‘productivity”
20
eliminates
21
sample.
22
have no clear justification
them almost guarantees
observations
Since these facilities
them from the analysis.
that actually exists over all postal facilities.
“scrub.”
of the
I therefore
For similar reasons, I would
scrub.
and reduces the representativeness
Finally, I would also recommend
In fact,
It systematically
of Bradley’s analysis
dropping the continuity “scrubs.’
They
in the context of the fixed effects model ;and even less in
the context of the cross-sectional
large numbers of observations
Econometric
analysis.
They also result in the elimination
of
from the analysis.
estimates of the volume variability
costs that are based upon these recommendations
of mail plrocessing labor
are presented
on page 7 above,
in Table 1.
6
7
8
D. Implementation
of These Changes Still Leaw?s
a Model That Fails to Consider Either Actual
Costs
-
9
Implementation
of all of the recommendations
presented
above will still
10
result in a model that (1) fails to include either a true measure of cost or a true
11
measure of volume and (2) relies heavily on a data source of uncertain and
12
unproven accuracy.
13
long-established
14
variable.
As a result, I recommend
determination
that the Commission
that mail processing
stand by its
labor costs are 100 percent
CONCLUSION
15
Bradley’s econometric
16
17
that labor is inherently
18
to reject its traditional
19
variable in favor of an alternative
20
mail processing
21
sweeping
analysis contradicts
a highly variable cost factor.
treatment of mail processing
the wmmon
sense conclusion
Bradley urges the Commission
labor costs as 100 percent
approach that would move significant
labor costs into the institutional
departure from precedent
cost category.
bears a substantial
-4-I-
portions of
A call for such a
burden to support the
1
assertion that such a departure
2
burden.
is in fact warranted.
Bradley has not met this
At the most basic level, Bradley has failed to address the questions
3
4
are really relevant to this proceeding.
5
either of cost or of volume.
6
accurately
His analysis fails to include a true measure
7
short-term view of volume variability;
8
adjustment
9
reflect only the changes that take place over that brief period.
He cannot, therefore,
the volume variability
that
of costs.
claim to have measured
His model also adopts an excessively
it is constructed
to allow for ai complete
of costs to a change in volume in two accounting
periods and can
The relevance
of
10
such short-run effects to the process of setting rates that will remain in effect for
11
years is highly questionable.
Moreover,
12
there are a number of problems with the implementation
13
Bradley’s econometric
14
accuracy.
15
highly subjective judgments.
16
character
17
equally defensible
18
volume variability
19
analyses.
of
He has relied heavily on data of uncertain
He has discarded from his analysis huge volumes of data on the basis of
These exclusions
of his empirical results.
have substantially
altered the
His empirical results do not appear to be robust;
changes in approach
of mail processing
lead to quite different estimates of the
labor costs.
For all of these reasons, Bradley’s results should be rejected.
reason for the Commission
20
offered a compelling
21
mail processing
labor costs.
In fact, a more appropriate
cross-sectional
22
suggests that the traditional,
common sense conclusion
that mail processing
23
costs are 100 percent volume variable is correct.
-4%
to reject its traditional
He has not
approach
to
analysis
labor
APPENDIX
1
A
The essence of Bradley’s errors-in-variables
2
of two alternative
analysis involves the
3
comparison
4
differing degrees of attenuation.
5
claims, allows him both to quantify the degree of measurement
6
handlings
7
showing the magnitudes
8
are shown in equations
9
testimony
and to determine
(USPS-T-14).
estimators of volume variability
The comparison
of the two estimates,
of the attenuation
Combining
these two expressions,
volume variability.
12
replacing the probability
13
values, combining
14
volume variability.
15
accuracy of equations
cited by Greene.
estimators
(19) and (21) on pages 81 and 82 of Bradley’s direct
11
17
The formulas
effects for the two alternative
formula shown in equation (22) on page 82 of USPS-T-14
Bradley derives the
for the true value of
Equation (22) is derived from equations
(19) and (21) by
limits shown there with the two estimated coefficient
the two resulting expressions,
Equation (22) therefore,
Both equation
Bradley
error in piece
the true value of volume variability.
10
16
that are subject to
and solving for the true value of
is derived from and depends upon the
(19) and (21).
(19) and equation
(21) demonstrate
the attenuation
Equation (19) takes the following form:
phlPr
=p
I
(T-VJ$
1-
TVar(x,t
A-l
- j?)
effects
’ d is the estimate of volume variability
where S
estimator,
g is the true volume variability,
generated
T is the number of time periods in the
dataset. c?~ is the variance of the measurement
variable, x*~ is the recorded
5
by the fixed effects
error in the piece-,handlings
number of piece handlings for site i in period t, and
;T,’ is the average across all time periods of the recorded piece handlings for site i.
This equation shows that the estimate of volume variability
6
provided by
7
the fixed effects estimator will be equal to the true volume variability
8
term that is equal to one minus the ratio of two variances.
9
positive, the ratio of the two variances will also be positive.
Hence, the estimated
10
volume variability will be equal to the true volume variability
multiplied by a number
11
less than one. In other words, the estimated variability will be less than the true
12
variability.
Since variances
by a
must be
Analysis of equation (21) yields a similar result. That equation takes the
13
14
multiplied
following form:
202
P/hr7pd = p l-
Var xh* - xi*t -1)
15
where p is the estimate of volume variability
16
estimator.
17
variability
generated
by the first difference
For the same reasons as those set forth above, the estimate of volume
provided by the first difference
estimator will be equal to the true volume
A-2
1
variability
multiplied by a number less than one. Again, the estimated variability
2
therefore be less than the true variability.
3
These mathematical
4
of the true volume variabilities
5
direct testimony.
6
for the true volume variability, 26 .6316 for the volume variability
7
by the fixed effects estimator,
8
produced
9
presents for the true volume variability
facts call into question the reliability of the estimates
presented
by Bradley in Table 17 on page 84 of his
For the manual letter sorting activity he presents values of .6048
by the first difference
estimate produced
and .7232 for the volume variability
estimator.
estimate
Note that the value that Bradley
is higher than either of the hvo attenuated
10
estimates.
If equations
11
conclusion
would be for the variance of the measurement
12
mathematically
(19) and (21) are correct, the only way to arrive at such a
impossible
error to be negative, a
result.
The reasons for these anomalous
13
will
14
Strictly speaking,
15
toward infinity.
16
of equation
17
available.
The source for equations
18
testimony
bases the derivation
19
assumptions
results are not completely
clear.
Bradley’s equation (22) holds only as his sample size grows
It is possible that these results reflect the small sample properties
(22) although,
as Bradley often points out, he has quits a large sample
(19) and (21) cited by Bradley in his direct
of these assumptions
about the structure of the data”
upon a numb’er of specific
- assumptions
that may not hold for
26.
This quantity is identified in the table as the “Errors-in-Variables
8.” Under
cross-examination
(Tr. 1 l/5575, 5576) Bradley stated that this quantity was
calculated from equation (22) and so it is identical to what I have called
here the true value of the volume variability coefficient.
27.
Cheng Hsiao, hlvsis
of Panel Da&
A-3
Cambridge
University
Press, 1986,
(continued...)
1
the data used by Bradley.
2
Bradley’s errors-in-variables
3
the manual letter sorting activity.
4
the “true” volume variability
5
error process with a negative variance.
6
Regardless
of the explanation,
analysis produces
nonsensical
however, it is clear that
results in the case of
We cannot place any credence
in an estimate of
that depends upon the existence of a measurement
This result is not limited to the manual letter sorting activity.
7
shows the results that are obtained when Bradley’s errors-in-variables
8
is applied to all of the MODS direct activities.
9
of a negative measurement
10
The mathematically
error variance appears in connection
activities.
27.(...continued)
pp. 63-65.
A4
Table A-l
methodology
impossible
result
with a number of
1
Table A-l
2
3
Results Obtained When Bradley’s Errors-in-Variables
Methodology
Is Applied to All MODS Direct Activities
4
Implied “Piece
Handiings”
Measurement
Error
“Errors-inVariables”
Activity
Variance
MODS Direct
5
6
BCS Sorting
1.0040
0.0026
7
OCR Sorting
0.9708
0.0035
8
LSM Sorting
0.8263
-0.0011
9
FSM Sorting
1.1359
0.0028
Manual Letter Sortina
0.5881
-0.0047
10
11
(Sortina
1
0.6967
I
0.5938
I
0.5539
13
Manual Parcel Sortina
I---~~
Manual Pnoritv Mail Sorting
I:-~~~~
14
SPBS - Priority Mail Sorting
0.9619
15
16
SPBS - Non Priority Mail
Sorting
0.7751
17
Cancellation
0.5470
12
18
and Mail Prep
Source: WP Ill.
A-5
1
I
I
-0.0002 1
0.0195
-0.0010
1
1