EPA/600/R-97/048
September 1997
Developing and Using Production-
Adjusted Measurements of
Pollution Prevention
by
Melissa Malkin and Jesse Baskir
Research Triangle Institute
Research Triangle Park, NC 27709
and
Timothy J, Greiner
Greiner Environmental
Gloucester, MA 01930
Cooperative Agreement No. CR 823018
Project Officer
N. Theresa Hoagland
Sustainable Technology Division
National Risk Management Research Laboratory
Cincinnati, OH 45268
National Risk Management Research Laboratory
Office of Research and Development
U.S. Environmental Protection Agency
Cincinnati, OH 45268
Printed on Recycled Paper
-------
Notice
The U S. Environmental Protection Agency through its Office of
Research and Development partially funded and collaborated in the
research described here under Cooperative Agreement No. CR
823018 to the National Risk Management Laboratory. It has been
subjected to the Agency's peer and administrative review and has
been approved for publications an EPA document.
11
-------
Foreword
The U.S. Environmental Protection Agency is charged by Congress with protecting the Nation's
The National Risk Management Research Laboratory is the Agency's center for investigation of
E. Timothy Oppelt, Director
National Risk Management Research Laboratory
in
-------
Abstract
This report describes research examining production-adjusted measures of pollution prevention (P2).
Under this research, a methodology was developed for applying statistical and graphical tools to
assess the accuracy of the factors (called units-of-product) used to adjust P2 measures. Graphical
analysis is used to qualitatively assess a unit-of-product, while regression analysis is used.to
quantitatively evaluate a unit-of-product. Researchers applied these statistical and graphical tools to
data from five case study facilities in different industrial sectors. The units-of-product currently being
used by the facilities were tested for correlation with key waste or chemical use streams It was found
that the methodology for applying statistical and graphical tools was usable with data routinely
collected at the five case study facilities. Researchers further found that the factors being used by
four of the facilities correlate with chemical usage for key input streams. This result indicates that
these factors accounted for enough of the variation in production that the factors could be used for
Curate production-adjusted P2 measurement. Data analysis from a fifth facility underlined the
challenges of obtaining data appropriate to the methodology, and conclusions were not drawn about
the unit-of-product used by that facility.
This report was submitted in fulfillment of Cooperative Agreement Number CR 823018 by Research
Triangte Institute and Greiner Environmental under the sponsorship of the United States
Environmental Protection Agency. This report covers a period from February 15,1995, to December
30,1996, and was completed as of February 11,1997.
IV
-------
Contents
Foreword
Abstract iii
Figures " ' iv
Tables '' vii
x
Executive Summary
ES.l Introduction ....." : 1
ES.2 Project Objectives .1. " ' " " 1
ES.2.1 Existing P2 Measurement Systems 3
ES.2.2 Evaluating Production-Adjusted P2 Measures '.'." 3
bb.2.3 Application of Methodology *
ES.3 Results By " ' 4
- 4
Section 1 Introduction
1.1 Background of P2 Measurement in General ' 7
1.1.1 Who Uses P2 Measurement and Why 7
1.1.2 Production-Adjusted Measures of P2: A More Detailed Look' " ' 8
Section2 Description of a Methodology for Application of Statistical and Graphical
Tools to Assess Accuracy of P2 Production-Adjusting Units 19
2.1 Evaluating a Unit-of-Product ' :^
. 2.1.1 The Unit-of-Product '.'.'.'.'.'.'.'." ' }?
2.1.2 Choosing a Unit-of-Product "7?
2.2 Analyzing the Unit-of-Product ' \*
2.2.1 Graphical Analysis . ! " '! 2
2.2.2 Statistical Analysis '.''.'.'.'.'.'.'.'.'.'.'.''.'.""" Jg
Section 3 Five Examples of Systems That Use Production-Adjusted P2 Measurement 24
3.1 Greene Manufacturing, Connorsville, Indiana ^rement ...24
31.1 Description of Facility P2 Measurement System .'.' 24
3.1.2 How the P2 Measurement System Is Used . oJ
^oC?nt Jechnologies, Merrimack Valley/Massachusetts'." " 27
3.2.1 Description of Facility P2 Measurement System., 27
a * t H?W the P2 Measurement System Is Used .. ' 78
3.3 .IBM, Burlington, Vermont f|
3.3.1 Description of Facility P2 Measurement System '.'. 29
c, A ? L H°W the P2 Measurement System Is Used . " " ' oo
3.4 Wyeth-Ayerst, Rouses Point, New York " OQ
3.4.1 Description of Facility P2 Measurement System .' 30
3.4.2 Uses of Facility P2 Measurement System . 30 '
3.5 Ervmg Paper, Erving, Massachusetts.. on
3.5.1 Description of Facility P2 Measurement System 30
-------
Contents (continued)
3.5.2 Uses for the P2 Measurement System at the Facility .. 31
4 Results Obtained by Correlating the Production-Adjusting Units Used
and Pollution or Chemical Use for the Five Case Study Sites 3Z
4.1 Greene Manufacturing Company, Inc " y>
4.1.1 Data Collection 32
4.1.2 Data Analysis 4Q
4.1.3 Findings ' 4Q
4.2 Lucent Technologies 42
4.2.1 Process 1 Analysis ._
4.2.2 Plot Time-Series and Moving Average ** .
4.2.3 Process 2 Data Analysis j
4.2.4 Findings 4§
4.3 IBM, Burlington, Vermont
4.3.1 Data Collection 49
4.3.2 Data Analysis 57
4.3.3 Findings , 59
4.4 Wyeth-Ayerst Analysis ' ~Q
4 4.1 Process Description/Prepare Process Row Chart ^
4.4.2 Identify and CollectData J*J
4.4.3 Graphical Analysis 62
4.4.4 Statistical Analysis :; 62
4.5 Results oTstatlstical and Graphical' Analysis on Data from Erving
Paper, Erving, Massachusetts ;
4.5.1 Process Description ,
4.5.2 Data Collection ^
4.5.3 Data Analysis 69
4.5.4 Findings
70
Section 5 one ^ons^. p-o^uct:o^justed P2 Measures .-.; ^ '
5.2 Methodology for Verification of Production-Adjusting Units /1
5.2.1 Assessment of Data ': 71
522 Using Chemical Use Data to Evaluate Umts-of-Product /1.
53 Units-of-Product Used by Case Study Firms £
5 3 1 Larger-Scale Production-Adjusted P2 Measurements 73
........74
Section 6 References ;
Appendices
A Selected Reports and Articles Dealing with Production-Adjusted
75
Measures of P2 '''
/if.
B Selected Statistical Resources
C Framework for Production-Adjusted Measurements of P2 '-77.
VI
-------
Figures
ES-1 Well-correlated unit-of-product relationship between waste and a related
P^ o £m -°!-Product before and after P2 improvements .. -
Plot of production and a waste that is not strongly correlated to production
No relationship can be detected . "u^uon.
ES-3 Five steps for unit-of-product analysis ...'.....'.'".'.'.'.'.'.' '\
2~l SofT^Kf °f-p"oduct rel*tionship between waste and a related
9 9 "mt;of-product before and after P2 improvements ... , o
2-2 Plot of production and a waste that is not strongly correlated to production "
No relationship can be detected "umon.
2-3 Five steps for unit-of-product analysis 13
2-4 Sulfuric acid use per ton of paper histogram
i~\ Barter P?ot showing paper produced per pound of sulfuric
2-6 Time series plot showing sulfuric acid use per ton of paper
Z-/ HtStnOTam clinYI71nrr »^~~1 Ji_.i...:i__. , ^ , . _ " ^
^
diStributi°n
use per unt-of-
2-8 Histogram showing bimodal distribution of chemical use per unit-of-''
product data
Histogram showing skewed ('exponentiao'dis'tribution'of chemical use per " " ^
2-9
2-10 Histogram showing uniform distribution of chemical'use pe^ni't-of-' " " " '' ^
product data r
2-11 Scatter plot showins relationshin K»tM,00« *«J«' 1^ ','"'' ^0
2-12 Residual plot showing random distribution of X variable residuals'.'.'.'.'.'.'.'.'.' .'22
4-1 Weekly pounds of sodium cyanide per 1,000 ft2 plated histogram (rack
W^eeklv nnnnHc nf ^in/-. ,,^^^1 1 r\r\r\ c,9. , '. ' ', '. '. 34
4-4
Weekly pounds of zinc used per i',666 ft'2 plated histogramVrack line)'
Weekly pounds of sodium cyanide used per 1,000 ft2 plated time series
Weekly pounds of zinc per 1,000 ft2 plated time series nfot
.^/"*O'f"f*/ai»" -f^l /~^4- r-iltx-vw».u^- !_* t i t: ^^ "...._
series
jjn_a
Weekly pounds of zinc per 1,000 ft2 plated time series plot
4-5 Scatter plot showing relationship between weekly pounds of sodium
4.6 Si.6^? lquare feet, Plated Orack line)
**""
35
35
square feet
36
4"? Sfne)°UndS °f S°diUm Cyanide Pei- '^e fo6t Plaied residual plot
4-8 Weekly pounds of zinc per square foot plated residual plot (rack line).'.'.'.'.'.' ]'.
Vll
-------
Figures (continued)
4-9 Monthly pounds ,of sodium cyanide per square foot plated scatter plot
(rack line) ................................... i ' ' / " i'r ' \ ....... " vt
4-10 Monthly pounds of zinc per square foot plated scatter plot (rack line) . . ...... 3 f
4- 1 1 Monthly pounds of sodium cyanide per square foot plated residual plot
(TS\C*\C lins^ ..**** ........ ....«
4-12 Monthly pounds of zinc per square foot plated residual plot (rack line) ....... 38
4-13 Monthly pounds of sodium cyanide per square foot plated histogram
(barrel line) .................................... ---- ; ,Y Y ..... " \Q
4-14 Monthly pounds of zinc per square foot plated histogram (barrel line) ........ 38
4- 1 5 Monthly pounds of sodium cyanide per square foot plated time series
plot (barrel line) .............................. .' ' ' Y ' Yu ' ' ' i i-' ' %' ao
4-16 Monthly pounds of zinc per square foot plated time series plot (barrel line) . . . &
4-17 Scatter plot showing relationship between monthly sodium cyanide use
and square foot plated (barrel line) . ....... .......... ' ' ' Y ' ........... 39
4-18 Scatter plot showing relationship between monthly zinc use and square
foot plated (barrel line) .................. "'"' ...... «\ ............. A*
4-19 Weekly glycol ether use (Ib) per substrate histogram (Process 1) ... ......... 43
4-20 Weekly glycol ether use per substrate time-series moving average plot
(Process 1) .............................. . .......... : ' ..........
4-2 1 Weekly glycol ether use per circuit time-series moving average plot
(Process 1) ................................. . ; ; ' ' ' ' ' ',' ...... ,
4-22 Monthly glycol ether use per unit-of-product time series plot (Process 1) ...-. . . 44
4-23 Monthly glycol ether use per circuit scatter plot (Process 1) ... ---- . . ....... 45
4-24 Monthly glycol ether use per substrate scatter plot (Process 1) .............. 45
4-25 Weekly glycol ether use per substrate histogram (Process 2) . . . ...... ---- 45
4-26 Glycol ether use per substrate time-series moving average plot (Process 2) ---- 40
4-27 Glycol ether use per circuit time-series moving average plot (Process 2) ...... 45
4-28 Glycol ether use versus substrates scatter plot (Process 2) ... .............. 4 /
4-29 Glycol ether use versus circuits scatter plot (Process 2) .......... ........ 4 /
4-30 Monthly IPA use per performance index unit histogram ................... j)
4-31 Monthly IP A use per million modules histogram .... .................... ^
4-32 Monthly IP A use per performance index unit time series plot .............. 5U
4-33 Monthly D?A use per million modules time series plot ........... ...... " ' «n
4-34 Monthly IPA use per performance index unit scatter.plot ---- .............. 5U
4-35 Monthly IPA use per million modules scatter plot ............ ..... ..... ->
4-36 Monthly PGMEA/cyclohexanone waste per performance index unit
1_ " * +>v\ ,." ^"^
4-37 Monthly PGMEA/cyclohexanone waste per million modules histogram ...... 52
4-38 Monthly PGMEA/cyclohexanone waste per performance index unit
time series plot ....... . ................ ..... ...........
4-39 Monthly PGMEA/cyclohexanone waste per million modules time
series plot ............................. ..... '; ''"'.'. .........
4-40 Monthly PGMEA/cyclohexanone waste per performance index unit
scatter plot .................................... ..................
Vlll
-------
Figures (continued)
4-4 1 Monthly PGMEA/cyclohexanone waste per million modules
scatter plot . . ....... ............... ...... 01
4-42 Monthly PGMEA/cyclohexanone waste per performance index unit ..........
histogram ............. ........... . <- .
Y?A ^^ PGMEA/cycl°hexanone waste per million' modules histogram ..... "54
4-44 Monthly PGMEA/cyclohexanone waste per performance index unit ......
time series plot with 1 month delay ...... 55
4-45 Monthly PGMEA/cyclohexanone waste per million' modules time ...........
series plot with 1 month delay ......... , ........... 55
4-46 Monthly PGMEA/cyclohexanone waste per performance index unit ..........
scatter plot with 1 month delay ................... 55
4-47 Monthly PGMEA/cyclohexanone waste per million modules scatter ......... ,
plot with 1 month delay ....................... 56
4-48 Solvent mixture use (kg) per kilogram of product histogram ....... ' ........ 61
4-49 Waste production (kg) per kilogram of product histogram ........... 61
4-50 Waste per product (kg) per kilogram time series plot ... ........ 61
4-51 Chemical use per product (kg) per kilogram time series plot .......... 62
4-52 Production vs chemical use scatter plot ... ...... ....... 62
4-53 Waste production (kg) per kilogram of product scatter plot ' .' ............... 62
4-54 Paper production process at Erving paper ........ ... ' 64
4-55 Daily caustic use (Ib) per ton of paper produced time series" plot ........ ' " " 64
4-56 Daily caustic use (lt>) per ton of paper produced time series plot with
Monday data removed ................. , _ «
4-57 Daily caustic use (Ib) per ton of paper produced scatter plot ............... 65
4-58 Weekly caustic use (Ib) per ton of paper produced time series plot ' .' ......... 66
4-59 Weekly caustic use (Ib) per ton of paper produced histogram " " 66
4-60 Weekly caustic use (Ib) per ton of paper produced scatter plot and
regression line ....... . .............. ....... 66
4-61 Weekly sulfuric acid use (Ib) per ton of paper produced histogram ........... 67
AK» eCHy SUlfuric acid use
-------
Tables
BS-1 How Well Units-of-Product Explained Variation in Chemical Use and
Waste Generation
2-1 Simple Linear Regression Output 21
3-1 Summary of Information about Five Case Study Sites -25
4-1 How Well Units-of-Product Explained Variation in Chemical Use and
Waste Generation ^
4-2 Glycol Ether Use per Unit-of-Product ^z
4-3 Process 1 Descriptive Statistics for Glycol Ether Use per Substrate 43
4-4 Process 2 Descriptive Statistics for Glycol Ether Use per Substrate 46
4-5 Chemical and Production Data Provided by IBM ..... 48
4-6 Results of Regression Analysis for IPA Use ^
4-7 Results of Statistical Analysis for PGMEA/Cyclohexanone Waste (Delayed) .. 56
4-8 R-Squared and P-Values for Chemical Use per Unit-of-Product -.... 57
4-9 Results of Regression Analysis for Waste and Chemical Use per
Unit-of-Product 63
x
-------
Executive Summary
ES.l Introduction
Accurate and meaningful measurement systems
are essential to the long-term success of pollu-
tion prevention (P2) in industrial settings. As
duction activity from those due to P2 measures
implemented at the firm. Box ES-1 presents
examples of production-adjusted P2 measures.
tion prevention (P2) m industrial settings. As For production-adjusted P2 measures aunit-of-
compamesmovebeyondshort-PaybackP2proj- product is the factor used to adju^oss quf -
ects to loneer-term. can tfll-infpnci* PO o~f;,,; *:*:__ .^ - . . . J^s11^ 4Udn
ects to longer-term, capital-intensive P2 activi-
ties, corporate management will rightly demand
accounting of the environmental and cost bene-
fits of these projects. In addition, many regu-
latory bodies and community groups are begin-
ning to ask individual facilities to demonstrate
that they are making progress in improving
environmental performance. Credible methods
of measuring P2 are key elements in any of
these requirements.
Accounting for varying levels of production is
one of the key issues in P2 measurement meth-
ods, If quantities of waste or chemical use
decrease after a P2 effort is made, the de-
crease may be attributed to the P2 effort.
However, other factors may also have in-
fluenced waste generation and chemical
consumption. For instance, if the number
of batches processed or quantity of pro-
duct produced has decreased during that
period, the change in waste may be re-
lated more to these external factors than
to the P2 efforts made.
J-"' ^ivuu VIM.MJ.J.
tities of waste or chemical use to infer the
amount of pollution prevention progress by
individual firms and groups of firms. If a firm
has made no pollution prevention improve-
ments, production-adjusted P2 measures should
show no change in waste generation per unit-of-
product. If pollution prevention changes have
been implemented, adjusted figures' should
show;a decrease in waste generation per unit-of-
product. Box ES-2 shows one example of how
a unit-of-product can be used to better assess P2
measurement data.
Production-adjusted measures of P2
account for changes in production activity
as well as for changes resulting from P2
efforts. In other words, production-ad-
justed measures of P2 allow a firm to
distinguish the components of waste
change that are due to changes in pro-
Box ES-1.
Typical Ways to Measure P2
Not Production-Adjusted
Change in quantity of emissions "Reduced
discharge of chromium by 20% last year"
Change in quantity of chemical or raw mate-
rials used "Reduced plating solution purchases
by 10% last year"
Production-Adjusted
Change in quantity of chemical used per unit
product "10% reduction in quantity of plating
solution used per part shipped last year"
Change in quantity of chemical used per unit
activity "Reduced solvent use by 15% for every
hour the degreaser ran last year"
-------
BoxES-2.
Using Unit-of-Product to Calculate P2 Improvements Can Filter out Effects
of Change in Production Activity
In 1993, Canton Circuits (a hypothetical firm) generated 22,000 pounds of trichlorethylene (TCE) waste
from a' vapor degreasing operation used to remove oil from the 16,000 metal circuit boxes it
manufactured. In 1994, after making several pollution-prevention changes, Canton generated 15,000
pounds of trichloroethylene waste in cleaning 20,000 circuit boxes. Under SARA, Canton could
measure P2 progress for the degreaser as follows:
,, . ._ , , ... ..
Un,t-of-Product Rate:
Boxes in 1994 _ 20,000 _
Boxes ,n 1993 " "
Using the Unit-of-Product Ratio
The production ratio is used to calculate the expected waste generation, given this year's level of
production, if no pollution prevention changes had been made during the past year. Expected waste
generation in 1 994 is calculated as follows:
(production ratio) (1 993 waste generation) = (1 .25) (22,000) = 27,500 Ib
1994 actual waste generation - 15,000 Ib, inferring 12,500 Ib waste reduction. This
measure of waste reduction filters out the effects of increased production at Canton.
Using Unit-of-Product to Assess P2 Changes on Efficiency
Another way to examine the effects of P2 is to assess whether the amount of waste per "widget"
produced has changed. Using Canton Circuits' data, the calculations would be as follows:
(TCE waste generated in 1 993)/(number widgets produced in 1993) =
(22,000)7(1 6,000) = 1 .38 Ib TCE per circuit box produced
(TCE waste generated in 1994)/(number widgets produced in 1994) =
(1 5,000)7(20,000) = 0.75 Ib TCE per circuit box produced.
The two waste1 efficiencies would then be compared to conclude that Canton had made substantial
waste reductions of 0.63 Ib TCE per circuit box produced.
Often pollution prevention activities are aimed
at reducing waste or emissions. However, P2
also includes the concept of usage of raw
materials, particularly hazardous raw materials.
Materials that are not introduced into a pro-
duction process cannot leave that process as
waste or emissions. Thus, reduction of materials
usage is an important part of the universe of
pollution prevention, and changes in materials
usage can be a measure of P2.
ES.2 Project Objectives
Three objectives were addressed in researching
production-adjusted measures of P2:
1. To describe different methods and systems
that firms are using to measure pollution
prevention; : .
-------
2. To develop methodology for application of
. statistical and graphical analysis for evalu-
ating production-adjusted measurements of
P2;and
3. To apply the statistical and graphical meth-
odology to "real world" data provided by
case study sites.
ES.2.1 Existing P2 Measurement Systems
The report describes five facilities that are cur-
rently using production-adjusted measures of
P2. These facilities were chosen to represent
both small and large facilities, as well as those
using complex and simple systems for P2 mea-
surement. The case study firms include a metal
finishing shop, two electronics firms, a phar-
maceutical firm, and a paper recycling facility.
ES.2.2 Evaluating Production-Adjusted P2
Measures
Under this research, a methodology was devel-
oped which applies statistical and graphical
tools to assess the accuracy of different units-
of-product used in P2 measurement. The pri-
mary focus of the methodology is to find a
unit-of-product that is closely related to the
waste being targeted.
The following example shows the importance
of finding a unit-of-product that is closely
related to the waste or chemical usage being
targeted. Imagine a production facility that has
modified its degreasing equipment to reduce
solvent loss. Suppose this facility finds that it
has reduced its purchase of solvent by X
gallons after the change is made, and that it has
cleaned Y parts in the month before the change
was made and Z parts in the month after the
change was made. If the loss of solvent has
more to do with the number of hours that the
'degreaser was running than with how many
parts were cleaned, then "solvent savings per
part cleaned" is a random number. "Solvent
saved per hour of operation," however, would
provide a good picture of the actual savings
resulting from the change. A unit-of-product
that is closely related to a target waste stream or'
chemical usage is said to be well-correlated
with the waste or chemical usage in question
(shown in Figures ES-1 and ES-2).
Waste
Line A: Before P2 Change
LineB: After P2 Change
PO PI
Production
Figure ES-1. Well-correlated unit-of-product
relationship between waste and a
related unit-of-product before and
after P2 improvements.
Waste
W,
Production
Figure ES-2. Plot of production and a waste that
is not strongly correlated to
production. No relationship can be
detected.
-------
The methodology for assessing the relationship
between a unit of product and a given waste
stream or chemical use stream is shown in
Figure ES-3.
ES.2.3 Application of Methodology
The research team tested the methodology for
applying statistical and graphical tools to assess
units-of-product by applying it to data supplied
by the five case study facilities. The case study
facilities consisted of manufacturers in metal
finishing, semiconductor fabrication, electron-
ics, Pharmaceuticals, and paper recycling. This
process allowed the research team to assess the
usability of the methodology in a practical
setting.
Using real-world data also allowed the research
team to make a preliminary assessment of how
different units-of-product might correlate with
key waste streams or key chemical inputs in
other firms in the same 'industries.
ES.3 Results
Use of Production-Adjusted P2 Measurement
Although the major driver for developing pro-
duction-adjusted measurements of P2 has been
regulatory requirements, firms have also found
these measures to be useful for other reasons.
The process of setting up a production-adjusted
P2 measurement system can have benefits
beyond those of fulfilling regulatory reporting
needs; conversely, some systems that have been
set up for other applications (e.g., statistical
process control, product pricing) can be used to
generate P2 measurement values.
In addition to providing a way to track pollution
prevention progress, production-adjusted mea-
sures of P2 provide firms with a more detailed
understanding of waste generation and chemical
use patterns. This insight can help firms fine-
tune their production processes to improve
efficiency.
Measuring P2 can be a resource-intensive pro-
cess. It is important to ensure that the resources
expended are in line with the benefits accrued.
It is counterproductive to spend many staff
hours to develop and implement a measurement
system if no resources will be left to actually
implement P2 projects. Likewise, a P2 mea-
surement system should be selected that is
appropriate to the production process or facility
being measured: if the process is constantly
changing, the measurement system should
Step 4:
Regression
Analysis
yes
Stepl:
Process
Description
Step 2:
Identify and
Collect Data
Step 3:
Graphical
Analysis
+ / Regres
^X^Analys
^v^-
Step 5: Periodically Repeat Analysis
Analysis
Complete
Figure ES-3. Five steps for unit-of-product analysis.
4
-------
accommodate changes. If the product in ques-
tion is being phased out, then a more rudi-
mentary measurement may be in order.
Application of Project Methodology
Testing the project methodology with facility
data showed that it was possible to use the
methodology to assess production-adjusted
measures of P2 at different manufacturing
facilities.
Finding data to use in the methodology for
verifying units-of-product for P2 measurement
requires extensive and thorough communication
among firm personnel, from production engi-
neers to accounting staff. Careful attention must
be paid to the sources and time frame of the
data.' For instance, it is important to know
whether the production data supplied by a
department refers to actual line production or to
shipments from inventory.
It may be difficult to obtain enough waste
generation data points to use this project's
methodology to directly assess a measurement
of P2 based on changes in waste generation.
Depending on the particular production process,
it may be possible to substitute chemical use
data as a surrogate for chemical waste.
Chemical use data can then be used to assess
the waste-based P2 measure or can be used to
construct a use-based P2 measure.
Assessing a unit-of-product used in a pro-
duction-adjusted P2 measurement system is an
iterative process. Users of the methodology
presented in this report must understand the
objectives of the analysis and periodically
assess how well the methodology fits the
available data.
Conclusions about Units-of-Product Used by
Case Study Facilities
Use of the case study facilities allowed re-
searchers to examine the workings of five dif-
ferent production-adjusted measures of P2 in
five different industries. These units-of-product
used by the case study facilities are summarized
in Table ES-1.
The research team detected a statistically
significant relationship between single units-of-
product (e.g., "square feet plated" or "kilograms
of product produced") and chemical usage at
the case study facilities. In the case study
facility for which waste data were analyzed,
correlation was also found between waste and a
single unit-of-product.
This finding is significant because there has
been some concern that single units-of-product
are inadequate to explain variation in waste
generation. If this were true, then it would be
much more difficult for firms to accurately
assess their P2 performance, as they would have
to account for many more variables than a
single, measurable output. The results of this
research, however, suggest that a carefully
chosen single variable unit-of-product can
account for enough of the variation in chemical
use or waste to be used in adjusting gross P2
measures.
-------
Table ES-1. How Well Units-of-Product Explained Variation in Chemical Use and
Waste Generation
Did unit-of-product explain
variations in
dustry
Unit-of-product used
for adjusting
pollution-prevention
measurement
Chemical use
for key inputs?
Waste genera-
tion for key waste
streams?
Facility or
company-wide
measure or
process
specific?
Metal finishing
Square feet substrate
plated or coated
Yes
Paper recycling Tons of paper produced Yes
Semiconductor
fabrication
Electronics
production
Combined unit-of-
product incorporates
number of memory
chips, logic chips, and
masks produced [as
surrogate for tech-
nological content of
product]; number of
module parts produced
Number of passes sub-
strate makes through
process ,
Combined unit-
of-product cor-
related for some
chemicals, not
for others;
module parts '
correlated for all
chemicals5
Yes
NAa
NA
Number of bits (a
component of the .
combined unit-of-
product) correlated
with one waste
stream; module
parts correlated.
with same waste
streamb
NA
Process-specific
Facility-wide
Facility-wide
Specific to each
product line
Pharmaceutical Kilograms of product
production produced
Yes Yes
Specific to
individual
department
8 NA = Not applicable.
b Results somewhat uncertain; see Section 4.3.3 for full discussion.
-------
Section 1
Introduction
Accurate and meaningful measurement systems
are essential to the long-term success of Pollu-
tion Prevention (P2) in industrial settings. As
companies move beyond short-payback P2 proj-
ects to longer-term, capital-intensive P2 activi-
ties, corporate management will rightly demand
accounting of the environmental and cost bene-
fits of these projects. In addition, many regula-
tory bodies and community groups are begin-
ning to ask individual facilities to demonstrate
that they are making progress in improving
environmental performance. Credible methods
of measuring P2 will be key elements in any of
these requirements. The U.S. Environmental,
Protection Agency's (EPA's) Office of Re-
search and Development (ORD), Research Tri-
angle Institute (RTI), and Greiner Environ-
mental undertook research to understand the
methods and structures that firms are using to
measure pollution prevention.
In particular, the objective of this research was
to investigate P2 measurement that reflects
changes in emissions, waste, or chemical usage,
and also reflects variations in production levels.
This kind of P2 measurement is referred to in
this report as "production-adjusted P2 measure-
ment." Other authors refer to it as "normalized"
or "indexed" measurement of P2 (Harriman et
al, 1991).
1.1 Background of P2 Measurement in
' General
P2 measurement issues come up along a spec-
trum of applications:
"-> Measuring effects of a single P2
project on one process line
"-+ Measuring P2 for a single facility
or company
"-» Measuring national, state, or in-
dustry sector P2 progress.
This research looks in detail at P2 measures for
a specific facility or production line. Others
have addressed the issues of measuring P2 on
larger scales. See, e.g., Tellus et al., 1991.
1.1.1 Who Uses P2 Measurement and Why
The users of P2 measurement are identified in
Box 1-1. It became clear as this research pro-
gressed that there is broad interest in P2 mea-
surement. It ties into many different areas of
environmental policy and regulation in this
country.
P2 as Measured by Change in Materials
Usage. Often pollution prevention activities are
aimed :at reducing waste or emissions.
However, P2 also includes the concept of usage
of raw materials, particularly hazardous raw
materials. Hazardous materials that are not
entered into a production process cannot leave
that process as waste or emissions. Thus,
reduction of hazardous materials use is included
in the universe of pollution prevention.
In this research data regarding changes in quan-
tity of raw materials were often used as a way
of assessing P2 progress. In this report, we use
the term "chemical usage" rather than "raw
-------
Box 1-1.
Who Can Use This Report?
Facility staff in industries examined in these
case studies. Areas of interest to them include:
» The effectiveness of the production-adjust-
ing units used by the facilities we visited.
» The aspects of measurement systems that
were successful and those that were not as
useful.
» How the measu rement system added value
to the process or improved quality of
product.
» What types of data are used by other facili-
ties to measure P2.
People in other industrial sectors who are
considering whether and how to measure P2.
Topics of interest to this audience include:
The characteristics of the P2 measurement
systems that seem to be effective.
« What types of data are used by other facili-
ties to measure P2.
« Information about how P2 measurement
has been valuable to companies.
EPA ORD staff. Topics of interest include:
Sources of good data at facilities (likely to
be of particular interest to people who are
working on P2-related software).
Information about factors that have led to
successful P2 measures at facilities.
Regulatory policy staff. Topics of particular
interest will be:
Information about the potential accuracy of
relationship between P2 measurements
that a facility generates and the kind of P2
that is actually occurring.
Information about what kinds of P2 data
can be generated at facilities and possible
overlaps with toxic release inventory (TRI)
information.
Information about uses for chemical use
data in measuring P2.
Citizens Groups/Environmentalists, particu-
larly those who want to find national tools to
measure P2. Information of particular relevance
includes:
General description of the issues involved
in developing an accurate measure of P2 at
a facility.
» Information about the limitations of various
approaches to P2 measurement as applied
to specific facilities.
material usage." This is because the raw mate-
rials in question were chemicals subject to
environmental regulation. Despite this use of
terminology, there is no reason that the method-
ology employed here could not be used to
assess measures of non-hazardous waste gen-
eration and nonchemical materials use.
1.1.2 Production-Adjusted Measures of P2:
A More Detailed Look
If quantities of waste or chemical use decrease
after a P2 effort is made, the decrease may be
attributed to the P2 effort. However, other fac-
tors may have also influenced waste generation
and chemical consumption. For instance, if the
number of batches processed or quantity of
product produced has decreased during that
period, waste reduction may be more properly
attributed to these external factors than to the
P2 efforts made. Production-adjusted measures
of P2 account for changes in production activity
as well as accounting for changes resulting from
P2 efforts. Another way of stating the same
concept is that production-adjusted measures of
P2 allow a firm to separate out the components
-------
of waste change that are due to changes in pro-
duction activity vs. those due to P2 measures
implemented at the firm. Box 1-2 presents
examples of production-adjusted P2 measures.
In addition, companies that report under the
Federal Superfund Amendments and Reauthori-
zation Act (SARA) Title m Section 313
(Toxics Release Inventory) arid under the
reporting acts of several states (e.g., Massa-
chusetts and New Jersey) are required to report
a production-adjustment factor along with infor-
mation about releases' or chemical use.
The Unit-of-Product. For production-adjusted
P2 measures, a unit-of-product is the factor
used for adjusting gross quantities of waste or
chemical use to infer the amount of pollution-
prevention progress by individual firms and
groups of firms. If a firm has made no pol-
lution-prevention improvements, production-
adjusted P2 measures should show no change in
waste generation per unit-of-product. If pollu-
tion prevention changes have been imple-
Box 1-2.
Typical Ways to Measure P2
Not Production-Adjusted
Change in quantity of emissions
"Reduced discharge of chromium by
20% last year"
Change in quantity of chemical or
raw materials used "Reduced plating
solution purchases by 10% last year"
Production-Adjusted
Change in quantity of chemical
used per unit product "10%
reduction in quantity of plating solution
used per part shipped last year"
Change in quantity of chemical
used per unit activity "Reduced
solvent use by 15% for every hour the
degreaser ran last year"
mented, adjusted figures should show a
decrease in waste generation per unit-of-
product. Box 1-3 shows one example of how a
unit-of-product can be used to better assess P2
measurement data.
Businesses use production-adjusted P2 mea-
sures for many reasons other thari reporting
requirements. Many businesses find production-
adjusted data useful for:
Gaining insight into chemical use and pro-
cess efficiency;
Setting P2 goals and measuring progress
against those goals;
Comparing corporate process, facility, and
division performance; and
Communicating P2 progress to stake-
holders.
Production-adjusted P2 Measurement Issues
Addressed in This Report. This research in-
vestigated three topics related to production-
adjusted P2 measurement:
1. Case studies providing a snapshot of
firms that use production-adjusted P2
measurement: How firms currently use
measures of P2. How they select the mea-
surement method they use. How valuable it
is to the firm to have a production-adjusted
measure of P2.
2. Develop methodology to apply graphical
and statistical tools for assessing the ac-
curacy of different production-adjusted
measures of P2.
3. Preliminary assessment of the accuracy
of how the case study facilities produc-
tion-adjusted their P2 measures.
In addressing the first topic, we identified five
facilities that are currently using production-
-------
Box 1-3.
Using Unit-of-Product to Calculate P2 Improvements Can Filter out Effects
of Change in Production Activity
In 1993, Canton Circuits (a hypothetical firm) generated 22,000 pounds of trichlorethylene (TCE) waste
from a vapor degreasing operation used to remove oil from the 16,000 metal circuit boxes it
manufactured. In 1994, after making several pollution-prevention changes, Canton generated 15,000
pounds of trichloroethylene waste in cleaning 20,000 circuit boxes. Under SARA, Canton could
measure P2 progress for the degreaser as follows:
Unit-of-Product Ratio: Boxes in 1994 = 20.000 = ^
Boxes in 1993 16,000
Using the Unit-of-Product Ratio
The production ratio is used to calculate the expected waste generation, given this year's level of
production, if no pollution prevention changes had been made during the past year. Expected waste
generation in 1994 is calculated as follows:
(production ratio) (1993 waste generation) = (1.25) (22,000) = 27,500 Ib
1994 actual waste generation -15,000 Ib, inferring 12,500 Ib waste reduction. This
measure of waste reduction filters out the effects of increased production at Canton.
Using Unit-of-Product to Assess P2 Changes on Efficiency
Another way to examine the effects of P2 is to assess whether the amount of waste per "widget"
produced has changed. Using Canton Circuits' data, the calculations would be as follows:
(TCE waste generated in 1993)/(number widgets produced in 1993) =
(22,000)7(16,000) = 1.38 Ib TCE per circuit box produced
(TCE waste generated in 1994)/(number widgets produced in 1994) =
(15,000)7(20,000) = 0.75 Ib TCE per circuit box produced.
The two waste efficiencies would then be compared to conclude that Canton had made substantial
waste reductions of 0.63 Ib TCE per circuit box produced.
adjusted measures of P2 and worked with them
to document their methods and results. This
information is presented in Section 3 of this
report. To address the second topic, we devel-
oped a methodology for applying statistical and
graphical tools to evaluate different units-of-
product used in production-adjusted P2 mea-
surement. This is presented in Section 2. We
applied this method to data that the case study
facilities shared with us. This allowed us to test
the usability of the methodology, as well as to
provide initial indications about the usefulness
of various potential production-adjusting "units-
of-product" for the industry sectors represented
by the case study facilities. These results are
presented in Section 4. Section 5 provides
conclusions from this work. Appendixes A and
B give relevant references.
10
-------
In addition, we developed a framework for
selection and use of production-adjusted mea-
sures of P2. The framework is based on the
information shared by the case study facilities
and the information obtained through the analy-
ses conducted during this research. The frame-
work is presented as Appendix C of this report.
11
-------
Section2
Description of a Methodology for Application of Statistical and Graphical
Tools to Assess Accuracy of P2 Production-Adjusting Units
A key component of the P2 measurement
framework is evaluation of the unit-of-product
used to adjust the P2 measurement to account
for variation in production. To create an accu-
rate measure of the effects of a P2 effort, it is
necessary to find a unit-of-product that is
closely related to the waste being targeted. To
illustrate, imagine a production facility that has
modified its degreasing equipment to reduce
solvent loss and finds that it has reduced its
purchase of solvent by X gallons after the
change is made. Suppose further that they have
cleaned Y parts in the month before the change
was made and Z parts in the month after the
change was made. If the loss of solvent had
more to do with the number of hours that the
degreaser was running, rather than how many
parts were run through it, then the "solvent
savings per part cleaned" is a random number,
whereas "solvent saved per hour of operation"
would provide a good picture of the actual
savings resulting from the change. This would
provide a comparison with which to measure
later P2 changes.
2.1 Evaluating a Unit-of-Product
Companies that file under the Federal Toxics
Release Inventory (TPJ) are required to report
a unit by which their reported levels of emis-
sions and releases can be adjusted. This is
known as the "production ratio" or "activity
ratio" or the "unit-of-product." The purpose of
a production ratio or an activity index is to
allow year-to-year comparisons of waste gener-
ation that are adjusted for the level of produc-
tion. In addition, many companies want to track
their P2 progress more accurately, assess their
P2 investments, and communicate their
achievements to stakeholders. Production-
adjusted measurement helps accomplish these
goals.
This section reviews a methodology for using
statistical and graphical tools for assessing a
unit-of-product. The methodology was devel-
oped for this project. It begins with an intro-
duction to the unit-of-product concept. Data
collection methods and requirements are then
presented. Next, the section presents three
graphical analysis tools arid an overview of a
regression analysis tool used to evaluate how
well a unit-of-product explains the variation in
key pollution or chemical use figures.
2.1.1 The Unit-of-Product
A unit-of-product is used to adjust the overall
measure of changes in chemical use or waste
generation. If a firm has made no pollution-
prevention improvements, adjusted P2 mea-
sures should show no change in waste genera-
tion per unit-of-product. If successful pollution-
prevention changes have been implemented,
adjusted figures should show a decrease in
waste generation per unit-of-product.
12
-------
2.1.2 Choosing a Unit-of-Product
The goal to keep in mind when choosing a unit-
of-product is to select one that is well cor-
related to chemical use or waste generation.
This means that waste per unit-of-product is
constant whatever the level of production, e.g:,
when production increases, generation in-
creases proportionally and waste per unit-of-
product remains constant. Line A in Figure 2-1
depicts this linear relationship between waste
and production data. Mathematically, the slope
of the line (W/P) is constant. Under this
assumption, if a P2 change were implemented,
the change would lead to a new relationship be-
tween production arid chemical datarepre-
sented as Line B in Figure 2-1.
A poorly correlated .unit-of-product will not
measure P2 progress adequately. For example,
when production doubles, waste generation
does not increase proportionally. This means
waste per unit-of-product is not constant but
depends on the level of production. As a result,
a poor unit-of-product will under- or over-
estimate P2 progress. Figure 2-2 represents a
poorly correlated unit-of-product where there is
a random relationship between waste and pro-
duction. The waste per unit-of-product ratio
(W/P) is different for most points. There is no
consistent, predictable relationship between
waste and the unit-of-product. Thus, variations
in the W/P ratio cannot be said to be attrib-
utable to P2 efforts.
Identifying a well-correlated unit-of-product
will be easiest in cases where:
There are few uses of a chemical at the site.
The greater the number of uses, such as the
case where a cleaning solvent is used in six
different sites around the plant, the more
difficult it is to find a measure of production
Waste
W
W0=W
Line A: Before P2 Change
LineB: After P2 Change
PO
; Production
Figure 2-1. Well-correlated unit-of-product
relationship between waste and a
related unit-of-product before and
after P2 improvements.
Waste
Wf
Production
Figure 2-2. Plot of production and a waste that
is not strongly correlated to
production. No relationship can be
detected.
that .correlates with the waste stream con-
taining this chemical.
There is little variation in the products pro-
duced using the chemical. Variation in
product types (such as printed circuit boards
and subassemblies) and attributes (such as
13
-------
surface area, geometric shape, or substrate
type) makes finding a unit-of-product more
complex since each attribute can affect
waste generation differently.
There is little change in processes. Pro-
cesses that are constantly changing make
measurement from year to year more
difficult. Firms with less variable produc-
tion find it easier to find a unit-of-product
since processes and products remain rela-
tively constant from year to year.
Choosing a well-correlated unit-of-product is
further confounded by one important con-
straintare the data available? Firms can only
choose among potential units-of-product for
those that the company has historical data or is
willing to collect new data. This is an obvious
but very real constraint since many candidates
are not tracked on a regular basis.
2.2 Analyzing the Unit-of-Product
How can an environmental professional choose
a unit-of-product that is well correlated to a
given chemical's use or waste generation? Two
analytical methods are presented here
graphical analysis and regression analysis.
Graphical analysis is used to qualitatively assess
a unit-of-product. Graphical analysis methods
include preparing histograms, time-series plots,
and scatter plots. Graphical analysis is also a
preliminary step when performing regression
analysis. See Figure 2-3.,
Regression analysis is used to evaluate a unit-
of-product quantitatively. Regression analysis
involves calculations to determine the degree of
correlation between chemical and production
data. Whether graphical methods alone are used
or graphical and regression methods are used
together, a multistep data collection and analy-
sis process should be followed when evaluating
a unit-of-product.
Step 1. Process Description,
The purpose of this step is to map out the pro-
cess under investigation. This step involves
drawing a flow diagram, tracing the chemical's
path through the process, and noting chemical
inputs, outputs, and conversions. The level of
complexity of the flow diagram will vary
depending on the level of accuracy one needs
for the analysis.
Step 4:
Regression
Analysis
Stepl:
Process
Description
»
Step 2:
Identify and
Collect Data
Stpn K
Step 3:
Graphical
Analysis
Periodically Repeat
^/ Regres
\Analy
\^
Analysis
Figure 2-3. Five steps for unit-of-product analysis.
14
-------
Step 2. Identify and Collect Tinie
Consistent Data u -
To analyze a unit-of-product, it is necessary to
have time-consistent chemical and production
data. The term "time consistent" means that the
chemical data and production data must cor-
respond to the same time period, e.g., daily
pounds of xylene used and daily square feet
painted. Analysis cannot be performed on data
from different times, e.g., daily square feet
painted and weekly pounds of xylene used.
Chemical data can be found in process engi-
neering records, materials accounting records,
or process control charts. Production data are
typically found in production logs. The data set
should cover an adequate number of time
periods to allow trends and relationships to be
apparent. We recommend attempting to have at
least 30 time periods (e.g., 30 days or 30 weeks)
in the analysis data set. More time periods are
preferable because more data points improve
the accuracy of the analysis.
Analysis will be improved where there is some
variation in production levels during the time
periods being investigated. This is because data
trends are easier to see when the data are not
entirely clustered around one set of values.
If regression analyses are to be used to analyze
the data, the data should be collected over a
time period during which there were no major
changes to the production process. For a regres-
sion analysis to be meaningful, it requires data
from a process that has performed consistently.
This consistency requirement makes the use of
quarterly or monthly data undesirable in regres-
sion analysis since it is likely that some major
change to the process would have occurred over
a 30-month or 30-quarter time period.
More often than not, firms find that they can
use chemical use data (as opposed to waste
data) to evaluate their unit(s)-of-product. Chem-
ical use data can be monitored on a real-time
basisbut waste volumes are difficult to moni-
tor in this way. Waste data are typically cal-
culated once a year for reporting purposes.
Waste data are also often estimated from mate-
rial balance calculations rather than measured
directly. For example, while it is difficult to
measure weekly waste generation (emissions)
from a solvent degreaser, directly measuring
solvent use is relatively straightforward. Fur-
ther, using waste inventory data for the pur-
poses of unit-of-product analysis can be prob-
lematic. This is because waste inventory data
often lag behind actual waste generation, and
data about offsite shipments often reflect more
information about the waste hauler's schedule
than about waste generation rates.
[Step 3. Graphical Analysis
Graphical analysis allows one to see data pat-
terns and is a relatively simple way to look at
the fit between measures of production and
chemical data. Specifically, plots of production
and chemical data allow one to see:
Distribution of the data (i.e., normal, bi-
modal, etc.) and trends in the data;
Extreme data points or outliers (e.g., very
high or very low values); and
Data entry errors (errors are easiest to spot
when they have extreme values).
Graphical analysis tools include histogram
plots, scatter plots, and time-series plots. These
tools are reviewed in detail in Section 2.2.1.
15
-------
Step 4. Regression Analysis
After completing a graphical analysis, firms can
choose to review the data further by performing
a regression analysis. Whereas graphical analy-
ses provide a qualitative sense of the correlation
between production and chemical data, regres-
sion analyses provide a quantitative measure of
the correlation between production and chem-
ical data. Whether a firm chooses to perform a
regression analysis depends on whether the
firm:
« Has the resources (expertise and software)
to analyze the data,
« Wants a quantitative measure of whether its
unit(s)-of-product are well correlated, and
« Finds the qualitative graphical analysis
results inconclusive.
If the company performs regression analysis, it
must determine whether to use simple linear
regression or multiple . regression methods.
Simple linear regression can be performed with
most hand-held calculators or spreadsheet soft-
ware programs. Simple linear regression is
appropriate when examining the correlation of
a single unit-of-product (e.g., square feet
plated). Multiple regression is used when exam-
ining whether some unit-of-product combina-
tion (e.g., square feet plated, amp hours, and
number of parts) correlates with chemical data.
In general, multiple regression analysis is much
more complex than simple linear regression.
Regression analysis is discussed in. Sec-
tion 2.2.2.
Step 5. Repetition
Once the analysis is complete, it should be
repeated periodically (especially after major
changes to the process) to make sure the
chemical and production data are still cor-
related. Figure 2-3 -depicts this multistep
method for analyzing a unit-of-product.
2.2.1 Graphical Analysis
This section reviews three graphical analysis
plotshistogram plots, scatter plots, and time-
series plots. When evaluating a unit-of-product,
one should prepare and examine each of these
plots.
Histograms. Histograms provide a picture Of
the frequency distribution of a data set. The
frequency is shown by drawing a rectangle
whose base is the "chemical data per unit-of-
product interval" (i.e., quantity sulfuric acid/
pound of paper) on the horizontal axis and
whose height is the corresponding frequency. In
this report, the x-axis of histograms is marked
in "bins." A bin is a range of values (i.e., values
falling between 10 and 15, 16 and 20, and so
on). The height of the bar shows how often
values from a given data set fall within that
range of values. Bell-shaped histograms are
indicative of a process undergoing normal
variation. Furthermore, bell-shaped histograms
are also indicative of a well-correlated unit-of-
product. If the histogram does not have a bell
shape, the ratio of chemical data to production
may be a poor choice. A histogram of the hypo-
thetical paper manufacturing data is shown in
Figure 2-4. Notice the normal distribution of the
data. While the plot indicates that tons of paper
produced is a good unit-of-product for sulfuric
acid, one should prepare scatter and time-series
plots before drawing conclusions.
Histograms also help the investigator to see
whether one or several "extreme" data points
are affecting the overall mean. Extreme data
points could also indicate particularly wasteful
or particularly efficient periods of operation that
warrant further examination. For example, the
16
-------
S 18
Chemical Use per Unit-of-Product
The scatter plot in Figure 2-5
depicts the hypothetical paper
manufacturing data set The plot
shows an increasing relationship
between production and chem-
ical useindicating that the two
are correlated. Taken together;
the scatter and histogram plots
strongly suggest tons of paper
produced would be a good unit-
of-product to measure sulfuric
acid P2 progress.
Figure 2-4. Sulfuric acid use per ton of paper histogram.
points representing the smallest values of
chemical use per unit-of7product are valuable
from a P2 perspective. The firm's engineers
could use them to identify optimal operating
conditions. If the paper manufacturer replicated
these operating conditions, the company would
significantly reduce sulfuric acid use, waste,
and raw material cost.
Scatter Plots. Scatter plots are used to examine
the relationship between chemical and produc-
tion data. If the two are perfectly correlated, the
points in a scatter plot would line up evenly and
one could draw a straight line through each
point (Figure 2-1). If chemical and'production
data are not correlated, the scatter plot would
have no discernible pattern-just a random
scatter of data points through which no line
could be drawn (Figure 2-2). Most scatter plot
data fall somewhere between these two ex-
tremes. After preparing a scatter plot, it is good
practice to draw a "best fit" line through the
data. The easier it is to draw such a line, the
stronger the correlation between chemical and
production data. The slope of this line repre-
sents the average chemical use per unit-of-
output.
Time-Series Plots. Time-series
plots are useful for data that have
been collected sequentially. When one plots the
observations in time sequence, trends and
cycles often become apparent. Data that either
consistently increase. or decrease should be
viewed with caution. Consistently increasing or
decreasing trends indicate that the process is
unstable and is not undergoing normal day-to-
day variation. Good normalization data should
have a random time-series plot. The time-
sequence plot of the paper manufacturing data
set shows a random trend (Figure 2-6).
Taking the paper manufacturing histogram,
scatter, and time-series plots together, it appears
that sulfuric acid use and tons of paper pro-
duced are correlatedhigh levels of sulfuric
acid use correspond to high levels of pro-
duction. This conclusion is derived from the
fact that
The histogram of sulfuric acid use per unit-
of-output is bell-shaped;
The scatter plot shows an increasing
trenda line depicting this trend can be
drawn through the data; and
The time series plot shows a random pattern
as opposed to a constantly increasing or
decreasing pattern.
17
-------
70.000
, 60,000
' 50.000
40.000
30,000
20.000
10,000
600 650 ,700 750 800 850 900
Chemical Use (Ib)
950
1000
Figure 2-5. Scatter plot showing paper produced per pound of
sulfuric acid.
so
80
[TO
jeo
so
«
30
so
10
ft |«
i ,
5t.j.H+M-i-)-i-tTi i rim nT
Day
Figure 2-6. Time series plot showing sulfuric acid use per ton of
paper.
The histogram and time-series plots also show
2 days where chemical use per unit-of-product
was abnormally low. These days should be
evaluated more closely since they represent
periods of greater chemical use efficiency. If
operating conditions on these 2 days could be
replicated, the company would significantly
reduce its waste generation and raw material
cost.
Descriptive Statistics. After viewing the data
graphically, it may become clear that there is an
outlier in the data, i.e., a data
value that is much larger or
smaller than the rest of the data
points. Compiling a set of
descriptive statistics for the full
data set and for the data set
without the outlier can help the
user understand the impact of the
outlier on the data set. Descrip-
tive statistics include values like
the mean, standard error, me-
dian, mode, standard deviation,
and confidence level. If the out-
lier is discovered to have a large
impact on the data set, then the
user may choose to exclude that
data point for purposes of the
unit-of-product analysis. Des-
criptive statistics are an analysis
option available in many spread-
sheet programs.
The confidence level descriptive
statistic deserves particular ex-
planation here. The 95% con-
fidence level statistic shows the
range around the calculated
mean in which the true mean is
likely to lie. Thus, if the des-
criptive statistics show that the
mean for the data set is 25 units,
and the 95% confidence level is 5.4 units, then
we can be 95% sure that the true mean will lie
within 5.4 units of the mean for the data set.
That is, we can be 95% confident that the true
mean is somewhere between 19.6 and 30.4.
2.2.2 Statistical Analysis
Unlike graphical analysis, regression methods
calculate the correlation between chemical data
and a unit-of-product. Regression tests can be
particularly helpful when choosing between two
possible units-of-product or when it is impor-
18
-------
tant to know the degree of
correlation .between chemical
and production data. While
regression methods quanti-
tatively determine a unit-of-
product's correlation, regression
analysis requires expertise and
either computer software or a
hand calculator with statistical
functions to analyze the data.
Before performing a regression,
it is important to check use or
waste per unit-of-product data to
see if the data are normally dis-
tributed. The best way to see if a
data set is normal or not is to
examine the histograms gener-
ated in Step 3. The histogram
should approximate a normal or
"bell-shaped" distribution (see
Figure 2-7). If the histogram is
bi-modal (two humps, shown in
Figure 2-8), skewed (most val-
ues high or low, shown in Figure
2-9), uniform (same frequency
for all bins, Figure 2-10), or in
some other way obviously non-
normal a regression analysis
should not be performed.
In cases where the data do not
appear to be normal, the data can
be mathematically "transformed"
to a normal shape. Data are
.transformed by multiplying each
data point by a factor such as log
x, In x, ex, 1/x, x 2, etc. The
choice of a specific factor de-
pends on the shape of the dis-
tribution (i.e., for skewed dis-
tributions, one would try a log-
arithmic transformation). There
are good reference materials
to o 10
r- CM (M
Bin
Figure 2-7. Histogram showing normal distribution of
chemical use per unit-of-product data.
w o to o m
T- OJ =OJ CO PJ
S 18 S S
Bin
o 10 o a>
at O) o {5
" E
Figure 2-8. Histogram showing bimodal distribution of
chemical use per unit-of-product data.
Figure 2-9. Histogram showing skewed (exponential)
distribution of chemical use per unit-of-product
data.
19
-------
where:
BO
X
y
x
error
Figure 2-10. Histogram showing uniform distribution of
chemical use per unit-of-product data.
available that provide guidance on how to trans-
form data to make it more normally distributed
(see Appendix B). Once one has reasonably
normal data, it is time to proceed with regres-
sion testing.
Regression analysis can be divided into two
related testssimple linear regression and
multiple linear regression. Which test one uses
depends on the questions one wants to answer
and the data one has in hand:
* Linear regression is u'sed to look at the cor- .
relation between chemical data and a single
unit-of-product (e.g., whether sulfuric acid
use and pounds of paper manufactured are
correlated).
Multiple regression is used to look at the
correlation between chemical data and more
than one unit-of-product (i.e., whether
xylene use and some combination of square
feet painted, part depth, and number of parts
per rack are correlated).
Simple Linear Regression. Mathematically,
the simple linear regression model is defined as:
y = BO + Bl X +error
y intercept
slope
chemical data
production data .
the error or devia-
tion of the actual y
value from the line
BO + BIX.
In a regression analysis, produc-
tion data are the independent
variable (x) and chemical data
are the dependent variable (y).
Simple linear regressions can be
run on spreadsheet programs such as Excel or
on a statistical software package. The general
procedure followed when using computer
packages is for the user to input the data (x and
y values) together with some instructions
concerning the types of analyses that are
required. The software package performs the
analysis and prints the results in an output
report. Output data that are useful for analyzing
normalization data include: (1) BO and.Bl
values, (2) the coefficient of determination (R-
squared), and (3) a P-value.
BO and Bl Values. Regression software pack-
ages generate an equation for a line that best fits
the data. Usually expressed as coefficients, the
regression produces intercept (BO ) and slope
(Bl) estimates,
R-Squared. While the output of different soft-
ware packages varies, all software regression
analyses calculate a value known as the "R-
squared" (r2) term or "coefficient of deter-
mination." The R-squared term is a measure of
the goodness-of-fit of the estimated regression
line. It ranges from 0.0 to 1.0. For P2 mea-
surement applications, R-squared values close
to one are indicative of a good unit-of-product.
However the R-squared term alone does not tell
20
-------
whether the relationship is statistically signifi-
cantfor this the regression "P-values" must be
known.
The R-squared term is a measure of the
goodness-of-fit of the estimated regression line.
It ranges from 0.0 to 1.0. In P2 applications, the
R-squared value estimates how much of the
variation in waste is explained by variation in
the chosen unit-of-product. The closer an R-
squared value comes to 1.0, the more of the
variation in waste is explained by variation in
that unit of product. However, the R-squared
term alone does not tell the user whether he/she
can be confident (in a statistical sense) that this
relationship between waste and unit-of-product
exists. In order to find out whether the rela-
tionship is statistically significant, the regres-
sion P-values must be calculated.
P-values. P-values indicate whether the values
computed for BO and B.1 are statistically sig-
nificant. A P-value of .05 for Bl indicates that
we can be 95% confident that the relationship
between pur x and y variables is not random.
The general rule for using P-values is as
follows:
P-values <0.05 indicate statistical signifi-
cance
P-values >0.05 indicate
statistical insignificance.
Table 2-1 depicts a typical
simple linear regression output.
In this case, the regression was
run on the paper manufacturing
data using the spreadsheet
program Excel.
Using Table 2-1, the equation for
the line for the paper manu-
facturing data is:
y = 1982+ 59X
intercept = 1982
slope =59.
The slope gives the average chemical used per
unit-of-product produced (59 Ib chemical/lb of
product). The R-squared value is .23a low
number for such an analysis (values near one
are indicative of a good unit-of-product). The P-
value for Bl is .0001 indicating 99.99% con-
fidence that chemical use and production are
correlated: The scatter plot, the line, the equa-
tion for the line, and the R-squared value are
presented in Figure 2-11.
Table 2-1. Simple Linear Regression
Output
Regression statistics
R-squared
Observations
0.2294
60.0
Coefficients
Coefficient
value
P-value
BO
Bi
1982
59
0.85794
0.00010
750 SOO . 850
Chemical Use (Ib)
Figure 2-11. Scatter plot showing relationship between
tons of paper produced and pounds of
sulf uric acid used.
21
-------
The regression results indicate that the average
sulfuric acid use (in pounds) per pound of paper
manufactured is equal to 59. This value is "sta-
tistically significant" because the P-value is
<0.05. The R-squared term is equal to 0.23.
This means that pounds of paper manufactured
and sulfuric acid use are correlated. The
quantity of paper produced explains 23% of the
variation in sulfuric acid use. Obviously there
are other factorsperhaps variation in raw
material quality, ambient temperature, or opera-.
tor factorsthat are contributing to the remain-
ing variation between the observed data and the
regression line. The following conclusions can
be drawn from regression analysis of the paper
manufacturing data:
* Sulfuric acid use and tons of paper pro-
duced are correlated with a high degree of
confidence (P value = .0001therefore, one
can be 99.999% confident the
two are correlated);
Since the correlation is strong,
tons of paper produced is a
good unit-of-product for sul-
furic acid; and
The amount of sulfuric acid
used each day is affected by
factors other than the amount of
paper produced (since r2 =
0.23).
ware programs will calculate and plot regres-
sion residuals.
The residuals in a residual plot should exhibit a
random pattern. For example, the residual plot
shown in Figure 2-12 has a random pattern. If
the residuals are clustered or display spreading
or narrowing patterns, the investigator should
reexamine his/her data set and modify the
regression model. The recommended way to
modify the regression is to transform the data
a procedure outlined in Section 2.2.2 and
described in greater detail in most regression
text books.
Multiple Regression Analysis. Multiple re-
gression analysis is used when one wants to
determine whether two or more measure(s) of
production are correlated with chemical data.
Analysis of Residuals. The analy-
sis of residuals plays an important
role in validating the regression
assumptions and results. For each
observation in a regression analy-
sis, there is a residual; it is the dif-
ference between the observed value
of the dependent variable (y) and
the value predicted by the regres-
sion equation. Most computer soft-
DC
fluuuu
15000 -
10000 -
5000 -
0 -
-5000 -
-10000 -
-15000 -
-20000 -
-QKnnn -
"' "*
* " ' ';*'
* 4 . W> ?
****** x * , ;;
' ,.'»'« ,.:-V-"-
* * +"+++* ;,* ,% f
* - * \ +<; :, '<
» % -"'*'
* "f ' ;? ^
* ,.f'r
600 700 800
X Variable 1
900
"1000
Figure 2-12. Residual plot showing random distribution of
X variable residuals.
22
-------
For example, in an electroplating process, do
pounds of cyanide waste correlate to square feet
plated, pounds of parts plated, number of parts
plated, or some combination of these three
measures of production?
In our sample data set we found that variations
in paper production explained only 23% of the
variation in sulfuric acid use. If understanding
where such variation comes from is important,
we could add other factors to our analysis'
such as variation in raw material quality, ambi-
ent temperature, and line speedby running a
multiple regression. Mathematically, the multi-
ple regression model can be expressed as:
y = BO + Bl XI + B2 X2 + B3 X3...
+ Bn Xn + error
where:
BO
Bn
y
x
n =
error =
y intercept of the line
the coefficient for Xn
chemical data
production data
the nth measure of production in
the model
the error or deviation of the actual
y value from the line BO + B1 X.
Multiple linear regressions are most often run
on statistical software packages such as Systat.
Running and interpreting multiple linear regres-
sion data requires more sophisticated under-
standing of regression techniques. Practitioners
should refer to regression analysis textbooks
and guides when conducting such analyses. See
Appendix B for references to text.
23
-------
Section 3
Five Examples of Systems That Use
Production-Adjusted P2 Measurement
Our research sought to investigate how produc-
tion-adjusted measures of P2 were being used at
the facility and process level in industry. We
identified five industrial facilities and analyzed
the production-adjusted P2 measurement meth-
ods that they use. Our objective was to select
case study facilities that represented larger and
small industry, as well as representing different
complexities of process. Given those objectives,
we identified a set of candidate firms and
invited them to participate in the case studies.
Prerequisites for participation also included
willingness to share data and host day-long site
visits.
The purpose of this section is to describe the P2
measurement methods used at the case study
facilities. In particular, this section focuses on:
» Measure of production-adjusted P2 used,
Data required for the measurement method,
and
» Function the measurement method serves at
the facility or corporate level.
Table 3-1 summarizes the P2 measurement
systems at the case study facilities.
3.1 Greene Manufacturing, Connorsville,
Indiana
Greene Manufacturing Company is a metal fin-
ishing job shop in Connorsville, Indiana.
Greene employs roughly 130 persons. Its parent
company is headquartered in Racine, Wiscon-
sin. The Connersville Division is a direct sup-
plier to Ford and an indirect supplier to GM and
Chrysler. The company plates automobile and
light truck tubes, heaters, and other automotive
and nonautomotive parts.
Greene's pollution-prevention measurement
system is an offshoot of the company's quality
tracking system. Greene Manufacturing tracks
its manufacturing operations by measuring the
square footage of every part it plates. Data are
recorded on log sheets and tallied daily, weekly,
monthly, and yearly. Greene uses these data to
price its products and control its plating baths.
The data are also used to track pollution-pre-
vention projects.
3.1.1 Description of Facility P2 Measure-
ment System
Greene measures P2 by tracking daily chemical
use, hazardous waste, daily off-spec parts and
daily production in logbooks. The charts are
then entered into spreadsheets by an adminis-
trative staff person. That same person then
prints out charts showing the following metrics:
Weekly change in plating sludge and haz-
ardous waste per square foot plated, and
"If we measure it, then we can fix it."
Brad Crowe, Greene Mfg.
24
-------
Table 3-1. Summary of Information about Five Case Study Sites
Green ManufacturingP2 measurement calculated by staff using data from logs.
Positive Attributes
» Weekly calculation and communication provide
incentive for greater worker efficiency
» Used to communicate with management
a Comprehensivemeasures raw material use,
waste generation and production on daily basis
Integral to business decisions: pricing,
improvement projects
Negative Attributes
Data collection is labor intensive
Paper-abasedall of the data used in
measurement come from paper records
rather than computerized information
systems
Lucent TechnologiesIntegrated software generates reports on demand. Tracks process, cost,
chemical use data.
Positive Attributes
System integrates existing data in various
databases: no new data collection required
Design allows manufacturing staff to improve
production
» Data automatically updated
Hazardous and nonhazardous materials tracking
Negative Attributes
Does not track waste
Data not used at corporate level
Labor-intensive installation and setup of
system; proprietary to Lucent
IBMP2 measurement calculated by EHS staff using data generated from various databases.
Positive Attributes
Takes into account changing nature of product;
makes cross-facility comparisons possible
» No new data requirement
Off-spec product is not counted in output, so quality
improvements are reflected in P2 measurement
Negative Attributes
Reactiveonly gives feedback at the end
of the year rather than providing feedback
to operations during the year
Tracks; only hazardous waste
reductionsnot improved efficiencies of
chemical use
Wyeth AyerstP2 Performance Tracking System.
Positive Attributes
Makes cross-facility comparisons possible
No new data requirements
Off-spec product is not counted in output, so quality
improvements will be reflected in P2 measurement
Negative Attributes
Paper-basedall of the data used in
measurement come from paper records
rather than computerized information
systems
Reactivesystem only generated data at
the end of the year rather than providing
feedbabk to production operations during
the year
Erving PaperStatistical Process Control-Type System.
Positive Attributes
Simple measure for straightforward process
Comprehensivemeasures hazardous and
nonhazardous raw material use, waste, and
production on a daily basis
Meets multiple environmental management
needse.g., Toxics Use Reduction Act (TURA)
and Reasonably Achievable Control Technology
(RACT) reporting
» Integral to business decisionsdata used to
diagnose production quality problems to track
high-cost materials
" Incorporates qualityoff-spec product not included
in output, so quality improvements will be reflected
in P2 progress measure
Negative Attributes
» None observed
25
-------
« Weekly change in number of rejects per
plating barrel or rack or per pieces plated.
Management at Greene started charting all
rejects by type or reason. Brad Crowe, manager,
states: "if we measure it, then we can fix it."
The quality figures are scrutinized by manage-
ment and posted at the entrance to the manu-
facturing floor. According to the staff person in
charge of generating the production charts,
"People on the line and the lab technicians are
sensitive to the [quality data]it's personal to
them. If they see a reject, they want to do
something about it... they almost take quality
problems personally." As a result, Greene has
improved its key quality metric (durability of
part as measured by salt spray hours).
In addition to tracking rejects, Greene's mea-
surement system tracks raw material use (haz-
ardous and nonhazardous), waste (hazardous
and nonhazardous), and daily production levels.
Careful tracking of plating bath life is also a
part of Greene's quality program. They are
concerned with obtaining the maximum number
of parts possible before they need to change a
bath, but they also are concerned with finding
the point at which the bath is so exhausted that
the reject rate increases. Greene has increased
its chromium bath life from 20,000 square feet
plated before dumping the bath to being able to
plate 65,000 square feet before dumping. This
accomplishment was achieved by carefully
tracking number of square feet plated and
quality of resultant parts to determine the
maximum bath life.
Greene staff said that even their chemical
vendor was shocked that the bath would last
that long. Since the chromium bath accounts for
85% of the chemical costs of the plating line,
this was a significant savings. Nevertheless,
Greene does not routinely calculate the savings
achieved by its P2 and quality control efforts.
3.1.2 How the P2 Measurement System Is
Used
Greene's P2 program began as a quality pro-
gram and still is motivated as much by quality
concerns as by concerns about costs of waste or
costs of raw materials. The pollution prevention
activities that Greene is most concerned with
are reducing its reject rate1 and reducing use of
costly plating chemicals. The P2 measurement
system is therefore used by manufacturing to
target and track efforts to reduce quality defects,
and chemical use, as well as tracking waste
reduction.
The results of P2 measurements at Greene are
broadly communicated. Weekly/monthly chem-
ical use and waste per unit-of-output are posted
for employees to see and are reviewed con-
tinuously by supervisors, the lab, and produc--
tion management. The P2 measurement system
is thus used as a driver for continuous improve-
ments as well as a way to track past efforts.
The P2 measurement system is used to estimate
pricing for different jobs. It allows Greene to
know costs associated with any given part that
they coat or plate.
When Greene's measurement system fails to
meet a new information need, the company
modifies the system. For example, when pro-
duction in the powder coating line increased
dramatically, waste also increased dramatically
due to changeovers between different colors. To
communicate the cost of lost raw material due
to color changeovers to production scheduling,
Greene instituted a measurement metric. The
new metric tracks powder changeover waste
Improved quality control and reduced reject rate is
sometimes not thought of as a P2 issue. But reductions of
off-specification product reduce the quantity of materials
that needs to be disposed of and reduce quantity of inputs
that a firm needs to purchase.
26
-------
pounds per total pounds of paint used. The pro-
duction scheduling group has used the metric to
set and achieve raw material cost and waste
generation reduction goals.
3.2 Lucent Technologies, Merrimack
Valley, Massachusetts
Lucent Technologies' Merrimack Valley site
(formerly AT&T) is a large manufacturer of
hybrid circuits, circuit packs, and other com-
puter equipment. For this case study, we fo-
cused on the semiconductor fabrication opera-
tions. These operations involve processing a
silicon substrate through a multistep process.
The basic process involves applying a pattern
onto the substrate by laying down the pattern
for the circuit and either etching the pattern into
the substrate or plating the circuit onto the
substrate. A resistant coating is used to define
the pattern of the circuit and is then stripped
from the substrate once the circuit is defined. A
substrate may make many passes through dif-
ferent etch and strip processes as layers of cir-
cuitry are built up on it.
3.2.1 Description of Facility P2
Measurement System
Lucent Technologies' Merrimack Valley Plant
began to develop a production-adjusted mea-
surement of facility P2 as a response to require-
ments to report under Massachusetts' Toxics
Use Reduction Act (TURA), which requires
that facilities report a unit-of-product along
with quantities of chemical use and waste
The process of developing and imple-
menting Lucent's P2 measurement
system had other benefits including
putting valuable information about costs
and chemical use into the hands of
process engineers.
reduced. They found that it was labor intensive
to manually calculate a unit-of-product and
consumption every year, so the plant had AT&T
Bell Labs work on a software package that
integrates data from production lines, corporate
systems, and facility-level systems to generate
the TURA-required measure. The resultant
software tracks the number of substrates that go
through different production process steps and
uses "number of substrates processed" as the
unit-of-product with which to adjust P2 mea-
sures. This unit-of-product measures the
number of passes that a given substrate makes
through process steps rather than merely
measuring the number of hybrid circuits that are
produced. Not all hybrid circuits require the
same number of passes through different
process steps. The Lucent software uses an
existing bar-code scanning system to calculate
the number of passes that substrates make
through production processes.
Lucent Technologies set up separate measure-
ments of P2 for each of its 10 production units.
In hybrid : circuit fabrication (the area we
focused on for this case study), Lucent uses the
number of substrates processed as the unit-of-
product with which to adjust P2 measurement.
They measure annual reductions in usage of
SARA chemicals and hazardous waste per
number of substrates processed.
The software that was developed to generate
Lucent Technologies P2 has the following
characteristics:
Data are automatically updated weekly.
This allows engineers and managers to keep
up with maintaining accurate measures of
P2 even when process lines and products
are changing often.
Data on materials. The UNIX-based soft-
ware is linked to both site and corporate
data tracking systems, including material
27
-------
safety data sheets (MSDSs) and production-
related data. This provides relatively simple
access to complex sets of information about
a particular product as well as a com-
prehensive view of waste reduction in dif-
ferent process lines.
» Product data. The P2 measurement soft-
ware tracks the number of substrates that go
through different production process steps
using the existing bar-code scanners at the
facility.
Cost data are obtained from an existing on-
site cost tracking system.
» Data collection from automated barcode
system allows for efficient measurements of
different product lines.
3.2.2 How the P2 Measurement System Is
Used
Merrimack Valley's P2 measurement system is
driven by the need to report to the State of
Massachusetts on reductions in hazardous waste
and chemical use under the State's TURA.
These measurements are calculated only once a
year.
However, Lucent found that the process of
developing and implementing their P2 measure-
ment system had other, more immediate, bene-
fits. Chief among these benefits was the new
ability of facility engineers to access the fol-
lowing types of information:
Production quantities,
« Withdrawal of chemicals from a central
storeroom,
Yield information,
Design information on the product,
« Information on how a specific hybrid circuit
moves through the product line, and
Cost of product at any stage of its pro-
duction.
Engineers can click on any data outlier (e.g.,
excessive use of a chemical) and find out what
products were going through the process at that
time and find design information on those
products. This is a powerful way of under-
standing process fluctuations and minimizing
variations in process conditions. Thus, the need
for P2 measurement resulted in installation of
a system that allowed better process control and
better understanding of process costs as well as
better tracking of P2 projects.
Not all of the process engineers take advantage
of this information, and the menu-driven soft-
ware was unfamiliar to some of the engineers.
But others have taken to the system "like ducks
to water," according to the environmental
engineering department. Three examples of
how the system has been used by process engi-
neers are for targeting chemical use reduction
efforts, to target change in operational strat-
egies, and to demonstrate P2 results on a given
process line.
3.3 IBM, Burlington, Vermont
The IBM Burlington facility employs 6,600
employees in manufacturing memory, logic, and
specialty chips using 1, 4, and 16 Mega-bit
technology. The manufacture of memory and
logic chips involves roughly 70 to 100 distinct
production steps. Wafers begin as raw silicon
and are processed through a variety of diffusion,
ion implantation, photolithography, etching,
metalization, and deposition steps before being
diced into individual chips. These chips are then
IBM's P2 measurement system is
important to stakeholder communication.
28
-------
mounted in modules and sold to internal (IBM)
and external customers.
3.3.1 Description of Facility P2 Measure-
ment System
IBM developed their P2 measurement system in
response to a desire at the corporate level to
track P2 progress and a concern that existing
measures of P2 might not accurately reflect
progress in the highly dynamic semiconductor
industry. The company's major concern in
developing a P2 measurement method was to
ensure that it not only accurately captured
reduction in waste but also captured the many
improvements in IBM's products from year to
year. This is important to a semiconductor
manufacturer because from year to year a
manufacturer produces products that remain
descriptively similar like "chips" and "mod-
ules," but the products nevertheless increase in
complexity so much that they are functionally
not the same product from, year to year. The
dynamic nature of IBM products makes mea-
surement of P2 progress for the company dif-
ficult. IBM therefore developed a combined
"performance index" that consists of a weighted
aggregate of the total number of bits, total
number of circuits, and total number of masks
produced. The numbers of bits, circuits, and
masks are weighted by the contribution they
made to sales from the facility. This per-
formance index is the unit-of-product with
which IBM adjusts its P2 measurements.
One criterion for the development of IBM's P2
system was that it must use existing data, but in
the past year a question has arisen as to whether
the necessary information for the current system
would continue to be available. This potential
problem arises from the fact that the IBM
system (like many of our case study systems)
relies on data collected by departments for
preexisting purposes. If the original purpose for
collecting the data is eliminated, then the data
will have to be collected exclusively for the
purposes of P2 measurement.
3.3.2 How the P2 Measurement System Is
Used
The primary purpose of IBM's P2 measurement
is stakeholder communication. IBM makes its
measurement of P2 on an annual basis and
provides that information to government regu-
lators. Since the P2 measurement system had
not been instituted IBM-wide in 1995, it was
not included in the IBM 1995 Corporate
Environmental Report.
3.4 Wyeth-Ayerst, Rouses Point, New York
Wyeth-Ayerst is a division of American Home
Products Corporation, an international Fortune
100 company manufacturing pharmaceuticals
and health jcare products. Our case study site
was Wyeth's Rouses Point facility in New
York. This site has approximately 1,200 em-
ployees. It focuses on production of both pre-
scription and over-the-counter pharmaceutical
products In addition, the Rouses Point site
handles some nonrecurring laboratory research
and development operations.
A major driver at Wyeth facilities is the desire
to be the lowest-cost producer of Wyeth prod-
ucts. Rouses Point has implemented a variety of
cost containment measures including efforts to
reduce cycle time and improve inventory man-
agement with a strong focus on production.
costs. !
Wyeth wanted to be able to assess the
success of their P2 programs. Gross
waste/emission data do not provide a
clear enough picture of the firm's
progress.
29
-------
3.4.1 Description of Facility P2
Measurement System
Wyeth's internal "P2 Performance Tracking
System" measures P2 at the division level. For
its manufacturing operations, Wyeth uses
kilograms of product as the unit-of-product with
which to calculate a production-adjusted mea-
sure of P2. For its laboratory operations, num-
ber of hours worked by staff is the unit used.
3.4.2 Uses of Facility P2 Measurement
System
At a corporate level, Wyeth wanted to be able to
assess the success of their P2 programs, and
gross waste/emission data do not provide a clear
enough picture of the facility's progress. There-
fore, they developed their internal "P2 Per-
formance Tracking System."
Annually, corporate environmental staff collect
hazardous waste data from Wyeth facilities in
the United States. They then calculate pro-
duction-adjusted measures of P2 for each divi-
sion and use these figures to write a corporate
annual P2 report. This is distributed to approx-
imately 100 people at all facilities, including
associate engineers, operations managers, en-
vironmental managers, and research managers.
The P2 Performance Tracking System is not
used by site personnel to improve operations on
a day-to-day basis.
3.5 Erving Paper, Erving, Massachusetts
Erving Paper is a privately owned manufacturer
of absorbent paper towels, tissues, wrapping
paper, and printed napkins. The company em-
ploys 300 people at three facilities in Miami,
Florida; Green Bay, Wisconsin; and Erving,
Massachusetts. The Erving facility employs 150
people and operates,? days per week, 52 weeks
per year.
Erving's measurement system evolved
from its quality assurance program. The
data help the firm spot process
problems as well as comply with
environmental regulations.
The Massachusetts facility operates continu-
ously. Their process involves pulping used
paper, and bleaching the pulp. During the pulp-
ing and bleaching process, sodium hydroxide
and sulfuric acid are used to modulate the pH of
the pulp for optimal results. The pulp is then
distributed onto screens and conducted through
rollers and driers to produce rolls of recycled
paper. These rolls undergo further finishing at
the Erving Paper facility or are sold directly to
customers.
3.5.1 Description of Facility P2
Measurement System
Erving's measurement system evolved from its
quality assurance program. Erving measures P2
by looking at chemical use reduction as well as
waste reduction. The production and chemical
use data that Erving uses to generate P2 mea-
surements were originally collected as part of
the company's statistical process control (SPC)
program. The data were used to determine pro-
cess trends, reduce process variation, and allow
for greater operator control over the process.
While no longer used for that purpose, the data
are still collected and used by the manufactur-
ing and environmental departments.
Chemical use is measured each morning by
either measuring levels inside tote tanks or
reading meters on pumps dedicated to specific
bulk tanks. Tote tank measurements are con-
verted from changes in the level in the tank in
inches to gallons used per day. Chemicals are
measured at the following intervals:
30
-------
Bulk Chemicalsuse measured daily
alum, sodium hypochloride, sodium hy-
droxide, sulfuric acid, and wet strength
resin;
Tote Chemicalsuse measured dailyanti-
foam chemical, continuous felt cleaner (pro-
prietary cleaneraliphatic hydrocarbon
with approximately 10% emulsifier), anti-
dusting agent, solvent, optical brightener;
and
Low-volume/low-cost chemicalsuse mea-
sured bi-weekly.
Production is measured each day in pounds of
paper. Erving's product may vary in brightness
or weight, but the qualities of paper are not sig-
nificantly different from the standpoint of
chemical use or waste production. Pounds of
off-spec paper are also measured daily.
3.5.2 Uses for the P2 Measurement System
at the Facility
The chemical use data collected under Erving's,
P2 measurement system are no longer used for
SPC. However, the use data are reviewed by
Erving's: Technical Director, who looks for
large daily variations. This allows Erving to
spot problems such as pump failure or operator
error and helps Erving target parts of their
process for cost control. Chemical use data are
also used to determine when to reorder chem-
icals to fill depleted storage tanks. The use data
also give them the information they need to
comply with their State air pollution reasonably
achievable control technology (RACT) require-
ments, State TURA reporting, and Federal TRI
reporting.
31
-------
Section 4
Results Obtained by Correlating the Production-Adjusting Units Used and
Pollution or Chemical Use for the Five Case Study Sites
The primary focus in this project was to identify
and verify how well the unit-of-product used in
pollution prevention measures at our five case
study facilities explained variation in key waste
streams or chemical usage at the facility. In
doing so, we were able to provide a preliminary
indication of the usefulness of that unit-of-
product in similarly situated facilities.
For each system we looked at, we applied the
method of analysis presented in Section 3. This
provided insight into the unit-of-product that
the facility uses to measure P2. This analysis
achieved two objectives:
1. Provided the case study facilities with a
better understanding of their measurement
accuracy; and
2. Tested the application of the statistical and
graphical verification method using real
rather than hypothetical data.
In this section, we present the results of the
verification of the different units-of-product
(i.e., how well they explained variations in
chemical use and chemical waste). The units-of-
product that we analyzed are summarized in
Table 4-1.
4.1 Greene Manufacturing Company, Inc.
RTI and Greiner Environmental examined the
correlation between Greene Manufacturing
Company's unit-of-product and two high-vol-
ume chemicalszinc and sodium cyanide. The
investigation was done for the firm's two major
metal plating operationsthe rack line and the
barrel line. Greene uses square feet plated as its
unit-of-product.
The analysis found that both zinc and sodium
cyanide were strongly correlated with square
feet plated for both the rack and the barrel line.
4.1.1 Data Collection
In November 1995, RTI and Greiner Environ-
mental researchers visited the Greene Connors-
ville site to observe the operation and collect
data for the study. Greene provided RTI with
data on daily chemical consumption and square
feet processed for the company's rack, barrel,
anodizing, and phosphatizing lines. We per-
formed unit-of-product analysis on two major
chemicals (sodium cyanide and zinc) used in
the rack and barrel electroplating lines.
Chemical Data. Greene provided chemical data
for the years 1994-95. The chemical data came
from quality control records of chemical addi-
tions to the process baths.
Production Data. Greene provided daily data
for the number of square feet plated on the rack
and barrel plating lines.
4.1.2 Data Analysis
The analysis for Greene data examined the cor-
relation between chemical use (zinc and sodium
32
-------
Table 4-1. How Well Units-of-Product Explained Variation in Chemical Use and Waste
Generation
Did unit-of-product explain
variations in
Industry
Unit-of-product used
for adjusting
pollution-prevention
measurement
Chemical use
for key inputs?
Waste genera-
tion for key waste
streams?
Facility or
company-wide
measure or
process
specific?
Metal finishing
Square feet substrate
plated or coated
Yes
Paper recycling Tons of paper produced Yes
NAa
NA
Process-specific
Facility-wide
Semiconductor
fabrication
Electronics
production
Pharmaceutical
production
Combined unit-of-
product incorporates
number of memory
chips, logic chips, and
masks produced [as
surrogate for tech-
nological content of
product]; number of
module parts produced
Number of passes sub-
strate makes through
process
Kilograms of product
produced
Combined unit-
of-product cor-
related for some
chemicals, not
for others;
module parts
correlated for all
chemicals'3
Yes
Yes I
'
Number of bits (a
component of the
combined unit-of-
product) correlated
with one waste
stream; module
parts correlated
with same waste
streamb
NA
Yes
Facility-wide
Specific to each
product line
Specific to
individual
department
a NA = Not applicable.
b Results somewhat uncertain; see Section 4.3.3 for full discussion.
cyanide) and Greene's unit-of-product (square
feet plated).
Rack LineZinc and Sodium Cyanide. The
histograms of weekly pounds zinc and pounds
sodium cyanide use adjusted by square feet
plated are presented in Figures 4-1 and 4-2.
Note that both figures have normally shaped
distributions. Time series plots were also pre-
pared to see the variation in chemical use per
1,000 square feet plated over time. All things
being equal, one would expect a random time
series patternas opposed to an increasing
pattern or decreasing pattern.2 The time series
plots in Figures 4-3 and 4-4 both have random
patterns.
Lastly, scatter plots of square feet plated on the
x and chemical use (pounds of sodium cyanide
and zinc) on the y-axis were prepared (Figures
4-5 and 4-6). Best-fit regression lines were
Constantly increasing or decreasing trends are indicative
of unstable processes. It would be next to impossible to
find a correlated unit-of-product for an unstable process.
33
-------
Figure 4-1. Weekly pounds of sodium cyanide per 1,000 ft2
plated histogram (rack line).
O O3
T
Figure 4-2. Weekly pounds of zinc used per 1,000 ft2 plated
histogram (rack line).
squared statistic equals ~0.74
inferring that square feet plated
accounts for 74% of the vari-
ation in sodium cyanide use.3
The equation of the line (y =
6.24x + 8.42) indicates that over
the 24-month time period 1994-
95, the average pounds of so-
dium cyanide use per 1,000
square feet plated equals 6.24.
Researchers found a disturbing
pattern among the regression
residuals for both sodium cya-
nide and zinc (see Figures 4-7
and 4-8). Standard statistical
practice requires that regression
residuals have a constant, non-
spreading pattern. A spreading
pattern can negate a regression
analysis' results. The residuals in
Figures 4-7 and 4-8 become
more negative and more positive
as the number of square feet
plated increases.
Researchers-felt that the spread-
ing pattern was due to the tre-
mendous variation in the weekly
production and chemical use
dataand decided to examine
monthly use and production data
to see if they also exhibited the
spreading pattern. Analysis of
monthly data eliminated the
residual spreading patternsee
added to each scatter plot as were R-squared
values and equation for the line. Greater R-
squared values, i.e., closer to 1.0, are indicative
of greater correlation. The scatter plots and
regression lines indicate that both sodium cya-
nide and zinc usage are correlated with square
feet plated. The Figure 4-5 linear regressions R-
3These regression results are statistically significant. The
P-value for the slope of the regression (pounds of NaCN
use per millions of modules) equals 1.1E-34. P-values
<0.05 are generally considered statistically significant.
The P-value tells us that we can be 99.99% confident that
the relationship between cyanide use and square feet
plated is not random.
34
-------
100
120
Figure 4-3. Weekly pounds of sodium cyanide used per 1,000
ft2 plated time series plot. '
120
Figure 4-4. Weekly pounds of zinc per 1,000 ft2 plated time
series plot.
ter f y=6.2x^8.4288
_ *>&/ N J -*Vi«A«^t f
y * " R2> 0.7387 \
square feet plated
Figure 4-5. Scatter plot showing relationship between weekly
pounds of sodium cyanide and square feet plated
(rack line).
scatter and residual plots for so-
dium cyanide and zinc in Figures
4-9 through 4-12.
In summary, square feet plated is
strongly correlated to both so-
dium cyanide and zinc. Monthly
data produced similar results as
weekly data but without the
problem of spread in the regres-
sion residuals.
Barrel Line Analysis. Graphical
methods were also used to exam-
ine the correlation between two
barrel line chemicals (sodium
cyanide and zinc) and square feet
plated. Histograms of monthly
use data per 1,000 square feet
plated are presented in Fig-
ures 4-13 and 4-14. Both histo-
grams appear to be bell-
shapedan early indication that
correlations for both chemicals
are likely to be strong.
Time series plots of monthly
barrel line sodium cyanide and
zinc were also prepared. Note
the large increase in both sodium
cyanide (Figure 4-15) and zinc
(Figure 4-16) per 1,000 square
foot plated in June 1995. The
June 1995 data points corre-
sponded to several weeks during
which Greene did little, if any,
plating, leading to low efficiency
of use.
Scatter plots with thousands of
square feet plated on the x-axis
and sodium cyanide and zinc on
the y-axis were also prepared
(Figures 4-17 and 4-18). Best-fit
regression lines were added to
35
-------
1600
square feet plated
Figure 4-6. Scatter plot showing relationship between pounds of
zinc and square feet plated (rack line).
250
200
150
100
50
I
s
I
s"
square feet plated
8
S"
Figure 4-7. Weekly pounds of sodium cyanide per square foot
plated residual plot (rack line).
SCO-
200-
1-200
-400-
-600-
L , ,
!:', ' .# ' * ;' ' * * \
*** * ' * v**1^ *«
,,t * * * ^ *
\ »
s s s §
s a" I"
* *" -,
* f /f*, J i * * ""/^ ^
I * * 1 * ^f* * 1 A
* * * "^ * ~
«. # ^ ^
% * '*..'
>f*fF'~ ^ #
^.. .,. ,!' - . . i
g g s g g
1 s" s 1 g
square feet plated
Figure 4-8. Weekly pounds of zinc per square foot plated
residual plot (rack line).
36
-------
. ,. square footage
Figure 4-9. Monthly pounds of sodium cyanide per square foot
plated scatter plot (rack line).
S
square footage
Figure 4-10. Monthly pounds of zinc per square foot plated
scatter plot (rack line).
500 -
400 -
300 -
200 -
100 -
0 -
-100 -
-200 -
-300 -
-400 -
-500 -
c
,<-^ ^t'afe' * - ' ' t 'f/? ^
» , / * ' ""
"5-1 "* " > X * *t V> * s ^
>** "^". ^ X: t'*' "^ i v
"^^^^ f f "~* I ,~ t
\ ',,*-+, '*** 'K^/'r" ,v:^ ^
=* 4~ " s \ 7" s ,.>»*_;
^ N " «; ^Wf? ^ A. ij-t.^* *
"- ."!*... . ...
, o o o o . o o
888 8 8 @
o o o Q^ o^ o"
10 ° ^ s a s
square footage
r '
x
**
^ ~~
5
W
Figure 4-11. Monthly pounds of sodium cyanide per square foot
plated residual plot (rack line).
37
-------
zuuu -
1500 -
1000 -
1 500 -
A .
-500 -
-1000 -
C
- : . . ' ' :
.
: :' : >..**»» ,-,,><
' f / / ^/ / / s/ ^ /fi *" / i
3OOQOOOC
§ 8.8 88 S . g
S S 8" § S 8 ' S
square footage
Figure 4-12. Monthly pounds of zinc per square foot plated
residual plot (rack line).
Figure 4-13. Monthly pounds of sodium cyanide per square foot
plated histogram (barrel line).
Figure 4-14. Monthly pounds of zinc per square foot plated
histogram (barrel line).
38
-------
10 15
Months
Figure 4-15. Monthly pounds of sodium cyanide per square
foot plated time series plot (barrel line).
10 15
Months
Figure 4-16. Monthly pounds of zinc per square foot plated
time series plot (barrel line). :
Square Feet Plated
Figure 4-17. Scatter plot showing relationship between monthly
sodium cyanide use and square foot plated (barrel
line).
39
-------
8
Square Feet Plated
Figure 4-18. Scatter plot showing relationship between
monthly zinc use and square foot plated (barrel
line).
used data from two of the semi-
conductor process lines.
Lucent Technologies provided
chemical use and production data
on two production processes that
remove the photo-resist from a
Lucent Technologies product.
Process 1 is Lucent's high-
volume line. This high-volume
line has three identical work-
stations. Process 2 is Lucent's
low-volume production line. It
comprises a single work station.
Lucent uses substrates produced
as the unit-of-product to adjust
measurements of change in gly-
col ether use in both processes
(Box 4-1).
each scatter plot as were R-squared values and
an equation for the line.
The scatter plots and regression lines indicate
that square feet plated is strongly correlated
with both sodium cyanide and zinc (R-squared
and P-values are 0.77 and 1.5E-08 for sodium
cyanide and 0.85 and 1.4E-10 for zinc). Resid-
ual plots of monthly data for both chemicals
have random patterns.
4.1.3 Findings
The unit-of-product, "square feet plated," is
strongly correlated to chemical use for the two
chemicals analyzed in this study. This is true for
both process lines examined.
4.2 Lucent Technologies
We followed the four-step data analysis pro-
cedure outlined in Section 3 for the data pro-
vided to us by Lucent Technologies. For the
purposes of examining the unit-of-product, we
Chemical Use Data. Weekly glycol ether data
were not available since use is not measured at
the process level. Instead Lucent Technologies
provided the withdrawal data for photo-resist
stripper which contains 55% glycol ether from
the chemical storage area for each process.
Because a small amount of chemical remains in
inventory either in the process or stored near the
process, weekly chemical withdrawals represent
an estimate of weekly chemical use.
Production Data. Weekly production data
(number of circuits and number of substrates)
were available for each process (and for each
Process 1 workstation). Lucent Technologies
defines a substrate as the number of passes a
"widget" makes through a workstation. If the
same widget makes three passes through a
workstation, throughput through the process
equals three substrates. Lucent Technologies
defines circuits differently. The number of cir-
cuits on a widget is constantno matter how
many times a widget passes through a work-
station, the. number of circuits on it remains
40
-------
Box 4-1.
Sources of Data for Analysis of Lucent Unit-of-Product
Process 1
1
work
tation
1
work
station
work
static
Process 2
chemical input data
measured as storeroom
withdrawals
,.'production data (substrates
and circuits measured at
each workstation)
unchanged. For example, if a widget has two
circuits on it and the widget passes through a
workstation three times, throughput through the
process equals two circuits.
Data Entry. Weekly data were entered into an
Excel spreadsheet in the following manner:
Week of year for 1994;
Gallons of photo-resist used each week
(taken out of chemical storage);
Process 1: number of substrates produced
each week for the three work centers;
Process 1: number of circuits produced each
week for the three work centers;
Process 2: number of substrates produced
each week; and
Process 2: number of circuits produced each
week.
Data Manipulation. Data were manipulated to
generate basic measures of process efficiency
(Table 4-2). The following manipulations were
performed:
Convert gallons of photo-resist stripper
used per week to pounds of glycol ether
used per week (stripper is 55% glycol ether
at9.1741b/gal);
Divide weekly chemical use by weekly
production (e.g., glycol ether pounds used
per substrate manufactured); and
Calculate 1994 average pounds of chemical
use per unit-of-output (e.g., number sub-
strates processed and number circuits pro-
duced).
How do Process 1 and Process 2 compare in
chemical use efficiency? Using number of sub-
strates processed to adjust glycol ether use,
Process 1 is more chemical use efficient than
Process 2: 23.4 Ib glycol ether/substrate versus
53.1 Ib glycol ether/substrate. Using circuits to
adjust the data, the two processes appear to
have equivalent chemical use efficiencies: 5.0
Ib glycol ether/circuit versus 6.3 Ib glycol ether/
circuit.
Although our two metrics give different results,
Lucent engineers believe that substrates are a
better unit-of-product than circuits since sub-
strates count the number of passes a product
might make through the operation. Furthermore,
Process 1 is a high-volume line while Process 2
41
-------
Table 4-2. Glycol Ether Use per Unit-of-Product
Process 1
1994 total
Pounds of glycol ether/1 ,000
units
Process 2
1994 total
Pounds of glycol ether/1 ,000
units
Circuits
produced
13,898,841
5.0
640,316
6.3
Substrates
produced8
2,970,872
23.4
75,499
53.1
Glycol ether
use
69,631
4,011
a Lucent defines substrates as the number of passes a "widget" makes through a workstation.
is a low-volume lineengineers believe the
high-volume line is more use efficient because
less waste is created during startup and shut-
down of the line.
Our analysis thus far raises two interesting
questions. First, which factor (substrates or
circuits) is a better unit-of-product to adjust P2
measures for Process 1? For Process 2? Second,
is Process 1 more use-efficient than Process 2?
To answer these questions, the data supplied
must be analyzed in greater depth.
4.2.1 Process 1 Analysis
The distribution of the data and descriptive sta-
tistics shed light on Process 1's unit-of-product
(substrates).
Histogram and Descriptive Statistics. To look
at the distribution of the data, a histogram was
prepared (Figure 4-19). All but two data points
fell into the range between 0 and 60. Two data
points were greater than 75. These points had
values of 120 and 3,144. Researchers checked
to see if the outlier (with a value of 3,144) was
due to a data entry error. It was not.
To examine the effect of this outlier, we com-
piled a set of descriptive statistics (an Excel
data analysis option). These are presented as
Table 4-3. The descriptive statistics were run on
the entire data set (51 weeks) and on the data
set with the outlier removed (50 weeks). Note
the large change in mean, standard error,
standard deviation, and 95% confidence level.
4.2.2 Plot Time-Series and Moving Average
Time series plots of glycol ether use per unit-of-
product were generated, using both substrates
and circuits as the unit-of-product (Figures 4-20
and 4-21).
Since some glycol ether is in inventory in the
workstation and in storage areas on the pro-
duction floor! a moving average was also cal-
culated and plotted. In this moving average plot,
each week's chemical use is the average of the
current week and the two preceding weeks. A
42
-------
Bin
Figure 4-19. Weekly glycol ether use (Ib) per substrate
histogram (Process 1).
Table 4-3. Process 1 Descriptive Statistics for Glycol
Ether Use per Substrate
Glycol Ether Use/Substrate (thousands)
Full data set
Mean
Standard error
Median
Mode
Standard deviation
Minimum
Maximum
Count
Confidence level (95.0%)a ,
86
61
20
0
437
0
3,143
51
123
Outlier removed
25
2.7 !
19
0
19
o :
' 120 '
50 j
5.4
a This confidence level indicates that the true mean is 95 percent likely 'to
be between 86±123. This is too large a range to be m'eaningfui. Once
the outlier is removed, the data become more manageable.
considerably off the scale. This
is the outlier (value = 3,144).
The plots indicate that, on the
whole, glycol ether use per sub-
strate is fairly constant over time
with two exceptions (week 1 and
week 36).
We aggregated the weekly data
into monthly data and prepared
time-series plots and performed
regression analyses. The time
series plot for both substrates
and circuits per unit-of-product
are presented as Figure 4-22.
Scatter plots of Process 1 depict-
ing glycol ether use per substrate
and glycol ether use per circuit
were prepared (Figures 4-23 and
4-24). Best-fit regression lines
were added to each scatter plot
as were R-squared values and
equation for the line. R-squared
values closer to 1.0 are indica-
tive of greater correlation.
The scatter plots and regression
lines indicate that substrates but
not circuits are correlated with
glycol ether usage. The Process
1 Substrate Plot linear regression
R-squared statistic equals -0.42
This infers that the number of
substrates processed accounts for
42% of the variation in sodium
cyanide use.4 The equation of
moving average plot tends to smooth out large
swings in a data set.
Figures 4-20 and 4-21 present the time series
plot (actual) and moving average plot (forecast)
of glycol ether use per substrate and per circuit.
Notice that in Figure 4-20 one data point is
''These regression results are statistically significant. The
P-value for the slope of the regression (pounds of glycol
ether use,per substrate equals 0.02). P-values <0.05 are
generally considered statistically significant. The P-value
infers that we can be 99.98% confident that the relation-
ship between glycol ether use and -substrates is not ran-
dom. ;
43
-------
Figure 4-20. Weekly glycol ether use per substrate time-series
moving average plot (Process 1).
5 3 3 C? 3 S
Figure 4-21. Weekly glycol ether use per circuit time-series
moving average plot (Process 1).
Use/100 Circuits
» Use/1,000 Substrates
Figure 4-22. Monthly glycol ether use per unit-of-product time
series plot (Process 1).
44
-------
the line (y = 0.0116x + 2,932)
indicates that over the 12
months in 1994 the average
number of pounds of glycol
ether use per substrate equals
0.0116 Ib (or 11.6 Ib per 1,000
substrates processed).
Process 1 Findings. It appears
that substrates processed track
glycol ether use well and cir-
cuits do not. This result is con-
sistent with the predictions of
the Lucent engineers who set up
the P2 measurement system.
Based on this analysis, using the
number of substrates produced
will provide Lucent Technolo-
gies with an accurate picture of
pollution-prevention progress
(as measured by change in quan-
tity of glycol ether used) in Pro-
cess 1.
4.2.3 Process 2 Data Analysis
The next step in the analysis is
to review Process 2 data more
carefully.
Histogram and Descriptive
Statistics. To look at the distri-
bution of Process 2 use/output
data, a histogram was prepared
(Figure 4-25). Data from Pro-
cess 2 have greater spread than
the data from Process 1. This is
because Process 2 is run inter-
mittently and nearly half of the
data points had values of zero.
Descriptive statistics were ap-
plied to the data set (Table 4-4).
The large difference between
the median, and mean indicate
6,000 -:,;.-; -. r,^z,5~t«
500,000 1,000,000 1,500,000 2,000,000
circuits
Figure 4-23. Monthly glycol ether use per circuit scatter plot
(Process 1).
50,000 100,000 150,000 ; 200,000 250,000 300,000 350,000 400,000 450,000
substrates
Figure 4-24. Monthly glycol ether use per substrate scatter plot
(Process 1).
10--
' 1M I
S-f [BiK^n'igB, , fJLj
M-JB-4J
o to
-------
the data are not normally dis-
tributed, an indication that a
correlation between chemical
use and unit-of-product will not
be found.
Plot Time-Series and Moving
Average. Figure 4-26 presents
the time series plots (actual) and
moving average plots (forecast)
for glycol ether use per sub-
strate. Notice that for the first
half of the year (first 26 weeks)
both actual and forecast glycol
ether use per substrate varies
from zero to 200 lb/1,000 units.
In the second half of the year,
however, the variation in use per
substrate increases dramatically.
One would expect no change in
average use per unit of output
over time (assuming no major
changes to the production pro-
cess or product runs through the
process). This large increase in
variation makes using substrates
as a normalization factor prob-
lematic.
Compare Substrates and Cir-
cuits as Adjusting Factors. To
examine the difference between
substrates and circuits as a unit-
of-product, a time series and
moving average plot of glycol
ether use per circuit was gen-
erated (Figure 4-27) in order to
compare it to the time-series
plot for Process 1 (Figure 4-21).
Process 2 substrate and circuit
plots exhibit some differences.
First, the first data point in the
circuit plot has an extremely
large value that was not seen in
Table 4-4. Process 2 Descriptive Statistics; for
Glycol Ether Use per Substrate
Mean
Standard error
Median
Mode
Standard deviation
Minimum
Maximum
Count
Confidence level (95.0%)
91
19
26
0
131
0
424
48
38
Figure 4-26. Glycol ether use per substrate time-series moving
average plot (Process 2).
100
Figure 4-27 Glycol ether use per circuit time-series moving
average plot (Process 2).
46
-------
the substrate plot. Second, while
circuit data for the second half
of the year vary more than the
first half of the year, the in-
crease in variation appears to be
less than that seen in the
substrate plot.
Scatter plots of Process 2 (low-
volume line) depict glycol ether
use per substrate and glycol
ether use per circuit (Figures 4-
28 and 4-29). Best-fit regression
lines, R-squared values, and the
equation for the line were added
to each scatter plot. R-squared
values closer to 1.0 are indica-
tive of greater correlation.
Process 2 Findings. Because
glycol ether use per substrate
per circuit changes significantly
halfway through the year,
neither metric could be char-
acterized as well-correlated nor-
malization factors.
4.2.4 Findings
2,000
4,000
6,000
substrates
8,000
10,000
12,000
Figure 4-28. Glycol ether use versus substrates scatter plot
(Process 2). ;
600
500
400
" 300
200
100
, s y = O.OQ24x*206.89j
R'2s
-------
4.3 IBM, Burlington, Vermont
RTI and Greiner Environmental worked with
staff at the IBM Burlington facility to assess
two possible units-of-product for the facility.
The first unit-of-product analyzed was IBM's
performance index which it uses for tracking P2
progress. The performance index consists of an
aggregate of bits,5 circuits,6 and masks data
weighted by percent revenue. We used IBM
data to construct a modified version of the
performance index, using number of bits and
number of circuits, weighted by percent reve-
nue. IBM's performance index is used to assess
P2 progress on an annual basis. The analysis
here uses monthly figures, since there were
inadequate annual data to conduct the analysis
(see Section 2.2 on the topic of number of data
points).
The second unit-of-product examined was num-
ber of modules. The analysis examined how
well number of modules explained chemical
usage and generation of one waste stream.
Modules are the final mounted chips, and there
are both bits and circuits on these mounted
chips. This alternative unit-of-product does not
account for the changing complexity of the IBM
products.
4.3.1 Data Collection
Chemical Data. IBM provided data for month-
ly SARA 313 chemical use over a two calendar-
year period 1993-1994 (Table 4-5). IBM uses a
computerized tracking system to monitor all
Table 4-5. Chemical and Production Data Provided by IBM
Chemical Dataa Production Datab
1. IPA (isopropyl alcohol) use
2. Xylene use
3. Ethylbenzene use
4. Cyclohexanone use
5. PGMEAuse
6. NBA (A/-butyl acetate) use
7. NMP (A/-methyl-2-pyrrolidone) use
8. Total of seven chemical uses listed above
9. PGMEA/cyciohexanone waste stream (IBM
internal waste stream #38)
1. Number of modules manufactured
2. Performance index
3. Number of bits manufactured (memory
product)
4. Number of circuits manufactured (logic
product)
a Chemical use data were Chemical Abstract System (CAS) number, monthly, pound totals for each chemical for
1993 and 1994. Chemical waste generation data were monthly shipment and beginning and ending inventory
data from the Chemical Distribution Center which manages both chemicals and waste.
b Production data were monthly totals for 1993 and 1994. The performance index is a combination of bits
manufactured and circuits manufactured weighted by the percent revenue from each product.
5Bits are the measure of production for memory products.
^Circuits are the measure of production for logic
products.
48
-------
chemical usage at the facility. Chemicals are
released to the production floor on an as-needed
basis. Little chemical inventory is held in
production areas except for some maintenance
and photo-resist chemicals. Since little inven-
tory is held on the production floor in pro-
duction tools or in storage, monthly chemical
withdrawals are a good proxy for monthly
chemical use at the IBM Burlington facility.
IBM provided RTI with 2 years of data on
monthly waste generation for a waste stream
known as PGMEA/cyclohexanone, and 1 year
of data for their general solvents waste stream.
Since the analytical method requires more than
12-data points; we were unable to use the
general solvents data. Other
waste data were also judged to
be inappropriate for this analysis
because they provided informa-
tion about waste inventory rather
than waste generation.
mance index, bits, circuits, and millions of
modules. Analyses for isopropyl alcohol (IPA)
use and PGMEA/cyclohexanone waste stream
are presented in the following sections. Sum-
mary results of analysis of the seven chemical
uses and the one waste stream following the
detailed analyses are presented later on page 57
(Table 4-8).
IPA Analysis. The histograms of IPA use
adjusted by the performance index (PI) and IPA
use adjusted by millions of module parts are
presented in Figures 4-30 and 4-31.
Time series plots were also prepared to see the
variation in IPA use per unit-of-product over
IBM's chemical waste data are
tracked in the hazardous-waste
storage and shipping area where
monthly inventory records are
maintained.
Production Data. IBM provided
researchers with monthly data
for the number of modules and
monthly data for bits and cir-
cuits. In addition, the company
provided information about the
percentage revenue attributable
to these products.
4.3.2 Data Analysis r
Analyses were performed on
chemical use data and the
PGMEA/cyclohexanone waste
stream data using four possible
units-of-product: IBM's perfor-
6--'
5---T ,
- ">i>,T. ^
4-F--X
g"
2--
1 --;
0
-4-
-P
30000 60000 90000 120000 150000 180000 More
Bin
Figure 4-30. Monthly IPA use per performance index unit
. histogram, i ,
Figure 4-31. Monthly IPA use per million modules histogram.
49
-------
time. All things being equal, one
would expect a random time
series patternas opposed to an
increasing pattern or decreasing
pattern.7 The time series plots in
Figures 4-32 and 4-33 have a
fairly random orderalthough a
somewhat cyclical pattern
emerges in months 13 to 22 in
Figure 4-32 (IPA use per perfor-
mance index unit).
Lastly, scatter plots of the unit-
of-product on the x-axis (per-
formance index and module
parts) and IPA use on the y-axis
were prepared (Figures 4-34 and
4-35). Best-fit regression lines
were added to each scatter plot
as were R-squared values and
equation for the line. R-squared
values closer to 1.0 are indi-
cative of greater correlation.
The scatter plots and regression
lines indicate that IPA use and
module parts are better cor-
related than IPA use and the
performance index. The Figure
4-35 linear regressions R-
squared statistic equals -0.58
inferring that the number of
module parts produced accounts
for 58% of the variation in IPA
use. The equation of the line (y =
7,014x + 37,033) indicates that
over the 24-month time period
1993-94, the average pounds of
IPA use per million module parts
equaled 7,014.
200,000 -p
180,000 --
_160,000 --
6:140,000 -
gj 20,000 --
<100,000 -
- 80,000 - -
60,000 --
40,000 --
20,000 -
0
10 15
Month
Figure 4-32. Monthly IPA use per performance index unit
time series plot.
10 15
Month
Figure 4-33. Monthly IPA use per million modules time series
plot.
itu.uuu -
120,000 -
g 100,000 -
£
g 80,000 -
^ 60,000 -
O
< 40,000 -
20,000 -
0 -
"?>
,, c"* / '", "r/> / * "
* ' ' '^ *A
** * * - " , i r-
* A ' X'
* I ' ' ^ ^
* ^ ^ x ^ ~~N
i-^ ^ ^ - '
« ' " , < yil724Qx4-581l7
. ^ _ .^ ( "\ R2 = 0.2327; ^ x ^
,-. /. - - , .-.-... .-.- ; ..,....-,- | . */
1 I r^ : I
0 0.5 1 1.5 2 2
Performance Index
Figure 4-34. Monthly IPA use per performance index unit
scatter plot.
'Constantly increasing or decreasing trends are indicative
of unstable processes. It would be next to impossible to
find a correlated unit-of-product for an unstable process.
50
-------
160,000
_ 140,000 -j;
]T 120,000 -j*
o 100,000 -- .
2
£ 80,000 --
g" 60,000 --
^ 40,000 -:, .^^
- 20,000 -*
0.0 2.0 4.0 6.0 . 8.0 10.0 12.0 14.0
Modules (M)
Figure 4-35. Monthly IPA use per million modules scatter plot.
monthly PGMEA/cyclohexa-
none waste (in pounds) per per-
formance index and per million
modules are presented in Figures
4-36 and 4-37. Note that neither
histogram has a bell-shaped ap-
pearance. In addition both histo-
grams appear skewed and have
extremely high valuesan early
warning that correlations for
both units-of-product are likely
to be weak.
We ran regression tests on the component parts
of the IBM performance index to see if either
bits or circuits alone were correlated with IPA
use. The tests showed that while bits were cor-
related, circuits were not (Table 4-6). Because
bits and circuits are weighted together using
percent revenues at the Burlington plant, the
lack of circuit correlation negatively affects the
correlation between the performance index and
IPA use.
In summary, module parts have a stronger
correlation than the performance index to IPA
usage. Of the two performance index com-
ponents, only bits were strongly correlated to
IPA use over the 24-month time-frame 1993-
1994.8 Depending on site conditions at IBM
Burlington, the analysis results for IPA usage
could be proxies for the correlations (or lack
thereof) for IPA waste streams.
PGMEA/Cyclohexahpne Waste Stream
Analysis. Graphical methods were also used to
examine the potential correlation between
PGMEA/cyclohexanone waste stream data and
different units-of-product. Histograms of
8It is worth pointing out here that circuits were not
correlated to any of the eight chemicals nor to PGMEA
waste. This result is discussed in detail later in this report.
Time series plots were also pre-
pared to see the variation in
PGMEA/cyclohexanone waste stream per unit-
of-product over time. All things being equal,
one would expect a random time series
patternas opposed to an increasing pattern or
decreasing pattern. The time series plots in
Figures 4-38 and 4-39 have a fairly random
pattern.
Scatter plots of the unit-of-product on the x-axis
(performance index and modules) and the waste
stream on the y-axis were also prepared (Fig-
ures 4-40 and 4-41). Best-fit regression lines
were added to each scatter plot as were R-
squared values and an equation for the line.
The scatter plots and regression lines indicate
that neither the performance index nor module
parts correlate with the PGMEA/cyclohexanone
waste stream generation. Both R-squared values
are below 0.1 and neither P-value was <0.05.
The regression test on bits and circuits data
produced similar resultsneither was cor-
related to the PGMEA/cyclohexanone waste
stream. One possible source of error in the
analysis is that all of the unit-of-product data
were registered on the IBM calendar while
waste inventory was tracked on a standard
calendar. :
51
-------
Table 4-6. Results of Regression Analysis for IPA Use
IPA (Ib) vs. Bits (trillions) IPA (Ib) vs. Circuits (millions)
Equation of line
R-squared value
P-value
y = 1,389x + 51,645
0.4989
0.0001
= 175x + 61,051
0.0974
0.13765
u_
5 * *
4.,
O . a
2 "
"I * *
0
1
2000
4000
6000 8000
Bin
10000 More
Figure 4-36. Monthly PGMEA/cycIohexanone waste per
performance index unit histogram.
Figure 4-37. Monthly PGMEA/cyclohexanone waste per million
modules histogram. .
Based on prior work analyzing
units-of-product,' RTJ and
Greiner Environmental thought
that a time-related function
might be at work in the data.
Because the waste generated
during the manufacture of a
given batch of modules, bits, or
circuits was likely to enter the
waste inventory storage room
sometime after actual manu-
facturing, we chose to represent
this time delay by adjusting the
waste data by a single month.
For example, rather than using
the January modules information
and January PGMEA/cyclohexa-
none information as a data pair,
RTI decided to use January mod-
ules with February PGMEA/
cyclohexanone data. IBM Bur-
lington staff agreed that such a
move was a reasonable approxi-
mation of site conditions.
Histograms of delayed monthly
PGMEA/cyclohexanone waste
(in pounds) per performance
index and per million modules
52
-------
^ 12,000
S" 10,000
8,000 --
6,000 -; -4.~"^
ul 4,000
2,000 --;
10 15
Month
Figure 4-38. Monthly PGMEA/cyclohexanone waste per
performance index unit time series plot.
,>*- "~AiF'^;^ ^V^rT""*-- -
-= ' ^ ^A-s*- -^ W^ w -^^ ^ ^ ^ ^ ^ * * '*
j«S 2,000 4- ^ '-.^ --- -^ » ' -. V
2 1,500 , _ .
2 1,000 -:',,v^> ;^4
500 -k ^*v"" .
0
0 0.5 1 1.5 2 2.5
Performance Index
Figure 4-40. Monthly PGMEA/cyclohexanone waste per
performance index unit scatter plot.
are presented in Figures 4-42
and 4-43. Both histograms
appear bell-shapeda sign that
correlations may have improved
due to the delay function.
Time series plots were also pre-
pared to see the variation in
delayed PGMEA/cycloexanone
waste per unit-of-product over
time. As in the prior PGMEA/
cyclohexanone plots without a
delay function, the time series
plots of delayed waste exhibit a
random pattern (Figures 4-44
and 4-45).
Finally, scatter plots of the unit-
of-product on the x-axis (perfor-
mance index and million mod-
ules) and PGMEA/cyclohexa-
none on the y-axis were prepared
(Figures 4-46 and 4-47). Best-fit
regression lines were added to
each scatter plot as were R-
squafed values and equation for
the line.
The scatter plots and regression
lines indicate that only module
parts are significantly correlated
to delayed PGMEA/cyclohex-
anone waste (R-squared = 0.48;
P-value = 0.0002). The perfor-
mance index was not correlated
with delayed PGMEA/cyclohex-
anone waste (R-squared = 0.04,
P-value = 0.40). Researchers ran
regression tests on the com-
ponent parts of the IBM perfor-
mance index to see if either bits
or circuits alone were correlated
with delayed PGMEA/cyclohex-
anone waste. The tests showed
53
-------
S"
i
CO
1
I
CL
a
4,000 -
3,000 -
2,000 -
1,000 -
0.
:}; -'. ....' * * V
* * * * __!'" "T
"^ ^ ' ^
'-, . . __5- -r~?~~~~~\
'T: , ".;. -;* * * *- * -
y=118.1x-f 2412.6
Ra = 0.0825 ,
0 2.0 4.0 6.0 8.0 ' 10.0 12.0
Modules (M)
i »
/T
14
Figure 4-41. Monthly PGMEA/cyclohexanone waste per million
modules scatter plot.
Figure 4-42. Monthly PGMEA/cyclohexanone waste per
performance index unit histogram.
Bin
Figure 4-43. Monthly PGMEA/cyclohexanone waste per million
modules histogram.
54
-------
Figure 4-44. Monthly PGMEA/cyclohexanone waste per
performance index unit time series plot with 1
month delay.
§ 1,200
Figure 4-45. Monthly PGMEA/cyclohexanone waste per million
modules time series plot with 1 month delay.
5,000
4,500
S 4,000 -fc;
V 3,500
~ 3,000
(3 1,500 + *£
°- 1,000
500 +
0
5 2,500 -y, - «»Vr ,-'- ; * * . '*'<>,"^.
-------
5,000
4.500
S" 4,000
oT 3,500
1 3,000
^ 2,500
Q 2,000
| 1,500
o- 1,000
500
0
,y=285.1x+1271.7
' R2 = 0,4821
0.0
2.0
4.0
6.0 8.0
Modules (M)
10.0
12.0
14.0
Figure 4-47. Monthly PGMEA/cyclohexanone waste per million
modules scatter plot with 1 month delay.
that while bits were significantly correlated,
circuits were not (Table 4-7). Since bits and
circuits are weighted together using percent
revenues at the Burlington plant, the poor
circuit correlation has a negative effect on the
overall correlation of the performance index.
Thus, only by delaying PGMEA/cyclohexanone
waste by 1 month were researchers able to find
a correlated unit-of-product. Bits, a component
of the performance index, also correlated with
delayed PGMEA/cyclohexanone waste, but
neither the performance index itself nor the
circuits exhibited any significant linear relation-
ship.
Analysis of Other Chemicals.
Researchers ran analyses on all
of the chemicals listed in Table
4-5. Table 4-8 presents a sum-
mary of these analyses. The top
number in each cell represents a
linear regression R-squared sta-
tistic. Values closer to one indi-
cate greater correlation. The
lower number in each cell repre-
sents a regression P-value. The
P-value determines whether or
not the correlation between the
unit-of-product and chemical are
statistically significant. P-values
<.05 indicate a statistically sig-
nificant correlation. The lower the P-value (e.g.,
P<.001), the stronger the correlation. In most
cases, those cells with high R-squared values
also have statistically significant P-values. The
text in cells with statistically significant P-
values have been highlighted to make the chart
easier to read.
Examining the chart, one can make the fol-
lowing observations:
The strongest correlations were seen where
modules was used as the unit-of-product;
Circuits did not correlate with chemical use
or waste data;
Table 4-7. Results of Statistical Analysis for PGMEA/Cyclohexanone Waste
(Delayed)
Delayed PGMEA waste
(Ib) vs. Bits (trillions)
Delayed PGMEA waste
(Ib) vs. Circuits (millions)
R-squared value
P-value
0.36
0.002
.005
0.75
56
-------
Table 4-8. R-Squared and P-Values for Chemical Use per Unit-of-Producta
B1TS(T) CIRCS (B) Per Index Modules
IPA
Ethyibenzene
PGMEA
Cyclohexanone
NBA
NMP
Xylene
Total of seven chemicals
PGMEA/cyclohexanone waste
PGMEA/cycIohexanone waste
.(1 mo delay)
0.50
<.001
0.47
<.001
0.47
<.001
0.19
>.05
0.03
>.05
0.02
>.05
0.47
<.001
0.04
>.05
0.04
>.05
0.36
<.01
0.09
>.05
0.04
>.05
0.07
>.05 ;
0.01
>.05
0.03
>.05
0.02
>.05
0.08
>.05
0.05 :
>.05 |
0.09
>.05
0.02
>.05
0.23
<.05
0.23
<.05
0.30
<.01
0.08
>.05
0.04
>.05
0.04
>.05
0.01 .
>.05
0.13
>.05
0.08
>.05
0.04
>.05
0.57
<.001
0.57
<.001
0.58
<.001
0.28
>.05
0.18
, >.05
0.01
>.05
0.32
<.01
0.22
>.05
0.05
>.05
0.48
<-001
a Summary of analysis chemicals listed in Table 4-5. The upper number in each cell represents the R-
squared statistic. The lower number in each cell represents a regression P-value. The text in cells with
statistically significant P-values are in bold type.
PGMEA/cyclohexanone waste data without
the delay function did not correlate with any
of the four units-of-product; and
PGMEA/cyclohexanone waste data cor-
related with bits and modules when waste
data were delayed 1 month from production
data.
4.3.3 Findings
A number of factors influenced the results of
the IBM data analysis. These factors made it
likely that that analysis would not detect a
correlation between the performance index and
chemical use or chemical waste, regardless of
whether such a correlation was actually present.
57
-------
After conducting the analysis, RTI and Greiner
Environmental discussed the results with IBM
staff. It became clear that, while the data
seemed to be time-consistent,9 they were not.
The issues with individual data sources are
detailed below. This lack of consistency
resulted in a lack of confidence in the results of
the analysis.
The analysis assumed that the production
data referred to the quantity of product
produced at the facility, rather than the
quantity of product shipped. This was an
erroneous assumption. The bits, circuits,
and modules may have been stored in
inventory as long as 2 months before being
shipped. Therefore, the information about
quantity of product is less related to the use
of chemical or generation of waste than it
would be if the production data referred to
the quantity of product coming out of the
production line in a given month.
» It was discovered that the revenue data used
to construct the performance index was
derived from revenue at the time the
product was shipped from inventory. Thus,
when we constructed a monthly perfor-
mance index for purposes of this analysis,
the revenue data were not time consistent
with production data, making it less likely
that any existing correlation would be
found. Note that IBM does not use a
monthly performance index, but rather con-
structs an annual index for its P2 report.
« Once it was determined that the figures
relating to production referred to products
"Section 2.2 of this report explains that the data for the
statistical and graphical tools must be time consistent.
That is, the waste or use data must correspond to the same
time period (e.g., daily chemical use and daily widget
production).
coming out of inventory, it became clear
that products shipped in month 2 would
likely be responsible for waste generated in
previous months (i.e., when that product
came off the production process). Thus, the
analysis that was run to assess the correla-
tion between production and PGMEA/
cyclohexanone waste with a 1-month time
lag should have actually been run with a 1-
month time lag on the production data.
The production cycle for IBM bits and
circuits complicated the analysis process. It
takes approximately 3 months to go from
components to a finished product. There-
fore, it was very difficult to associate a
given batch of product with a given quantity
of chemical usage or waste generation. Thus
it is possible that we failed to detect correla-
tions that are actually present in the manu-
facturing environment.
After applying a 1-month delay function to
the waste data, we found a correlation be-
tween PGMEA/cyclohexanone waste and
bits. However, IBM staff reviewing the
results report that these two variables
' should not be correlated, since production
of bits does not generate a discrete
PGMEA/cyclohexanone waste. A counter-
intuitive result such as this indicates that
there may be other indirect relationships
between the waste and the product. It may
also indicate a false positive result.
When these data were collected, IBM was
using its own internal calendar. The cal-
endar allocates as few as 14 and as many as
49 days to the 12 months of the year (for
example, there are 14 days in January, 28 in
February, 35 in March, and 49 in Decem-
ber). Since both production and chemical
use data were tracked using the same cal-
endar, the calendar should not impact on the
58
-------
data analysis. We confirmed this by exam-
ining the data using both the IBM calendar
and a standard calendar..
However, waste data are tracked on a stan-
dard calendar rather than the IBM internal
calendar. The difference between the cal-
endar on which waste is tracked and the
calendar on which production is tracked
make it difficult to match chemical use or
waste with its associated production. IBM
no longer uses a nonstandard internal cal-
endar.
These problematic issues illustrate some key
points about verifying accurate units-of-
product:
Staff who are doing the analysis must
understand clearly the sources of the data
they examine. This requires an emphasis on
communication between the staff conduct-
ing the study and the staff of the facility that
provides the information.
Data must be carefully assessed in light of
the requirements of the statistical and
graphical tool use methodology, described
in Section 2.2 of this report.
Assessing a unit-of-product is an iterative
process. If the analysis provides a counter-
intuitive result, then this is an indication
that the data sources should be reassessed,
as well as the unit-of-product. Depending
on how important it is to get an accurate
unit-of-product, staff may want to test out
more than one unit-of-product and more
than one source of data.
4.4 Wyeth-Ayerst Analysis
Wyeth-Ayerst conducted its own data analysis
for the unit-of-product used to measure P2 for
production of a major pharmaceutical product,
referred to here as "product X" at their Rouses
Point, New York, facility. Confidentiality con-
cerns were the impetus behind Wyeth's desire
to handle production data in-house. They
followed the four-step analysis procedure, pre-
sented in Section 2 of this report, and worked
with Greiner Environmental and RTI to com-
plete the. analysis. Thus, Wyeth provided a
"field test" of the methodology outlined in this
report. ,
This particular production line was selected for
analysis because it is the major hazardous-waste
producing process at Rouses Point. We elected
not to examine the nonrecurring pilot plant
operations at the facility- due to resource
constraints. Nonrecurring operations present a
particularly difficult challenge for evaluation
unit-of-product and adjusted P2 measurement.
The challenge arises from the fact that non-.
recurring:operations tend to generate waste and
use chemicals in a way that may be unique to
each given operation. For instance, a chemical
development pilot plant may make a batch of
clinical trial material and not make that com-
pound again or it may make it at some time in
the future. The chemical development activities
typically involve many different chemical in-
puts, batch sizes and other variables. This
results in varied types and quantities of waste
per unit input and output.
4.4.1 Process Description/Prepare Process
Flow Chart
Product X is produced in two independent pro-
cess lines at Rouses Point. The process consists
of wet granulation steps, drying, and com-
pression into tablets. The final step in the pro-
cess is product inspection. The quantity of raw
materials incorporated into each batch varies
according to the dosage of active ingredient for
that batch (i.e., if the final product will be an x
milligram dose pill, the quantities of raw
59
-------
materials used in that batch are different than if
the final product will be a y milligram dose
pill).
4,4.2 Identify and Collect Data
RTI and Greiner Environmental worked with
' Wyeth-Ayerst staff to identify waste streams
and chemicals used in the pharmaceutical
products process that would be suitable for
analysis. Because the objective was to test the
accuracy of the unit-of-product used to adjust
P2 measures for the entire product line (kilo-
grams of product), we tried to identify a
material that was used in a way that did not vary
dramatically for different batches of the prod-
uct. It was also necessary to find materials and
a waste stream for which data were already
collected in an accessible form.
Chemical Waste Data. The most easily avail-
able waste data are through monthly shipping
records maintained as part of the facility's
hazardous waste management system. Those
figures, however, were calculated as truckloads
of waste and, therefore, did not accurately
reflect the amounts of waste generated in the
prior month. Originally, plant personnel thought
that they would not be able to obtain data about
hazardous waste from the individual production
lines. However, persistent investigation re-
vealed that such data were being collected by
the Technical Services Department for an
unrelated special project. These data may not be
available for future analyses.
Chemical Use Data. In order to avoid having to
make separate calculations for each batch of
different formulation product, we chose to
analyze use of a solvent mixture ("wetting
agent"). This mixture is used as part of the
granulation wetting process, and therefore the
quantities used are not dependent on which
formulation of product X is being produced in
a given batch. As described earlier, data about
wetting agent usage were found in records
collected by the Technical Services Department
for a special project.
Unit-of-Product Data. Wyeth's P2 perfor-
mance tracking system uses kilograms of prod-
uct as the unit-of-product by which they adjust
P2 measurements for product X production.
Data showing weekly kilograms of product
produced were obtained from the Technical
Services Department.
Data Entry. Weekly data from 1996 were
entered in spreadsheets by facility staff as
follows:
Solvent mixture (wetting agent) waste gen-
erated,
Solvent mixture (wetting agent) used, and
Kilograms of product produced.
4.4.3 Graphical Analysis
Histogram and Descriptive Statistics. Histo-
grams were prepared for chemical use data and
waste data (Figure 4-48 and 4-49). The histo-
grams show that chemical usage and waste
generation are normally distributed. There are
no significant outliers that would indicate data-
collection inaccuracies or irregularities.
Plot Time-Series and Moving Average.
Wyeth generated time series plots of wetting
agent use per kilogram of product and waste
generation per kilogram of product (Figures 4-
50 and 4-51). These are used to show trends and
cycles in the data. In both cases, the time series
plots show random data trends. This indicates a
process undergoing normal day-to-day varia-
tion. Therefore, the data are suitable for statis-
tical analysis.
Prepare Scatter Plot Diagram. Scatter plot
diagrams were prepared for chemical use and
60
-------
12
10
|- 8
§ 6
f 4
" .2
0
i i tH
I INPUT per kg PRODUCT
Figure 4-48. Solvent mixture use (kg) per. kilogram of product
histogram.
ue
15
10
| 5
£
0
i
ocjcicicio
» c\j co
">"?"?
ooo
kg WASTE per kg PRODUCT
Figure 4-49. Waste production (kg) per kilogram of product
histogram.
t>
BATCH
Figure 4-50. Waste per product (kg) per kilogram time series
plot.
61
-------
..... T r r t t t i i t
BATCH
Figure 4-51. Chemical use per product (kg) per kilogram time
series plot.
540-r
520 - -
500-
480-r
460-i
440-1
420-^
400 4-
y = 0.5043X + 62.956
ff = 0.9156
Production
Figure 4-52. Production vs chemical use scatter plot.
440
y=0.4523x +19.806"
# = 0.8554
Production
Figure 4-53. Waste production (kg) per kilogram of product
scatter plot.
chemical waste (Figures 4-52 and 4-53). The
scatter plots show increasing relationship be-
tween chemical use and production as well as
between waste generation and kilograms of
production. It was clear that it would be easy to
fit a line to the data. This indi-
cated that kilograms of product
are a good unit-of-product to use
in adjusting measures of P2.
4.4.4 Statistical Analysis
Wyeth ran regression tests on the
data for chemical use vs. pro-
duction and waste generation vs.
production. Since the scatter
plots clearly indicated that there
was a straight-line correlation
between these, the regression
tests were something of a for-
mality. The results are presented
in Table 4-9.
4.4.5 Findings
Wyeth uses "kilogram of product *
produced" as the unit-of-product
to create a production-adjusted
measure of P2 for its production
lines. The firm thus calculates
change in hazardous waste per
kilogram of product produced to
use in assessing P2 progress. The
analysis method presented in
Section 2 of this report, and
applied by Wyeth's staff, in-
dicates that the unit-of-product
used by the Rouses Point facility
is well-correlated with hazardous
waste production from a major
production line at the facility.
In addition, the analysis showed
that there is a correlation be-
tween the quantity of wetting
agent solvent mixture that is used and kilo-
grams of product X produced. This correlation
allowed Wyeth to identify usage of wetting
agent as another potential P2 progress indicator.
62
-------
Table 4-9. Results of Regression Analysis for Waste
and Chemical Use per Unit-of-Product
Waste per kg
product
Wetting agent use
per kg product
Equation of y = 0.4523x + 19,806 y = 0.5043x + 62.956
line
R-squared
value
P-value
0.8554
0.000
0.9156
0.000 :
They also were able to use the Section 2
methodology to identify another potential P2
progress indicatorusage of wetting agent.
4.5 Results of Statistical and Graphical
Analysis on Data from Erving Paper,
Erving, Massachusetts
Erving provided us with data on usage of
caustic and bleach for this analysis. The infor-
mation came from Erving's daily process con-
trol data-collection process. During a site visit
in January 1996.to the company's Massachusetts
facility, Erving Paper provided chemical use
and production data from the company's Massa-
chusetts paper manufacturing facility. The data
were analyzed to determine whether sulfuric
acid, caustic, and bleach usage (chemical data)
are correlated with paper production. Tons of
paper produced is the unit-of-product that
Erving uses to adjust its P2 measurements.
Erving Paper staff expected that usage of all
three chemicals is strongly related to the level
of paper production.
4.5.1 Process Description
The continuous manufacturing operations at the
Massachusetts facility use a variety of acids,
bases, and paper finishing chemicals to produce
its products (Figure 4-54). We examined the
correlation between production
and the use of three process
chemicals. Caustic or sodium
hydroxide is added to a wet
paper slurry to raise its pH;
bleach or sodium hypochloride is
added as a de-inking or white-
ning agent; and sulfuric acid is
added to lower the wet paper
slurry pH to 7.5.
4.5.2 Data Collection
Erving Paper measures the ma-
jority of its chemical use dailytaking readings
each weekday morning at 7 a.m.
4.5.3 Data Analysis
We analyzed Erving's P2 measurement data to
determine whether chemical use and paper
production were correlated. Analysis for each of
the three chemicals is presented below.
Caustic (Sodium Hydroxide). There was ab-
normally high use of caustic recorded each
Monday. This results from the way data are
collected: Monday data points represent 3 days
of production (Friday at 7 a.m. to Monday at 7
a.m.). The time series plot (Figure 4-55) shows
this pattern (notice the data peaks every fifth
data point).
We therefore removed Mondays from the data
set and examined chemical use and production
for Tuesday through Friday data (Figure 4-56).
Removing the Monday data eliminated the
pattern noted above.
Next, we prepared a scatter plot and linear
regression of the Tuesday-Thursday data. Al-
though one would expect caustic use and paper
production to be correlated, a scatter plot of the
data depicts no such correlation. The simple
linear regression generated an R-squared value
63
-------
Waste Paper
Add bleach to
de-ink pulp
Add caustic to
raise pH to 11
Add sulf uric to
lower pH to 7.5
Product
Figure 4-54. Paper production process at Erving paper.
Figure 4-55. Daily caustic use (Ib) per ton of paper
produced time series plot.
equal to .01meaning that caustic use de-
scribed only 1% of the daily variation in paper
production (Figure 4-57).
However, closer examination of the data
revealed several abnormalities that indicate
probable data-collection errors. These abnor-
malities include the following:
" On several days caustic use
per ton of paper was unchar-
acteristically low. These
days were followed by days
with uncharacteristically
high caustic use per ton of
paper.
On several days when deliv-
eries were accepted to fill the
caustic bulk tank, caustic use
per ton of paper was unchar-
acteristically high or low
given the level of produc-
tion.
A lack of caustic measure-
ment resolution may intro-
duce a significant variability
into the data. This variability
would make it difficult to
observe a relationship be-
tween daily caustic use and
daily production.
To minimize these ambiguities,
we conducted the analysis using
weekly data (as opposed to daily
data). Weekly data have the ad-
vantage of smoothing out these
measurement problems. Using
weekly data, a time series plot,
histogram plot, and a scatter plot
were prepared. The time series
, plot (Figure 4-58) depicts a
process undergoing normal .variation as opposed
to exhibiting a consistently increasing or
decreasing pattern.
The histogram of caustic per ton of paper
(Figure 4-59) has a bell-shape, again indicating
normal process variation.
64
-------
45.00
Figure 4-56. Daily caustic use (Ib) per ton of paper produced
time series plot with Monday data removed.
80 90 100
Production
Figure 4-57. Daily caustic use (Ib) per ton of paper produced
scatter plot.
A simple linear regression generated an R-
squared statistic equal to 0.6269inferring that
the level of paper production accounts for
62.7% of the variation in caustic use'(Figure 4-
60). The equation of the line (y=25.531x -
5090.6) indicates that the average pound of
caustic used per ton of paper equals 25.53. The
three data points (in Figure 4-60) showing the
lowest weekly caustic use are
especially significant in the
regression. These data points
represent weeks with holidays
(Thanksgiving, Christmas, and
New Years). Were these data
points removed from the data
set, the relationship between
caustic use and paper production
would not be as apparent.
Sulfuric Acid. Erving Paper's
daily sulfuric acid data have
many of the same issues as the
company's daily- caustic data.
We ran the same diagnostic
checks on the sulfuric acid data
that we ran on the caustic data.
The diagnostic checks found that
the lack of weekend data and the
poor precision of chemical use
measurement made the use of
daily data problematic. However,
we found that weekly data could
be analyzed.
Using weekly sulfuric acid data
and paper production data, a
times series plot, histogram plot,
.and scatter plot were prepared.
The histogram of sulfuric acid
per ton of paper has a bell shape,
again indicating normal process
variation (Figure 4-61). The time
; series plot depicts a process
"undergoing normal variation as
opposed to exhibiting a consistently increasing
or decreasing pattern (Figure 4-62).
The scatter plot of production versus sulfuric
acid shows an increasing relationship, and a
"best-fit" line is easily drawn through the data
points (Figure 4-63). This observation confirms
Erving Paper's expectation that higher levels of
65
-------
6 8 10 12
Caustic per ton-of -paper
18
Figure 4-58. Weekly caustic use (Ib) per ton of paper produced
time series plot.
cr
&
Bin
Figure 4-59. Weekly caustic use (Ib) per ton of paper produced
histogram.
20,000 --
18,000 -
^ 16,000 -
=- 14,000 --
B 12,000 --
| 10,000 -
^ 8,000 --
2 6,000 -
g 4,000 -
2,000 -:
y = 25.531X - 5090.6
R2 = 0.6269
300 400 500 600 700
weekly production (tons)
800
900
Figure 4-60. Weekly caustic use (Ib) per ton of paper produced
scatter plot and regression line.
66
-------
300 330 XO 390 420 450 480 .510 540 More
Figure 4-61. Weekly suit uric acid use (Ib) per ton of paper
produced histogram.
week
Figure 4-62. Weekly sulfuric acid use (Ib) per ton of paper
produced time series plot.
sulfuric acid use correspond to higher levels of
production.
A simple linear regression produced an R-
squared statistic equal to 0.681, inferring that
the level of paper production accounts for
68.1% of the variation in sulfuric acid use. The
equation of the line (y=504x - 67311) indicates
that the average pounds of sulfuric acid used
per ton of paper equal 504. As in the caustic
case, three data points (in the lower left of
Figure 4-63) are especially significant in the
regression. These data points represent weeks
with holidays (Thanksgiving, Christmas, New
Years). If these data points were
removed from the set, the rela-
tionship between sulfuric acid
and paper production would not
be as apparent.
In summary, tons of paper pro-
duced could serve as a unit-of-
product to adjust measures of
change in sulfuric acid use.
Pounds of caustic used could
also serve as a surrogate for
production in adjusting measures
of change in sulfuric acid use.
Bleach (Sodium Hypochlor-
ide). Unlike caustic and sulfuric
acid patterns, Erving's bleach
data do not display a Monday-
spiking effect (Figure 4-64).
This was puzzling, since the data
are collected at the same times as
the other chemical usage data
are.
Using daily data on bleach
usage, we prepared a scatter plot
of daily bleach use versus paper
production (Figure 4-65). The
: plot shows no discernible
relationship between bleach use and paper
production. The regression R-squared term is
equal to 0.0116, indicating that the tons of
paper produced explains only 1.16% of the
variation in daily bleach use.
In addition to analyzing daily bleach use.data,
we prepared time series and scatter plots of
weekly bleach use and tons of paper. The time
series plot (Figure 4-66) is typical of a process
undergoing normal process variation.
However, the scatter plot (Figure 4-67) shows
no relationship between the two factors. The
data are so scattered that no "best-fit" regres-
67
-------
I
CD
CL
fi
&
1
400000
350000
300000--
250000
200000 -
150000 -
100000--
50000
y = 504x- 67311
R2 = 0.6818
sulfuric
Linear (sulfuric)
-4-
-t-
350
450
550
650
750
850
tons of paper (per week)
Figure 4-63. Weekly sulfuric acid use (Ib) per ton of paper
produced scatter plot with regression line.
o
I
1
.£
m
80
100
Figure 4-64. Daily bleach use (Ib) per ton of paper produced
time series plot.
CD
CO
.c
o
a
CD
CD
1 OjUUIJ
14,050 -
12,050
10,050
8,050
6,050 -
4,050 -
2,050 -
sn -
"'""*»'* ''' ' ^^
[ " ', . * ~\*ri'~ v '- -A
:' - - rL^l^-^^r'
<. « » * * , ^ *» *
* # ^ * *
t *
*. ** »
y = 31.528x + 5610.4 * * .
R2 = 0,0116 -Vs ,H*' /,:/,'
L_: 1 - y i_jii 4_ -L^ i_£ i
60 70 80 90 100 110
Tons of Paper
120
130,
Figure 4-65: Daily bleach use (Ib) per ton of paper produced
scatter plot with regression line.
68
-------
4.5.4 Findings
10
Week
Figure 4-66. Weekly bleach use (Ib) per ton of paper produced
time series plot.
60,000 j 75:
50,000
40,000 --
~ , v," '«#* ' > ^,. ; ;. -, -«f *_ -^ , ^
-" - t r ^r -,;,; ^:; ^ :- w-' »; ^. * -
**.-,* . * - : VH " ,,-' _L4 ^ -
300 400 500 600 700 800
Production (tons/week)
900
Figure 4-67. Weekly bleach use (Ib) per ton of paper produced
scatter plot with regression line.
sion line can be drawn with any certainty. The
regression term for weekly bleach versus tons- . :
of-paper produced is only 0.1338. This implies
that only 13.38% of the variation in the weekly ;
bleach use is explained by variations in pro-
duction. Regressions examining bleach use
versus production on Tuesday through Thursday
produced similar results. .
Since 1990, Erving Paper has
used tons of paper produced as.
the unit-of-product to adjust
measures of chemical use of the
level of production. The present
analysis determined that this
practice is effective for two
major chemical uses (caustic and
sulfuric acid) but not for bleach.
These two chemicals are good
indicators of Erving's P2 pro-
gress.
The lack of correlation between
bleach use and paper production
runs counter to what one might
expect. Conventional wisdom
holds that bleach usage is di-
rectly proportional to tons of
paper produced. However, based
on this analysis, it would be far
better to measure P2 by looking.
at changes in caustic use or
sulfuric acid usage per ton of
paper produced. Alternatively,
there might be units-of-product
at Erving paper that have a
stronger relationship with bleach
use (e.g., tons of pulp rolled out,
or number of boxes shipped).
69
-------
Section 5
Conclusions
In this research we conducted three tasks:
" Described the use of production-adjusted
measures of P2 at five different facilities;
" Developed a method to apply statistical and
graphical tools to analyze the accuracy of
factors used for production-adjusting P2
measurement; and
« Analyzed the factors used for production-
adjusted P2 measurement at five case study
facilities.
This section elaborates on conclusions drawn
from the results of the three task areas.
5.1 Use of Production-Adjusted P2
Measures
While the major driver for developing pro-
duction-adjusted measurements of P2 has been
regulatory requirements, firms have also found
nonregulatory uses for production-adjusted P2
measures. Specific applications include:
« Process control,
« Quality control,
* Internal communications, and
External communications.
Production-adjusted measures of P2 can be used
to assess yearly P2 progress, or they can be
generated more frequently to provide insight
into day-to-day functioning the process line.
This insight can help firms fine-tune their
production processes to improve efficiency.
The process of developing and verifying pro-
duction-adjusted measurements of P2 can be
valuable to the facility. The process of
identifying monthly or weekly data for a
production process provides insight into the
kinds of data that are available at the facility for
purposes of measurement and for process
control. This is a different perspective than that
gained while identifying yearly data for report-
ing purposes.
A dilemma is presented to environmental staff
who are selecting indicators of P2 progress.
They must either choose to base the measure-
ment on existing data or they must collect new
data to feed into the measurement. Collection of
new data can take significant resources and may
not be a desirable option. But if the measure-
ment is based on existing data, then all future
measurements will be dependent on the con-
tinued collection of that data. Two of our case
study facilities found that collection of some of
the data on which they based their mea-
surements was scheduled to be phased out. This
' experience indicates that it is desirable to obtain
institutional support for P2 measurement in
order to ensure continued collection of the
necessary data.
Measuring P2 can be a resource-intensive pro-
cess. It is important to ensure that the resources
expended are in line with the benefits that are
expected. It is counterproductive to spend many
staff hours to develop and implement a mea-
surement system if no resources will be left to
actually implement P2 projects. Likewise, a P2
70
-------
measurement system should be selected that is
appropriate to the production process or facility
being measured: If the process is constantly
changing, the measurement system should re-
flect that. If the product in question is being
phased out, then a more rudimentary measure-
ment may be in order.
5.2 Methodology for Verification of
Production-Adjusting Units
This research developed a methodology for
applying statistical and graphical tools for
analyzing units-of-product that allow a facility
to assess how well correlated a unit-6f-product
and a target waste stream or chemical use are.
The user assesses this by using tools to calcu-
late how much of the variation in a waste
stream or chemical use is due to variation in the
unit-of-product. The user also calculates
whether this result is statistically significant. If
a unit-of-product and target waste stream or
chemical use are well correlated, then a P2
measurement using that unit-of-product will
accurately reflect production-adjusted P2 for
that process or facility.
5.2.1 Assessment of Data
Assessing a production-adjusted P2 measure-
ment system using the tools recommended in
this report is an iterative process. In most of the
analyses conducted during this research, several
attempts were needed to identify data suitable
for analysis. Issues that arose during the
analysis include the following:
Accurate monthly or weekly hazardous
waste generation data may not be available.
Often the available hazardous waste data
are from shipping records rather than from
the production process. The data may, there-
fore, reveal more about the schedule of the
hazardous waste hauler and the capacity of
their trucks than about the generation of
waste during a particular period. In cases
where waste generation data appear not to
reflect production well, chemical use data
may be a good substitute.
Chemical use or waste data may lag behind
actual production. In some cases, the avail-
able data come from sources that are not
directly associated with the production line.
Chemical use data may come from with-
drawals from inventory. Hazardous waste
data may come from transfers to hazardous
waste storage. In either of these cases, the
data from month 2 may actually have
resulted from production in month 1. In
these cases, it may be possible to add a
delay function in the analysis.
Data must be verified to ensure that they
reflect what the user thinks they are sup-
posed to be showing. For instance, periodic
data peaks every Monday in data at the
Erving Paper case study (given in Section
4.5) turned out to represent 3 days' worth of
data rather than just 1 day of data. Use of a
time-series plot should provide insight into
data anomalies. These should be followed
up with facility personnel to assess the
sources of the data.
5.2.2 Using Chemical Use Data to Evaluate
Units-of-Product
In many cases during the analysis of units-of-
products used by case study firms, RTI and
Greiner Environmental worked with chemical
usage as an indicator of P2 progress. This
occurred in two different situations:
1. In some cases, the facility was using change
in chemical .usage per unit-of-product as a
P2 indicator. The analysis then served to
verify whether the unit-of-product that the
facility was using was correlated with
chemical usage (i.e., served to verify the
71
-------
existing production-adjusted P2 measure
that the facility had in place).
2. In other cases, facilities were tracking waste
per unit-of-product but were unable to
obtain adequate waste data to conduct an
analysis. In these cases, we conducted
analyses based on chemical use data. This
tested whether variation in chemical usage
was correlated with variations, in unit-of-
product. The results of the analysis provided
the firms with information about a potential
new production-adjusted indicator of P2
progress (change in chemical usage per
unit-of-product). In addition, chemical
usage may be a good surrogate»for certain
waste streams. This is discussed further in
the remainder of this section.
Chemical usage will be a good surrogate for
waste generation in situations where the two
are strongly related. This will often be true
where the chemical is used in a process but
not incorporated into the product. If facility
personnel determine that this is the case,
then a unit-of-product that is correlated to
changes in chemical use will also be cor-
related to change in waste streams for that
chemical.
Analysis of Data. Analysis of data to assess the
correlation between waste or chemical use data
and unit-of-product is not a "black-box" pro-
cess. 'It requires extensive communication
among firm personnel, from production engi-
neers to accounting staff. Users of the method-
ology presented in this report must understand
the objectives of the analysis and should
periodically assess how well the methodology
fits the data that are available through the
analysis process. Section 4 of this report
describes the analysis process for data from the
five case study sites. Some of the major issues
encountered during these analyses include the
following:
Where weekly data gave indicators of being
inappropriate for analysis based on the
criteria explained in Section 2, sometimes
aggregating the data into monthly data made
it more amenable to analysis.
Where extreme outliers were present in the
data, we assessed whether the outlier
affected the descriptive statistics for the
measurement unit to the point where it
could not be used. In that case, the analysis
was repeated without the outlier to assess
whether the unit-of-product and waste or
chemical use were correlated under normal
circumstances.
Where it appeared that data were npt time-
consistent as required by the methodology,
we introduced time lag functions. The time
lag is intended to allow the analyst to look
for the correlations between a given unit-of-
product and the waste or chemical usage
actually associated with that batch of pro-
duct, rather than a batch of product pro-
duced in the following or previous month.
In performing the analysis and working
around data issues, it was often necessary to
consult with facility and production staff to
ensure that the proposed change in the
analysis was still consistent with the way
the production process is run.
After the analysis was completed, it was re-
viewed by production staff to ensure that the
results were consistent with common sense.
Results that seemed counter-intuitive were re-
examined for further insights and for accuracy.
72
-------
5.3 Units-of-Product Used by Case Study
Firms
Examination of case study facilities allowed us
to examine the workings of five different pro-
duction-adjusted measurements of P2 in five
different industries. These industries use pro-
duction-adjusted measures of P2 for very
different purposesfrom process control to,
stakeholder communication to regulatory re-
quirements. Further, the units-of-product they
use in their P2 measurement schemes vary from
the simple to the complex.
We found that single units-pf-product (like
"square feet plated" or "kilograms of product
produced") generally correlated well with
chemical use and, in some cases, with waste
generation. This finding is important because
there has been some concern that single units-
of-product are inadequate to explain variation in
waste generation. If this were true, then it would
be much more difficult for .firms to accurately
assess their P2 performance, since they would
have to account for many more variables than a
single measurable output. Our results, however,
suggest that a carefully chosen single-variable
unit-of-product can account for enough of the
variation in chemical use or waste to be used in
adjusting gross P2 measures.
In other words, we found that there was a statis-
tically significant relationship between a single
unit-of-product and chemical usage in all five of
the case study facilities. This is somewhat
surprising, since complex processes might be
expected to have several variables that explain
variation in chemical usage (e.g., operational
conditions, quantity of product, quality of
inputs). However, production levels of a given
product affected chemical use enough that a
statistically significant linear relationship
between the two could be detected.
5.3.1 Larger-Scale Production-Adjusted P2
Measurements
This research investigated only one instance of
production-adjusted P2 measurement across
multiple product lines. It did not investigate
production-adjusted P2 measurement across
multiple facilities. Further investigation of how
to appropriately aggregate measures of P2 is an
important next step in understanding the
impacts of efficiency and environmental pro-
tection efforts by firms.
73
-------
Section 6
References
Greiner, Timothy J. 1994-95. Normalizing P2
data for TRI reports. Pollution Prevention
Review. Winter, pp. 65-75.
Harriman, Elizabeth, Jay Markarian, Jay
Naparstek, James Stolecki, and Anne Marie
Desmarais. 1991. Measuring Progress in
Toxics Use Reduction. Department of Civil
Engineering, Tufts University. Prepared for
Massachusetts Department of Environ-
mental Protection, Boston, Massachusetts.
August.
Tellus Institute, Sound Resource Management
Corporation, CCA, Inc., and Matrix Man-
agement Group. 1991. P2 Measurement
Project: Normalization Measures: A Report
to Washington Department of Ecology.
Olympia, Washington. June.
74
-------
Appendix A
Selected Reports and Articles Dealing with
Production-Adjusted Measures of P2
Behmanesh, Nasrin, Julie A. Roque, and David
T. Allen. An analysis of normalized mea-
sures of pollution prevention. Pollution
Prevention Review. Spring, 1993, pp. 161-
166.
Butler, Craig. Ohio Waste Minimization Mea-
surement Pilot Project: An Analysis of
Pollution Prevention Measurement Options
for Ohio. Columbus OH: Ohio Environ-
mental Protection Agency, Office of Pollu-
tion Prevention, March 1996.
INFORM, Inc., Toxics Watch, 1995. New York.
1995.
Malkin, Melissa, Jesse N. Baskir, and Jordan
Spooner. Issues in facility-level pollution
prevention measurement. Environmental
Progress 14(4):240-246.
Washington State, Department of Ecology.
Pollution Prevention in Washington State,
Task 2: Testing the Utility of Pollution
Prevention Measurement Methods and Data
at the Facility Level. Washington State
Department of Ecology, Hazardous Waste
and Toxics Reduction Program, Olympia,
WA, Publication Number 94-191, August
1994.
75
-------
Appendix B
Selected Statistical Resources
Most basic statistics and data analysis text
books review graphical and regression analysis
techniques. The three books listed below are
good starting points for those wishing to inves-
tigate these methods.
Anderson, David R., Dennis J. Sweeny, and
Thomas A. Williams. Introduction to
Statistics: Concepts and Applications.
St. Paul, MN: West Publishing Com-
pany. 1991.
Box, George E.P., William G. Hunter, and
J. Stuart Hunter. Statistics for Experi-
ments. New York: John Wiley and
Sons. 1978
Hogg, Robert V., and Johannes Ledolteer.
Applied Statistics for Engineers and
Physical Scientists. New York: Mac-
millian Publishing Company. 1992.
76
-------
Appendix C
Framework for Production-Adjusted Measurement of P2
Use of production-adjusted measures of P2 can
ensure that a facility is measuring emissions or
waste changes that are the result of P2 or other
factors besides mere fluctuations in production
levels. Our case studies also found that the
process of setting up a production-normalized
measure of P2 can provide valuable insights to
facility management and staff.
This appendix provides a "step-by-step" frame-
work for developing normalized P2 measures at
the corporate, facility, or process level.
C.1 Production-Adjusted P2
Measurement Framework
This framework asks a series of questions to
lead facility staff through the steps needed to
select and verify a production-adjusted measure
of P2; The framework for production-adjusted
P2 measurement, shown in Figure C-l, provides
guidance on identifying the overall scope of a
P2 measurement system, collecting data, and
selecting a unit-of-product with which to adjust
the P2 measurement.
This section outlines the questions that a P2
professional would ask as he/she develops or
upgrades a P2 measurement system.
Step 1. ^hat is the goal of the P2
measurement? "'*'"'
Is the goal to measure a particular waste
stream? Overall facility reduction goals? The
answer to this question will provide information
about what kind of data is necessary.
For instance:
"Ours is an extremely large facility. How
can weiget the individual process engineers
to be responsible for pollution prevention
for their areas?" - We need data and
metrics for each main production area, plus
product data for each of these production
areas. !
"We have a facility-wide target of reducing
waste by 50% in 5 years. How much of that
target have we achieved?" - We need data
for the chemicals of concern at the facility,
plus production data for product from the
major production lines.
"We have reduced our discharges of arsenic
to the publicly owned treatment works
(POTW), but we're not sure if this is due to
P2 efforts or due to drop-off in activity" -
We need data on arsenic discharges, plus a
unit-of-product related to arsenic discharge.
"We want to develop an accurate unit-of-
product for use in annual reporting under
TURA or TRI" - We need information
about annual discharge of regulated chemi-
cals plus a unit-of-product that is repre-
sentative of the entire facility.
77
-------
Analyze strength of
relationship between
unlt-of-product and
waste or chemical usa
graphical I I regression
anifytli 1 1 analysis
1
r
Ensure data are
colloctod regularly
1
r
<
Examples
one process
line
entire
facility
one
chemical
product
line
1 all
1 TRI waste
^
Examples
withdrawals
from
stockroom
yield
Information
percent
rejects
process
waste
generation
sales
figures
lab technician
records
Identify
unlt-of-product
to adjust
P2 measure
^
r
^
^
Example
engineering guesses
about what
unit-of-product is
closer related
to variations in waste'
or chemical use
Review occasionally
Figure C-1. Framework for production-
adjusted P2 measurement.
Step 2. What data are available to answer
the questions during the periods in
which we want them answered
(e.g., daily, weekly, or monthly
data for internal measurements;
annual data for regulatory
reporting purposes)?
Examples of data to measure P2:
Monthly, quantity of waste shipped offsite -»
Resource Conservation and Recovery Act
(RCRA) manifests; hazardous-waste track-
ing system;
Periodic quantity of inputs withdrawn from
stockroom or otherwise purchased - from
accounting department, from quality assur-
ance records;
Quantity of chemicals added to plating bath
weekly - from the lab technicians' records;
Mass balance or materials balance informa-
tion -» collected during more detailed P2
audits;
Efficiency data
Product yield -»
ment; and
Production data
department.
- process engineering;
quality assurance depart-
production control
Step 3. What units-of-product can
logically be used to adjust the P2
measurement to account for
production levels?
As described in this report, production-adjusted
measures of P2 give a more detailed picture of
P2 than gross measures of change in waste or
change in chemical use do. At this stage of
developing a P2 measure, the facility staff must
identify possible units-of-product that can be
78
-------
Table C-1. Examples of Unit-of-Product Used by Case Study Facilities
Type of operation
Unit-of-product
Metal finishing facility
Paper recycling ,
Electronics production
Semiconductor fabrication
Pharmaceutical production
Square feet of substrate plated
Tons of paper produced
Number of passes substrate makes through
process
Combined unit-ofrproduct incorporating bits,
circuits, and masks
Kilograms of product produced
used to adjust measures of P2. The objective is
to find those units-of-product that are likely to
explain the variation in emissions or chemical
use for the facility or process line. Examples
from case study facilities are given in Table C-
1. Process flow diagrams and conversations
with process engineers or line managers can be
valuable resources in choosing potential units-
of-product.
Step 4/ Of the possible uhits-of-pfoduct
identified in Step 3, for which are
> data available?
As described in Section 2, it is preferable to
have at least 30 data points in order to verify
that the unit-of-product that is chosen is related
to the variation in chemical use or waste
generation.
At the facilities examined in this research,
engineers often found that sources for the data
they needed, while not immediately on hand,
were typically available. This is because many
different data sets are generated at facilities for
many different purposes; generally there is not
one central place to go for process information.
For instance, at one facility, when we first asked
the environmental health and safety staff
whether weekly production and chemical use
data were available, they thought that the infor-
mation was not available for a single process on
a weekly basis. They later found that the infor-
mation did exist, and they were able to use it to
analyze their P2 measurement system.
Where a facility is developing a measure of P2
that measures a full year period, it will
obviously be impossible to get 30 data points,
each representing the annual measure of P2. For
such a measure, the facility should try to find 30
data points for daily, weekly, or monthly data
and Use these to assess the accuracy of the unit-
of-product. If it is found that the unit-of-product
chosen explains variation in emissions or
chemical use, then the facility can go on and use
the available yearly data to construct its annual
measures of P2.
Step 5. Which otthe possible units of
normalization are strongly related
"c to variation in chemical use or
S S/s /
'- - waste-generation? -
Use statistical and graphical analysis to assess
the accuracy of unit-of-product options.
Section 2 of this report describes a statistical
and graphical method for testing the accuracy of
possible unitsrof-product used at a facility. It is
79
-------
important to test the unit-of-product identified
in Step 3 to ensure that it actually explains the
variation in chemical use or waste generation.
In this step, staff should take the data collected
in Step 4 and conduct the statistical and graph-
ical analysis on it. If the unit-of-product is
found to explain variation in waste or chemical
use, then it should be used in measuring P2.
If the unit-of-product does not explain variation
in waste or chemical use, then other units-of-
product should be analyzed.
Once a unit-of-product that explains.variation in
emissions or chemical use is identified, then the
facility can calculate the production-adjusted P2
measurement as often as is necessary for their
purposes (anywhere from annually for TRI
reporting purposes to daily for process control
purposes).
80
GOVERNMENT PRINTING OFFICE: 1997-549-001/60158
------- |