Int J Life Cycle Assess (2010) 15:511–520

DOI 10.1007/s11367-010-0158-5

SENSITIVITY ANALYSIS

Sensitivity coefficients for matrix-based LCA

Reinout Heijungs

Received: 26 May 2009 / Accepted: 12 January 2010 / Published online: 20 February 2010
# The Author(s) 2010. This article is published with open access at Springerlink.com

Abstract Recommendations and perspectives Although the formulas

Background, aim, and scope Matrix-based life cycle derived are not simple, they are straightforward to implement
assessment (LCA) is part of the standard ingredients of in software for LCA. Once this is done, the use of these
modern LCA tools. An important aspect of matrix-based formulas can become routine practice, enabling a key issue
LCA that is straightforward to carry out, but that requires analysis and speeding up perturbation and uncertainty
a careful mathematical handling, is the inclusion of analysis.
sensitivity coefficients based on differentiating the
matrix-based formulas. Keywords Derivatives . Life cycle interpretation .
Materials and methods We briefly review the basic Matrix-based LCA . Sensitivity . Taylor series expansion .
equations for LCA and the basic theory of sensitivity Uncertainty
coefficients.
Results We present the complete set of sensitivity coef-
ficients from inventory to weighting through characteriza- 1 Introduction
tion and normalization. We show the specific formulas for
perturbation analysis, uncertainty analysis, and key issue
analysis. We also provide an example using the ecoinvent 2 Background, aim, and scope
data.
Discussion The limitations of the present approach include By now, the matrix approach to life cycle assessment
the restriction to small changes and uncertainties and the (LCA) has received wide recognition. Proposed in the early
ignorance of correlation between input uncertainties. In 1990s (Möller 1992; Heijungs et al. 1992), it took a decade
contrast to common thinking, there is no restriction to or so before this idea was embraced as a general accepted
normally distributed uncertainties: Every uncertainty distri- way of doing LCA (see Heijungs and Suh 2006 for a short
bution for which a variance can be defined can be history of the matrix approach to LCA).
submitted to the analytical uncertainty analysis. The matrix method has been developed to solve the
Conclusions This paper provides a useful set of tables for a inventory problem in LCA. With the inventory problem, we
number of purposes related to uncertainty and sensitivity refer to the task of scaling all unit processes in the system in
analysis. such a way that they exactly produce the reference flow (or
functional unit) and to use this scaling such that the
inventory table can be computed. Formulas such as

Responsible editor: Ralph Rosenbaum g ¼ BA1 f ð1Þ

R. Heijungs (*) (see Tables 1 and 2 for an explanation of the symbols
Institute of Environmental Sciences (CML), Faculty of Science,
involved) are now common in the specialized literature on
Leiden University,
PO Box 9518, 2300 RA Leiden, The Netherlands life cycle inventory (LCI; Heijungs and Suh 2002; Peters
e-mail: heijungs@cml.leidenuniv.nl 2007; Tan 2008).
512 Int J Life Cycle Assess (2010) 15:511–520

Table 1 Overview of symbols

representing input data in the Symbol Name Dimension (rows × columns) Defined in
LCA
f Final demand vector Economic flows×1 Goal and scope definition
A Technology matrix Economic flows × processes Inventory analysis
B Intervention matrix Environmental flows × processes Inventory analysis
Q Characterisation matrix Categories × environmental flows Impact assessment
g Intervention totalsa Environmental flows×1 Impact assessment

h Category totalsa Categories×1 Impact assessment
a
See the note below on w Weighting factors 1×categories Impact assessment
normalization

The matrix approach has also been extended in various Moreover, the impact assessment adds additional uncer-
directions. For instance, research has been devoted to add tainty to the already uncertain results of the LCI. Likewise,
details on: not only the sensitivity of inventory results is of interest
but also (or perhaps even more so) the impact assessment
& the treatment of allocation and cutoff (Heijungs and
results.
Frischknecht 1998);
The extension of the sensitivity coefficients from LCI to
& how to connect a process-based LCI to an input–output
life cycle impact assessment was “left as an exercise” to the
table (Suh and Huppes 2005);
LCA practitioner. For instance, Heijungs and Suh (2002,
& how to efficiently compute an answer to the inventory
p. 144) write that “In this way, all equations of LCA may be
problem (Peters 2007);
processed,” but they do not pursue this. It is a task that is
& how to analyze the feedback structure of the system
not so often carried out, we guess, at least we have never
(Suh and Heijungs 2007);
seen the explicit results of such an exercise. This paper
& how to calculate sensitivity coefficients (Heijungs
therefore aims to carry out this exercise and to make the
1994).
results available. The formulas obtained can easily be
Most of these details indeed refer to exclusively inventory- implemented in matrix-based software for LCA. We have
oriented questions. The impact assessment follows the done so in CMLCA, a program for doing LCA, and some
inventory results, so all issues related to cutoff, allocation, screenshots are shown at the end of the paper.
IO-based LCA, efficient algorithms, and structural path In this paper, we first review the basic equations of LCA.
analysis are only interesting from an LCI point of view. For Then, we proceed to review the theory of sensitivity
the last one mentioned, the sensitivity coefficients, this is coefficients in general and their form in LCI. This finally
different, however. leads to a derivation and coherent presentation of the
Sensitivity coefficients are important for both uncertainty sensitivity coefficients for the entire LCA process.
analysis and sensitivity analysis (Heijungs 1994). In the
context of uncertainty analysis, they serve to establish
essential information for a Taylor series expansion, and for 3 Materials and methods
sensitivity analysis, they provide the multipliers that enable
one to distinguish sensitive from non-sensitive parameters, 3.1 Basic equations for LCA
the so-called key issues for refined data collection (Heijungs
1996). But uncertainty and sensitivity analyses are not only The basic equations for LCA have been discussed in a
important in LCI but in impact assessment as well. consistent notation by Heijungs and Suh (2002). Below, we

Table 2 Overview of symbols

representing output results in the Symbol Name Dimension (rows × columns) Equation
LCA and the formulas with
which these results are obtained s Scaling factors processes×1 s ¼ A1 f
g Inventory results Environmental flows×1 g ¼ Bs
h Characterization results Categories×1 h ¼ Qg
Λ Intensity matrix Environmental flows × Λ ¼ BA1
economic flows
e
h Normalization results Categories×1 8k : ehk ¼ hhk
 
h ¼ Q g
k
h Category totalsa Categories×1
W ¼ we
a
See the note below on W Weighted index 1×1 h
normalization
Int J Life Cycle Assess (2010) 15:511–520 513

present two tables of symbols and equations connecting the 4 Results

fundamental concepts of LCA.
With respect to normalization, there are two situations 4.1 Sensitivity coefficients for LCA
which require separate treatment.
& A vector of intervention totals, g, can be defined for Table 2 provides a specification of the functions f in the
context of LCA. So we are to insert the equations of
the reference situation. For instance, one can collect
Table 2 into the general equation for calculating a
data on the emissions of CO2, SO2, etc., which then
represent g 1 , g 2 , etc. These then can be processed by
sensitivity coefficient (Eq. 4). Example calculations have
been provided by Heijungs and Suh (2002, Eqs. (6.21),
the same characterization model to yield the vector of
 (6.26), and (6.29)). For the scaling factors, s, we have:
category totals, h. This then forms the basis of the

normalization. Changing g will induce a change in h,  
 @sk
but h itself will not be changed directly by the LCA ¼  A1 ki sj ð5Þ
@aij
practitioner. We will refer to this as normalization
case 1. and for the inventory results, g, we have:

& Alternatively, the vector of category totals h can be
known without a detailed specification of the underly- @gk
 ¼ lki sj ð6Þ
ing interventions. In that case, h is not an output result @aij
(in the sense of belonging to Table 2), but input data
 for the dependence on the elements of A, and
(in the sense of belonging to Table 1). Thus, h can be
changed directly, and it will affect the normalization
@gk
results and the weighted index. We will refer to this as ¼ sj d ik ð7Þ
normalization case 2. @bij

Both approaches in fact appear in practice and are for the dependence on the elements of B, where

therefore elaborated below as separate cases. 1 if i ¼ k represents the Kronecker delta.
d ¼
ik
0 otherwise

3.2 General theory of sensitivity coefficients It is clear that the higher one moves in the sequence
inventory–characterization–normalization–weighting, the
There are various situations in which the stability of the results larger the number of sensitivity coefficients there will be.
in terms of sensitivity for perturbations of the input data is of Scaling factors only depend on the technology matrix, but
interest. In general, we may explore this issue as follows. the inventory results depend on the technology matrix and
Suppose that an output variable z depends on two input on the intervention matrix.
variables x and y and that the dependence is expressed by a Table 3 gives a tabular overview of the derivatives for all
function f: expressions of matrix-based LCA.
One remark on these coefficients. Each of the formulas
in Table 3 contains A−1 or a symbol that depends on A−1,
z ¼ f ðx; yÞ: ð2Þ
such as s, Λ or h. Thus, we need to go through the process
The crucial elements in determining sensitivity is the change of matrix inversion to evaluate the sensitivity coefficients.
of the result (Δz) caused by a marginal change in x (Δx) and Moreover, although Ciroth et al. (2004) argue that one can
by a marginal change in y (Δy). This is expressed using the solve LCAs without calculating a full matrix inverse,
partial derivatives: Table 3 shows that we need the full matrix inverse when
we wish to extend the analysis to include sensitivity and
@z @z analytical uncertainty studies.
Δz ¼ Δx þ Δy: ð3Þ
@x @y
4.2 Perturbation analysis
@z @z
Coefficients such as @x and @y are referred to as sensitivity
coefficients in the present context. Their evaluation requires Heijungs and Kleijn (2001) and Sakai and Yokoyama
a specification of the function f. Thus, with f specified, the (2002) describe perturbation analysis as a way of investi-
sensitivity coefficient of f with respect to input parameter x is gating which input data are most decisive for the results in
defined as: terms of their relative sensitivity. That is, given a system
with the prototypical form
@z @f ðx; yÞ
¼ : ð4Þ z ¼ f ðx; yÞ ð8Þ
@x @x
514 Int J Life Cycle Assess (2010) 15:511–520

Table 3 Overview of the sensitivity coefficients that express how LCA output results (columns) change if LCA input data (rows) data change
.
@sk =    @gk =    @hk =    @e
hk    @W =   
   P P P wk P
 A1 ki sj
s
   @aij −λkisj sj qkl lli  hj qkl lli sj 
hk
qkl lli
k
 l
qki sj
l
P k l
   @bij 0 sjδik qkisj  sj wk
 qki
k ðhk Þ
2
hk
  
 gj h g g h g
   @qij 0 0 gjδik   k j2 d ik wi hj  i j2
hk ðhk Þ i ðhi Þ
P wk hk qki
   =@ġi (normalization case 1) 0 0 0  hkqki2   2
ðhk Þ k ðhk Þ

   =@ hi (normalization case 2) 0 0 0  h k 2 dik  wi hi2
   =@wi 0 0 0 0
ðhk Þ hi ð i Þ
e h

@h
Example: row 3 and column 4 tell us that @bk ¼ qki sj
ij  
Note that depending on the situation in normalization, either row 5 (   @qij ) or row 6 (   =@ hi ) applies

one investigates dimensionless multipliers, such as

@z x
: ð9Þ
@x z
The idea is that a small change of either of the input Table 4 Overview of the relative sensitivity coefficients (multipliers)
parameters (say in x, hence a change Δx) leads to a change that express how small changes in input data propagate into changes
Δz=z in output results
in the result (z, hence Δz), and that the relative change Δx=x
@z=z
can be approximated by @x=x. The results of Tables 2 and 3 Multiplier Definition Formula
can be combined to yield a complete overview of these @sk =sk a  
multipliers (see Table 4). σk(aij) @aij =aij  skij A1 ki sj
@gk =gk a
γk(aij) @aij =aij  gijk lki sj
4.3 Uncertainty analysis @gk =gk bij
γk(bij) @bij =bij gk sj d ik
Although the Monte Carlo method is right now the most @hk =hk a P
ηk(aij) @aij =aij  hijk sj qkl lli
frequently applied method for studying the propagation of bij l
@hk =hk
input uncertainties into output uncertainties (Lloyd and Ries ηk(bij) @bij =bij hk qki sj
2007), it has been recognized that this method may be too @hk =hk qij
ηk(qij) @qij =qij hk gj dik
computationally intensive for application to large systems   @e
hk =e P
(Ciroth et al. 2004; Heijungs et al. 2005; Hong et al. 2008). ehk aij hk
@aij =aij 
aij sj
ehk
 qkl lli
 
hk
l
Indeed, the present release of the ecoinvent data (v2.0, @e
hk =e
ehk bij hk
@bij =bij
bij qki sj
ehk hk

comprising almost 4,000 processes) has not been processed   
@eh=e hk gj
with Monte Carlo analysis, while the previous release ehk qij hk qij gj
  d ik
@qij =qij eh hk ð hk Þ
2
k
(v1.3, 2,500 processes) contained such results, precisely for @e
hk =e 
hk ð g i Þ (normalization case 1)
e hk
@ gi = gi  gi hk qki2
this reason. One way to address this is by means of smarter ehk ð hk Þ
  @e
hk =e 
algorithms. Peters (2007) has proposed power series e
hk hi (normalization case 2)  
hk
 hi h k 2 d ik
@ hi = hi ehk ð hk Þ
expansion, also in connection to Monte Carlo analysis. @W=W a P wk P
Moreover, there are more sophisticated sampling strategies ω(aij) @aij =aij  Wij sj 
hk
qkl lli
than the naive Monte Carlo method, such as Latin @W=W bij P k wk l
ω(bij) @bij =bij W sj  2 qki
hypercube sampling and response surface methods (e.g., k ð hk Þ 
@W=W qij gj hi gj
Morgan and Henrion 1990). In this paper, we address it ω(qij) wi  
@qij =qij ð hi Þ
W 2
hi
@W =W  P wk hk qki
with a solution on the basis of analytical error propagation. wð g i Þ (normalization case 1) @ gi = gi  Wgi  2
The theory of analytical error propagation, using Taylor   @W =W  k ð hk Þ
series expansion, has been proposed in LCA (Heijungs w hi (normalization case 2)    Whi w i hi
@ hi = hi ð hi Þ
2

1994; Ciroth et al. 2004; Hong et al. 2008). Taylor series @W =W wi e

ω(wi) @wi =wi W hi
expansions are based on the approximation formula for
calculating the variance of a stochastic result using Example: σk(aij) tells how many percent sk will change if aij changes 1%
Int J Life Cycle Assess (2010) 15:511–520 515

stochastic data (Bevington and Robinson 1994; Morgan interpretation of these indices, we see summations over all
and Henrion 1990). For a system of the form economic flows and twice over all processes. Take the case
of ecoinvent v2.0 where the number of processes and
z ¼ f ðx; yÞ ð10Þ
economic flows is almost 4,000; this variance is the result
it assumes the form of almost 12,000 terms. Each of these terms is positive, so
 2  2 we may really consider (sj1ki)2var(aij) and (sj1ki)2var(bij) as
@f @f the contribution that one individual var(aij) or var(bij)
varðzÞ ¼ varðxÞ þ varðyÞ ð11Þ
@x @y makes to the total var(gk). We therefore define a number
@f @f of dimensionless coefficients ζ, namely:
þ2  covðx; yÞ
@x @y  2  
  sj lki var aij
where var(x) is the variance of the variable x and cov(x,y) z gk ; aij ¼ ð14Þ
varðgk Þ
represents the covariance between the stochastic variables x
and y.
In most cases, no covariance data are available, or the and
covariance can be assumed to be negligible as the
 2  
uncertainties are in many cases independent. In those cases,   sj d ik var bij
we set covariance to zero and obtain z gk ; bij ¼ ð15Þ
varðgk Þ
 2  2
@f @f
varðzÞ ¼ varðxÞ þ varðyÞ ð12Þ as the relative contributions by each var(aij) and var(bij) to
@x @y
the total var(gk). Naturally,
Such equations are for a general function f (x,y), and they
require the evaluation of the derivatives @f @f
@x and @y . In the
X   X  
present case, we have specified the function (Table 2), and z gk ; aij þ z gk ; bkj ¼ 1 ð16Þ
i;j j
we have found equations for the derivatives (Table 3). So
all ingredients are available to complete the structure of the
uncertainty analysis. so these indeed represent relative contributions to the
Heijungs and Suh (2002, Eqs. (6.73) and (8.87)) variance of the total.
elaborate this only for the inventory vector, and then even Table 6 provides an overview of the expressions for the
only partially, and with a typo. The complete and correct different types of key issues.
expression for this is
4.4.1 Example
X 2   X  2  
varðgk Þ ¼ sj lki var aij þ sj var bkj : ð13Þ We have programmed these equations into CMLCA
i;j j
(http://www.cmlca.eu/). As an example, we loaded the
The remaining expressions for the uncertainty of the ecoinvent v1.3 (2,630 unit processes) and calculated the
impact assessment results are elaborated in Table 5.  system for a reference flow of 1-kWh Swiss electricity,
Again, mind that there are two expressions for var e
hk low voltage, at grid. In the example below, we have
and two expressions for var(W) for both normalization restricted the analysis to the inventory analysis and to the
variants. emission of Carbon dioxide, fossil, emitted to air, low
population density.
4.4 Key issue analysis We performed a perturbation analysis, concentrating
on the perturbations of the technology matrix. Figure 1
Key issue analysis has been defined by Heijungs (1996) as shows a screenshot of the results, hence of the coefficients
the decomposition of the uncertainty of a result in terms of γk(aij) for k is CO2 and for all i (economic flows) and all j
the contribution of the uncertainties of the input data. (processes). A not-too modern computer was able to
Morgan and Henrion (1990) refer to it as uncertainty complete the calculations within a few minutes.
importance, while Saltelli et al. (2000) use the (perhaps We also performed an uncertainty analysis. Figure 2
confusing) term sensitivity analysis. All expressions for the shows a screenshot of these results, so tabulating var(gk)
variance in Table 4 are the result of the weighted (along with some other statistics) for k is CO2. Again,
aggregation of a large number of variances of input data. results are obtained within a few minutes.
For instance, the expression for var(gk) comprises two We finally performed a key issue analysis. Figure 3
summations, one over i,j and one over j. Adding the shows a screenshot of these results, so tabulating ζ(gk,aij)
Table 5 Overview of the variance of output results as a function of the variance of the input data
516

Uncertainty Equation
P  2  1  2  
var(sk) sj A ki var aij
i;j

P 2   P  2  
var(gk) sj lki var aij þ sj var bkj
i;j j

 2
P P   P 2   P  2  
var(hk) sj qkl lli var aij þ sj qki var bij þ gj var qkj
i;j l i;j j

8 "  2 #
> P P   P 2   P  hk gj 2   P hk qki 2
>
> 1
>  varð g i Þ ðnormalization case 1Þ
 < ð hk Þ2 i;j sj qkl lli var aij þ i;j sj qki var bij þ j gj  hk var qkj þ i hk
l
var e
hk "  2 #
>
> P P   P 2    2
>
> 1
:  2 sj qkl lli var aij þ sj qki var bij þ hhk varð h k Þ ðnormalization case 2Þ
k
ð hk Þ i;j l i;j

8  2  2
> P  2 P wl P   P  2 P wl   P wi 2 hig j 2   P  w i hi 2 2
>
> s j  qlk ki
l var aij þ s j  q var bij þ þ  gj   var qij þ  varð h i Þþ
P e
hi varðwi Þ ðnormalization case 1Þ
< hl hl li hi hi hi hi
i;j l k i;j l i;j i i
var(W)  2  2
2 2 2
>
> P  2 P wl P   P  2 P wl   P wi hi g j 2   P wi hi   P
>
: sj  qlk lki var aij þ sj  q var bij þ þ  gj   var qij þ   var hi þ ehi varðwi Þ ðnormalization case 2Þ
hl hl li hi hi hihi
i;j l k i;j l i;j i i

Example: var(gk) represents the variance of the inventory result for the kth environmental flow
Int J Life Cycle Assess (2010) 15:511–520
Int J Life Cycle Assess (2010) 15:511–520 517

Table 6 Overview of the contributions to the variance of output obtain a more reliable, less uncertain result by trying to find
results by the individual variance of the input data
more accurate data for these four coefficients.
Contribution Formula
ððA1 Þki sj Þ varðaij Þ
2

ζ(sk,aij) 5 Discussion
varðsk Þ
ðsj lki Þ varðaij Þ
2

ζ(gk,aij) varðgk Þ Three important restrictions must be born in mind.

ðsj dik Þ varðbij Þ
2

ζ(gk,bij) First, the formulas are for the “normal” LCI. This means
varðgk Þ
 2 that the equations become more complicated once we
P
sj qkl lli varðaij Þ incorporate other features and developments, such as
ζ(hk,aij)
l
varðhk Þ allocation and IO-based LCA. Heijungs et al. (2006) show
@gk
ðsj qki Þ varðbij Þ
2
how the formula for @a changes when the inventory is done
ζ(hk,bij) varðhk Þ
ij
using a hybrid method, combining process-based and IO-
based LCA. This modification may be propagated to the
ðgj Þ varðqkj Þ
2

ζ(hk,qij) impact assessment level as well.

varðhk Þ
 2 Second, some of the formulas (namely those in Tables 4,
P
 sj qkl lli varðaij Þ 5, and 6) are based on a first-order Taylor series
z e
hk ; aij l
 
ð hk Þ var ehk approximation. That means that they are correct for small
2

 changes, uncertainties, and perturbations, but not neces-

ðsj qki Þ varðbijÞ
2

z e
hk ; bij sarily for larger ones. To fix this, one may either include
ð hk Þ var ehk
2

 2
second-order terms (or even further). For instance, for the
hk gj
 gj   var ðqkj Þ perturbations, we may improve by going from
z e
hk ; qij  
hk

ð Þ  2
hk var hk e
 2
 hk qki
 varð gi Þ
z e   @z
hk
hk ; ġi (normalization case 1)
ð hk Þ var ehk Δz ¼ Δx ð17Þ
2

 2
@x
 hk 
varð hk Þ
 
z e  
hk
hk ; hi (normalization case 2)
ð hk Þvar ehk
2
2
P wl P into
ð jÞ varðaij Þ
2
s 
hl
q l
lk ki

ζ(W,aij)
l k
varðW Þ
 2
P wl
ðsj Þ varðbij Þ @z 1 @2z 1 @3z
2
q
hl li

ζ(W,bij)
l Δz ¼ Δx þ ðΔxÞ2 þ ðΔxÞ3 þ    ð18Þ
varðW Þ @x 2 @x2 6 @x3
 2  2
hi g j
varðqij Þ
wi

hi
gj  
hi
ζ(W,qij) varðW Þ @z
This requires the evaluation of not only @x but also of the
 2 @2z
second derivative @x2 or even beyond. Table 3 might be

varð hi Þ
wi hi

extended to include expressions for @ sk 2 and similar.

  2
hi hi
z ðW ; ġi Þ (normalization case 1) varðW Þ @ ðaij Þ
 2
Alternatively, the approach by Sherman and Morrison

varð hi Þ
wi hi

  hi hi (1950) can be used to provide an exact answer to this
z W ; hi (normalization case 2)
 varðWÞ question.
ehi varðwi Þ
2

ζ(W,wi) Third and finally, the formulas for uncertainty (Tables 5 and
varðW Þ
6) are based on the ignorance of the covariance between input
Example: ζ(gk,aij) represents the contribution to the variance of gk by variables. As we noted above, the complete first-order
the variance of aij expression includes an additional term for the covariance
between input data:

   2
and ζ(gk,bij) for k is CO2 and for all i (economic flows) @f 2 @f
varðzÞ ¼ varðxÞ þ varðyÞ ð19Þ
and all j (processes). Again, a few minutes suffice. @x @y
Here, we see that only four coefficients make up 80% of @f @f
þ2  covðx; yÞ
the uncertainty of the CO2 emission. In other words, we can @x @y
518 Int J Life Cycle Assess (2010) 15:511–520

Fig. 1 Screenshot of CMLCA

showing the results of a pertur-
bation analysis, the sensitivity of
an output result for small
changes of the input data. Only
multipliers smaller than −0.3
and larger than 0.3 have been
shown

It is possible to carry out the program of this paper A question that always arises in connection to analytical
including covariance, but the expressions become much error propagation is whether it only works for normally
more complicated. For instance, for var(gk), we obtain distributed uncertainties. This is not the case. The theory of
analytical error propagation through Taylor series expansions
X 2   X  2   (Morgan and Henrion 1990, p. 183 ff.) nowhere contains the
varðgk Þ ¼ sj lki var aij þ sj var bkj assumption of normally distributed distributions. The only
i;j j
X assumption is that for a first-order approximation, the
 
þ2 sj lki sm lkl cov aij ; alm function should be sufficiently close to a linear function
i;j;l:m within the range of uncertainty. This point has been addressed
X above. A practical issue is of course that the formulas require
 
þ2 sj sl cov bkj ; bkl a specification of the variance, whereas most distributions are
j;l specified without an explicit variance. For instance, a uniform
X   distribution is often specified in terms of its width and a
þ2 sj lki sl cov aij ; bkl : ð20Þ lognormal distribution in terms of the geometric standard
i;j;l
deviation (or its square, as in ecoinvent). Heijungs and
Frischknecht (2005) provide formulas to easily calculate a
The last three terms requires us to specify cov(aij,alm),
variance from the standard parameters of a normal, lognor-
cov(akj,bkl ), and cov(aij ,bkl ) for all combinations of
mal, uniform, and triangular distribution. As we can see in
processes, economic flows, and environmental flows.
Table 5, the propagated variances are the sum of a large
Although some uncertainties will be definitely correlated
number of terms. Following the central limit theorem (e.g.,
(for instance, the fuel input of a combustion process and
Morgan and Henrion 1990), such a propagated uncertainty
the CO2 emission of the same process may have a
will tend to become normally distributed provided that the
correlation close to 1), most uncertainties will be uncor-
input uncertainties are independent.
related, or any information on such correlations is lacking.
The infrastructure needed (lots of memory for storing a
number of covariance matrices, much more complicated
formulas, much more data collection and estimation) will 6 Conclusions
probably not offset the relatively limited gain of having a
slightly more accurate computation. In the end, there is This paper has carried out the “exercise” that was left over by
always something perverse about knowing the uncertainty Heijungs and Suh (2002) and related work on deriving the
with certainty. complete set of sensitivity coefficients for matrix-based LCA.
Int J Life Cycle Assess (2010) 15:511–520 519

Fig. 2 Screenshot of CMLCA

showing the results of an
analytical uncertainty analysis
consisting of a number of
characteristic parameters for the
uncertainty of an output result

7 Recommendations and perspectives We hope that the availability of these equations will
stimulate developers of software, commercial or not, to
For some, the formulas in Tables 4, 5 and 6 are intimidating implement analytical, Taylor series-based approaches
and may offer little insight. But they are straightforward to toward uncertainty and sensitivity analysis. In particular
implement in computer code. Once implemented, doing an for the key issue analysis, no good Monte Carlo
uncertainty analysis is as easy as doing a Monte Carlo approach is available, and the analytical solution using
analysis: click a button and wait for the results. Likewise, Table 6 provides an extremely powerful way of reducing
doing a key issue analysis is as easy as doing a classical the uncertainties in LCA. For perturbation and uncertainty
contribution analysis. analysis, numerical approaches can be used as well, but

Fig. 3 Screenshot of CMLCA showing the results of a key issue analysis, the decomposition of the uncertainty of an output result in terms of
contributions by the uncertainties of the input data. Contributions below 1% are not shown
520 Int J Life Cycle Assess (2010) 15:511–520

these are extremely time-consuming for large LCA Heijungs R, Kleijn R (2001) Numerical approaches to life-cycle
interpretation: five examples. Int J Life Cycle Assess 6(3):141–148
systems.
Heijungs R, Suh S (2002) The computational structure of life cycle
As noted above, Monte Carlo analyses for large LCA assessment. Kluwer, Dordrecht
systems may be unfeasible. This paper develops sensitivity Heijungs R, Frischknecht R (2005) Representing statistical distri-
coefficients that serve to derive the formulas for analytical butions for uncertain parameters in LCA. Relationships
between mathematical forms, their representation in EcoSpold,
error propagation based on first (or higher)-order Taylor
and their representation in CMLCA. Int J Life Cycle Assess
series approximation. This has other limitations, for 10(4):248–254
instance, relating to a more restricted range of validity. Heijungs R, Suh S, Kleijn R (2005) Numerical approaches to life
Therefore, we welcome the simultaneous development of cycle interpretation. The case of the Ecoinvent’96 database. Int J
Life Cycle Assess 10(2):103–112
more sophisticated sampling methods such as Latin Heijungs R, Suh S (2006) Reformulation of matrix-based LCI: from
hypercube sampling and response surface methods, as well product balance to process balance. Int J Life Cycle Assess
as the development of more efficient algorithms such as 14:47–51
those based on a power series expansion. Yet, even when Heijungs R, de Koning A, Suh S, Huppes G (2006) Toward an
information tool for integrated product policy: requirements for
this would overtake the analytical error propagation, we
data and computation. J Ind Ecol 10(3):147–158
still see a role for the sensitivity coefficients in perturbation Hong J, Rosenbaum R, Jolliet O (2008) Analysis of uncertainty
analysis and in key issue analysis. propagation in life cycle inventory and impact assessment:
application to an aluminum front panel. SETAC Europe 18th
Annual Meeting, Warsaw, p 326
Lloyd SM, Ries R (2007) Characterizing, propagating, and analyzing
Open Access This article is distributed under the terms of the uncertainty in life-cycle assessment. J Ind Ecol 11(1):161–179
Creative Commons Attribution Noncommercial License which per- Möller F-J. (1992) Ökobilanzen erstellen und anwenden. Entwicklung
mits any noncommercial use, distribution, and reproduction in any eines Untersuchungsmodells für die umweltverträglichkeit von
medium, provided the original author(s) and source are credited. Verpackungen. München: Ecobalance Applied Research
Morgan MG, Henrion M (1990) Uncertainty: a guide to dealing with
uncertainty in quantitative risk and policy analysis. Cambridge
University Press, Cambridge
References
Peters G (2007) Efficient algorithms for life cycle assessment, input–
output analysis, and Monte-Carlo analysis. Int J Life Cycle
Bevington PR, Robinson DK (1994) Data reduction and error analysis Assess 12(6):373–380
for the physical sciences. McGraw-Hill, New York Sakai S, Yokoyama K (2002) Formulation of sensitivity analysis in
Ciroth A, Fleischer G, Steinbach J (2004) Uncertainty calculation in life cycle assessment using a perturbation method. Clean
life cycle assessments: a combined model of simulation and Technology and Environmental Policy 4:72–78
approximation. Int J Life Cycle Assess 9(4):216–226 Saltelli A, Chan K, Scott EM (2000) Sensitivity analysis. Wiley,
Heijungs R, Guinée JB, Huppes G, Lankreijer RM, Udo de Haes HA, Chichester
Wegener Sleeswijk A, Ansems AMM, Eggels PG, van Duin R, Sherman J, Morrison WJ (1950) Adjustment of an inverse matrix
de Goede HP (1992) Environmental life cycle assessment of corresponding to a change in one element of a given matrix. Ann
products. Backgrounds—October 1992. CML, Leiden Math Stat 21(1):124–127
Heijungs R (1994) A generic method for the identification of options Suh S, Huppes G (2005) Techniques for life cycle inventory of a
for cleaner products. Ecol Econ 10(1):69–81 product. J Clean Prod 13:687–697
Heijungs R (1996) Identification of key issues for further investigation in Suh S, Heijungs R (2007) Power series expansion and structural
improving the reliability of life-cycle assessments. J Clean Prod 4 analysis for life cycle assessment. Int J Life Cycle Assess 12
(3–4):159–166 (6):381–390
Heijungs R, Frischknecht R (1998) A special view on the nature of the Tan RR (2008) Using fuzzy numbers to propagate uncertainty in
allocation problem. Int J Life Cycle Assess 3(6):321–332 matrix-based LCI. Int J Life Cycle Assess 13(7):585–592