How Many Environmental Impact Indicators Are ... - ACS Publications

Mar 10, 2016 - the indicators we used the “trim.matrix” function on the rank ..... Corresponding author address: P.O. Box 9010 (box 89), NL-. 6500...
0 downloads 0 Views 733KB Size
Subscriber access provided by ORTA DOGU TEKNIK UNIVERSITESI KUTUPHANESI

Article

How many environmental impact indicators are needed in the evaluation of product life cycles? Zoran J.N. Steinmann, Aafke M. Schipper, Mara Hauck, and Mark A.J. Huijbregts Environ. Sci. Technol., Just Accepted Manuscript • DOI: 10.1021/acs.est.5b05179 • Publication Date (Web): 10 Mar 2016 Downloaded from http://pubs.acs.org on March 21, 2016

Just Accepted “Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides “Just Accepted” as a free service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are accessible to all readers and citable by the Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore, the “Just Accepted” Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these “Just Accepted” manuscripts.

Environmental Science & Technology is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.

Page 1 of 23

Environmental Science & Technology

1

How many environmental impact indicators are needed in the

2

evaluation of product life cycles?

3 4

Zoran Joeri Nikolaj Steinmann, MSc a,*, Dr. Aafke Margaretha Schippera, Dr. Mara Hauckb, Prof. Dr.

5

Mark Antonius Jan Huijbregtsa

6

a) Department of Environmental Science, IWWR, Radboud University, 6525 AJ Nijmegen, The

7

Netherlands

8

b) Climate, Air and Sustainability, TNO, Princetonlaan 6, 3584 CB Utrecht, The Netherlands

9

* Corresponding author: P.O. Box 9010 (box 89) NL-6500 GL Nijmegen, Visiting address:

10

Heyendaalseweg 135 (room HG02.632), NL-6525 ED Nijmegen, The Netherlands, Tel: +31-(0)24-

11

3652393, e-mail: [email protected]

12

Keywords: PCA, life cycle impact assessment, methods, indicators, selection

13

Abstract

14

Numerous indicators are currently available for environmental impact assessments, especially in the

15

field of Life Cycle Impact Assessment (LCIA). Because decision-making on the basis of hundreds of

16

indicators simultaneously is unfeasible, a non-redundant key set of indicators representative of the

17

overall environmental impact is needed. We aimed to find such a non-redundant set of indicators

18

based on their mutual correlations. We have used Principal Component Analysis (PCA) in

19

combination with an optimization algorithm to find an optimal set of indicators out of 135 impact

20

indicators calculated for 976 products from the ecoinvent database. The first four principal

21

components covered 92% of the variance in product rankings, showing the potential for indicator

22

reduction. The same amount of variance (92%) could be covered by a minimal set of six indicators,

ACS Paragon Plus Environment

Environmental Science & Technology

23

related to climate change, ozone depletion, the combined effects of acidification & eutrophication,

24

terrestrial ecotoxicity, marine ecotoxicity and land use. In comparison, four commonly used resource

25

footprints (energy, water, land, materials) together accounted for 82% of the variance in product

26

rankings. We conclude that the plethora of environmental indicators can be reduced to a small key

27

set, representing the major part of the variation in environmental impacts between product life

28

cycles.

29 30

Introduction

31

Over the course of the last two decades a wide variety of methods for quantifying environmental

32

impacts have been introduced into the area of Life Cycle Impact Assessment (LCIA).1 There are clear

33

differences in coverage and complexity of the various LCIA methods available. Damage-based

34

indicators, also called endpoint indicators, provide insight in the impact of emissions and resource

35

use through the full cause-and-effect chain, which ultimately leads to ecosystem and/or human

36

health damage.2 Midpoint indicators provide a proxy for environmental damage by modeling only

37

part of the cause-impact pathway. An example is the well-known global warming potential to

38

quantify the contribution of various greenhouse gas emissions to climate change in terms of

39

integrated radiative forcing.3 Finally, resource footprints estimate the resources required in a life

40

cycle without further impact or damage modeling. Cumulative energy demand is an example of such

41

an indicator.4-7 Additional variation in indicators results from the fact that there are several

42

competing methods available that quantify the same type of effect, but with different models and

ACS Paragon Plus Environment

Page 2 of 23

Page 3 of 23

Environmental Science & Technology

43

data. For example, the TRACI method8 and the EDIP method9 both provide midpoint indicators for

44

the impacts due to acidification.

45

A key question that this wide variety of available methods raises is whether the outcome of a

46

Life Cycle Assessment (LCA) will change among different indicators and methods.10-12 Furthermore, it

47

is unpractical to base decisions for environmental product optimization or environmental policy on

48

dozens of indicators simultaneously. Endpoint damage models solve the problem of having too many

49

indicators, but this goes at the expense of increased uncertainty, as uncertainties involved in damage

50

modelling can be very large.13-15

51

Therefore, as an alternative to the endpoint damage models, there is a clear need to provide

52

a limited set of indicators that is sufficiently small for efficient communication and decision-making,

53

but at the same time representative of the overall environmental impact. Several attempts have

54

been made to find a representative subset of relevant indicators. In a study commissioned by the

55

Joint Research Centre of the European Commission (JRC), various midpoint and endpoint indicators

56

and methods were qualitatively compared in terms of scientific and stakeholder acceptance.1

57

However, with 14 different recommended midpoint indicators, the set is still quite large and some of

58

the methods have since then undergone quite significant changes. Moreover, the JRC study did not

59

consider any redundancy among the recommended indicators caused by similarity in the underlying

60

processes that cause the emissions and environmental impacts. Others have suggested using a

61

“resource-based indicator family”, consisting of a set of resource footprints covering the cumulative

62

demand of land, water, carbon (or fossil energy) and, potentially, materials.16 However, while these

63

resource footprints are easy to use and communicate17, they may not account for the full range of

64

impacts employed in the damage-based methodologies. Finding a set of indicators that is optimal in

65

both its size and coverage requires a more systematic and quantitative way of reducing the number

66

of indicators. A few studies have attempted to achieve this via Principal Component Analysis (PCA).18-

67

22

68

by PCA. For example, only two out of eleven indicators were needed to cover more than 95% of the

If correlations between indicators are high, substantial dimensionality reduction can be achieved

ACS Paragon Plus Environment

Environmental Science & Technology

69

total variance of the environmental impacts within a set of nine household electronic products.23 A

70

more recent study, covering 17 impact indicators for two product categories (electricity and oil),

71

employed a slightly different and novel technique to demonstrate that the number of impact

72

indicators can be greatly reduced.24 However, while these studies provided valuable insights for one

73

or a few method(s) and/or product group(s), their coverage has been too small to answer questions

74

about which set of indicators is optimal in a more general sense. The search for an optimal set of

75

indicators using PCA has, to the best of our knowledge, not yet been applied over a large range of

76

products and impact assessment methods.

77

Here, we aim to find an optimal set of environmental indicators to cover the variance in the

78

rankings of a large number of products. We selected 976 products and 135 environmental indicators

79

from the ecoinvent 3.1 database25 as input for our analysis. We combined PCA with multiple

80

regression to arrive at the minimum set of indicators explaining the major part of the variance in the

81

product rankings. Apart from this minimum set, we also evaluated the extent to which four

82

commonly used resource footprints (fossil energy, water, land and materials) were representative of

83

the variation in product rankings.

84

85

Methods & Materials

86

Products

87

All products and corresponding environmental impacts were taken from the ecoinvent database

88

(version 3.1). Environmental impacts of all products are expressed per kilogram of product. To avoid

89

overlap among the products, we applied the following selection criteria:

90



Aggregated product categories (e.g., ‘inorganic chemicals’) were removed from the dataset.

91



In case of identical products, we preferred the global or ‘rest-of-the-world’ market mix over

92

specific sub-types based on particular production methods or regions. For example, we

93

selected global market rye grain and excluded organic rye grain and Swiss rye grain.

ACS Paragon Plus Environment

Page 4 of 23

Page 5 of 23

94

Environmental Science & Technology



In case of products with almost identical production chains (e.g., butanol and iso-butanol),

95

we selected the products with the largest overall amount of emissions or resource

96

extractions.

97

After applying the first two criteria, there were 1002 products remaining. To apply the third criterion,

98

we then searched for products with (nearly) identical product chains. This was done by dividing

99

every emission/resource extraction of a product by the same emissions/resource extractions of all

100

other products. This yielded 1001 sets of 1869 ratios for each product. If these ratios were (almost)

101

equal for all emissions and resources considered (coefficient of variation 1) was retained. Our final selection contained

105

976 products (Table S1), which we grouped into seven categories: Chemicals (435), Plastics (64),

106

Ores, minerals & fuels (91), Building materials (72), Processed bio-based products (80), Agricultural &

107

forestry products (106), Metal products & electronics (128).

108 109

Indicators

110

Ecoinvent25 reports a total of 692 impact indicators. Our method and the algorithm that we

111

employed to find the optimal set of indicators is not suitable for use on sets of data with (almost)

112

perfectly correlated data (i.e. a correlation coefficient of approximately 1) because two perfectly

113

correlated indicators are indistinguishable.26 Therefore we excluded impact indicators that met at

114

least one of the following criteria:

115



Newer versions of the same impact indicator were assumed to fully replace the older ones.

116

Therefore we only used the most recent version of each indicator. For example, the endpoint

117

indicator calculated with the Ecoscarcity 201327 method was retained, while the same

118

endpoint indicators from Ecoscarcity 2006 and 1997 were excluded.

ACS Paragon Plus Environment

Environmental Science & Technology

119



For most environmental impact indicators multiple versions are provided, either including or

120

excluding long-term emissions. These long-term emissions can stem from landfill sites, long

121

after the product is disposed. For most indicators there is no difference between the versions

122

including and excluding the long-term emissions. For toxicity indicators however, excluding

123

the emissions on a long-term leads to an underestimation of the damage. Therefore only the

124

indicator versions including the long-term emissions were retained for our analysis.

125



After these first two steps, the remaining set of 161 indicators included several indicators

126

that are identical for different impact assessment methods. For example, the Ozone

127

Depletion Potential (ODP) for an infinite time horizon is calculated in the same way in the

128

CML200128, EDIP20039 and Impact200229 methods. Similarly, there are several indicators for

129

global warming that are identical for different methods. In order to check whether there was

130

complete correlation between the indicators we used the “trim.matrix” function on the rank

131

correlation matrix of the 976 products in the R package “subselect”. 26 If two or more

132

indicators were found to be indistinguishable, only the indicator in the first method (based

133

on alphabetical order) was retained. This resulted in the removal of 26 indicators that all

134

showed a correlation of 1 to one or more of the other indicators.

135

After this procedure we retained a set of 135 environmental indicators for 976 products. The 135

136

environmental indicators belong to 13 different methods (Table 1).

137

Table 1. Impact assessment methods, numbers of indicators and corresponding main categories as included in

138

the analysis.

Impact assessment method CML2001 28 EDIP 2003 9 Impact 2002 29 ReCiPe 2 TRACI 8 EcoIndicator99 30 Ecological scarcity2013 27 EPS2000 31, 32 Impact 2002 29

Number of indicators 49 21 14 31 9 3 1 1 3

Category

Midpoint indicators

Endpoint indicators

ACS Paragon Plus Environment

Page 6 of 23

Page 7 of 23

Environmental Science & Technology

139

ReCiPe 2 3 Fossil energy footprint * 1 Water footprint * 1 Resource footprints Material footprint * 1 Land footprint * 1 * Calculated from the inventory outcomes of the ecoinvent database

140

Data analysis

141

We employed PCA combined with regression analysis to find the minimum set of indicators

142

necessary to adequately describe the variance in product rankings among the full set of impact

143

indicators. In PCA the total variation in a dataset is rewritten as a set of non-correlated linear

144

combinations of the original variables (in our case, the indicators), which are called principal

145

components. The first principal component describes the maximum amount of variance and

146

subsequent principal components explain decreasing amounts of variance. If there is a large

147

correlation between variables, the first couple of principal components will cover the majority of the

148

variance in the dataset and these components can be used as a highly parsimonious summary of the

149

total dataset.33

150

First, we performed a PCA on the rank correlation matrix, i.e. the correlation matrix of the

151

dataset after ranking the products on each of the 135 indicators from 1 to 976 (see Table S2). This

152

transformation to rank scores was done to give equal weight to each product in the dataset and to

153

each impact indicator (i.e. each indicator now has the same mean and standard deviation). It is

154

desirable to give equal weight to each product because the analysis of impact per kilogram of

155

product is an arbitrary choice that does not reflect differences in production volumes or value to

156

society. Without this transformation the correlation structure would be dominated by a few products

157

that have very high impacts per kilogram (for example the precious metals gold and platina).

158

Second, we determined the number of non-trivial components with the Avg-Pa (Average

159

parallel analysis) stopping rule as described by Peres-Neto et al..34 This stopping rule is a criterion

160

which tells whether a component describes more or less variance than one would expect if it was

161

based on purely randomized data. Because we work with product rankings rather than impact scores,

ACS Paragon Plus Environment

Environmental Science & Technology

162

we adjusted the procedure to work with randomized product rankings rather than normally

163

distributed random data. All integers from 1 to 976 were randomly sampled to create a dataset with

164

the same number of rows (976) and columns (135) as the original data. This was repeated 1000

165

times. A PCA was then performed on each of the 1,000 randomized datasets and the average

166

variance explained by each component was calculated. If a component explained more variance in

167

the original data than in the randomized data, the component was considered non-trivial.

168

Third, to interpret the non-trivial components, the loading of each indicator as well as the

169

principal component scores per product (and product group) were plotted. Biplots of the loadings

170

show the association between the indicators and the principal components: the more extreme (high

171

or low) the loading, the more representative the indicator is of that component. The principal

172

component scores show the associations between the observations and the principal components,

173

thus revealing the similarity of the indicators and products.

174

Fourth, to find a subset of the original indicators most closely associated with the non-trivial

175

principal components, we calculated the amount of variance in principal components scores that

176

could be explained by subsets of the original indicators. Subsets of any number of indicators can be

177

tested, whereby a set comprising all indicators would cover 100% of the variance and smaller subsets

178

would cover less. The best set, in terms of explained variance, for each number of indicators was

179

found by using the “improve” algorithm from the package “subselect” 26 in the statistical program R.35

180

The improve algorithm uses multiple linear correlation to determine the amount of explained

181

variance.

182

Fifth, we defined the optimal size of the indicator set by using the explained variance of the

183

so called non-trivial principal components as a benchmark. The indicator set with the lowest number

184

of indicators and an explained variance equal to or higher than the explained variance of the non-

185

trivial components was considered to be the preferred set of indicators.

186 187

Finally, we tested the extent to which a set of four resource footprints (fossil energy, land, water and material use) covered variance in our dataset by using the multiple linear correlation

ACS Paragon Plus Environment

Page 8 of 23

Page 9 of 23

Environmental Science & Technology

188

analysis on all principal component scores, as mentioned above. Note that these four resource

189

footprints were not part of the set of 135 indicators.

190

191

Results

192

Principal components and indicators

193

The number of non-trivial components was four, which together covered 92.0% of the total variance

194

in the dataset. The majority of the variance (83.3%) was covered by the first component, with 3.9%,

195

3.1% and 1.9% covered by the three consecutive components (Figure 1). Since all correlations in the

196

database were positive, the first component can be seen as an overall indicator of environmental

197

impact. Indicators that are most similar to other indicators (i.e. have the highest mutual correlations)

198

have the highest absolute loadings on this component. For the first component these were the

199

indicators related to freshwater and marine ecotoxicity. However, most other indicators showed very

200

similar loadings (Figure S1, see SI1 for a more extensive analysis of the principal components).

201

Indicators with the lowest absolute loadings on the first component were related to land use, which

202

means that indicators related to land use are least correlated to all other indicators. On the second

203

principal component, high loadings were found for indicators related to land use and terrestrial

204

ecotoxicity while indicators related to fossil energy use and global warming had low loadings. This

205

means that after taking into account the fact that impact-intensive products are generally impact-

206

intensive according to all impact indicators; products that use relatively large amounts of land are

207

likely to use a relatively limited amount of fossil energy and vice versa. Indicators related to ionizing

208

radiation (i.e. from the use of nuclear energy) and indicators of ozone depletion were further

209

distinguished from the rest by components 3 and 4. See Figures S1 and S2 for the loadings of all

210

indicators on the first four components.

ACS Paragon Plus Environment

Environmental Science & Technology

211 212

Figure 1. Variance explained per principal component based on our dataset (blue), per principal component

213

based on random data (purple), and by the best set of indicators (orange). Solid lines with dots indicate the

214

cumulative variance explained.

215

Products

216

The scores per principal component (Figure 2) indicate how the products are related to each of the

217

components. Products with the highest and lowest overall ranks showed the most extreme scores on

218

the first principal component. On a per kg basis, we found the highest impacts (lowest scores on the

219

first principal component) for metal products and electronics (which include very high-priced and

220

energy-intensive products such as gold), and the lowest impacts for ores, minerals & fuels and

ACS Paragon Plus Environment

Page 10 of 23

Page 11 of 23

Environmental Science & Technology

221

building materials. On the second component a clear distinction could be seen between the products

222

on a biological base (agricultural & forestry products and processed bio-based products) and the rest

223

of the products. This coincided with the high loadings for (agricultural) land use on the second

224

principal component. A number of plastics showed the lowest scores on the third component. These

225

plastics have relatively small emissions of ozone depleting substances as a result of their production

226

process compared to the other emissions. The highest scores are found among the inorganic

227

chemicals.

228 229

Figure 2. Boxplots of principal component scores per product and product group: colored boxes cover the

230

interquartile distance, black bars inside the boxes denote the medians, whiskers extend to the 5 and 95

231

quantiles.

th

ACS Paragon Plus Environment

th

Environmental Science & Technology

Page 12 of 23

232

Best set of indicators

233

A set of six indicators was needed to cover slightly more variance than the first four principal

234

components (i.e. 92.3% and 92.0% respectively; Figure 1). The best set of six indicators contained

235

indicators of climate change, ozone depletion, terrestrial ecotoxicity, the combined ecosystem

236

effects of acidification & eutrophication, marine ecotoxicity and land use (Table 2).

237

The four resource footprints together accounted for an explained variance of 82.0% (Table

238

2). The results suggest that the fossil energy indicator, with an explained variance of 72.9%, is a

239

reasonable indicator of overall impact. The explained variance can be raised to 76.8% by adding

240

material use and then to 80.1% by adding land use. The water footprint explained only 1.9% of

241

additional variance; this is due to the fact that water consumption is related to both the energy-

242

intensive process of electricity generation and the land-intensive process of crop production.

243

Table 2. Best sets of impact indicators and the variance explained by sets up to six indicators, and the variance

244

explained by the resource footprints.

#

Impact indicator

Method: Full name of indicator

1 2

Marine ecotoxicity Climate change Marine ecotoxicity Land use Human toxicity Climate change Climate change Ozone depletion Land use Marine ecotoxicity Climate change Ozone depletion Terrestrial ecotoxicity Acidification & eutrophication Marine ecotoxicity Climate change Land use Ozone depletion Acidification & eutrophication Marine ecotoxicity Terrestrial ecotoxicity

ReCiPe: METP100a (I) CML2001: upper limit of net GWP100a* ReCiPe: METPinf (H) Impact2002: land occupation ReCiPe: HTPinf (E) ReCiPe: GWP100a (H) CML2001: upper limit of net GWP100a CML2001: ODP40a Impact2002: land occupation ReCiPe: METPinf (E)

3

4

5

6

Explained variance 78.4% 84.0%

87.8%

90.1%

CML2001: GWP20a CML2001: ODP40a CML2001: TAETP20a Impact2002: terrestrial acidification & eutrofication ReCiPe: METPinf (E)

91.3%

CML2001: GWP20a CML2001: Land use: competition** CML2001: ODP40a Impact2002 ReCiPe Midpoint E: METPinf ReCiPe Midpoint I: TETP100a

92.3%

ACS Paragon Plus Environment

Page 13 of 23

Environmental Science & Technology

Resource-based sets 1 Fossil energy Land Water Material 2 Fossil Energy, Material 3 Fossil Energy, Land, Material 4 Fossil Energy, Land, Material and Water

72.9% 37.9% 58.1% 54.3% 76.8% 80.1% 82.0%

245

Abbreviations:

246

METP: Marine EcoToxicity Potential

247

GWP: Global Warming Potential

248

HTP: Human Toxicity Potential

249

ODP: Ozone Depletion Potential

250

T(A)ETP: TerrestriAl EcoToxicity Potential

251

inf/100a/40a/20a: Time Horizons of infinite duration and 100, 200 and 40 years respectively

252

(I): Individualist perspective, (E): Egalitarian perspective, (H): Hierarchist perspective

253

* Upper limit of GWP100a: a method where the upper values of several uncertain GHGs are used rather than

254

the most likely value

255

** Like the land footprint this consists of an unweighted summation of different land uses, however not all

256

types are included

257

Discussion

258

Our analysis revealed a large amount of redundancy in a set of 135 impact indicators, since a

259

single indicator already covered 78.4% of the variance in the ranking of 976 products. This result is in

260

line with earlier research in which large correlations between a single indicator (energy demand) and

261

different impact indicators were found.36, 37 However, even though one indicator might be able to

262

cover a large proportion of the damage over a wide range of products, within certain product groups

263

it may be less useful to distinguish products from each other.38 As the scores of the individual

264

products demonstrate, several indicators are needed to distinguish products within specific groups.

265

For example, bio-based products are mostly separated by the amount of land use, while some

266

plastics have relatively high emissions of ozone depleting substances; therefore an indicator that

ACS Paragon Plus Environment

Environmental Science & Technology

267

takes this into account is necessary to distinguish products within this group. The requirement of a

268

limited set of indicators is in line with the findings from Gutierrez et al. 23 who used a PCA-based

269

approach for data reduction. These authors found that three indicators (respiratory inorganics, ozone

270

depleting substances and minerals) were needed to cover the majority of the variance in

271

environmental impacts of a small set of household electronics. Pascual-Gonzalez et al. 24 applied a

272

combination of multiple regression and an approach called MILP (Mixed-Integer-Linear-

273

Programming) to reduce the number of impact indicators for a large number of oil and electricity

274

products. Fossil energy demand, respiratory inorganics and ozone depletion constituted the best set

275

for oil, while for electricity the different types of primary energy carriers (water, wind, nuclear and

276

fossil) and an indicator of climate change were needed. Our study, adds to these insights, by

277

providing an application of a data reduction technique on a dataset that is both larger and more

278

diverse than used up to now.

279

According to our results, the best indicator set includes indicators of global warming (which is

280

strongly correlated to all other indicators), land use and ozone depletion, supplemented by indicators

281

for the combined effect of acidification and eutrophication (these two impacts are combined into

282

one indicator in the IMPACT 2002 methodology), and terrestrial and marine ecotoxicity. While this

283

set maximizes the amount of explained variance, these six indicators are not necessarily the most

284

preferable according to additional criteria, such as the RACER (Relevant, Accepted, Credible, Easy and

285

Robust) criteria.39 For example, the problem of ozone depletion is becoming less relevant due to

286

successful emission reduction policies40 and the indicators of toxicity (including the statistically best

287

indicator for marine ecotoxicity) are not always robust, due to the large uncertainties involved in

288

their calculations.13-15 As the loadings of the indicators (figure S1) on the first component (which

289

covers 83.3% of the total variance) show, there are several indicators with approximately the same

290

explanatory power. This means that alternative sets of indicators can be defined which are only

291

marginally worse in terms of explained variance compared to the set of six indicators derived from

292

our analysis. The resource footprints are relatively easy and robust to calculate. Being indicators of

ACS Paragon Plus Environment

Page 14 of 23

Page 15 of 23

Environmental Science & Technology

293

resource use rather than indicators of environmental impact, it was not known whether they can be

294

used as a proxy for environmental impacts. Our results indicate that they cover the variance in

295

impact indicators quite well (82% explained variance with four indicators), which suggests that these

296

indicators are relevant for use in the area of LCIA. Using simple resource footprints would reduce the

297

need for the complicated mid- and endpoint damage models.

298

In this study we have used product rankings calculated on a 1 kg basis, to give insight in the

299

variation between very diverse products. It can be argued that other bases of comparison, for

300

instance, based on products with the same functional unit, are more in line with one of the main

301

applications of LCA: product optimization. Comparing products with the same functional unit was

302

done by Sabio et al.21, who found four relevant environmental indicators for hydrogen supply chains,

303

thereby also demonstrating that a small set of indicators was a sufficient base for the comparison of

304

similar product alternatives. The products used in our study are mostly quite simple commodities,

305

with only the electronics category covering more complex consumer products, such as LCD screens.

306

Furthermore, the analysis was based on cradle-to-gate data, whereby the product’s use and disposal

307

are not taken into account. The use phase of products can be energy-intensive, meaning that, if

308

included, products with long use times would generally rank higher on indicators related to energy

309

use, such as indicators of climate change. It is difficult to predict how the correlation structure, and

310

therefore the indicator selection, would change when more complex products or use phase data are

311

taken into account. The correlation between indicators could either increase or decrease, resulting in

312

more or less explained variance for indicator sets of each size. However, in the case of taking into

313

account the use phase, it is likely that there will be more explained variance with the same number

314

of indicators, since all indicators related to energy use will increase simultaneously for products with

315

long use times.

316

Additionally, the coverage of the ecoinvent database may not be complete in terms of the

317

inventory flows included (i.e. the extent to which relevant emissions and resource extractions are

318

included for all products) and in terms of (novel) impact pathways covered. Limiting the inventory

ACS Paragon Plus Environment

Environmental Science & Technology

319

flows to a limited number of well-studied pollutants may result in an overestimation of the

320

correlations between the impact indicators that are eventually calculated from these flows, for

321

example not all of the toxic chemicals covered by the USEtox model13 are available in the ecoinvent

322

database. Similarly, leaving out potentially relevant damage pathways (for example ecosystem

323

damage caused by tropospheric ozone formation), may also cause an overestimation of the amount

324

of variance that can be covered with a limited number of indicators. Despite these limitations, it

325

should be noted that the ecoinvent database is one of the largest databases available for life cycle

326

assessment, with clear guidelines on how to ensure completeness and consistency of the data as

327

much as possible.41

328

Finally, spatially explicit impact assessment methods, as recently developed for several

329

impact categories, including water, land, acidification and eutrophication42-45, were not included in

330

the assessment. This is due to the fact the coupling of spatially explicit inventory and impact

331

information is still not common practice in LCA databases, including ecoinvent 3.1. Spatially explicit

332

methods use a different amount of impact for the same environmental intervention, which can

333

reorder the ranking of the impact indicator, depending on the location where the intervention takes

334

place. Whether this results in more or less correlation between the different impact indicators

335

cannot be known a priori. Regionalization may increase the amount of variance to be explained by a

336

small set of indicators and therefore limit the coverage of our recommended set, however

337

regionalization can potentially also increase the correlations between indicators. This would be the

338

case if regionalization results in several impact indicators that are consistently higher than average in

339

one country. Products that rely heavily on that country as a sourcing region will then have a higher

340

rank for all these indicators of impact, resulting in a higher correlation between the indicators than

341

one would observe without regionalization of the impacts. It is recommended to redo the PCA-

342

regression procedure after the introduction of spatially explicit impact assessment methods in

343

common LCA practice.

ACS Paragon Plus Environment

Page 16 of 23

Page 17 of 23

Environmental Science & Technology

344

Summarizing, we introduced a procedure for systematically selecting a set of impact

345

indicators that is optimal in covering the maximum amount of variance with a minimal number of

346

indicators. With this method we have demonstrated that reducing the number of indicators in our

347

dataset from 135 to six indicators is feasible without losing more than 8% of the total variance.

348

Although it is unlikely that any practitioner would have used 135 indicators simultaneously, there

349

was no reason to discard any of the indicators a priori other than for reasons of convenience. Our

350

results can be seen as comforting for LCA practitioners, since we showed a clear overlap between

351

and within methods. For the first time, we have demonstrated a large dimensionality reduction to be

352

feasible over a wide range of indicators and products. This suggests that the plethora of indicators

353

available for product assessments is largely superfluous and that a small set of key impact indicators

354

is sufficient for product optimization and environmental policy. Alternatively, a set of 4 resource

355

footprints covering 82% of the total variance may be used.

356

Author information

357

Corresponding author

358

*Tel: +31-(0)24-3652393, e-mail: [email protected]

359

Notes

360

The authors declare no conflict of interest

361

Acknowledgements

362

This project was funded by the European Commission via the WP7 project Desire: number 308552.

363

Associated content

364

Supporting information

ACS Paragon Plus Environment

Environmental Science & Technology

365

The supporting information contains additional information about the indicator loadings, including

366

figure S1 and S2 which display the loadings of all indicators. A full list of all indicators is also included

367

in table S1, and the correlation matrix (table S2) is included as a separate Excel file. This information

368

is available free of charge via the Internet at http://pubs.acs.org/.

369

References

370

1. Hauschild, M. Z.; Goedkoop, M.; Guinée, J.; Heijungs, R.; Huijbregts, M.A.J.; Jolliet, O.; Margni, M.;

371

De Schryver, A; Humbert, S.; Laurent, A.; Salla, S.; Pant, R. Identifying best existing practice for

372

characterization modeling in life cycle impact assessment. Int. J. Life Cycle Assess. 2009, 18(3), 683-

373

697.

374

2. Goedkoop, M.J.; Heijungs; R. Huijbregts M. A. J.; De Schryver A.; Struijs J.; van Zelm, R. ReCIPe

375

2008: A life cycle impact assessment method which comprises harmonised category indicators at the

376

midpoint and the endpoint level. Report 1: Characterisation. 2013, http://www.lcia-recipe.net.

377

3. IPCC. Climate Change 2013: The Physical Science Basis. Contribution of Working Group I to the

378

Fifth Assessment Report of the Intergovernmental Panel on Climate Change [Stocker, T.F., D. Qin, G.-

379

K. Plattner, M. Tignor, S.K. Allen, J. Boschung, A. Nauels, Y. Xia, V. Bex and P.M. Midgley (eds.)].

380

Cambridge University Press, Cambridge, United Kingdom and New York, NY, USA, 1535 pp. 2013

381

4. Hirst, E. Food-related energy requirements. Science 1974, 184(4133), 134-138.

382

5. Arvesen, A., Hertwich, E.G. More caution is needed when using life cycle assessment to determine

383

energy return on investment (EROI). Energy Policy 2015, 76(0), 1-6

384

6. Arvidsson, R., Fransson, K., Fröling M., Svanström M., Molander, S. Energy use indicators in energy

385

and life cycle assessments of biofuels: review and recommendations. J. Clean. Prod. 2012, 31, 54-61

386

7. Frischknecht, R., Wyss, F., Büsser Knöpfel, S., Lützkendorf, T., Balouktsi, M. Cumulative energy

387

demand in LCA: the energy harvested approach. Int. J. Life Cycle Assess. 2015, 20(7), 957-969

ACS Paragon Plus Environment

Page 18 of 23

Page 19 of 23

Environmental Science & Technology

388

8. Bare, J. TRACI 2.0: the tool for the reduction and assessment of chemical and other environmental

389

impacts 2.0. Clean Technol. Environ. Policy 2011, 13(5), 687-696.

390

9. Hauschild M.; Potting J. Spatial differentiation in life cycle impact assessment—the EDIP2003

391

methodology. Environmental News no. 80, 2005, The Danish Ministry of the Environment,

392

Environmental Protection Agency, Copenhagen

393

10. Bueno, C.; Hauschild, M.Z.; Rossignolo, J.A.; Ometto, A.R.; Mendes, N.C. Sensitivity Analysis of the

394

use of Life Cycle Impact Assessment Methods: A case study on building materials. J. Clean. Prod.

395

2015, doi:10.1016/j.jclepro.2015.10.006.

396

11. Dreyer, L.C., Niemann, A.L., and Hauschild, M.Z. Comparison of three different LCIA methods:

397

EDIP97, CML2001 and Eco-indicator 99. Does it matter which one you choose? Int. J. Life Cycle

398

Assess. 2003, 8(4), 191-200.

399

12. Owsianiak, M.; Laurent, A.; Bjørn, A.; Hauschild, M.Z. IMPACT 2002+, ReCiPe 2008, and ILCD's

400

recommended practice for characterization modelling in life cycle impact assessment: a case study-

401

based comparison. Int. J. Life Cycle Assess. 2014, 19(5), 1007-1021.

402

13. Rosenbaum, R.K.; Bachmann, T.M.; Gold, L.S.; Huijbregts, M.A.J.; Jolliet, O.; Juraske, R.; … ;

403

Hauschild, M.Z. (2008). USEtox—the UNEP-SETAC toxicity model: recommended characterisation

404

factors for human toxicity and freshwater ecotoxicity in life cycle impact assessment. Int. J. Life Cycle

405

Assess. 2008, 13(7), 532-546.

406

14. van Zelm, R., Huijbregts, M. A.J., Harbers, J. V., Wintersen, A., Struijs, J., Posthuma, L., van de

407

Meent, D. Uncertainty in msPAF-based ecotoxicological effect factors for freshwater ecosystems in

408

life cycle impact assessment. Integr. Enviro. Assess. Manage. 2007, 3(4), e6-e37.

ACS Paragon Plus Environment

Environmental Science & Technology

409

15. van Zelm, R., Huijbregts, M.A.J., Posthuma, L., Wintersen, A., van de Meent, D. Pesticide

410

ecotoxicological effect factors and their uncertainties for freshwater ecosystems. Int. J. Life Cycle

411

Assess. 2009, 14(1), 43-51.

412

16. Galli, A.; Wiedmann, T.; Ercin, E.; Knoblauch, D.; Ewing, B.; Giljum, S. Integrating ecological,

413

carbon and water footprint into a “footprint family” of indicators: definition and role in tracking

414

human pressure on the planet. Ecol. Indic. 2012, 16, 100-112.

415

17. European Resource Efficiency Platform. Manifesto & policy recommendations. 2014 European

416

Commission, Brussels, Belgium.

417

18. Pozo, C.; Ruiz-Femenia, R.; Caballero, J.; Guillén-Gosálbez, G.; Jiménez, L. On the use of Principal

418

Component Analysis for reducing the number of environmental objectives in multi-objective

419

optimization: Application to the design of chemical supply chains. Chem. Eng. Sci. 2012, 69(1), 146-

420

158.

421

19. Brunet, R.; Guillén-Gosálbez, G.; Jiménez, L. Cleaner design of single-product biotechnological

422

facilities through the integration of process simulation, multiobjective optimization, life cycle

423

assessment, and principal component analysis. Ind. Eng. Chem. Res. 2011, 51(1), 410-424.

424

20. de Saxcé, M.; Rabenasolo, B.; Perwuelz, A. Assessment and improvement of the appropriateness

425

of an LCI data set on a system level–application to textile manufacturing. Int. J. Life Cycle Assess.

426

2014, 19(4), 950-961.

427

21. Sabio, N.; Kostin, A.; Guillén-Gosálbez, G.; Jiménez, L. Holistic minimization of the life cycle

428

environmental impact of hydrogen infrastructures using multi-objective optimization and principal

429

component analysis. Int. J. Hydrogen Energ. 2012, 37(6), 5385-5405.

430

22. Li, T.; Zhang, H.; Yuan, C.; Liu, Z.; Fan, C. A PCA-based method for construction of composite

431

sustainability indicators. Int. J. Life Cycle Assess. 2012, 17(5), 593-603.

ACS Paragon Plus Environment

Page 20 of 23

Page 21 of 23

Environmental Science & Technology

432

23. Gutiérrez, E.; Lozano, S.; Adenso-Díaz, B. Dimensionality reduction and visualization of the

433

environmental impacts of domestic appliances. J. Ind. Ecol. 2010, 14(6), 878-889.

434

24. Pascual-González, J.; Pozo, C.; Guillén-Gosálbez, G.; Jiménez-Esteller, L. Combined use of MILP

435

and multi-linear regression to simplify LCA studies. Comput. Chem. Eng., 2015, 82, 34-43

436

25. Moreno Ruiz E.; Weidema B. P.; Bauer C.; Nemecek T.; Vadenbo C. O.; Treyer K.; Wernet G.

437

Documentation of changes implemented in ecoinvent Data 3.0. Ecoinvent Report 5 (v4) 2013. St.

438

Gallen: The ecoinvent Centre

439

26. Orestes Cerdeira, J.; Duarte Silva P.; Cadima J. Minhoto M. subselect: Selecting Variable Subsets.

440

R package version 0.12-5. http://CRAN.R-project.org/package=subselect, 2015

441

27. Frischknecht R.; Büsser Knöpfel S. Swiss Eco-Factors 2013 according to the Ecological Scarcity

442

Method. Methodological fundamentals and their application in Switzerland. Environmental studies

443

no. 1330. 2013, Federal Office for the Environment, Bern: 254 pp.

444

28. Guinée, J.B.; Gorrée, M.; Heijungs, R.; Huppes, G.; Kleijn, R.; Koning, A. de; Oers, L. van; Wegener

445

Sleeswijk, A.; Suh, S.; Udo de Haes, H.A.; Bruijn, H. de; Duin, R. van; Huijbregts, M.A.J. Handbook on

446

life cycle assessment. Operational guide to the ISO standards. I: LCA in perspective. IIa: Guide. IIb:

447

Operational annex. III: Scientific background. Kluwer Academic Publishers, ISBN 1-4020-0228-9,

448

Dordrecht, 2002, 692 pp.

449

29. Jolliet O.; Margni M.; Charles R.; Humbert S.; Payet J.; Rebitzer G.; Rosenbaum R. IMPACT 2002+:

450

a new life cycle impact assessment methodology. Int. J. Life Cycle Assess. 2003, 8(6),324–330

451

30. Goedkoop M.J.; Spriensma R. Eco-indicator 99, a damage oriented method for lifecycle impact

452

assessment, methodology report, (update June 2001) http://www.pre-sustainability.com/scientific-

453

publications

ACS Paragon Plus Environment

Environmental Science & Technology

454

31. Steen B. A systematic approach to environmental priority strategies in product development

455

(EPS). Version 2000-general system characteristics; CPM report 1999:4, 1999, Chalmers University of

456

Technology, Gothenburg, Sweden

457

32. Steen B .A systematic approach to environmental priority strategies in product development

458

(EPS). Version 2000-models and data of the default method; CPM report 1999:5, 1999, Chalmers

459

University of Technology, Gothenburg, Sweden

460

33. Cadima, J. F., & Jolliffe, I. T. (2001). Variable selection and the interpretation of principal

461

subspaces. J. Agric. Biol. Environ. S. 2001, 6(1), 62-79.

462

34. Peres-Neto, P. R.; Jackson, D. A.; Somers, K. M. How many principal components? Stopping rules

463

for determining the number of non-trivial axes revisited. Comput. Stat. Data An. 2005, 49(4), 974-

464

997.

465

35. R Core Team. R: A language and environment for statistical computing. R Foundation for

466

Statistical Computing, Vienna, Austria. http://www.R-project.org/, 2015

467

36 Huijbregts, M.A.J.; Rombouts, L. J.; Hellweg, S.; Frischknecht, R.; Hendriks, A. J.; van de Meent, D.;

468

Ragas, A.M.J.; Reijnders, L.; Struijs, J. Is cumulative fossil energy demand a useful indicator for the

469

environmental performance of products?. Environ. Sci. Technol. 2006, 40(3), 641-648.

470

37. Huijbregts, M.A.J.; Hellweg, S.; Frischknecht, R.; Hendriks, H.W.; Hungerbühler, K.; Hendriks, A.J.

471

Cumulative energy demand as predictor for the environmental burden of commodity production.

472

Environ. Sci. Technol. 2010, 44(6), 2189-2196.

473

38. Laurent, A.; Olsen, S.I.; Hauschild, M. Z. Limitations of carbon footprint as indicator of

474

environmental sustainability. Environ. Sci. Technol. 2012, 46(7), 4100-4108.

475

39. Lutter, S.; Giljum S. Development of a methodology for the assessment of global environmental

476

impacts of traded goods and services. Development of RACER Evaluation Framework. EIPOT

ACS Paragon Plus Environment

Page 22 of 23

Page 23 of 23

Environmental Science & Technology

477

Workpackage 2, 2008 http://seri.at/wp-content/uploads/2010/06/EIPOT-RACER-evaluation-

478

framework-final-07Oct08.pdf

479

40. WMO Scientific assessment of ozone depletion: 2010, Global Ozone Research and Monitoring

480

Project-report no.52., 2011, World Meteorological Organization, Geneva

481

41. Weidema, B.P., Bauer C.; Hischier, R.; Mutel, C.; Nemecek, T.; Reinhard, J.; Vadenbo, C.O.;

482

Wernet, G. Overview and methodology. Data quality guideline for the ecoinvent database version 3.

483

Ecoinvent Report 1 (v3) 2013. St. Gallen: The ecoinvent Centre

484

42. Verones, F.; Saner, D.; Pfister, S.; Baisero, D.; Rondinini, C.; Hellweg, S. Effects of consumptive

485

water use on biodiversity in wetlands of international importance. Environ. Sci. Technol. 2013, 47(21),

486

12248-12257.

487

43. De Baan, L.; Mutel, C. L.; Curran, M.; Hellweg, S.; Koellner, T. Land use in life cycle assessment:

488

global characterization factors based on regional and global potential species extinction. Environ. Sci.

489

Technol. 2013, 47(16), 9281-9290.

490

44. Roy, P. O.; Azevedo, L. B.; Margni, M.; van Zelm, R.; Deschênes, L.; Huijbregts, M. A.

491

Characterization factors for terrestrial acidification at the global scale: A systematic analysis of spatial

492

variability and uncertainty. Sci. Total Environ. 2014, 500, 270-276.

493

45. Azevedo, L. B.; Henderson, A. D.; van Zelm, R.; Jolliet, O.; Huijbregts, M. A. Assessing the

494

importance of spatial variability versus model choices in life cycle impact assessment: the case of

495

freshwater eutrophication in Europe. Environ. Sci. Technol. 2013, 47(23), 13565-13570.

ACS Paragon Plus Environment