Technical Annex IV:

Scope 3 Overview and Modelling

CDP Full GHG Emissions Dataset 2016

Authors:

Hugh Sawbridge

Dr. Paul Griffin

Contributors:

Ian van der Vlugt

Mallika Sharma

Additional Project team members:

Jacopo Peirano

Rebecca Shannon

Tom Crocker

Holly Taylor

Andrew Clapper

29 June 2017 Page 1 of 30@cdp | www.cdp.net @cdp | www.cdp.net

Contents

1 Introduction .......................................................................................................................... 3

1.1 Scope 3 overview ............................................................................................................ 3

1.2 Purpose of this dataset .................................................................................................... 3

1.2.1 Scope 1, 2 and 3 dataset comparisons ..................................................................... 4

1.2.2 Suggested uses for Scope 3 data ............................................................................. 4

1.2.3 As a tool for engagement .......................................................................................... 5

1.3 Scope 3 Data Reported to CDP ....................................................................................... 6

1.4 Driving factors behind disclosure ..................................................................................... 7

1.4.1 Ease of calculation .................................................................................................... 7

1.4.2 Relevance ................................................................................................................. 7

1.4.3 Demand from investors ............................................................................................. 8

1.4.4 Demand from other stakeholders .............................................................................. 8

1.5 Issues of comparability with reported Scope 3 data ......................................................... 9

1.5.1 Incomplete data ........................................................................................................ 9

1.5.2 Example: Different business models ......................................................................... 9

1.5.3 Example: Differences in calculation methodologies ..................................................10

1.5.4 Example: Different interpretations of category definitions .........................................11

1.5.5 Example: Different consolidation approaches ...........................................................12

1.6 Cleaning Scope 3 data .................................................................................................. 13

2 Statistical Models and Application to Scope 3 ................................................................ 14

2.1 Statistical Models ........................................................................................................... 14

2.2 Model assumptions ........................................................................................................ 14

2.2.1 Activity revenue as the independent variable ...........................................................14

2.2.2 Relationship to FTE and CapEx models .......................................................................14

2.2.2 Companies use similar calculation methodologies ...................................................15

2.3 CDP Activity Classification System ................................................................................ 16

2.3.1 Classification hierarchy ............................................................................................16

2.3.2 Climate change hybrid classification system ............................................................16

2.4 Three levels of model granularity ................................................................................... 16

2.5 Model selection ............................................................................................................. 18

2.5.1 AIC single model selection .......................................................................................18

2.5.2 Stepwise activity based model selection ..................................................................18

2.5.3 Applicable Scope 3 categories .................................................................................19

29 June 2017 Page 2 of 30@cdp | www.cdp.net @cdp | www.cdp.net

3 Bottom-up ‘Use of Sold Product’ modelling .................................................................... 20

3.1 Data sources ................................................................................................................. 20

3.2 Methodology .................................................................................................................. 20

3.2.1 Oil and Gas ..............................................................................................................20

3.2.2 Coal .........................................................................................................................22

3.2.3 Automotive Manufacturing ........................................................................................22

4 Results ................................................................................................................................ 23

4.1 General Results ............................................................................................................. 23

4.2 Use of Bottom Up Estimates for ‘Use of Sold Products’ ................................................. 25

4.3 Use of mixed level Gamma GLM ................................................................................... 25

4.4 Success of stepwise model selection approach ............................................................. 26

Appendix 1 Statistical Modelling References ...................................................................... 27

Appendix 2 Bottom-up Modelling References ..................................................................... 28

Appendix 3 Common Comparability Issues ......................................................................... 29

29 June 2017 Page 3 of 30@cdp | www.cdp.net @cdp | www.cdp.net

1 Introduction

1.1 Scope 3 overview

The purpose of this document is to provide CDP’s investor signatories with an overview of Scope 3 emissions, a methodology of the models used to fill gaps in reported data and examples of how the data can be used. The first section explains how and why Scope 3 data is reported and highlights some of the key issues of comparability. Section 2 describes how these issues have informed the development of the statistical models. Section 3 describes how company activity data is used to estimate the ‘Use of Sold Products’ emissions in key sectors and the concluding section describes how this dataset is compiled and assesses the results.

By filling gaps in reported data with modelled estimates this project helps to overcome one of the main hurdles that prevent investors using Scope 3 data in their analysis.

For many companies, the indirect emissions caused by their business can far outweigh their direct emissions. Scope 3 represents the indirect GHG emissions of a company from all sources excluding purchased energy, accounted for under Scope 2.

The GHG Protocol splits Scope 3 emissions into 15 different categories, grouped into Upstream and Downstream – the Scope 3 question (CC14.1) in the CDP Climate Change Questionnaire is based on this standard. The GHG Protocol provides guidance on how the emissions for each category may be calculated. While this guidance is widely used, it is less prescriptive than the Scope 1 & 2 guidance and differing interpretations result in varied response values.

Companies may account for their Scope 3 emissions in several ways, with more than one correct way to compile a GHG inventory. Companies involved in similar activities can have very different corporate structures, resulting in different emissions profiles. This presents a problem when trying to compare the emissions profiles across companies. While these difficulties exist in Scope 1 & 2 reporting, they are magnified by the lack of consensus on Scope 3 reporting.

Scope 3 emissions by their very nature occur outside of the reporting company’s control boundary, it is often difficult for companies to collect sufficient primary data to be able to calculate their Scope 3 emissions to the same level of accuracy as scope 1 & 2. Simplifying assumptions can be made to overcome this lack of primary data, different assumptions can drastically alter the final figure. Each Scope 3 category has its own issues to do with data collection, behavioral assumptions and boundary settings, which are summarized in Appendix 1.

1.2 Purpose of this dataset

Increasingly investors are recognizing the importance of Scope 3 accounting and exploring means to integrate these emissions into their corporate assessments. Often representing upwards of 90% of a company’s total GHG footprint, excluding these emissions results in an incomplete representation of a company’s impacts.

In order to support investors and other stakeholders in their work, this dataset provides a reflection of current Scope 3 accounting practices. It is a first step in improving Scope 3 accounting and disclosure while providing users with emissions profiles that cover the full corporate value chain; most Scope 3 datasets only cover supply chain emissions. As reporting practises become more consistent, the uncertainty in these models will decrease.

29 June 2017 Page 4 of 30@cdp | www.cdp.net @cdp | www.cdp.net

1.2.1 Scope 1, 2 and 3 dataset comparisons

This data set has been compiled independently from the Scope 1 & 2 data set and should be used as an extension to the core Scope 1 & 2 dataset. Whilst the statistical approach is being used for Scopes 1, 2 & 3 is the same and many of the modelling assumptions are similar, the cleaning of the Scope 1 & 2 data has been far more thorough and each Scope has been viewed independently.

Please see table 1 below for a brief overview of the high-level aspects and differences between the Scope 1, 2 and 3 data sets.

Table 1 Scope 1, 2, and 3 Dataset comparisons

Scope 1 & 2 Dataset Scope 3 Dataset

Statistical Models

Estimates based on Activity average revenue intensity.

Estimates for each Scope 3 category based on either Activity/Sector/Industry average revenue intensity.

Bottom Up Estimates

Available for Oil & Gas Power Generators, Cement Producers, Steel Producers & Coal Extractors.

Only available for Oil & Gas, Coal Extraction & Automobile Manufacturers.

Applicability Estimates used to fill all gaps in S1 & Location-based S2 data.

Scope 3 estimates only fill gaps for categories that are applicable to each company (see section 2.5.3).

Other Reported

data

Any other data reported to CDP is included in same sheet as modelled data.

Other reported data kept in a separate sheet because the Scope 3 inventory is incomplete for these companies is and therefore incomparable.

1.2.2 Suggested uses for Scope 3 data

This dataset provides users with complete Scope 3 inventories for the companies in the sample providing investors with a comprehensive breakdown of each company’s climate change impacts. There are many issues with Scope 3 data (discussed later in section 1.4), this dataset attempts to resolve the issue that reported Scope 3 data is too sparse to use by filling gaps with modelled estimates however, the other issues persist.

In providing users with complete Scope 3 inventories this dataset helps overcome the most significant barrier to incorporating Scope 3 data into any qualitative or quantitative analysis. For any analysis, it is important to understand the issues of with reported Scope 3 data and how these issues manifest themselves in the statistical models.

1.2.2.1 In Portfolio Carbon Footprinting

Within a single company’s GHG inventory. However, ‘Use of Sold Products’ emissions from a company will also be included in many companies Scope 1 emissions. There is even some double counting across the 15 different Scope 3 categories (e.g. a coal company’s ‘Use of Sold Products’ emissions will be accounted in a steel producer’s ‘Processing of Sold Products’ emissions through the manufacturing of steel). According to Kepler Chevreaux’s Carbon Compass the risk of double counting increases with the sectoral diversification and size of the portfolio. One way to quantify this double counting across supply chains would be to use an Input-Output model. Regional diversification could help reduce double counting of emissions sources as there are likely to be fewer linkages across supply chains.

Scope 3 emissions data is crucial for a comprehensive assessment of a company’s climate change impacts. In the Carbon Compass paper, the authors argue that inclusion Scope 3 emission data provides a better picture of a portfolio’s impacts despite the associated uncertainty.

29 June 2017 Page 5 of 30@cdp | www.cdp.net @cdp | www.cdp.net

1.2.3 As a tool for engagement

In any diversified portfolio, it is likely that the majority of emissions will come from a minority of companies. Any efforts to engage with companies should be concentrated on this minority. This dataset can also be used to prioritise companies in a portfolio for engagement. By using Total Scope 1 + 2 + 3 emissions (weighted or absolute) to rank companies in a portfolio investors can prioritise which companies should be engaged with.

When focussing on these high emitters, first look at whether they report:

Scope 1 and both Location Based and Market Based Scope 2 (CDP Questionnaire CC8)

All applicable Scope 3 categories (CDP Questionnaire CC14.1)

Set targets for emissions reductions (CDP Questionnaire CC3.1)

Report Emissions Reduction Activities (CDP Questionnaire CC3.2)

Have a member of the board who is responsible for climate change (CDP Questionnaire CC1.1)

Have integrated Climate change in their strategy (CDP Questionnaire CC3.1)

Provide Low Carbon products/Services (CDP Questionnaire CC3.2)

Engage with Their Customers/Suppliers/Other Stakeholder on Climate Change (CDP Questionnaire CC14.4)

Companies that report these data points are more likely to reduce their emissions in line with a 2-degree compliant pathway (Niehues publication forthcoming 2017) and are more likely to be prepared for both the threats posed by climate change and the opportunities presented by a low carbon transition.

Where companies do not report these data points it is important to first check whether the company is ranked highly because of a modelled data point and, if so consider the reliability of that model. Consider why they do not report these data points and whether they are relevant for the company.

In 2017 CDP’s investor Members will have access to a tool that helps explore these data points more easily.

29 June 2017 Page 6 of 30@cdp | www.cdp.net @cdp | www.cdp.net

1.3 Scope 3 Data Reported to CDP

The CDP Climate Change Questionnaire provides companies with the opportunity to disclose data for all 15 categories plus two ‘Other’ categories for additional up- & downstream emissions.

Alongside the emissions figures broken down by Source (or category), there are additional data points that show how companies have reported their emissions.

CC14.1 in the CDP Questionnaire provides companies with the following table to complete.

Figure 1: Scope 3 response table taken from CDP Climate Change Questionnaire Guidance

29 June 2017 Page 7 of 30@cdp | www.cdp.net @cdp | www.cdp.net

1.4 Driving factors behind disclosure

There are a number of different driving factors that influence whether companies disclose data for a particular scope 3 category, summarised across the following sections. Figure 2 shows that the response rate for Business Travel emissions is on a par with Scope 1 & 2 but the response rate across the other scope 3 categories is variable. (LB & MB stand for Location Based & Market Based respectively).

Figure 2: Response rate by emissions Scope (MSCI ACWI constituents reported in 2016)

1.4.1 Ease of calculation

Business Travel is so commonly reported because the calculations for estimating the emissions are relatively straightforward1. The calculations can become significantly more complex for other Scope 3 categories and require generalising assumptions for simplification. According to the explanations provided by companies, one of the main barriers preventing the disclosure of a full GHG inventory is a lack of reliable data. Many companies are uncomfortable making assumptions when reporting data and so choose not to disclose at all. It is better to have rough Scope 3 estimates than no data; an estimate base on primary data calculated by the reporting company will invariably be better than a sector average even if the company’s methodology is very rough.

1.4.2 Relevance

The second column of CC14.1 shown in Figure 1 is headed ‘Evaluation Status’. Companies are asked to declare whether they consider a Scope 3 category relevant to their business and whether they have carried out the calculation. Many companies perform a rough calculation to gain a sense of scale before deciding whether a category is relevant to their GHG inventory. If they decide the figure is irrelevant then they may still provide the value with the evaluation status ‘Not relevant, calculated’ as can be seen in figure 3.

1 The 2016 Clean and Complete data set work showed there is a high correlation between revenue and emissions in the Airlines industry. The, the emissions intensity is approximately 1 tCO2e per thousand USD of revenue and so the crudest form of this calculation would be 𝐵𝑢𝑠𝑖𝑛𝑒𝑠𝑠 𝑇𝑟𝑎𝑣𝑒𝑙 𝐸𝑚𝑖𝑠𝑠𝑖𝑜𝑛𝑠 = 𝐴𝑖𝑟 𝑇𝑟𝑎𝑣𝑒𝑙 𝑠𝑝𝑒𝑛𝑑 ×1𝑡𝐶𝑂2𝑒/1000 𝑈𝑆𝐷 .ground transportation emissions tends to be negligible compared to ground transportation.

29 June 2017 Page 8 of 30@cdp | www.cdp.net @cdp | www.cdp.net

Figure 3: Counts of the reported relevance of each Scope 3 category

Despite the extensive guidance in the GHG Protocol’s Scope 3 Standard, similar companies rarely agree upon which categories are relevant to them. The ‘Relevance’ of a Scope 3 category is defined by the company according to 8 criteria outlined in the GHG Protocol Scope 3 Guidance. CDP has defined which categories are ‘Applicable’ to each activity according to a set of rules discussed in section 2.5.3. Gaps in reported data of any company in the sample will be filled for all applicable Scope 3 categories. Different Scope 3 categories are more/less significant for different business activities, the Scope 3 categories that are Applicable to each activity are shown in the ‘Applicable’ column in the ‘Model Appendix’ tab.

1.4.3 Demand from investors

For many companies, Scope 3 emissions far outweigh their direct impacts. Activist investors increasingly apply pressure to companies to calculate and disclose their Scope 3 emissions. There has also been a growing focus on transparency and resiliency in the supply chain over recent years, spurring increased investor interest in upstream Scope 3 reporting. Niehues (publication forthcoming 2017) concludes that companies need to report and set targets for Scope 3 in order to differentiate themselves from their peers -- only then is their stock able to gain the full benefits of a ‘green premium’.

1.4.4 Demand from other stakeholders

There is a growing body of research into why companies disclose their GHG emissions. Aside from investors, a number of other stakeholder groups have been identified to have a demonstrable influence on the level of disclosure by companies. Guenther (2015) identifies 5 key stakeholder groups; media, government, the general public, employees and customers - that have a significant influence on carbon disclosure and carbon performance.

29 June 2017 Page 9 of 30@cdp | www.cdp.net @cdp | www.cdp.net

1.5 Issues of comparability with reported Scope 3 data

There are a number of common issues that arise when comparing the GHG inventories of different companies. The most significant are outlined in this section with an accompanying table in Appendix 3.

1.5.1 Incomplete data

The main problem with Scope 3 data is that it is sparsely reported, the reasons for disclosing are complex and many companies can’t or won’t disclose scope 3 data. This patchy data means that it is difficult to compare the GHG inventories of two similar companies and any sort of analysis using reported data is likely to be flawed. In publishing estimates and reaching out to companies to highlight inconsistencies, this project also hopes to help improve the Scope 3 data reported to CDP in the future.

1.5.2 Example: Different business models

On the surface, many would assume that Apple Inc. and Samsung Electronics have similar emissions profiles because of the similarity of their products. Figures 6 & 7 illustrate that Samsung has much higher Scope 1 & 2 emissions than Apple. This is because it manufactures components whereas Apple has outsourced its manufacturing to other companies (including Samsung), so these emissions are accounted for in its Scope 3 ‘Purchased Goods and Services’.

Figure 4: Reported emissions profile of Samsung Electronics

29 June 2017 Page 10 of 30@cdp | www.cdp.net @cdp | www.cdp.net

Figure 5: Reported emissions profile of Apple Inc.

Even though these two companies are typically be considered to be peers their emissions profiles are different because Apple Inc. have made the strategic decision to outsource their manufacturing. This case study highlights the importance of using Scope 3 data in GHG footprinting, the table below compares the revenue intensity of the two companies.

Table 2: Comparison of revenue intensity between Apple Inc. and Samsung Electronics

Company Scopes 1 & 2 Only Scopes 1, 2 & 3

Apple Inc. 0.14 156

Samsung Electronics

57 274

values shown are Revenue Intensity (tCO2e / million USD)

Ignoring Scope 3 data would lead one to believe that Apple’s Climate Change impact was ~400 times smaller than Samsung’s.

1.5.3 Example: Differences in calculation methodologies

Another example of two companies that make similar products are Johnson Controls and United Technologies Corporation. They both manufacture electrical equipment and engines with 80% of their operations based in the US and the remainder in Europe.

Johnson Controls collects the emissions data from its direct suppliers and uses this information to calculate their equity share emissions from the goods and services procured. United Technologies

29 June 2017 Page 11 of 30@cdp | www.cdp.net @cdp | www.cdp.net

Corp. explains that they have over 25,000 suppliers and have therefore opted to use the Carnegie Mellon Economic Input-Output Life Cycle Assessment (EIO-LCA) model to estimate emissions.

Whereas Johnson Controls considers only the emissions from its direct suppliers, the EIO-LCA model used by United Technologies considers the cradle-to-gate emissions of all products purchased. This explains why the estimate of United Technologies’ ‘Purchased Goods and Services’ emissions is 11 times greater. Weighting these emissions by their respective revenues still results in the EIO-LCA model emissions being seven times larger.

These two options, both acceptable under the GHG Protocol Scope 3 Standard, yield substantially different results, providing a challenge to any type of comparative analysis of the ‘Purchased Goods and Services’ emissions for these companies.

Table 3: Illustration of2 Illustrating the differences in methodologies used by two similar companies

Company Name Methodology

Purchased Goods and Services

Emissions (tCO2e)

Total Revenue

(millions USD)

Revenue Intensity (tCO2e / million

USD)

Johnson Controls

Tier 1 Suppliers 13,200,000 37,179 355

United

Technologies Corporation

EIO-LCA Model 143,789,320 56,098 2563

1.5.4 Example: Different interpretations of category definitions

There are six companies that disclose emissions to CDP that sell Coca-Cola2 in some form around the world. One would expect similarity in the reporting of the product use phase. However, of those six companies:

Two companies did not evaluate the emissions from the ‘Use of Sold Products’ without providing an explanation;

One company explained that they considered these emissions ‘immaterial’ compared to the other sources of their emissions;

One company understood the ‘Use of Sold Products’ to mean the CO2 released by users in opening and drinking their carbonated drinks; and

Two companies understood ‘Use of Sold Products’ to refer to the emissions associated with the refrigeration of their products before they are drunk.

This example highlights the differences in interpretation of the meaning of these Scope 3 categories.

2 Coca-Cola Femsa Sab-Ser l; Coca-Cola HBC AG; Coca-Cola West Co., Ltd.; Coca-Cola Içecek A.S.; Coca-Cola European Partners; The Coca-Cola Company.

Some produce the syrup whilst others bottle/can the finished product, in both cases the ‘Use’ phase of the life cycle is still the same. https://en.wikipedia.org/wiki/The_Coca-Cola_Company

29 June 2017 Page 12 of 30@cdp | www.cdp.net @cdp | www.cdp.net

1.5.5 Example: Different consolidation approaches

The majority of the emissions from an electricity generator like SSE is in their Scope 1, as shown in figure 6. SSE reports on an Operational Control boundary and the only material Scope 3 category is ‘Fuel-and-Energy-Related-Activities’ from the upstream production of their fuel stock.

Figure 6: Reported emissions profile of SSE

29 June 2017 Page 13 of 30@cdp | www.cdp.net @cdp | www.cdp.net

Korea Electric Power Corp (KEPCO), another power generator, has a very different emissions profile. It reports on an Operational Control boundary which excludes emissions from its six power generating subsidiaries3 and affiliated companies. A few of these operating companies disclose their emissions independently to CDP. However, within KEPCO’s GHG inventory the related emissions are reported under the Scope 3 ‘Investments’ category, not under Scope 1.

Both companies are reporting their emissions profile correctly according to the guidelines, but the different consolidation approaches result in substantially different emissions profiles.

Figure 7: Reported Emissions profile of KEPCO

1.6 Cleaning Scope 3 data

Despite the nuances of Scope 3 accounting discussed in the previous sections, CDP has reviewed the reported methodologies employed by companies and flagged values that are either incomplete or at odds with other companies in the sector. Data points may have been flagged if:

The company has indicated it omitted key parts of its business, activities or products;

The calculation methodology appears suspect;

CDP’s analysts suspect that the data has been entered incorrectly;

The value is an outlier and the methodology used to derive the emissions is not clear; or

The emissions figure has been entered in the wrong category.

Only 280 data points have been flagged out of over 10,000 reported Scope 3 data points in 2016.

3 https://en.wikipedia.org/wiki/Korea_Electric_Power_Corporation

29 June 2017 Page 14 of 30@cdp | www.cdp.net @cdp | www.cdp.net

2 Statistical Models and Application to Scope 3

2.1 Statistical Models

Technical Annex III: Statistical Framework provides an overview of the statistical framework used for modelling the Scope 1 & 2 emissions for the CDP 2016 Full GHG Emissions Dataset. The Scope 3 data reported to CDP is treated in a similar manner, applying the same multi-variable Gamma family Generalised Linear Model (Gamma GLM) using revenue and activity information.

The Scope 3 data reported to CDP is considerably less consistent and the samples for each category are much smaller than with Scopes 1 & 2. Despite this, the data is still positive and heteroskedastic much in the same way as the Scope 1 data. For these reasons, the Gamma GLM model is still appropriate.

In the simplest terms the model coefficients (also called predictors or estimators) can be thought of average revenue intensities4 based on the data reported to CDP. The more details of the modelling approach can be read in Technical Annex III and will not be repeated here.

2.2 Model assumptions

In the same way that Scope 1, SHEC, and fuel each had their own independent multi-variable regression model, so too does each of the 15 Scope 3 categories. There are some new assumptions that are made in the Scope 3 modelling and a couple of the original assumptions are revised here.

2.2.1 Activity revenue as the independent variable

The revenue earned by activity segment is used as the basis of the regression model. This approach assumes that revenue is directly proportional to production and therefore proportional to emissions. For more detail on these basic assumptions, please review Technical Annex III: Statistical Framework.

2.2.2 Relationship to FTE and CapEx models

In the previous iteration of this modelling work the emissions associated with ‘Employee Commuting’ were estimated using the number of full time-equivalent employees (FTE) and the emissions associated with ‘Capital Goods’ were estimated using capital expenditure (CapEx). These models were built using a single sector classification of each company, ignoring the fact that most companies have more than one activity. There is revenue breakdown by activity data available through Bloomberg whereas there are no similar breakdowns for number of FTE or CapEx. The benefits from using the revenue breakdowns to estimate these emissions are deemed to outweigh the benefits gained from the marginally stronger relationships between CapEx and Scope 3 Capital Goods or FTE and Scope 3 Employee Commuting. The Following subsections explore the validity of this assumption.

2.2.1.1 Constant earnings per employee

The earnings per employee is used by analysts to compare personnel productivity. This ratio varies drastically across industries but here it is assumed to be reasonably consistent for companies engaged in similar activities.

There is an assumption made in the statistical modelling framework that companies undertaking the same activity do so in a similar manner using similar processes. This can be extended to apply

4 Revenue Intensity (tCO2e / million USD)

29 June 2017 Page 15 of 30@cdp | www.cdp.net @cdp | www.cdp.net

to the number of employees as well; if this assumption is true then companies should employ a similar number of people per unit of production/revenue.

This means that the earnings per employee could be assumed to be reasonably consistent within any given activity group. This effect can be modelled using the same multivariable Gamma GLM5 applied for Scope 1 estimation by simply substituting FTE for Scope 1 emissions. The results of this FTE estimation model are reasonably similar to the revenue/activity models before they were cleaned. The R2 of the Scope 1 Gamma GLM was 0.74 before cleaning (0.85 after), while the R2 for the FTE model is 0.77 before any cleaning.

2.2.1.2 Constant capital expenditure ratio

Analysts use this ratio to indicate potential future growth; the ratio is much more cyclical in nature. The same assumptions could be applied to CapEx: Companies engaged in any given activity could be assumed to have made similar capital expenditures per unit of production assuming they use the same equipment for production.

A regression model to calculate the activity average capital expenditure ratio can be fitted using a Gamma GLM in exactly the same way as before. This time the R2 is 0.71, which again is similar to the revenue/activity based emissions models.

Given that these fundamental ratios can be shown to be at least reasonably constant within activity groups, then the revenue/activity approach should still provide a reasonably good model for these emissions.

2.2.2 Companies use similar calculation methodologies

This assumption was made about the Scope 1 & 2 data reported to CDP in the Statistical Framework. It is not reasonable to use this assumption in Scope 3 modelling due to the different interpretations and estimation approaches of the Scope 3 categories introduced previously.

2.2.2.1 Estimates reflect a mixture of calculation methodologies

In any given sector, the model estimates will reflect a mixture of the calculation methodologies of the reporting companies. In many cases these differing methodologies and assumptions will yield roughly similar results but in some cases the difference between approaches can be significant.

Taking the two approaches used for calculating Scope 3 ‘Purchased Goods Services’ emissions in the manufacturing industry (explained in Table 3) as a basic example, some companies use an input-output model and others use primary data gathered from first-tier suppliers. If 60% of the companies used an input-output model and 40% used the primary data approach, then the emissions intensity for that category would be 1,680 tCO2e per million USD (0.6 ∗ 2,563 + 0.4 ∗355 = 1,680 𝑡𝐶𝑂2𝑒/𝑚𝑖𝑙𝑙𝑖𝑜𝑛𝑈𝑆𝐷). This is markedly different from the 355 tCO2e per million USD from the first-tier suppliers and the 2,563 tCO2e per million USD from the input-output model.

The regression models calculate an average revenue intensity of the companies’ data, which represents a mixture of these methodologies. This means that the estimates should be thought as a weighted average of the various methodologies employed by companies.

This also means that CDP does not choose an approach for the predictions. If the reported Scope 3 ‘Purchased Goods and Services’ data were split into two sets and one model was constructed using the reported data calculated through the EIO-LCA and another using the reported data that only included first tier suppliers, then a choice would have to be made as to which model to employ for predicting for non-disclosers.

5 Keener mathematicians may point out that the number of FTE is a not a continuous variable but a count, so a Poisson distribution would be more appropriate. However, given that the number of FTE are generally large and that one part-time employee could be counted as ½ an FTE then the data could be considered continuous and the Gamma family distribution can still be used.

29 June 2017 Page 16 of 30@cdp | www.cdp.net @cdp | www.cdp.net

This is the most significant source of uncertainty in these Scope 3 models and can be rectified by more harmonized reporting between peer companies.

2.3 CDP Activity Classification System

Whereas the Scope 1 & 2 modelling work relied on Company revenue split up by the most granular level of the CDP Classification Activity System (CDP-ACS), the Scope 3 modelling work uses all three levels.

2.3.1 Classification hierarchy

These revenue-based statistical models are dependent upon how companies are grouped together. The CDP-ACS has been developed to provide a framework for quantifying a company’s environmental impacts connected to its activities; impacts across the Forests, Water and Climate Change programs have been considered. The classification hierarchy has three levels:

Activity: A company’s environmental impacts are the result of their activities and so the most granular level of the CDP hierarchy is the Activity. Companies may have many different activities. There are 180 Activities in the CDP-ACS.

Sector: Typically a sector is defined as a group of companies with common activities and so, the second level of the CDP hierarchy is the Sector. There are 47 Sectors.

Industry: These Sectors make up an Industry which is the highest or the least granular level of the hierarchy. There are 13 Industries.

Each grouping has been created to try to ensure that the environmental impacts across the three programs are as consistent as possible.

2.3.2 Climate change hybrid classification system

The CDP-ACS has been adapted to focus on climate change related impacts. Some of the Sector and Activity groupings have been combined where the main distinction between them relates to their Forests or Water Impacts.

For example, hydro power has significantly different impacts on both local water systems and forests so a distinction made between hydro power and other renewable electricity sources. In the climate change hybrid, this distinction is ignored because all types of renewable energy have similarly small emissions intensities. In the climate change hybrid activity classification system, there are 100 Activities, 38 Sectors and 13 Industries. The climate change hybrid has been developed primarily for Scopes 1 & 2 adjustments were made for Scope 3 for overall model improvement.

2.4 Three levels of model granularity

Each of the statistical models used for the Scope 1, fuel & Scope 2/SHEC estimations used the revenue broken down into the activity groups, resulting in a multi-variable regression model with 100 different independent variables and as such, 100 model coefficients representing the average revenue intensity for each activity. The model sample used to fit these coefficients needs to contain enough data points for each activity to be able to find a good fit. This is not simply an issue of how many data points are in the sample but also the consistency of the reported data.

29 June 2017 Page 17 of 30@cdp | www.cdp.net @cdp | www.cdp.net

There is enough Scope 1 & 2 GHG data available to be able to build regression models using the activity revenue. For many of the Scope 3 categories, the disclosure rate is much lower (see Figure 1). This means that for some activities where the data is either too sparsely reported or too inconsistent for the model to find a good fit, a different level of aggregation is necessitated. Using the revenue broken down by sector as opposed to activity results in 38 variables instead of 100. This means that the model is more likely to be able to find a good fit but that the coefficients of the model will be the sector average revenue intensities. This aggregation results in the estimates being less specific but more robust.

If there is insufficient data in the sample to fit a model using the revenue breakdown by sector, then the industry level is selected. Grouping together the variables into sectors and industries has the effect of making the model estimates more reliable due to enhanced stability, but the model is more generic and therefore less useful for predicting a single company’s emissions.

For each Scope 3 category there are three models generated using the Activity, Sector and Industry groups to break down each company’s revenue.

Level Number Granularity Robustness

CDP Activity 100 Most Worst

CDP Industry 38 - -

CDP Sector 13 Least Best

It should be obvious that the more granular Activity level models should yield better estimates than the more general Sector or Industry level models. To estimate the emissions of an aluminium refiner, it would be best to use the activity-specific estimator based on data disclosed by other aluminium refiners as opposed to using a more generic average intensity based on data disclosed by companies from across the metals processing sector.

The figure below shows a portion of the CDP classification tree which can be ‘pruned’ to improve the robustness of the estimates if there is insufficient data.

Based on the above diagram, there may not be enough companies involved in textiles or clothing manufacture reporting their Scope 3 ‘Capital Goods’ emissions to CDP to be able to calculate a reliable average intensity for those activities, and so it may be necessary to group them into the clothing & textile manufacturing sector to get a reliable sector average intensity.

29 June 2017 Page 18 of 30@cdp | www.cdp.net @cdp | www.cdp.net

2.5 Model selection

2.5.1 AIC single model selection

Equipped with these three models, testing was carried out using the Akaike Information Criterion (AIC)6 to choose which level would be the most appropriate for each Scope 3 category. For well reported categories like ‘Business Travel’ the activity model could be used but for poorly reported categories like ‘Franchises’, the more general industry level would be more appropriate. However, this one-size-fits-all approach for each Scope 3 category ignores the fact that there are some categories that are more relevant to different sectors than others.

2.5.2 Stepwise activity based model selection

2.5.2.1 Rationale

To overcome this problem, a bespoke revenue breakdown matrix is required for each Scope 3 category that combines Activity, Sector & Industry revenue columns. There are plenty of feature selection techniques available for focussing the model by reducing the importance of less useful variables as this is a common problem in data science. Generally, these techniques are designed for selecting between numbers of independent variables. In this project, there are approximately 150 different revenue variables (99 Activities, 38 sectors, 13 Industries) which are not independent due to the hierarchical relationship between them. This means that some columns are linear combinations of other columns (i.e. each sector column is equal to the sum of the revenue from its

related activity columns (𝑟𝑒𝑣𝑒𝑛𝑢𝑒𝑆𝑒𝑐𝑡𝑜𝑟 = ∑ 𝑟𝑒𝑣𝑒𝑛𝑢𝑒𝐴𝑐𝑡𝑖𝑣𝑖𝑡𝑦𝑖𝑁𝑖 ). This renders many of the more

common machine learning techniques unusable because they can’t incorporate this information7.

Reducing the number of activity dimensions (using techniques like Principal Component Analysis (PCA)) tends to result in groupings not dissimilar to the Sector groupings defined in the CDP-ACS (this should not be surprising because the Sector groupings are designed to group similar companies together). This means it makes little sense to use PCA or any other dimensionality reduction techniques to groups of variables because these sector/industry groupings already exist. In any case the revenue breakdown data is largely orthogonal, the activity matrix is sparse since companies tend to earn revenue in a few of the many possible activities and in most cases, these are in the same Sector/Industry. Using a regularisation method like LASSO on a matrix of all available variables would not guarantee that the final revenue breakdown matrix is consistent in rank (in other words, the sum of the breakdowns may not equal the total revenue of the company).

Whilst widely considered to be inferior, using a stepwise selection technique remains the only viable means of automatically selecting variables. The revenue breakdown matrices each must contain exactly one of every Industry/Sector/Activity combination to be able to predict for the full universe of companies. The bestglm package in R has been developed for performing stepwise criterion based model selection and provides a more flexible set of options for establishing the type of search and the criterion used than the ‘step’ function from the ‘stats’ packages. This uses a ‘branch and bound’ algorithm that can find a solution efficiently provided the number of variables is less than 25, whereas we start with ~100 different revenue variables. These issues have meant that a manual process has had to be developed. This process draws on the most commonly used stepwise variable selection approaches used whilst explicitly accounting for the relational constraints between variables.

6 The AIC is a measure of goodness of fit and model complexity that is used as a relative measure to compare between two similar models. https://en.wikipedia.org/wiki/Akaike_information_criterion

7 Even if they were appropriate, the R package for performing LASSO for GLMs called ‘glmnet’ doesdo not have solutions for Gamma family GLMs.

29 June 2017 Page 19 of 30@cdp | www.cdp.net @cdp | www.cdp.net

2.5.2.2 Two step pruning algorithm

An iterative process is used to generate a subset of possible models based on different groupings of variables. The AIC is used at each iteration to compare the test model with the current best-known model (the one with the lowest AIC). 8The algorithm starts with the full classification tree and ‘prunes’ branches by grouping the Activity/Sector variables into the respective Sector/Industry. This method can be thought of as a Backwards Elimination approach.

There are two loops that cycle through different combinations of models. The first cycles through each Sector, grouping the Activities together and recalculating the model’s AIC while comparing it with the best-known model. If the test model yields a lower AIC than the current best-known model, then the test model becomes the best-known model. The second loop cycles through each Industry, regrouping columns and comparing the test model with the current best-known model.

The algorithm starts by comparing the Industries/Sectors which are least Applicable (see subsequent section) to the Scope 3 category under consideration, grouping those first before examining the effects of grouping the more Applicable sectors. This considerably reduces the number of combinations to test by excluding less relevant sectors ex-ante.

A third stage prunes away any undesirable model coefficients left after this process. These undesirable coefficients are either negative or implausibly high; when they do occur, they are absorbed into their Sector/Industry.

2.5.3 Applicable Scope 3 categories

There are many combinations of activity and category where it is not clear what the source of emissions would mean. For example, there would be no emissions from the ‘Use of Sold Products’ for clothing manufacturers. Similarly, a health care provider would have no ‘Processing of Sold Products’ emissions. In order to capture this logic, CDP has defined an applicability categorisation.

For each activity/category combination, CDP has decided whether it is either Applicable / Not

Applicable. Estimates will not be shown for activity/category combinations that were deemed Not

Applicable. The applicability of each category for each Scope is hard coded for each activity. These

definitions are largely based on work carried out by Niehues (publication forthcoming 2017) in

collaboration with CDP. As part of this work several rules have been formulated, based on the data

reported, to define whether a category is applicable. CDP has made adjustments to these

definitions based on the reliability of the reported data.

NB: Under this terminology ‘Applicable’ is distinct from ‘Relevant’. An emissions source for a

company may not be deemed ‘relevant’ by the reporting company because it is negligible compared

its other emissions, however the emissions source may be real and calculable so it is still

‘applicable’. For a coal extractor, the emissions from Scope 3 ‘Employee Commuting’ may not be

deemed ‘relevant’ by the company reporting because they are many times smaller than the

emissions from ‘Use of Sold Products’, but these emissions are still ‘applicable’ because the

company still has employees who commute to work. The emissions from ‘Upstream Transportation

and Distribution’ would generally neither be considered applicable nor relevant for a coal extractor.

The Model Appendix worksheet in the dataset is a table of the Applicable categories for each

Activity as well as the hierarchical classification level used to estimate for that activity/category

combination. These estimates are calculated on an activity basis, meaning companies with more

than one activity may have some activities that are/aren’t applicable for the same Scope 3 category,

some that are modelled at an industry level and some that are not, etc. This makes it cumbersome

to present this information alongside the estimates, but for the purpose of transparency it is included

in a separate sheet for reference purposes.

8 The AIC is preferred over the Bayesian Information Criterion (BIC) because of the BIC’s tendency to penalise models with larger variable sets.

29 June 2017 Page 20 of 30@cdp | www.cdp.net @cdp | www.cdp.net

3 Bottom-up ‘Use of Sold Product’ modelling

Bottom-up modelling is used to estimate Scope 3 ‘Use of Sold Products’ emissions in three sectors: Oil & Gas, Coal, Automotive Manufacturing.

3.1 Data sources

The bottom-up approach requires activity data in the form of physical or energetic output, e.g. barrels of oil and number of vehicles, along with emission and energy related factors that combine with the activity data in the estimation of company emissions.

For extraction in the Oil and Gas sector, oil density and sulphur content is sourced from Globaldata Upstream Analytics (GD 2016) and converted to carbon content using an equation formulated from an IPCC survey (IPCC 2001). Global average emission factors applied to natural gas, NGL, Bitumen, and refinery products are taken from API (2009), EPA (2014), and IPCC (2006a). Oxidisation levels are calculated from combustion oxidation factors, non-energy use fractions, and non-combustion sequestration rates, which are sourced, respectively, from IPCC (2006a), Eurostat (2016a, 2016b) and IEA (2016), and UNFCCC (2015). In the absence of information on refinery product slate, typical Scope 3 emission factors of refinery facilities (hydrocracking, topping, coking, etc.) are derived from the PRELIM model, which was built for the Oil Climate Index project of the Carnegie Endowment for International Peace (CEF 2014).

For the Coal sector, global emission factors and oxidisation factors for coal grades lignite, sub-bituminous, bituminous, coking, and anthracite, are taken from IPCC (2006a). Where data on coal calorific value is available in company filings, this is also used.

For the Automotive Manufacturing sector, data on production and sales volumes is sourced from company filings and responses to the Automotive Manufacturing module of the CDP (2016) Climate Change Questionnaire. Data on vehicle carbon intensity is taken from datasets published under national and regional emissions standards regulation (NHTSA 2015 and EU 2016) and the CDP Questionnaire. Information on typical vehicle carbon intensity of different vehicle classifications are sourced from ICCT (2014) and used where data on company-level product emission intensity is unavailable. Data on typical vehicle mileage or kilometrage is informed from responses on Scope 3 estimation methodology in the CDP (2016) Climate Change Questionnaire. This is combined with ICCT (2014) data to inform for the mileage between different vehicle classifications.

3.2 Methodology

The generalized methodological approach for Scope 3 is unchanged from Scope 1 and 2 and is detailed in the IPCC (2006b) Guidelines for National Greenhouse Gas Inventories. It is described here by Equation (1):

Emissions = Activity data • Emission factor (1)

3.2.1 Oil and Gas

Scope 3 ‘Use of Sold Products’ typically accounts for around 90% of total inventory emissions. The calculation of Scope 3 category 11 ‘Use of Sold Products’ company emissions is described by Equation (2), where ES3.11 is Scope 3 category 11 emissions, EF is emission factor, EO is the effective oxidation rate, P is net production, subscripts o, n, g, and r denote oil, NGL, natural gas, and refinery product respectively.

ES3.11 = Po • EFo • EOo + Pn • EFn • EOn + Pg • EFg • EOg + ∑ Pr • EFr • EOr

N

r

(2)

29 June 2017 Page 21 of 30@cdp | www.cdp.net @cdp | www.cdp.net

To avoid double counting the Scope 3 emissions of integrated companies, the larger of upstream and downstream oil-related emission is selected to represent total oil-related Scope 3 emission. The preferable approach is to account for the company’s total production of upstream and downstream oil products while deducting for self-produced refinery feedstock. However, data on the proportion of a company’s intake deriving from its own upstream production is unavailable9.

Net production is production that is net of losses, stock-changes, self-consumption, and royalties or entitlements to third parties. These deductions are independent of the company’s organizational boundary. Net production is chosen to represent ‘Use of Sold Products’ because sales data reported by companies can often include flows between entities inside the organizational boundary, which would lead to double counting.

Oil and gas products are not completely oxidised over their lifetime. Imperfect combustion is accounted for by the product’s oxidation factor (OF), which is typically between 0.99 and 1. The OF is applied to the fraction of a product amount that is used for energy purposes. Within the non-energy use (NEU) fraction of a product amount, a proportion of carbon is expected to be sequestered. This proportion is accounted for by the product’s storage factor (SF). Together these factors amount to a product’s effective oxidation rate (EO). The EO is defined here as the ultimate proportion of a product that is emitted over its lifetime. Average global EO is about 0.9 for oil and 0.99 for gas. Equation (5) describes the calculation of EO using the factors described above, where p is the hydrocarbon product.

EOp = OFp • (1 - NEUp) + NEUp • (1 - SFp) (3)

Oil incorporates crude oil, condensate, bitumen, and any other designation applied to the full range of oil densities. Due to the wide range of oil densities, the carbon contained in any volume varies between companies. Chapter 2 of the IPCC (2001) publication ‘Good Practice Guidance and Uncertainty Management in National Greenhouse Gas Inventories’ contains a method for estimating oil carbon content from API10 gravity and sulfur content. The method is expressed in Equations (4) and (5), where C is carbon content (%), SC is sulfur content (%), SG is specific gravity, API is the API gravity at 60 degrees Fahrenheit. The sector’s other upstream products, natural gas and natural gas liquids, are not as variable in carbon content and are therefore represented by global average emission factors.

where,

C = 76.99 + 10.19 • SG - 0.76 • SC

SG = 141.5

(API + 131.5)

(4)

(5)

Uncertainty in ‘Use of Sold Product’ emission estimations can derive from variability in oil API, which equates to a variability in oil carbon content, although this is expected to be small. The main area of uncertainty in this estimation comes from the oxidation rate, for which a global figure is used. Oil oxidation rate is dependent on the refinery product slate and application for which the product is destined.

9 For more detail on Scopescope 3, including how to define it for integrated extraction and refining companies, see CDP’s ‘Guidance methodology for estimation of Scope 3 category 11 emissions for oil and gas companies’‘Guidance methodology for estimation of scope 3 category 11 emissions for oil and gas companies’.

10 API gravity is an arbitrary scale for oil density devised by the American Petroleum Institute.

29 June 2017 Page 22 of 30@cdp | www.cdp.net @cdp | www.cdp.net

3.2.2 Coal

‘Use of Sold Products’ emissions in the Coal sector typically account for 90-95% of total company emissions. The calculation of ‘Use of Sold Products’ company emissions is described by Equations (8) and (9), where ES3.11 is Scope 3 category 11 emissions, EF is emission factor, EFce is emission factor relating to energy, HV is heating value, EO is the effective oxidation rate, P is net production, subscript p is coal product type or grade.

or,

ES3.11 = ∑ Pp • EFp • EOp

N

p

ES3.11 = ∑ Pp • HVp • EFpce

• EOp

N

p

(6)

(7)

P and EO are defined as in section 3.2.1, with average EO equalling approximately 0.98 for coal. Net production or sales differ by +/-2% in most cases and may therefore be used interchangeably, except for Chinese companies where larger differences are observed. However, due to the greater availability of production data over sales data, net production is prioritized over sales to help ensure consistency between companies. Sales data is used only when production data is unavailable.

Coal grades include anthracite, bituminous, sub-bituminous, and lignite. Product descriptions are also used by companies and may include coking, metallurgical, thermal, and so on. To ensure that emission factors best represent the company’s product mix, the analysis attempts to disaggregate by grade or product type as far as data is available. Further to this, coal heating value is sought from company filings. Energy content has a more linear relationship with carbon content, thereby increasing accuracy.

3.2.3 Automotive Manufacturing

The Automotive Manufacturing sector is defined for this analysis as including companies that manufacture and sell: passenger cars; light, medium, and heavy trucks; busses; two wheelers, three wheelers, and quads; construction vehicles, and engines.

The calculation ‘Use of Sold Products’ company emissions is described by Equations (19) and (20), where ES3.11 is Scope 3 category 11 emissions, EF is emission factor, S is sales volume, D is average annual distance travelled, L is vehicle lifetime in years, and subscript C denotes vehicle classification.

ES3.11 = ∑ SC • DC • LC • EFC

N

C

(19)

The classifications distinguished in the analysis are: cars, sports cars, utility vehicles, light trucks, medium trucks, heavy trucks, busses, two wheelers, three wheelers and quads, tractors, construction vehicles, and engines. Companies reporting to CDP on the Scope 3 emissions of cars typically assume an average annual kilometrage of 15,000km and a lifetime of 10 years. These figures are applied and adjusted for other vehicle types using data from ICCT (2014).

29 June 2017 Page 23 of 30@cdp | www.cdp.net @cdp | www.cdp.net

4 Results

The final dataset combines the self-reported values with estimates from the methods described in this document.

4.1 General Results

Where a data point has been flagged, CDP has made provided a comment along with an estimate for that value.Figure 8 shows the emissions profile of the MSCI ACWI combining modelled and reported data. The ‘Use of Sold Products’ emissions of 25GtCO2e would amount to around half of global emissions if this figure didn’t include a significant double counting. The emissions from the ‘Use of Sold Products’ for an automobile manufacturer overlap with the ‘Use of Sold Products’ emissions for oil extractors and refiners because they both count the emissions from the petrol burned in their vehicles.

Figure 7: MSCI ACWI emissions breakdown by Scope

The emissions from ‘Purchased Goods & Services’ comes in as the second largest Scope 3 category, which also includes a degree of double counting when adding the emissions of companies in the same product value chain i.e. the emissions from farming will be included in the ‘Purchased Goods and Services’ emissions of a supermarket and any food processors. The companies that report for their first-tier suppliers have the effect of mitigating this by including only directly attributable emissions.

29 June 2017 Page 24 of 30@cdp | www.cdp.net @cdp | www.cdp.net

Figure 9: Emissions breakdown by Scope of the MSCI ACWI (reported & estimated data)

The manufacturing sector includes companies producing automobiles, electrical equipment, and transportation equipment. It is therefore unsurprising that the ‘Use of Sold products’ is large. There is obvious overlap here between the ‘Use of Sold Products’ in the fossil fuel industry and the ‘Use of Sold Products’ in the manufacturing industry. Figure 10 shows that within this Industry a good deal of the emissions are reported directly by companies. Furthermore, the estimates for the automobile manufacturing sector are bottom-up model estimates, helping to further reduce uncertainty in the figures. The Scope 3 emissions of the chemicals sector and the electrical/electronic equipment sector are generally less well reported, yet the total Scope 3 emissions is almost on par with the automobile manufacturing sector when the estimated data is included.

29 June 2017 Page 25 of 30@cdp | www.cdp.net @cdp | www.cdp.net

Figure 10: Emission breakdown by Sector for the manufacturing Industry

4.2 Use of Bottom Up Estimates for ‘Use of Sold Products’

The ‘Use of Sold Products’ emissions from these sectors numerically outweigh the other sources of emissions and so improving the comparability and accuracy of the estimates has gone a long way to improving the overall accuracy of any GHG footprint based on this data.

The application of a consistent methodology across all companies in a sector using verified publicly available data helps overcome many of the issues described in the first section of this report. These estimates use similar methodologies to many of the reporting companies and in some cases these estimates go into more detail. Where as an individual company’s reported data may or may not be more accurate, these estimates are more comparable across companies than reported data.

4.3 Use of mixed level Gamma GLM

The Gamma family GLM does not accommodate for companies that have zero emissions from any of these scope 3 categories. The modelling code effectively treats the zeroes as NA values. This is problematic for categories where a company may legitimately have zero emissions from a particular category. For Scope 1 & 2 each company was assumed to have at least some Scope 1 and location-based Scope 2 emissions, even if they are small. With Scope 3 categories, companies may not generate emissions, e.g. if a company has no franchises, no investments or no leased assets then it will have zero emissions for these categories. These models only consider non-zero observations and so any estimates will be based off the non-zero values only.

After the model coefficients have been calculated for each Scope 3 category, the estimates are

only provided for Applicable categories, meaning that this issue is dealt with ex-post. The

Applicability of each category has been hard coded by CDP's analysts for each Activity. There are

very few companies that have reported a complete GHG inventory according to these criteria: Only

76 of the companies in the data set have reported all the Scope 3 categories applicable to their

29 June 2017 Page 26 of 30@cdp | www.cdp.net @cdp | www.cdp.net

business activities. This mean that the vast majority of reporting companies will still have estimated

data for some Scope 3 categories. The data has been presented in such a way that users of the

data can see the evaluation status and explanation provided by the company, when disclosed, next

to the estimates provided by CDP.

4.4 Success of stepwise model selection approach

The current approach used for selecting variables is relatively crude and lacks the elegance of some of the popular machine learning solutions. It applies a brute force approach to the problem, taking advantage of the advances in modern processing power to churn through calculations rather than using data specific knowledge of each sector. This top-down approach does ignore a lot of the other information in CC14.1 and it should be possible to take a more bottom-up approach by carrying out a comprehensive assessment of each sector and each Scope 3 category and basing the decision on what level of aggregation to use. With approximately 100 activities and 15 Scope 3 categories to model this decision needs to be made roughly 1,500 times.

The approach employed does not test every single possible combination of possible models (there are roughly 1011 different combinations), which means that the optimal model could have been missed. Another commonly cited issue with these types of approaches is that that the model significance can be overstated (Judd & McClelland 1989) and so estimates can seem more reliable than they actually are.

Despite these shortcomings, the approach taken does provide a way to automatically select the variables incorporating the constraints on which levels can be used. The table below summarises the different levels chosen to model each Scope 3 category. The proportion of Activity variables used for each category can be taken to mean that the data is well reported for that category. This mirrors the information in Figures 2 & 3, where Business Travel and Employee Commuting result in more of the granular Activity level variables than other categories. The data in this table also indirectly accounts for the consistency of reporting for these categories, which may explain why the ‘Purchased Goods and Services’ relies so heavily on Industry level data even though it is one of the more widely reported and relevant categories. This is likely to be due to the different calculation approaches that can yield very different results (see Table 1table 2).

Emissions Source Activity Sector Industry Total

Business travel 62 29 9 100

Capital goods 23 38 27 88

Downstream leased assets 1 0 0 1

Downstream transportation and distribution 44 24 5 73

Employee commuting 53 29 17 99

End of life treatment of sold products 18 5 13 36

Fuel-and-energy-related activities 79 16 0 95

Investments 2 0 0 2

Processing of sold products 12 4 0 16

Purchased goods and services 20 23 24 67

Upstream transportation and distribution 61 18 13 92

Use of sold products 15 4 0 19

Waste generated in operations 60 25 4 89

Total 449 215 112 776

29 June 2017 Page 27 of 30@cdp | www.cdp.net @cdp | www.cdp.net

Appendix 1 Statistical Modelling References

GHG Protocol 2011, Corporate Value Chain (Scope 3) Accounting and Reporting Standard

KeplerChevreux 2015, Carbon Compass. IIGCC 2015, Carbon Compass.

Economic Input Output Life Cycle Assessment, Carnegie Mellon University

Business & Society, March 2015, Stakeholder Relevance for Reporting - Explanatory Factors of Carbon Disclosure, Edeltraud Guenther; Thomas W. Günther; Frank Schiemann; Gabriel Weber

CDP 2016, Technical Annex III: Statistical Framework

Investopedia, Operating Performance Ratios: Sales/Revenue Per Employee

Investopedia, Cash Flow to Capital Expenditures - CF to CAPEX

Brian Ripley Lecture Notes, University of Oxford Statistics Department

A survey on feature selection methods, Girish Chandrashekar, Ferat Sahin, Electrical and Microelectronic Engineering, Rochester Institute of Technology

Advanced Methods in Biostatistics 1 (Chapter 10, lecture notes), Ingo Ruczinski. Johns Hopkins University, Bloomberg School of Public Health

CRAN R Package Vignettes ‘bestglm’

Why companies disclose GHG emissions, Nils A. Niehues, Andreas Dutzi, Hugh Sawbridge Critical Management Studies (Forthcoming 2017)

Judd, C. M., & McClelland, G. H. (1989). Data analysis: A model comparison approach. New York: Harcourt Brace Jovanovich.

29 June 2017 Page 28 of 30@cdp | www.cdp.net @cdp | www.cdp.net

Appendix 2 Bottom-up Modelling References

API 2009, ‘Compendium of Greenhouse Gas Emissions Methodologies for the Oil and Natural Gas Industry’, American Petroleum Institute, Washington DC, 2009.

CDP 2016, Climate Change Questionnaire 2016 responses. CDP, London.

CEF 2014, PRELIM: The Petroleum Refinery Life Cycle Inventory Model. Life Cycle Assessment of Oil Sands Technologies Project, University of Calgary, and the Carnegie Endowment for International Peace.

EPA 2014, Emission Factors for Greenhouse Gas Inventories, US Environmental Protection Agency, United States.

EU 2016, Monitoring of CO2 emissions from passenger cars - Data 2014 - Provisional data/ Monitoring of CO2 emissions from passenger cars – Regulation 443/2009

Eurostat 2016a, "Supply, transformation and consumption of gas - annual data," in (nrg_103a).

Eurostat 2016b, "Supply, transformation and consumption of oil - annual data," in (nrg_102a).

GD 2016, GlobalData Oil & Gas: Upstream Analytics, GlobalData. [Accessed: 2016].

NHTSA 2015, Manufacturer Fuel Economy Performance Report. CAFE Public Information Center, National Highway Traffic Safety Administration, United States.

IEA 2016, ‘World Energy Statistics dataset,’ OECD/IEA, Paris.

ICCT 2014, Global Transportation Roadmap Model: ICCT Roadmap model v1.0. The International Council on Clean Transportation.

IPCC 2001, ‘Chapter 2 Energy’, ‘Good Practice Guidance and Uncertainty Management in National Greenhouse Gas Inventories’, IGES, Japan, 2001.

IPCC 2006a, ‘2006 IPCC Guidelines for National Greenhouse Gas Inventories. Volume 2: Energy’, IGES, Japan.

IPCC 2006b, ‘2006 IPCC Guidelines for National Greenhouse Gas Inventories. Volume 1: General Guidance and Reporting’, IGES, Japan.

UNFCCC 2016, ‘Table 1.A(b) Sectoral Background Data for Energy: CO2 from Fuel Combustion Activities - Reference Approach’ and ‘Table 1.A(d) Sectoral Background Data for Energy: Feedstocks and Non-Energy Use of Fuels’, National Inventory Submissions 2016: Common Reporting Format.

29 June 2017 Page 29 of 30@cdp | www.cdp.net @cdp | www.cdp.net

Appendix 3 Common Comparability Issues Scope 3 Category Common Issues with Data reported to CDP

Business travel Best responded category, calculations are very sensitive to different emission factors and assumptions.

Capital goods Companies’ capital investments are not necessarily consistent year on year because companies do not make consistent capital investments. Many companies choose to account for these emissions by depreciation but many do not.

Downstream transportation and distribution

Calculations are very sensitive to the assumptions about mode of transport and so similar calculation methodologies may result in different values.

Downstream leased assets

The decision to lease or purchase assets depends on the company’s business strategy more than on size or activity and so any data reported in this category is not well explained by the variables used.

Employee commuting

Different assumptions about employee behaviour and emission factors from public transport can lead to different results. Variables used are often site specific.

End of life treatment of sold products

Calculations depend on assumptions about behaviour of users or clients which can affect the calculations.

Franchises Depends on the company’s reporting boundary and business model.

Fuel-and-energy-related activities

This Scope 3 category often confusion amongst companies and the calculation methodologies vary considerably.

Investments Dependent on Scope 1 & 2 reporting boundary, if a company excludes Scope 1 & 2 emissions from assets that it does not operate because it is reporting on an Operational Control boundary then the emissions from these assets should be included in their Scope 3 ‘Investments’. The emissions from these assets would be included in Scope 1& 2 if the company reported on an Equity Share basis.

Processing of sold products

Companies often differ on which parts of their value chain constitutes ‘Processing’ and which parts constitute ‘Use’.

Purchased goods and services

Companies either use Life Cycle Analysis, which considers the emissions of the emissions from the full value chain, whereas other companies only consider the emissions of their direct suppliers, ignoring the rest of the value chain.

Companies may not include all raw materials, goods, or services they purchase; many only account for paper or water purchases.

Upstream leased assets

The decision to lease or purchase assets depends on the company’s business strategy more than on size or activity and so any data reported in this category is inconsistent.

Upstream transportation and distribution

Calculations are very sensitive to the assumptions about mode of transport and so similar calculation methodologies may result in different values.

Use of sold products Calculations are sensitive to behavioural assumptions made about end users.

Waste generated in operations

Calculation methodologies vary and emissions from waste depend on method of disposal which may have a much stronger regional variation due to differences in regulations.