Divergence of Esg Ratings
Divergence of Esg Ratings
https://doi.org/10.1093/rof/rfac033
Advance Access Publication Date: 23 May 2022
Abstract
This paper investigates the divergence of environmental, social, and governance
(ESG) ratings based on data from six prominent ESG rating agencies: Kinder,
Lydenberg, and Domini (KLD), Sustainalytics, Moody’s ESG (Vigeo-Eiris), S&P Global
(RobecoSAM), Refinitiv (Asset4), and MSCI. We document the rating divergence and
map the different methodologies onto a common taxonomy of categories. Using this
taxonomy, we decompose the divergence into contributions of scope, measurement,
and weight. Measurement contributes 56% of the divergence, scope 38%, and weight
6%. Further analyzing the reasons for measurement divergence, we detect a rater
effect where a rater’s overall view of a firm influences the measurement of specific
categories. The results call for greater attention to how the data underlying ESG
ratings are generated.
* We thank Alex Edmans and an anonymous referee for helpful comments; Suraj Srinivasan and three
anonymous referees at Management Science for helpful comments on an earlier version of this
manuscript; and all the ESG rating agencies that provided their data to this project. We also thank
Andrew King, Eric Orts, Jason Jay, Kathryn Kaminski, Lisa Goldberg, Rene Bohnsack, Robert Eccles,
Stefano Ramelli, Thomas Lyon, Timo Busch, and Yannick Le Pen, as well as participants at the AFA
conference, the EFA conference, the GRASFI conference, the Gronen conference, the Frankfurt
School of Finance ESG seminar, the JOIM seminar, the Fondazione Eni Enrico Mattei webinar, the
COIN Seminar at the JRC of the European Commission, the Harvard Kennedy School M-RCBG sem-
inar, the Wharton Business Ethics seminar, the Geneva Sustainable Finance Summit, the
Liechtenstein Sustainable Finance Workshop, and the Brown Bag Lunch Finance Seminar at the
University of Zurich for their inputs. Jonathan Hsieh, Armaan Gori, Andrew Lu, Erin Duddy, Sanjana
Rajaram, and Nadya Dettwiler provided excellent research assistance. All remaining errors are ours.
R.R. and F.B. are grateful to Massachusetts Pension Reserves Investment Management Board, AQR
Capital Management, MFS Investment Management, AssetOne Asset Management, and Qontigo—
members of the Aggregate Confusion Project Council—for their generous support. J.F.K. gratefully
acknowledges financial support of the BMW Foundation Herbert Quandt.
C The Author(s) 2022. Published by Oxford University Press on behalf of the European Finance Association.
V
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (https://creativecom-
mons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the origin-
al work is properly cited.
1316 F. Berg et al.
1. Introduction
Environmental, social, and governance (ESG) rating providers1 have become influential
institutions. A total of 3,038 investors representing over $100 trillion in combined assets
have signed a commitment to integrate ESG information into their investment decisions
(PRI, 2020). Sustainable investing is growing quickly and mutual funds that invest accord-
ing to ESG ratings experience sizable inflows (Hartzmark and Sussman, 2019). Due to these
trends, more and more investors rely on ESG ratings to obtain a third-party assessment of
1 ESG ratings are also referred to as sustainability ratings or corporate social responsibility ratings.
We use the terms ESG ratings and sustainability ratings interchangeably.
Aggregate Confusion 1317
only a small subset of the underlying indicators that make up the different ESG ratings. To
advance on this front, this paper provides a quantitative decomposition of ESG rating diver-
gence, relying on six ESG ratings along with the complete set of 709 underlying indicators.
We identify three distinct sources of divergence. “Scope divergence” refers to the situ-
ation where ratings are based on different sets of attributes. One rating agency may include
lobbying activities, while another might not, causing the two ratings to diverge.
“Measurement divergence” refers to a situation where rating agencies measure the same at-
tribute using different indicators. For example, a firm’s labor practices could be evaluated
2 Scope and weight combined are equivalent to theorization and one could argue that an attribute
that is out of scope has a weight of zero. However, in practice, rating agencies do not collect data
for attributes that are beyond their scope. There is thus a qualitative difference between an attri-
bute with zero weight and an attribute for which data are not available.
1318 F. Berg et al.
not reflect the real situation. Thus, we use a taxonomy that matches indicators by attribute.
To rule out that our subjective judgment drives the results, we sort the indicators according
to an alternative taxonomy provided by the Sustainability Accounting Standards Board
(SASB).3 The results based on this alternative taxonomy are consistent with those based on
our own taxonomy. Second, our linear aggregation rule is not industry-specific, while most
ESG rating agencies use industry-specific aggregation rules. However, this approximation
seems to be relatively innocuous because a simple linear rule achieves a very high quality of
fit. More sophisticated non-linear estimators, such as neural networks, do not yield better
3 Founded in 2011, SASB works to establish disclosure standards on ESG topics that are comparable
across companies on a global basis.
Aggregate Confusion 1319
addition, the rater effect raises questions about the economics of the ESG ratings market.
Structural reasons or incentives that need to be better understood may influence how certain
companies or categories are rated. Beyond improving measurement, the divergence itself begs
the question of how uncertainty in ESG ratings affects asset prices, a topic that is gaining at-
tention in the literature (Avramov et al., 2021; Gibson Brandon, Krueger, and Schmidt,
2021). Finally, our results raise the question of how companies respond to being scored differ-
ently by different raters (see also Chatterji, Levine, and Toffel, 2009), which will inform the
effects of sustainable finance on the real economy.
2. Data
ESG ratings first emerged in the 1980s as a way for investors to screen companies on environ-
mental, social, and corporate governance performance. The earliest ESG rating agency, Eiris
(merged with Vigeo in 2015), was established in 1983 in France, and 7 years later, Kinder,
Lydenberg, and Domini (KLD) was established in the USA. While initially catering to a highly
specialized investor clientele, including faith-based organizations, the market for ESG ratings
has widened dramatically, especially in the past decade. Because ESG ratings are an essential
basis for most kinds of sustainable investing, the market for ESG ratings grew in parallel to
sustainable investing. As sustainable investing transitioned from niche to mainstream, many
early ESG rating providers were acquired by established financial data providers. For ex-
ample, MSCI bought KLD in 2010, Morningstar acquired 40% of Sustainalytics in 2017,
Moody’s bought Vigeo-Eiris in 2019, and S&P Global bought RobecoSAM in 2019.
ESG rating agencies allow investors to screen companies for ESG performance, like
credit ratings allow investors to screen companies for creditworthiness. However, at least
three important differences between ESG ratings and credit ratings exist. First, while credit-
worthiness is relatively clearly defined as the probability of default, the definition of ESG
performance is less clear. It is a concept based on values that are diverse and evolving.
Thus, an important part of the service that ESG rating agencies offer is an interpretation of
what ESG performance means. Second, while financial reporting standards have matured
and converged over the past century, ESG reporting is in its infancy. There are competing
reporting standards for ESG disclosure, many of which are voluntary or limited to single
jurisdictions, giving corporations broad discretion regarding whether and what to report.
Thus, ESG ratings provide a service to investors by collecting and aggregating information
1320 F. Berg et al.
from across a spectrum of sources and reporting standards. These two differences explain
why the divergence between ESG ratings is so much more pronounced than the divergence
between credit ratings, the latter being correlated at 99%.4 Third, ESG raters are paid by
the investors who use the ratings, not by the companies that are rated, as is the case with
credit raters. As a result, the problem of ratings shopping, which has been discussed as a po-
tential reason for credit ratings diverging (see, e.g., Bongaerts, Cremers, and Goetzmann,
2012), does not apply to ESG rating providers.
We use data from six different ESG rating providers: KLD5, Sustainalytics, Moody’s
3. Divergence
Our point of departure is that ESG ratings diverge. In this section, we establish that the di-
vergence in our sample is substantial and consistent with prior studies. First, we compute
4 Because credit ratings are expressed on an ordinal scale, researchers usually do not report corre-
lations. For the sake of illustration, we use the data from Jewell and Livingston (1998) and calculate
a Pearson correlation by replacing the categories with integers.
5 KLD, formerly known as Kinder, Lydenberg, Domini & Co., was acquired by RiskMetrics in 2009.
MSCI bought RiskMetrics in 2010. The data set was subsequently renamed to MSCI KLD Stats as a
legacy database. We keep the original name of the data set to distinguish it from the MSCI data
set.
6 All raters, except for KLD, in our sample are featured in the 2019 and 2020 investor survey, “Rate
the Raters,” performed by the SustainAbility Institute (see https://www.sustainability.com/globalas
sets/sustainability.com/thinking/pdfs/sustainability-ratetheraters2020-report.pdf). Other raters
included in this list are CDP, Bloomberg, ISS, and FTSE. CDP and Bloomberg were ruled out as CDP
focuses only on environmental issues and Bloomberg mostly on the quality of disclosure. ISS was
unable to provide granular data. FTSE Russel merged with Refinitiv and is no longer an independent
rating. Calvert is not mentioned in the survey and we thus did not include them contrary to Chatterji
et al. (2016).
7 The KLD data set does not provide an aggregate rating; it only provides binary indicators of
“strengths” and “weaknesses.” We created an aggregate rating for KLD by following the proced-
ure that is chosen in most academic studies—namely, summing all strengths and subtracting all
weaknesses (see, e.g., Lins, Servaes, and Tamayo, 2017).
Aggregate Confusion 1321
Descriptive statistics of the aggregate ratings (ESG level) in 2014 for the six rating agencies.
Panel A shows the data for the full sample and Panel B for the common sample.
Correlations between ESG ratings at the aggregate rating level (ESG) and at the level of the en-
vironmental dimension (E), the social dimension (S), and the governance dimension (G) using
the common sample. The results are similar using pairwise common samples based on the full
sample. SA, SP, MO, RE, KL, and MS are short for Sustainalytics, S&P Global, Moody’s ESG,
Refinitiv, KLD, and MSCI, respectively.
2
Values of other ratings
Rater
S&P Global
Moody’s ESG
0 KLD
Refinitiv
MSCI
−2
−4
−2 −1 0 1 2 3
Value of benchmark rating (Sustainalytics)
þ1.5, which places it among the top 10% of companies rated by Sustainalytics. Yet, other
ratings at x ¼ 1.5 score the company below zero, placing the company below the sample
average. In other words, the extent of the divergence is such that it is difficult to tell a leader
from an average performer. This issue becomes even more pronounced when using other
ratings as a benchmark or when looking at rankings. For an illustration, refer to Online
Appendix Figure A.1.
The purpose of ESG ratings is to assess a company’s ESG performance. Yet, ESG rat-
ings disagree to an extent that leaves observers with considerable uncertainty as to how
4.1 Scope
The decomposition of ESG rating divergence is not trivial because at the granular level,
the structures of different ESG ratings are incompatible. Each rater chooses to break
down the concept of ESG performance into different indicators and organizes them in
different hierarchies. For example, at the first level of disaggregation, Moody’s ESG,
S&P Global, MSCI, and Sustainalytics have three dimensions (E, S, and G), Refinitiv
has four, and KLD has seven. Below these first-level dimensions, there are between one
and three levels of more granular sub-categories, depending on the rater. At the lowest
level, our data set contains between 38 and 282 indicators per rater, which often, but
not always, relate to similar underlying attributes. These incompatible structures make
it difficult to understand how and why different raters assess the same company in dif-
ferent ways.
We impose our own taxonomy on the data to perform a meaningful comparison of these
different rating systems. We develop this taxonomy using a bottom-up approach. First, we
create a long list of all available indicators and their detailed descriptions. In cases where
the descriptions were not available (or were insufficient), we interviewed the data providers
for clarification. In total, the list contains 709 indicators. Second, we group indicators that
describe the same attribute in the same “category.” For example, we group all indicators
related to the category Water, as shown in Table III. Third, we iteratively refine the tax-
onomy, following two rules: (a) each indicator is assigned to only one category and (b) a
new category is established when at least two indicators from different raters both describe
an attribute that is not yet covered by existing categories. For example, indicators related to
Forests were taken out of the larger category of Biodiversity to form their own category.
The classification is purely based on the attribute that indicators intend to measure, regard-
less of the method or data source used. Indicators that are unique to one rater and could
not be grouped with indicators from other raters were labeled “unclassified” and each
given their own rater-specific category.
The resulting taxonomy, shown in Table IV, assigns the 709 indicators to a total of
sixty-four distinct categories. Refinitiv has the most individual indicators with 282,
1324 F. Berg et al.
followed by Sustainalytics with 163. KLD, S&P Global, and MSCI have seventy-eight,
eighty, and sixty-eight, respectively, and Moody’s ESG has thirty-eight. Some catego-
ries—Forests, for example—contain only one indicator from two raters. Others, such as
Supply Chain, contain several indicators from all raters. Arguably, Forests is a much
narrower category than Supply Chain. The reason for this difference in broadness is
that there were no indicators in Supply Chain that together represented a more narrow
common category. Therefore, the Supply Chain comparison is at a more general level
and it may seem obvious that different raters take a different view of this category.
Nevertheless, this broad comparison represents the most specific level possible given the
data.
Table IV shows how many indicators each rater provides per category. On the one
hand, some categories are considered by all six raters, indicating that these are com-
monly accepted core ESG issues. These are Biodiversity, Employee Development,
Energy, Green Products, Health and Safety, Labor Practices, Product Safety,
Remuneration, Supply Chain, and Water. On the other hand, many empty cells show
Aggregate Confusion 1325
This table shows the indicators from different rating agencies assigned to the category Water.
that far from all categories are covered by all ratings. Gaps exist both for categories
that could be described as specialized, such as Electromagnetic Fields and also for the
Taxes category, which could be viewed as a fundamental concern in the context of ESG
issues. Also, the considerable number of unclassified indicators shows that many ESG
aspects are only measured by one out of six raters. Refinitiv has, with forty-two, the
most unclassified indicators, almost all of which stem from Refinitiv’s economic di-
mension. This dimension contains indicators, such as net income growth or capital ex-
penditure that other rating agencies do not consider. MSCI has thirty-four unclassified
indicators; these are what MSCI terms “exposure scores.” Next to scores that evaluate
how well a company manages an issue, MSCI has scores that measure how relevant the
issue is for the specific company. None of the other raters have indicators that explicit-
ly measure this.
The taxonomy imposes a structure on the data that allows a systematic comparison.
Obviously, results may be sensitive to the particular way we built it. To make sure our
results are not driven by a particular classification, we created an alternative taxonomy as a
robustness check. Instead of constructing the categories from the bottom up, we produce a
top-down taxonomy based on SASB. SASB has identified twenty-six general issue categories
based on a comprehensive stakeholder consultation process. As such, these categories repre-
sent the consensus of a wide range of investors and regulators on the scope of relevant ESG
categories. We map all indicators against these twenty-six general issue categories, again
requiring that each indicator can only be assigned to one category. This alternative tax-
onomy, along with results that are based on it, is provided in the Online Appendix. All our
results also hold for this alternative taxonomy.
4.2 Measurement
We can study measurement divergence using our taxonomy by comparing the assessments
of different raters at the level of categories. We create category scores (C) for each category,
firm, and rater. Category scores are calculated by taking the average of the indicator values
assigned to the category. Let us define the notations:
1326 F. Berg et al.
This table shows how many indicators are provided by the different rating agencies per cat-
egory. Categories that are covered by all raters are printed in bold.
Ozone-Depleting Gases 1 1
Packaging 1 1 1
Philanthropy 3 1 2 1 1
Privacy and IT 1 3 1 2
Product Safety 2 2 13 3 2 6
Public Health 1 3 1 2
The following variables and indexes are used throughout the paper: The category score is
computed as
1 X
Cfkj ¼ Ifki (1)
nfkj i2N
fkj
for firm f 2 ð1; 924Þ, rating agency k 2 ð1; 6Þ, and category j.
Category scores represent a rating agency’s assessment of a certain ESG category. They
are based on different sets of indicators that each rely on different measurement protocols.
It follows that differences between category scores stem from differences in “how” rating
agencies choose to measure, rather than what they choose to measure. Thus, differences be-
tween the same categories from different raters can be interpreted as measurement diver-
gence. Some rating agencies employ different sets of indicators for different industries. Such
Table V. Correlation of category scores
1328
Correlations of category scores for pairs of rating agencies. Category scores are an average of the indicators for firm f and rater k assigned to
the same category. Empty cells reflect differences in scope or missing data. SA, SP, MO, RE, KL, and MS are short for Sustainalytics, S&P
Global, Moody’s ESG, Refinitiv, KLD, and MSCI, respectively.
Category
F. Berg et al.
Category
Table V. Continued
1330 F. Berg et al.
industry-specific considerations about measurement are also reflected in the category scores
because those scores take the average of all available indicator values.
Table V shows the correlations between the categories. The correlations are calculated
on the basis of complete pairwise observations per category and rater pair. The table offers
two insights. First, correlation levels are heterogeneous. Environmental policy, for instance,
has an average correlation level of 0.55. This indicates at least some level of agreement
regarding the existence and quality of the firms’ environmental policy. However, even cate-
gories that measure straightforward facts that are easily obtained from public records do
(continued) The second insight is that correlations tend to increase with granularity. For
example, the correlations of the Water and Energy categories are on average 0.36 and 0.38,
respectively. This is substantially lower than the correlation of the environmental dimen-
sion, with an average of 0.53 reported in Table II. This implies that divergences compensate
for each other to some extent during aggregation. Several potential reasons can explain this
observation. One reason might be that category scores behave like noisy measures of an
underlying latent quality so that the measurement disagreement on individual categories
cancels out during aggregation. It may also be the case that rating agencies assess a firm
relatively strictly in one category and relatively leniently in another. A concern might be
that the low correlations at the category level result from misclassification in our tax-
onomy, in the sense that highly correlated indicators were sorted into different categories.
While we cannot rule this out completely, the alternative taxonomy based on SASB criteria
mitigates this concern. It is a much less granular classification, which, therefore, should de-
crease the influence of any misclassification. However, the average correlation per rater
pair changes only a little and not systematically when using this alternative taxonomy. This
provides reassurance that the observed correlation levels are not an artifact of misclassifica-
tion in our taxonomy. The correlations with the taxonomy based on SASB criteria can be
seen in Table A.3 of the Online Appendix.
4.3 Weight
We can proceed with an analysis of weight divergence based on the category scores. To do
so, we estimate the aggregation rule that transforms the category scores Cfkj into the rating
Rfk for each rater k. Category scores, as defined in Section 4.2, serve as independent varia-
bles. When no indicator values are available to compute the category score for a given firm,
the score is set to zero. This is necessary to run regressions without dropping all categories
with missing values, which are numerous. Of course, this entails an assumption that miss-
ing data indicate poor performance. Categories for which no values are available for any
firm in the common sample are dropped. After this treatment, category scores are normal-
ized to zero mean and unit variance, corresponding to the normalized ratings. Each unclas-
sified indicator is treated as a separate rater-specific category.
We perform a non-negative least squares regression, which includes the constraint that
coefficients cannot be negative. This is because we know a priori the directionality of all
Aggregate Confusion 1331
indicators and can thus rule out negative weights in a linear function. Thus, we estimate the
weights (wkj) with the following specification:
P
Rfk ¼ j2ð1;mÞ Cfkj wkj þ fk
(2)
wkj 0:
Because all the data have been normalized, we exclude the constant term. Due to the
non-negativity constraint, we calculate the standard errors by bootstrap. We focus on the
R2 as a measure of the quality of fit.
5. Decomposition
So far, we have shown that scope, measurement, and weight divergence exist. In this sec-
tion, we decompose the overall ratings divergence into the contributions of scope, measure-
ment, and weight divergence. We perform an arithmetic decomposition that relies on the
taxonomy, the category scores, and the aggregation weights estimated in Section 4. Scope
divergence is partialed out by considering only the categories that are exclusively contained
in one of the two ratings. Measurement divergence is isolated by calculating both ratings
with identical weights so that differences can only stem from differences in measurement.
Weight divergence is what remains of the total difference.
We assume that all ESG ratings are linear combinations of their category scores, based
^ fk (where k 2 a; b) be the rating provided
on the quality of fit of the linear estimations. Let R
by rating agency a and rating agency b for a common set of f companies. R ^ fk denotes the fit-
ted rating and w^ kj the estimated weights for rater k and category j based on the regression
in Table VI. Thus, the decomposition is based on the following relationship:
^ fk ¼ Cfkj w
R ^ kj : (3)
Common categories included in the scope of both raters are denoted as Cfkjcom . Exclusive
categories included by only one rater are denoted as Cfaja;ex and Cfbjb;ex . We separate the rat-
ing based on common and exclusive categories as follows:
^ fk;com ¼ Cfkj w
R ^ kjcom
com
R^ fk ¼ R ^ fk;com þ R^ fk;ex :
On this basis, we can provide terms for the contributions of scope, measurement, and
weight divergence to the overall divergence. Scope divergence Dscope is the difference be-
tween ratings that are calculated using only mutually exclusive categories.
Measurement divergence Dmeas is calculated based on the common categories and identical
Aggregate Confusion 1333
Non-negative linear regressions of the ESG rating on the categories of the same rater. The con-
stant term is excluded because the data have been normalized. The standard errors are boot-
strapped. Non-existent categories are denoted by dashes. The three most important categories
per rater, indicated by coefficient size, are printed in bold.
Significance levels: *P < 0.05, **P < 0.01, ***P < 0.001.
weights for both raters. The identical weights w ^ are estimated jointly for two ratings, as
specified in Equation (7). This is a non-negative pooling regression of the stacked
ratings on the stacked category scores of the two raters. Because the regression optimizes
the fit with w ^ , we can attribute the remaining differences to measurement divergence.
Weight divergence Dweights is simply the remainder of the total difference, or, more explicit-
ly, a rater’s category scores multiplied with the difference between the rater-specific weights
w ^ . It must be noted that all these calculations are performed using the fitted rat-
^ ajcom and w
^
ings R and the fitted weights w ^ because the original aggregation function is not known
with certainty.
Aggregate Confusion 1335
^ fa R
Dfa;b ¼ R ^ fb ¼ Dscope þ Dmeas þ Dweights : (5)
The terms for scope, measurement, and weight are given as follows:
Dscope ¼ Cfaja;ex w
^ aja;ex Cfbjb;ex w
^ bjb;ex
Dmeas ¼ ðCfajcom Cfbjcom Þ w ^ (6)
Dweights ¼ Cfajcom ðw^ ajcom w^ Þ Cfbjcom ðw ^ Þ;
^ bjcom w
!
^ fa;com
R Cfajcom fa
¼ w þ : (7)
^ fb;com
R Cfbjcom fb
We analyze the variance of Da;b over the sample of firms to calculate the cross-sectional con-
tribution of scope, weight, and measurement. Taking variances in Equation (5), we obtain
Because all ratings are normalized to a zero mean, the total difference between two ratings
is equal to zero. Yet the firm-specific differences are different from zero and the variance of
Da;b provides a summary statistic of these differences. Using the variance as a measure of di-
vergence, Equation (8) yields a straightforward way to calculate the contributions of scope,
measurement, and weight to this divergence.
Comparison of the quality of fit in terms of R2 for the estimation of rater-specific aggregation
functions using different specifications. NNLS stands for non-negative least squares and OLS
for ordinary least squares. NN stands for neural network with linear activation function and NN
ReLU for a neural network with a non-linear ReLU activation function. RF stands for random for-
est. The symbol * indicates that the R2 is reported for a testing set consisting of a randomly
chosen 10% of the sample. The three last lines report results from the original method with dif-
ferent underlying data. For NNLS SASB, the category scores were calculated based on the
divergence. This suggests that those two raters have very similar views on what ESG is
about. However, differences remain and these are mainly due to measurement divergence.
Panel B highlights differences between raters. MSCI stands out as the only rater
where scope instead of measurement contributes most to the divergence. This result is
driven by MSCI’s exposure scores. As described in Section 4, these scores essentially
set company-specific weights for each category. As these scores have no equivalent in
the other rating methods, they increase the scope divergence of MSCI with respect to all
other raters. At the same time, the contribution of weight is negative for MSCI due to a
negative covariance between scope and weight divergence. In other words, the effects of
scope divergence and weight divergence tend to compensate for each other in the case of
MSCI. For all other raters except MSCI, the contribution decreases from measurement to
scope to weight.
Our analysis also allows us to identify the categories that are most consequential for
measurement divergence across all raters. To this end, we average the absolute measure-
ment divergence by category. It turns out that some categories for which there is pro-
nounced measurement disagreement ultimately do not matter much for rating divergence
because they tend to have a small weight in the aggregate ratings. These include
Environmental Fines, Clinical Trials, Employee Turnover, HIV Programs, and Non-GHG
Air Emissions. On the other end of the spectrum are categories where measurement diver-
gence is very consequential for overall divergence, namely, Climate Risk Mgmt., Product
Safety, Corporate Governance, Corruption, and Environmental Mgmt. System. These latter
categories are priority targets in terms of addressing measurement divergence.
Aggregate Confusion 1337
Results from the arithmetic decomposition that implements Equation (8) and relies on the cat-
egory scores and estimated weights from Section 4. Panel A reports the relative contribution of
scope, measurement, and weight to the ESG rating divergence. For convenience, Panel B
reports averages per rater based on the values shown in Panel A.
Since the decomposition is our main result, we subject it to several robustness checks
that are available in the Online Appendix. First, we perform the decomposition on the basis
of the SASB taxonomy (Online Appendix Table A.5). In this case, the contribution of meas-
urement divergence is even higher (60%). This is consistent with the expectation that as cat-
egories become broader, measurement divergence within those categories increases.
Second, we perform a regression-based decomposition as an alternative (Online Appendix
Section A). Using this method, scope is as important as measurement, but weight continues
to play a minor role. Third, we repeat the decomposition with data from 2017 (Online
Appendix Table A.9), confirming that measurement divergence is the dominant source of
ESG rating divergence in 2017 as well.
1338 F. Berg et al.
0.5
Rating
0.0
Residual
−1.0
Other
Other
Community and Society
Remuneration
Biodiversity
Difference to fitted rating
Toxic Spills
6. Rater Effect
To further investigate the underlying reasons for measurement divergence, this section tests
for the presence of a “rater effect.” The rater effect describes a bias, where performance in
one category influences perceived performance in other categories. This phenomenon is
also called the “halo effect” and related biases have been extensively studied in sociology,
management, and psychology, especially in performance evaluation (see Shrout and Fleiss,
1979). The process of evaluating firms’ ESG attributes seems prone to a rater effect.
Evaluating firm performance in the Human Rights, Community and Society, and Labor
Practices categories requires rating agencies to use some degree of judgment. The rater ef-
fect implies that when the judgment of a company is positive for one particular indicator, it
is also likely to be positive for another indicator. We evaluate the rater effect using two pro-
cedures. First, we estimate fixed-effects regressions comparing categories, firms, and raters.
Aggregate Confusion 1339
where 1f are dummies for each firm, 1f k is an interaction term between firm and rater-
fixed effects, and 1f k is an interaction term between firm- and category-fixed effects. The
vector Cfkj stacks the scores for all common categories across all raters and firms. We drop
pure category and rater-fixed effects because of the normalization at the rating and cat-
egory scores level. We only use the intersection of categories from all raters and the com-
mon sample of firms to reduce sample bias.
The baseline regression in Equation (9) explains category scores with firm dummies.
The second regression adds the firm-rater-fixed effects, that is, a dummy variable for each
firm-rater pair. The increment in R2 between the two regressions is the rater effect. The
third and fourth regressions repeat the procedure with the additional inclusion of category-
firm-fixed effects. The results of these regressions are shown in Table IX.
We detect a clear rater effect. Firm dummies alone explain 0.22 of the variance of the
scores in Equation (9). However, when including firm-rater dummies, the R2 increases to
0.38, an increase of 0.16. Similarly, the difference in R2 between Equations (11) and (12)
yields an increase of 0.15. Therefore, the rater effect explains about 0.15–0.16 of the vari-
ation in category scores. The rater effect is relevant in comparison to the other dummies.
Comparing the estimates of Equations (11) and (9), we find that including firm-category
dummies improves the fit by 0.25. Similarly, comparing the outcomes of Regressions 12
and 10 yields an increase of 0.24. Thus, firm dummies explain 0.22, firm-category dummies
0.24–0.25, and firm-rater dummies 0.15–0.16. This means that after controlling for which
firm is rated and in which category, the rater itself has a substantial effect on the category
score.
This table reports R2 values of different dummy regressions. The dependent variable in each re-
gression is a vector that stacks the scores of all raters and firms for all categories that are com-
mon across raters. The independent variables are firm, firm-rater, and firm-category dummies.
The difference in R2 between regression 1 and 2 as well as 3 and 4 represents the rater effect.
Dummies R2
Firm 0.22
added one after another. This implies that one could replicate an overall rating with less
than the full set of categories.
We test this by estimating the linear aggregation rules with a LASSO regression. The
LASSO estimator adds a regularization to the minimization problem of ordinary least
squares. The objective is to reduce the number of wkj ¼ 6 0 and find the combination of regres-
sors that maximizes the explanatory power of the regression. The optimization is as follows:
X X
min ðRfk Cfkj wkj Þ2 þ k jwkj j; (13)
wkj
j j
where k controls the penalty. When k ¼ 0, the estimates from OLS are recovered. As k
increases, the variables with the smallest explanatory power are eliminated. In other words,
the first category that has the smallest marginal contribution to the R2 is dropped from the
regression (or its coefficient is set to zero). When k continues to increase, more and more
coefficients are set to zero, until there is only one category left.
Figure 4 shows the increase in R2 for each rating agency. The last part of the curve to
the right coincides with an unrestricted OLS estimate where all variables are included. KLD
and MSCI have the smallest cross-category correlation, judging by the slope in Figure 4(a)
and (f). In contrast, the slopes for Sustainalytics, Moody’s ESG, Asset 4, and S&P Global
suggest that only a few categories already explain most of the ESG rating. For S&P Global,
10% of the categories explain 75% of the rating.
The rater effect suggests that measurement divergence is not only randomly distributed
noise. Instead, a part of the divergence follows a pattern that suggests structural reasons. A po-
tential explanation for the rater effect is that rating agencies are organized so that analysts spe-
cialize in firms rather than indicators. A firm that is perceived as good in general may be seen
through a positive lens and receive better indicator scores than a firm that is perceived as bad
in general. In discussions with S&P Global, we learned about another potential cause for such
a rater effect. Some raters make it impossible for firms to receive a good indicator score if they
do not give an answer to the corresponding question in the questionnaire. This happens re-
gardless of the actual indicator performance. The extent to which the firms answer specific
questions may be correlated across indicators. Hence, the rater effect could also be due to
rater-specific assumptions that systematically affect assessments. There could also be economic
incentives that affect measurement. For example, Cornaggia, Cornaggia, and Hund (2017)
Aggregate Confusion 1341
suggest that credit raters may have incentives to inflate certain ratings. An interesting avenue
for future research is whether ESG raters have similar incentives to adjust their ratings.
7. Conclusions
The contribution of this article is a decomposition of ESG ratings divergence. Chatterji
et al. (2016) have taken an important first step by distinguishing two aspects that matter,
first, how ESG raters define what they intend to measure, and second, how they measure it.
(a) (b)
Refinitiv Sustainalytics
(e) (f)
Figure 4. LASSO regressions. (a) KLD. (b) S&P Global. (c) Refinitiv. (d) Sustainalytics. (e) Moody’s
ESG. (f) MSCI.
Notes: The plots show the R2 values of a series of LASSO regressions, regressing the aggregate ESG
rating of the different rating agencies on the categories of the same rater. The x-axis shows how many
indicators are used as covariates and the y-axis indicates the corresponding R2 value.
with raw data that can be independently verified. If that is not feasible, researchers should
carefully examine how the data are generated and remain skeptical of data where the data
generation process is not entirely transparent. When high-quality data are not available,
researchers should also consider collecting ESG data themselves and sharing the data set. In
short, given the ESG rating divergence, any research using ESG ratings or metrics needs to
pay special attention to the validity of the data used.
Turning to investors, our methodology enables them to understand why a company has
received different ratings from different rating agencies. The example in Figure 3 illustrates
Supplementary Material
Supplementary data are available at Review of Finance online.
Funding
R.R. and F.B. are grateful to Massachusetts Pension Reserves Investment Management Board,
AQR Capital Management, MFS Investment Management, AssetOne Asset Management, and
Qontigo—members of the Aggregate Confusion Project Council—for their generous support.
References
Albuquerque, R., Koskinen, Y., and Zhang, C. (2019): Corporate social responsibility and firm
risk: theory and empirical evidence, Management Science 65, 4451–4469.
Avramov, D., Cheng, S., Lioui, A., and Tarelli, A. (2021): Sustainable investing with ESG rating
uncertainty, Journal of Financial Economics, https://doi.org/10.1016/j.jfineco.2021.09.009.
Bongaerts, D., Cremers, K. J. M., and Goetzmann, W. N. (2012): Tiebreaker: certification and
multiple credit ratings, Journal of Finance 67, 113–152.
Chatterji, A. K., Durand, R., Levine, D. I., and Touboul, S. (2016): Do ratings of firms converge?
Implications for managers, investors and strategy researchers, Strategic Management Journal
37, 1597–1614.
Chatterji, A. K., Levine, D. I., and Toffel, M. W. (2009): How well do social ratings actually meas-
ure corporate social responsibility?, Journal of Economics & Management Strategy 18,
125–169.
Christensen, D. M., Serafeim, G., and Sikochi, A. (2022): Why is corporate virtue in the eye of the
beholder? The case of ESG ratings, The Accounting Review 97, 147–175.
Cornaggia, J. N., Cornaggia, K. J., and Hund, J. E. (2017): Credit ratings across asset classes: a
long-term perspective, Review of Finance 21, 465–509.
Flammer, C. (2015): Does corporate social responsibility lead to superior financial performance?
A regression discontinuity approach, Management Science 61, 2549–2568.
Gibson Brandon, R., Krueger, P., and Schmidt, P. S. (2021): ESG rating disagreement and stock
returns, Financial Analysts Journal 77, 104–127.
Hartzmark, S. M. and Sussman, A. B. (2019): Do investors value sustainability? A natural experi-
ment examining ranking and fund flows, Journal of Finance 74, 2789–2837.
Heinkel, R., Kraus, A., and Zechner, J. (2001): The effect of green investment on corporate behav-
ior, Journal of Financial and Quantitative Analysis 36, 431–449.
Jewell, J. and Livingston, M. (1998): Split ratings, bond yields, and underwriter spreads, Journal
of Financial Research 21, 185–204.
Krippendorff, K. H. (2004): Content Analysis: An Introduction to Its Methodology, 2nd edition,
Sage Publications Thousand Oaks, USA.
Liang, H. and Renneboog, L. (2017): On the foundations of corporate social responsibility,
Journal of Finance 72, 853–910.
Lins, K. V., Servaes, H., and Tamayo, A. M. (2017): Social capital, trust, and firm performance:
the value of corporate social responsibility during the financial crisis, Journal of Finance 72,
1785–1824.
PRI. (2020): PRI Annual Report. Available at: https://www.unpri.org/annual-report-2020/
Servaes, H. and Tamayo, A. (2013): The impact of corporate social responsibility on firm value:
the role of customer awareness, Management Science 59, 1045–1061.
Shrout, P. E. and Fleiss, J. L. (1979): Intraclass correlations: uses in assessing rater reliability,
Psychological Bulletin 86, 420–428.