Mandating the Sustainability Disclosure in Annual Reports—Evidence from the United Kingdom

Abstract

This study investigates the sustainability disclosure effects of the introduction of the Companies Act 2006 Regulations 2013 in the United Kingdom. The regulation mandates the disclosure of information on greenhouse gas emissions, gender distribution and human rights issues. We examine two research questions: first, whether firms increased disclosure on the mandated topics after the regulation became effective relative to a control group, and second, whether a potential increase in disclosure is moderated by firms’ reporting incentives, namely, firms’ capital market visibility, growth orientation, governance structure, prior voluntary sustainability disclosure levels and critical media coverage. Our sample consists of the FTSE-350 firms and a matched control group of US firms. We use textual analysis to assess the disclosure of the mandated sustainability topics in firms’ annual reports. Specifically, we examine two types of disclosure, namely, the disclosure of the mandated key performance indicators and the narrative disclosure. Our results reveal a significant increase for both types of disclosure relative to the control group. Overall, this treatment effect tends to be smaller for firms with higher reporting incentives, i. e., reporting incentives mitigate the regulatory effect. Taken together, our results suggest that both standards and reporting incentives shape firms’ sustainability disclosure level.

Appendices

Appendix A

1.1 Details On the Construction of the Disclosure Measures

1.1.1 General Procedure Analogous to Hoberg and Maksimovic (2015) and Hummel et al. (2017)

The construction of KPI disclosure measures:

• 1. For each topic, we query according to predefined logical expressions across all documents. In particular, with respect to disclosure of

  • a. GHG emissions, the following logical expression is used for the search query:

    (‘tonne’ OR ‘ton’ OR ‘numeric’) AND (‘GHG’ or ‘*CO2*’ OR ‘carbon’ OR (‘greenhouse’ AND ‘gas’))

  • b. gender distribution, the following logical expression is used for the search query:

    ((‘female’ OR ‘gender’ OR ‘woman’ or ‘sex’) AND (‘board’ OR ‘director’ OR ‘executive’ OR ‘manager’ or ‘employee’) AND ‘numeric’) OR ((‘gender’ AND ‘distribution’) OR (‘gender’ AND ‘split’) OR (‘gender’ AND ‘breakdown’) AND ‘numeric’)

• 2. ghg_KPI and gender_KPI take on the value of “1” if the report loads on the search query and “0” otherwise.

The construction of topic-specific narrative disclosure measures:

• 1. Let N denote the number of unique words in the entire corpus.

• 2. For each topic, we query according to predefined logical expressions across all documents. In particular, with respect to disclosure on

  • a. GHG emissions, the following logical expression is used for the search query:

    (‘ghg’ AND ‘emission’) OR (‘*CO2*’ AND ‘emission’) OR (‘carbon’ AND ‘dioxide’) OR (‘greenhouse’ AND ‘gas’) OR (‘climate’ AND ‘change’) OR (‘kyoto’ AND ‘protocol’) OR (‘global’ AND ‘warming’)

  • b. gender distribution, the following logical expression is used for the search query:

    (‘gender’ AND ‘split’) OR (‘gender’ AND ‘diversity’) OR (‘gender’ AND ‘distribution’) OR (‘gender’ AND ‘breakdown’) OR (‘female’ AND ‘manager’) OR (‘woman’ AND ‘manager’) OR (‘female’ AND ‘management’) OR (‘woman’ AND ‘management’) OR (‘female’ AND ‘director’) OR (‘woman’ AND ‘director’) OR (‘female’ AND ‘executive’) OR (‘woman’ AND ‘executive’) OR (‘female’ AND ‘board’) OR (‘woman’ AND ‘board’)

  • c. human rights, the following logical expression is used for the search query:

    ‘human’ AND ‘right’

• 3. For each topic, we aggregate all retrieved ten-word windows into a topic-specific vocabulary list. The vocabulary list includes all words that appear in all retrieved ten-word windows for each topic.

• 4. For each topic, we define an N-vector search that is filled with the term-frequency-inverse-document-frequency (tf-idf) of each word in the topic vocabulary corresponding to each of the N elements.

• 5. For each firm i in each year t, we define an N-vector text_{i, t} that is filled with the tf-idf for each word in firm i’s annual report in year t corresponding to each of the N elements.³⁸

• 6. For each element of the N-vector, the inverse-document-frequency (idf) is calculated according to:

$$idf=\log _{2}\frac{n}{f}$$

(1)

where n::

number of all documents

f::

number of documents in which the word appears

• 7. For each element of the N-vector search, the tf-idf is calculated as the product of the number of times the word appears in the training set and the idf.

• 8. For each element of the N-vector text_{i, t}, the tf-idf is calculated as the product of the number of times the word appears in the annual report of firm i in year t (i. e., the term frequency) and the idf.

• 9. To neutralize the impact of the document length, we normalize the N-vector search according to:

$$\textit{search}\_ norm=\frac{\textit{search}}{\sqrt{\textit{search}\cdot \textit{search}}}$$

(2)

• 10. Similarly, we normalize the N-vector text_{i, t} according to:

$$\mathrm{norm}_{i,t}=\frac{\mathrm{text}_{i,t}}{\sqrt{\mathrm{text}_{i,t}\cdot \mathrm{text}_{i,t}}}$$

(3)

• 11. To obtain the similarity between firm i’s annual disclosure in year t and the topic vocabulary, we calculate similarity_{i, t} as the cosine similarity (i. e., the dot product) between norm_{i, t} and search_norm.

$$\text{similarity}_{i,t}=\mathrm{norm}_{i,t}\cdot \textit{search}\_ norm$$

(4)

• 12. For conventional reasons, the cosine similarity is multiplied by 100.

1.1.2 A Simple Example for the Calculation of the Cosine Similarity (Analogous to Hummel et al. (2017))

• 1. Consider three texts that, after application of the preprocessing methods, can be described according to the following word lists:

  text_1 = [‘employee’, ‘educate’, ‘women’]

  text_2 = [‘engage’, ‘board’, ‘gender’, ‘composition’, ‘women’]

  text_3 = [‘board’, ‘composition’, ‘engage’, ‘educate’, ‘women’]

• 2. Consider the following training set (as a result of the search query):

  search = [‘gender’, ‘board’, ‘women’]

• 3. The corpus is given by:

  corpus = [‘gender’, ‘board’, ‘women’, ‘composition’, ‘engage’, ‘employee’, ‘educate’]

• 4. The inverse-document-frequency for each word corresponds to:

  w_gender = 1.5850

  w_board = 0.5850

  w_women = 0.0000

  w_composition = 0.5850

  w_engage = 0.5850

  w_employee = 1.5850

  w_educate = 0.5850

• 5. The tfidf-vector for the training set and each text corresponds to:

  search = [1.5850, 0.5850, 0.0, 0.0, 0.0, 0.0, 0.0]

  text_1 = [0.0, 0.0, 0.0, 0.0, 0.0, 1.5850, 0.5850]

  text_2 = [1.5850, 0.5850, 0.0, 0.5850, 0.5850, 0.0, 0.0]

  text_3 = [0.0, 0.5850, 0.0, 0.5850, 0.5850, 0.0, 0.5850]

• 6. The normalized tfidf-vector for the training set and each text corresponds to:

  norm_search = [0.9381, 0.3462, 0.0, 0.0, 0.0, 0.0, 0.0]

  norm_text_1 = [0.0, 0.0, 0.0, 0.0, 0.0, 0.9381, 0.3462]

  norm_text_2 = [0.8426, 0.311, 0.0, 0.311, 0.311, 0.0, 0.0]

  norm_text_3 = [0.0, 0.5, 0.0, 0.5, 0.5, 0.0, 0.5]

• 7. The cosine similarity for each text corresponds to:

  similarity_text_1 = norm_search ∙ norm_text_1 = 0.0000

  similarity_text_2 = norm_search ∙ norm_text_2 = 0.8981

  similarity_text_3 = norm_search ∙ norm_text_3 = 0.1731

Appendix B

2.1 Examples of Incorrect Classifications with Regard to ghg_KPI and gender_KPI

We manually checked the validity of the disclosure measures, particularly with regard to the disclosure of key performance indicators. The results indicate that the textual analysis might not correctly identify the disclosure of the key performance indicators in some cases. Table 10 provides examples of incorrect classifications.

Table 10 Examples of incorrect classifications