Background & Summary

Under the goal of carbon neutrality, building green and low-carbon transformation systems and facilitating high-quality development of manufacturing industries are of great importance to achieving ambitious decarbonization targets1. Affected by the developing stage, China’s manufacturing industries are exposed to the dilemma of reducing carbon emissions and promoting economic development2, which requires governments’ support and stimulation through low-carbon policies. However, owing to the data limitation, what types of low-carbon policies are effective in facilitating low-carbon transformation and industrial upgrading are still unclear. Therefore, it is imperative to construct a comprehensive low-carbon policy intensity index based on policy texts from national- to prefecture-level.

Policy intensity is an index reflecting the stringency and importance of policy, which weighs through policy objectives and policy instruments3,4,5,6,7. While policy objectives mainly focus on issues that low-carbon policies formally aim to address, policy instruments pay attention to general norms that guide implementation preferences4,7. Based on this “policy objective–policy instrument” pattern, the meaning of policy intensity has been interpreted from different perspectives. Some studies coded objectives in terms of emission reduction and renewable energy production8. Other researchers divided policy instruments into demand-side, environmental and supply-side instruments9. Meanwhile, six measures containing objectives, scope, budget, implementation, and monitoring were also adopted to build policy intensity8.

However, a majority of policy intensity indices are constructed and discussed at the national level, ignoring the contribution of policy effects from sub-national regions. For example, an environmental policy stringency (EPS) index developed by OECD measured 13 policy instruments for 40 countries10,11. By investigating countries’ policy portfolios from 1998 to 2010, an index of climate policy activity was constructed for the energy production sector in Austria, Germany, and the United Kingdom8. To be more specific, policy level is an essential factor influencing the intensity of policy instruments. In general, national policies are more overarching, which provide developing directions and guidance for the whole country but promulgate few region-specific actions. Sub-national policies (e.g., provincial-level, prefecture-level instruments) are more focused and capable of devising appropriate measures suitable to local conditions. Hence, some policies from lower administration levels could have a higher intensity than their counterparts from higher policy levels, which needs to be quantified through policy intensity.

Moreover, studies pay less attention to the policy intensity of low-carbon policies. Previous literature mainly used dummy variables or ordinal variables to quantify the individual effect of pilot12,13,14,15, short-term16, regional and industrial17 low-carbon policies, which are not able to evaluate the effects of low-carbon policies as a whole. Meanwhile, when it comes to different types of low-carbon policy instruments, although effects of tax measures, subsidies, investment incentives, bidding systems, voluntary programs, quantity obligations, and environmental protection laws have been discussed by several researchers18,19, no consistent conclusion has been made for command-and-control regulation and market-based policies.

From the methodology perspective, owing to the immense progress made in machine learning and artificial intelligence, it is possible to construct policy indicators based on the meaning of policy texts and these methods, which could have limited human biases and increase the effectiveness of policy indices20. However, this research area has not been sufficiently developed, and only a few studies built low-carbon or environmental indicators based on machine learning. On the one hand, a semi-automated policy analysis tool with a labelling process and sentence-BERT (SBERT) model was used to identify forest policies and classify policy instruments21. On the other hand, China’s national-level environmental policy intensity for 1978–2019 has been constructed by policy text analysis and machine learning algorithms (e.g., random forest, support vector machine, Ridge), which shows the policy-intensity evolution towards environmental issues5.

Therefore, this paper aims to build a low-carbon policy intensity index from 2007 to 2022 by prompt learning for China’s manufacturing industries, which contributes to the literature in three aspects. First, the meaning of low-carbon policy intensity is deepened by adding the factor of policy level, which is based on 7282 low-carbon policies promulgated from the nation, 31 provinces and all 334 prefecture cities for manufacturing industries. Hence, following the “policy objective-policy instrument” pattern, the low-carbon policy intensity in this paper is quantified by multiplying each policy’s objective, instrument, and level. Second, the construction process of low-carbon policy intensity combines phrase-oriented NLP algorithm and text-based prompt learning, which is a few-shot learning suitable for policy texts with smaller samples and higher manual labelling costs. This new paradigm in NLP field avoids the need for large samples in pre-train and fine-tuning paradigms. Meanwhile, it also reduces human biases and the non-repeatability brought by traditional 100% manual policy scoring. Lastly, this study further aggregates sub-intensity indices for three policy levels (i.e., national-, provincial-, and prefecture-level), four policy objectives (i.e., carbon reduction, energy conservation, capacity utilization, and technology), and three policy instruments (i.e., command-and-control, market-based, composite instruments). It is more convenient for future research to merge this dataset with macro- and micro-data for extended analysis and discuss the impact of low-carbon policies from different perspectives.

Methods

The research framework of this paper includes six modules, which are displayed in Fig. 1. Based on the constructed low-carbon policy inventory with 7282 policies over 2007–2022, this paper extends the “policy target-policy instrument” pattern and uses prompt learning to quantify the low-carbon policy intensity, thereby calculating the policy intensity via multiplying policy level, objective and instrument.

Fig. 1
figure 1

The research framework of building low-carbon policy intensity.

Full size image

Data preparation

The section for data preparation lays the foundation for building low-carbon policy intensity, which is achieved through three steps: constructing a low-carbon policy inventory, structuring policy texts, and disaggregating texts into policy objectives and policy instruments.

Step 1: Constructing a low-carbon policy inventory

This paper collects low-carbon policies from PKULaw.com (https://pkulaw.com/), which is a policy database containing not only national policies but also policies from provincial- and prefecture-level. This paper focuses on different administration levels’ low-carbon policies from 2007 to 2022. 2007 is of great importance for China’s climate change issue when a national leading group headed by the Premier of the State Council and with officials from 30 ministries and commissions as members was founded in response to climate change, energy conservation, and emission reduction. Since then, China’s low-carbon policy has gradually changed from being attached to various environmental policies to an important policy category with more independent policies.

When selecting low-carbon policies, this paper defines a generalized scope of “low-carbon”. Keywords such as “carbon reduction”, “greenhouse gases”, “energy conservation”, “energy efficiency”, “energy consumption”, and “overcapacity” have been adopted to search policies, since policies in these fields are able to directly or indirectly mitigate climate change and influence low carbon development. Meanwhile, we put more emphasis on targeted and implementable policies, which have relatively clear low-carbon policy objectives and instruments with carbon reduction potential. To be more specific, despite higher legal forces, laws and outlines for quality development related to low-carbon and climate change are excluded from the policy inventory. Meanwhile, policies related to publicity have not been considered, which are more likely to raise awareness rather than taking action.

Finally, this paper builds a low-carbon policy inventory with 7282 policy texts from national-, provincial- and prefecture-level by careful selection. Policies from central and local departments related to development and reform, ecology and environment, finance, industry and information technology, and taxation are included in the inventory. Table 1 displays departments at different administration levels that have promulgated low-carbon policies.

Table 1 Departments from different levels that have promulgated low-carbon policies.
Full size table

Step 2: Structuring policy texts

It is known that policy texts have a high degree of fixed features, which contain structural information such as policy title, policy background, policy objectives, policy instruments, issuing institution, and publication year8. Hence, it is more convenient to grasp the meaning of whole policy texts by breaking it into different parts.

During the structuring process, headings and subheadings are key marks of each section. Headings of policy texts usually include sections for guiding ideology, working principle, target and goals, and supporting measures, which are led by phrases such as “Chapter X” and “Chinese number、”. Subheadings of policy texts are detailed sub-sections of headings, led by “(Chinese number)” or “Arabic figure、”. Thus, by locating lines with those special phases, headings, subheadings as well as contents for each part in the text file are able to be extracted respectively and sequentially to each cell of Excel in columns “title” and “content”.

Step 3: Disaggregating policy texts into the policy-objective file and policy-instrument file

Based on the Excel file generated in Step 2, we refine texts for policy objective and policy instrument into two separate files by keywords. On the one hand, “target”, “tasks”, “indices”, and “policy guidance” are keywords in headings, subheadings and contents for selecting policy objectives. Those cells are selected and sorted by subheadings, headings, and contents before being saved as a new text file for policy objective. Because the description of policy objectives in subsections tends to be more specified with clear expression than the counterparts. For some policy texts without headings or subheadings, sentences related to policy objectives cannot be located and have to be extracted from the whole content. On the other hand, policy instruments have diverse keywords and need to adopt an exclusion method. Hence, after dropping parts for policy objective and background (e.g., status, meaning, guidance, principle, general thoughts, requirements), the remaining parts of the policy are saved as a new text file for policy instrument.

Policy text classification

Having followed the “policy objective-policy instrument” pattern and refined text files for policy objective and instrument5,8, we consider three factors influencing low-carbon policy intensity, namely, policy level, policy objective, and policy instrument. Based on the low-carbon policy inventory we built, this paper classifies policies into national-, provincial-, and prefecture-level, which are promulgated by departments of the nation, provinces, autonomous regions, or cities. As for policy objectives, we follow the scope of “low-carbon” defined in this paper, and include policy objectives for carbon reduction, energy conservation, capacity utilization, and technology. Carbon reduction is the composite objective directly related to low carbon, while energy conservation is the composite objective for energy consumption and efficiency indirectly associated with low carbon. The remaining two objectives focus on specific fields of easing overcapacity, technological innovation, and industrial upgrading, which show an indirect connection with low carbon. Finally, based on existing studies5,22, we not only include policies with single command-and-control or market-based instruments, but also define policies with composite instruments (i.e., having command-and-control and market-based instruments in one policy at the same time).

Apart from the policy level, policy objectives and instruments are classified through high-frequency phrases extracted from policy texts and titles. First, during the process of data preparation, a lexicon has been constructed for policy classification, containing keywords and detailed expressions for titles and contents. Following Tian et al.23, keywords related to different policy objectives and instruments are constructed by extensively checking through low-carbon literature, which include 38 seed words and are shown in Table 2. Furthermore, detailed expressions for titles and contents are extracted by using word2vec to acquire the semantic similarity between keywords and tokenized policy contents. To be more specific, having calculated the similarity between the word vectors of each phrase in policy contents and keywords, we selected 90 synonym words with the highest similarity to each keyword. After manually checking and excluding phrases with inappropriate meanings to low-carbon, a lexicon for policy classification is constructed and applied to calculate phrase frequency for policy text classification.

Table 2 Selection of keywords in different dimensions for policy text classification.
Full size table

Second, in the policy classification, one policy only belongs to a single category of policy objectives and policy instruments. For policies with the same number of phrase frequency in multiple categories, it is necessary to set the priority for classification. In this paper, composite objectives (i.e., carbon reduction, energy conservation) take precedence over specific fields (i.e., capacity utilization, technology), and direct policy objective (i.e., carbon reduction) is given priority over the indirect counterpart (i.e., energy conservation). For policy instruments, on the basis of using high-frequency phrases to divide command-and-control and market-based instruments, it is necessary to define the grouping criteria for composite instruments. The average phrase frequency for market-based instruments is 10 phrases higher than the counterpart for command-and-control instruments. Hence, policies are classified into the composite instrument if the phrase frequencies for both instruments are greater than 10 and the difference is within 10.

Quantifying low-carbon policy intensity

Having classified policies into different categories, we add the factor of policy level into “policy objective-policy instrument” pattern and calculate the low-carbon policy intensity for each policy using Eq. (1):

$$P{I}_{r,a,t}={L}_{r,a,t,l}\times {O}_{r,a,t,o}\times {I}_{r,a,t,i}$$
(1)

where \(P{I}_{r,a,t}\) is the low-carbon policy intensity for policy a in region r and year t. Region r could be a nation, a province ρ or a prefecture-level city c. \({L}_{r,a,t,l}\) is the intensity of policy level l for policy a in region r and year t. Or,a,t,o is the intensity of policy objective o for policy a in region r and year t. \({I}_{r,a,t,i}\) is the intensity of policy instrument i for policy a in region r and year t.

Figure 2 shows the components of low-carbon policy intensity. To be more specific, the policy level l could be national-, provincial-, or prefecture-level. Policy objective o has four categories (i.e., carbon reduction, energy conservation, capacity utilization, technology). There are three types of policy instrument i, namely, command-and-control, market-based, and composite instruments.

Fig. 2
figure 2

Process of aggregating low-carbon policy intensity.

Full size image

As shown in Fig. 2, \(P{I}_{r,a,t}\) includes 36 types of policies and can be aggregated in different patterns by Eqs. (2–4).

$$P{I}_{r,t}={\sum }_{a}P{I}_{r,a,t}$$
(2)
$$P{I}_{r,t}=PI\_C{R}_{r,t}+PI\_E{C}_{r,t}+PI\_C{U}_{r,t}+PI\_{T}_{r,t}$$
(3)
$$P{I}_{r,t}=PI\_C{C}_{r,t}+PI\_M{B}_{r,t}+PI\_C{M}_{r,t}$$
(4)

Calculated by Eq. (2), \(P{I}_{r,t}\) is the sum of policy intensity promulgated by region r in year t. From the perspective of policy level, \(P{I}_{r,t}\) could be \(PI\_{N}_{t}\), \(PI\_{P}_{\rho ,t}\) or \(PI\_{C}_{c,t}\). \(PI\_{C}_{c,t}\) is the sum of prefecture-level low-carbon intensity promulgated by prefecture-level city c in year t. \(PI\_{P}_{\rho ,t}\) is the sum of provincial low-carbon intensity promulgated by province ρ in year t. \(PI\_{N}_{t}\) is the national level low-carbon intensity in year t. In Eq. (3), \(P{I}_{r,t}\) is summed by four different policy objectives. \(PI\_C{R}_{r,t}\), \(PI\_E{C}_{r,t}\), \(PI\_C{U}_{r,t}\) and \(PI\_{T}_{r,t}\) are policy intensities respectively aiming at carbon reduction, energy conservation, capacity utilization, and technology promulgated by region r in year t. Finally, \(P{I}_{r,t}\) can also be aggregated by three different policy instruments in Eq. (4). \(PI\_C{C}_{r,t}\), \(PI\_M{B}_{r,t}\) and \(PI\_C{M}_{r,t}\) are respectively policy intensities for command-and-control, market-based and composite instruments for region r in year t.

Manual labelling

In order to create a supervised prompt learning model and predict intensities for policies’ objective and instrument, a manual annotation of 3334 policy texts in 16 provinces is used to train the model. The intensity of policy level \({L}_{r,a,t,l}\), policy objective Or,a,t,o, and policy instrument \({I}_{r,a,t,i}\) are respectively rated on a scale of 1 to 3 (Label 1: weak intensity, Label 2: medium intensity, Label 3: high intensity, Label 4: highest intensity (for energy conservation objective only)).

For policy objectives, quantitative policy objectives with specific industrial, regional, or year-by-year supporting goals will be given a higher score. On the contrary, if the expression of policy objective is vague without quantitative or clear targets, the intensity will be relatively low. Unlike the scoring system for carbon reduction, capacity utilization, and technology goals, the policy objective for energy conservation has a 4-level scoring system and will be normalized within 3 (i.e., 3, 2.25, 1.5, 0.75) after scoring. Because the maximum level of energy conservation objective has been intensified after 2015, which changed from controlling the intensity of energy consumption to restricting the total amount and intensity of energy consumption at the same time. Hence, the energy conservation objective with level 4 is mainly used for scoring policies with “dual control of energy consumption”.

Meanwhile, the scoring criteria for three policy instruments are quite different. For command-and-control instruments, Label 3 will be given to policies for shutting down enterprises, eliminating excess production capacity and the target responsibility system for local government officials, which is related to their promotion, awards and punishment. Label 2 is mainly for strictly prohibiting additional production capacity, which has less impact on existing capacity. The intensity for market-based instruments is related to the clarity of policies’ economic supports. For composite instruments, since it includes command-and-control and market-based instruments at the same time, this type of policy will be respectively scored on the basis of two instruments. The maximum value of two labels will be taken as the intensity of composite instrument.

Detailed criteria and labelling examples for each objective and instrument are presented in Tables 3, 4.

Table 3 Labelling criteria of the intensity for different policy objectives.
Full size table
Table 4 Labelling criteria of the intensity for different policy instruments.
Full size table

Prompt learning and prediction

Based on the dataset of manual labelling, this paper trains models for each objective and instrument, and predicts the intensity by prompt learning, which is a few-shot learning suitable for policy texts with a smaller sample and higher manual labelling costs. Before prompt learning, the dominant pre-train and fine-tune paradigm in NLP requires a fixed architecture pre-trained as a language model (LM) and adapts the LM model to different downstream tasks by adding additional parameters and fine-tuning24. Hence, it is inefficient due to the need for parameters for every single project, and the gap between large-scale unsupervised learning and downstream tasks. Compared with the paradigm mentioned above, prompt learning is able to reformulate downstream tasks, makes it similar to the original pre-trained LM, and predicts desired output without additional training for tasks with the help of a textual prompt24.

To be more specific, prompt learning requires three steps to make a text input x to predict label output y through learning the parameters θ of model \(P\left(y| x;\theta \right)\) in the task for text classification24. First, a template is designed by the prompting function \(prompt\;x{\prime} ={f}_{{\rm{prompt}}}\left(x\right)\) with an input slot x and an answer text z, which will be mapped to y. Then, based on the pre-trained LM and small word subset Z for classification, the highest-scoring text \(\widehat{z}\) is searched using Eq. (5), which could maximize the score of LM. Finally, the highest-scoring answer \(\widehat{z}\) is used to predict the highest-scoring label output \(\widehat{y}\).

$$\widehat{z}=\mathop{{\rm{search}}}\limits_{z\in Z}P\left({f}_{fill}\left(x{\prime} ,z\right);\theta \right)$$
(5)

Therefore, this paper applies ERNIE 3.0 to conduct the task of classifying the intensity of texts for policy objectives and instruments through prompt learning. ERNIE 3.0 (https://wenxin.baidu.com/wenxin/modelbasedetail/ernie3/) is a large-scale knowledge enhanced pre-training model for language understanding, which has better performance on few-shot learning25. Having trained the model with 3334 labelled policy texts, we conduct a model evaluation to evaluate the training result. Then two strategies of data augmentation (TrustAI) are used to deal with data sparsity and class-imbalance. Then, models with the best accuracy are selected and used to predict the intensity of policy objectives and instruments for the remaining unlabelled policy texts.

Low-carbon policy intensity integration

After receiving all the intensity of policy objectives and instruments through manual labelling and prompt learning, the low-carbon policy intensity of each policy is multiplied by Eq. (1). Descriptive statistics of the low-carbon policy intensity are provided in the Supplementary Information. Figure 3 further shows the heterogeneity of low-carbon policy intensity from the perspective of policy level, time, and space. Figure 3a,b indicate strong cyclicality for national-level low-carbon policy intensity, which had stronger intensity in 2011 and 2016 (i.e., the first year of China’s 12th and 13th Five-Year Plan). Policies for energy conservation had a better continuity, while carbon reduction started to have independent policies in 2010 after China first proposed the carbon emission reduction target in 2009. Although command-and-control policies dominated the intensity of national low-carbon policies before 2015, market-based and composite instruments made greater contributions after 2015. The yearly average low-carbon policy intensity at the provincial- and prefecture-level in Fig. 3c,d show that prefecture-level policy intensity was relatively lower compared with the provincial-level counterpart. Shanghai (east), Shandong Province (east), and Shaanxi Province (northwest) had the largest average provincial low-carbon policy intensity. Among 334 prefecture-level cities, 21 cities’ prefecture-level policy intensities were in the highest level, including 11 cities in the east of China (e.g., Wuxi, Jiangsu Province), 3 cities in the north of China (e.g., Shijiazhuang, Hebei Province), 3 cities in the central of China (e.g., Luoyang, Henan Province), and other 4 cities from southwest, northwest, and south of China.

Fig. 3
figure 3

Heterogeneous low-carbon policy intensity from the perspective of policy level, time and, space. (a) and (b) show the national-level policy intensity divided by policy objectives and instruments. (c,d) are yearly average low-carbon policy intensity at the provincial- and prefecture-level.

Full size image

Data Records

Not only the low-carbon policy inventory and intensity for each policy but also the aggregated low-carbon policy intensity for national-, provincial- and prefecture-level are organized in the dataset. We provide two different format datasets for multidiscipline researchers. The first format of the dataset is stored in Stata files (.dta), while the second format is in Excel files (.xlsx). Lexicons for policy classification and definitions of variables are also attached to the dataset. Detailed information for this dataset is listed in Table 5. These files are all accessible via figshare26.

Table 5 Detailed information for data files.
Full size table

The link of each policy has been provided in folder “(For each policy) low-carbon policy inventory and intensity” of the dataset, which could locate to policy texts in the Laws & Regulations of PKULAW Database. Full policy texts in this database are only available for subscribers. Hence, researchers need to make sure that the academic institutions or individuals have subscribed to this database if you want to reproduce the process based on this paper.

Technical Validation

Model performance

In this paper, we use prompt learning to train models and predict the intensity for each objective and instrument. Among 7282 policies in our low-carbon policy inventory, 3334 policies with labels are used to train and evaluate the prompt learning model. 80% of the labelled dataset is used for training, while the remaining 20% is for validation. Two data augmentation strategies (i.e., sparse data identification and training data augmentation) are applied to improve prompt learning models by dealing with data sparsity and class-imbalance. Table 6 summarizes the accuracy of each model, which reflects the ratio of how many predicted labels are equal to actual labels in the sample. A detailed explanation of accuracy is provided in the Supplementary Information. Table 6 shows that the policy objective for carbon reduction has the best performance through “Prompt learning + Sparse data identification”, while “Prompt learning + Training data augmentation” is more suitable for the remaining three policy objectives and two policy instruments.

Table 6 Accuracy of prompt learning models.
Full size table

Comparison with existing studies

Due to the absence of a publicly available dataset for low-carbon policy intensity with three administrative levels, we cannot directly compare with existing datasets. However, there are present studies constructing environmental policy indices for China. On the basis of the “policy objectives-policy measures” pattern, Zhang et al. constructed China’s national-level environmental policy intensity for 1978–2019 by policy text analysis and machine learning algorithms5. OECD built a country-specific and internationally comparable environmental policy stringency index (EPS) on the basis of second-hand data for 13 policy instruments, which covers 40 countries from 1990 to 202010,11. These two policy indices have subcategories or sub-indices (e.g., energy conservation and emission reduction, industrial upgrading, market-based policies, non-market-based policies, technology support), which are closely related to low-carbon policies. Hence, the trend similarity among policy indices mentioned above could be used to validate the low-carbon policy intensity index constructed in this paper.

Following Zhang et al., we apply dynamic time warping (DTW) to measure the similarity between two series that vary in trend. Having normalized three indices, three pairs of DTW distance are calculated by Euclidean distance, and paths minimizing DTW distance between three national indices are shown in Fig. 4. The distance between this paper and Zhang et al. is 2.53, being the smallest distance among three pairs. Its DTW minimum paths in the white line are small and almost parallel with the main diagonal. This shows that there is a highest degree of trend similarity between national-level indices of this paper and Zhang et al.’s environmental policy intensity, both of which are based on policy text analysis but with different training methodologies. That is to say, the low-carbon policy intensity built in this paper is able to reflect the changing trend of policy intensity.

Fig. 4
figure 4

Path diagram with DTW distance between our national-level low-carbon policy intensity and existing studies. The horizontal axis of (a,b) are national-level low-carbon policy intensity in this paper. The vertical axis in (a) and (b) respectively represent environmental policy intensity from Zhang et al. and environmental policy stringency index from OECD. (c) compares indices between Zhang et al. and OECD.

Full size image

Manual checking

To further verify the accuracy of prompt learning prediction, we randomly select 500 policies through stratified random sampling based on policy objectives. Hence, each type of policy objective contains 125 policies to be manually labelled. From the perspective of policy instruments, these policies include 169 command-and-control policies, 276 market-based policies and 55 composite policies. Because composite policies are scored on the basis of two instruments, 224 command-and-control policies and 331 market-based policies need to be manually labelled. Table 7 presents the accuracy between results of prompt learning prediction and manual labelling, ranging from 0.85 to 0.93.

Table 7 Manual checking for prompt learning results.
Full size table

Usage Notes

Usage of low-carbon policy intensity

For the usage of this policy intensity index, the understanding of policy intensity needs to change from “policy-maker” to “policy-taker”. Hence, the meaning of \(PI\_al{l}_{r,t}\) is the total low-carbon policy intensity received by region r in year t, which are \(PI\_al{l}_{c,t}\) for city c in year t and \(PI\_al{l}_{\rho ,t}\) for province ρ in year t. It is known that China is a centralized nation with the central government at the top and provincial, prefecture governments below it27,28. Under the planning- and goal-based governance, the central government usually disaggregates low-carbon targets among provinces, while provincial governments assign targets to prefecture-level cities through agreements and performance measures to achieve their targets. Hence, China has a multi-layered policy structure29, and governments not only implement policies from their own levels but also receive policies from higher administrative levels. To be more specific, when the research object is the performance of prefecture-level cities or firms, the received total low-carbon policy intensity \(PI\_al{l}_{c,t}\) in Eq. (6) contains not only the intensity of prefecture-level policies but also those above-city policies (i.e., national- and provincial-level policies from which province it affiliates). If the study focuses on the provincial level, the total policy intensity \(PI\_al{l}_{\rho ,t}\) in Eq. (7) received by provincial policy-takers includes both national and provincial policies.

Hence, the impact of policy intensity on city-level or firm-level policy-takers needs to be based on \(PI\_al{l}_{c,t}\). If studies want to further discuss the impact of the policy intensity of different policy levels on carbon performances, the city-level intensity (\(PI\_{C}_{c,t}\)) and the above-city intensity \(PI\_aboveCit{y}_{c,t}\left(=PI\_{P}_{\rho ,t}+PI\_{N}_{t}\right)\) in Eq. (6) can be added to the regression at the same time.

$$\begin{array}{ccc}PI\_al{l}_{c,t} & = & PI\_{C}_{c,t}+PI\_aboveCit{y}_{c,t}\\ & = & PI\_{C}_{c,t}+PI\_{P}_{\rho ,t}+PI\_{N}_{t}\end{array}$$
(6)
$$PI\_al{l}_{\rho ,t}=PI\_{P}_{\rho ,t}+PI\_{N}_{t}$$
(7)

The starting and ending years of each policy

For user convenience, this dataset provides policies’ starting and ending years. Unlike the certain starting year of each policy, ending years of most policies in China are not clearly stated. Hence, policies’ end years have been verified in four criteria. First, the ending year of a few policies has been mentioned in the policy text (e.g., the planning period is 2012–2015, valid for 5 years, valid until 2021). Then, some policies have been abolished or replaced by new policies. Moreover, for policies without a clear ending year or abolishment, the policy end-year is verified on the basis of policy objectives. For example, although a policy was released in 2007 and hasn’t been abolished yet, expressions such as “at the end of 11th five-year plan period” and “until 2010” could be seen in the section of policy objective. Hence, the ending year of this policy is verified to be 2010. Lastly, for policies that cannot be confirmed using criteria mentioned above, the ending years of those policies are marked as “-” for uncertainty.

Limitations and future work

This study has some possible limitations. On the one hand, limited by the availability of policy texts, this paper constructs the low-carbon policy intensity primarily based on policies that are publicly available. Hence, the intensity of those policy texts that have not been publicly released or only partly released in public cannot be fully evaluated. On the other hand, the release of policy texts and the execution from relevant departments are key factors affecting low-carbon policy effects. This paper mainly focuses on policy texts from the perspective of policy release, but pays less attention to policy execution. Future studies can further investigate the gap between the intensity of released low-carbon policy texts and strength of execution from different policy levels, which are likely to influence the progress of carbon reduction and low-carbon transformation.