Does Public Environmental Affect Influence the World’s Largest Electric Vehicle Market? A Big Data Analytics Study of China

Abstract

The transition from combustion vehicles to electric vehicles (EVs) is critical for mitigating climate change. As the global leader in the EV market, China has been propelled by government policies, market dynamics, and public awareness. As sentiment is fundamental to human communication, however, existing research lacks a systematic examination of the extent to which public environmental affect influences EV adoption at the macro level, particularly in the presence of government interventions and market strategies. To address this gap, we construct a novel affect index using a CNN deep learning model to extract environmental affect (positive, negative, and neutral) from Weibo posts between 2014 and 2023. Employing monthly EV market share as the dependent variable, we incorporate affect indices as key independent variables, alongside control variables such as government subsidy reductions, price levels, and other marketing strategies. A time-series cointegration model is applied to assess the long-term impact of environmental affect on EV sales. Empirical results reveal that only positive environmental affect has a significant and positive impact on EV adoption, whereas key industry factors, including subsidies, charger availability, patent activity, and price disparities, also play crucial roles. These findings highlight the growing influence of public awareness in shaping the EV market transition from government-driven to market-driven growth. Our study reconciles conflicting findings in prior research and provides actionable insights for policymakers and marketers seeking to foster sustainable EV adoption.

1. Introduction

Transportation has long contributed to fossil energy consumption and greenhouse gas emissions, and it has become one of the fastest-growing sources of greenhouse gas emissions in China in recent years [1]. Globally, the transportation sector faces significant carbon emission challenges. In regions such as the United States and Europe, this sector is a major contributor to CO₂ and overall greenhouse gas emissions. In the United States, transportation accounts for a substantial share of carbon emissions, and its reliance on petroleum makes reducing emissions from this sector a critical challenge in addressing climate change. Similarly, Europe faces pressure to reduce emissions but has distinct needs and strategies for decarbonizing transportation due to differences in geography and economic structure [2,3]. Against the backdrop of promoting global energy conservation and emission reduction, the popularization of electric vehicles (EVs) integrates elements such as green energy consumption, intelligent travel, and innovative mobile internet platforms, while catalyzing industrial convergence through cutting-edge technologies, advanced materials, and renewable energy. Among EVs, fuel cell vehicles (FCVs) have emerged as a promising alternative, offering advantages such as zero emissions, long driving ranges, and rapid refueling times compared to traditional battery electric vehicles [4,5]. Thus, EVs have become a pivotal instrument for carbon decoupling efforts worldwide [6]. A transition from combustion to EVs not only has the potential to significantly reduce climate change impacts but has also been recognized as an inevitable choice for realizing structural upgrading and sustainable development of China’s auto industry.

Therefore, the Chinese government has introduced various guiding policies to promote the development of EVs. Since 2010, the Chinese government has launched a series of policies aimed at accelerating EV adoption, particularly emphasizing preferential road access rights and financial subsidies [7,8]. Internationally, many countries have implemented policies to promote electric vehicles (EVs). The United States has enforced strict emission standards and offered tax incentives to encourage EV adoption, with some states setting clear sales targets. European countries have supported EVs through strict CO₂ regulations, consumer subsidies, and investments in charging infrastructure [3]. Among these measures, tax exemptions on vehicle purchases for EVs have attracted increased academic attention, with empirical studies widely verifying their positive impact on consumer purchases [9,10]. Consequently, China has achieved remarkable growth in EV production, sales, and stock. According to official data released by the MIIT of China, EV production and sales volumes reached 12.89 million and 12.87 million, respectively, in 2024, with EV sales accounting for 40.9% of total automotive sales and maintaining China’s position as the global EV leader for ten consecutive years. Up to now, China’s EV industry has achieved ‘one breakthrough and three upgrades’, being the first to break through the 10 million EV sales milestone globally. Besides that, product performance, industrial system maturity, and user accessibility were all upgraded with measurable improvements.

To avoid low-quality, indiscriminate industry expansion and to transition EV development toward market-driven mechanisms, China has gradually phased out national and local subsidy policies since 2016. The immediate impact of subsidy reductions on EV market dynamics was evident, leading to intensified market competition, declining profit margins for battery manufacturers and automakers, and the accelerated elimination of lower-tier enterprises. As the development of China’s EV industry has transitioned from policy-led to market-oriented, it is necessary to understand consumer decision-making patterns regarding low-carbon consumption behaviors.

Recent academic attention has increasingly focused on low-carbon consumption behaviors in China, and governmental agencies have also carried out extensive communication and education campaigns to promote sustainable consumption practices. Despite these efforts, empirical evidence demonstrates suboptimal effects on actual carbon-reduction behaviors [11]. As sentiment is fundamental to human communication, affective experiences generated alongside cognition fundamentally affect subsequent behavior.

Existing research predominantly examines correlations between environmental awareness and low-carbon consumption behaviors by employing cross-sectional data at the micro-individual level. Gu and You applied the theory of planned behavior, using questionnaires to analyze consumers’ psychological attributions towards the purchase of new energy hybrid vehicles, emphasizing how individual-level factors affect purchase intention and behavior. Jesus et al. collected data from European electric vehicle drivers, exploring the factors influencing drivers’ satisfaction and continuance intention from the perspectives of task-technology fit and green self-identity [10,12].

In contrast, this study innovatively investigates whether environmental affect (EA) influences sales of EVs at the macro level, controlling for policy interventions and major market strategies, aiming to unravel the underlying unique mechanisms driving the world’s largest EV market.

The remainder of the paper is structured as follows. We first review the relevant literature on environmental affect, as well as its mechanisms, in the context of sustainable behavior. Following that, we describe our data collection and the methodological approach used to investigate the underlying research question. We then innovatively employ a CNN deep learning model to extract the topics and associated sentiments in Weibo text from 2014 to 2023. Using the monthly market share of EVs as the dependent variable, the three types of EA indices (positive, negative, and neutral) as independent variables, and adding government subsidy reductions, price, and other marketing strategy indices as control variables, we apply the time-series cointegration model to evaluate the long-term effects of environmental emotions. The results show that the direct association of environmental affect with EV sales is found for only one of the three sentiment variables: the Positive Environmental Affect Index. Followed by a general discussion of the key findings, the last part also includes theoretical contributions, managerial implications, and future research directions.

2. Literature Review

2.1. Definition and Development of Environmental Affect

Environmental affect refers to individuals’ transient, valenced affective responses (e.g., love, worry, or guilt) toward environmental issues or behaviors, distinct from enduring attitudes, which are stable evaluations rooted in beliefs and cognition [13]. Unlike environmental attitudes, which reflect stable, evaluative beliefs about the environment, environmental affect captures immediate, visceral emotional responses that can independently shape behavior. This distinction is critical, as emotions function as “cognitive filters” that bias perception and decision-making in ways that rational evaluations alone may not predict [14].

Systematic research on the independent role of environmental affect in shaping pro-environmental behaviors began in the 1970s. While early studies often treated it as a dimension of environmental attitude or ecological consciousness, later work has established it as a distinct construct with unique behavioral effects. Schultz (2000) subsequently elaborated that environmental affect originates from individuals’ perceived connectedness to nature, thereby influencing both cognitive evaluations and affective responses [15]. Kanchanapibul et al. (2014) further examined the influence of environmental affect and knowledge on green purchasing behavior and demonstrated the reliability and validity of the measured variables [16]. While environmental affect was often treated as a dimension of environmental attitude or ecological consciousness in early research, it has been progressively constructed as a distinct variable and increasingly gained theoretical recognition of its distinct influence on behavioral outcomes.

Triandis’s (1977) Interpersonal Behavior Theory (TIB) is a foundational framework for understanding affect and how individuals make behavioral decisions [14]. IBT integrates cognitive, affective, and social–cultural factors to explain interpersonal dynamics, The affective component refers to emotional responses, acting as ‘cognitive filters’ and biasing perception and decision-making. Notably, emotional responses operate distinctly from rational evaluation and may manifest varying intensities of positive or negative valence.

2.2. Components and Effect of Environmental Affect

Emerging research has enriched the components of EA and demonstrated its role in consumer decision-making. Wang and Wang (2011) constructed the ‘Consciousness-Situation-Behavior Integrated Model’ using grounded theory, which can effectively explain how the public adopts low-carbon consumption behavior [17]. Low-carbon awareness is an internal driver for the motivation of low-carbon consumption, serving as a predisposing factor and promoting behavior by influencing the public’s psychological preference for low-carbon consumption behavior. Different sources of low-carbon psychological awareness affect the consistency between awareness and behavior.

Both positive and negative emotions in promoting pro-environmental behaviors have been documented. Kao et al. (2020) demonstrated through a quasi-experimental design that positive emotions exhibit strong advertising effectiveness in encouraging sustainable practices [18]. Mallett (2012) demonstrated the potential of eco-guilt to increase engagement in pro-environmental behavior [19]. White et al. (2019) further synthesized research on how both positive and negative emotions influence pro-environmental behaviors by considering information processing, eco-labeling, and framing mechanisms [20].

The relationship between emotion and behavior is more specific rather than being in a broad sense positive or negative. Harth et al. (2013) compared the in-group-focused emotions of pride, guilt, and anger as predictors of three environmental intentions [21]. They found that guilt predicted intentions to repair damage, anger predicted intentions to punish wrongdoers, and pride predicted intentions for in-group-favored environmental protection.

2.3. Environmental Affect in EV Adoption

In the context of EV adoption, environmental affect also plays a crucial role. A nationally representative U.S. survey (N = 2302) shows that fuel economy consistently ranks among the most important vehicle attributes. Notably, consumers who express early interest in adopting EVs are typically highly educated, environmentally sensitive, and in alignment with the “Environmental Views Index” [22]. Okada et al. (2019) conducted a large-scale survey (N = 246,642) in Japan, revealing that environmental awareness directly influences non-EV users’ purchase intentions while affecting EV owners’ post-purchase satisfaction [23]. Cohen et al. (2019) found that the most commonly researched determinant that researchers assume to have an impact on the EV market is environmental awareness, and its impact is frequently explained with the theory of planned behavior (TPB) established by Ajzen in 1991 [24].

Existing studies have proved the importance of environmental awareness including affect, but they generally focus on intention rather than actual adoption behavior [25]. Surveys are the most frequently used method for data collection, involving directly asking respondents questions. In conclusion, the predominant survey method, which focuses on the micro-individual level and primarily collects cross-sectional data, is exposed to certain limitations.

To understand the macroeconomic drivers of EV sales, Austmann and Vigne (2021) bridged the research gap and applied Twitter as an awareness barometer to investigate effect of environmental awareness on the electric vehicle market in 29 European countries covering a period of 8 years between 2010 and 2017 [26]. Contradictorily to previous micro-level studies, the results show that environmental awareness does not significantly influence the electric vehicle market. China undoubtedly takes the lead in the EV market with its unique features in terms of government regulations, market strategies, and public awareness. Extant research inadequately addresses whether environmental affect impacts China’s EV market in the macro-level and this remains to be further explored.

3. Data and Methodology

3.1. Data

Given that the large-scale adoption of EVs in China began after 2011 and that Sina Weibo launched its IPO in 2014, this study employs monthly data from January 2014 to June 2023, based on the availability of EV market share data alongside data on Weibo posts and active user counts.

Table 1. Variable definitions and data sources.

Variable Type	Variable Name	Variable Definitions	Data Sources
Independent variables	Pos_EA	Positive Environmental Affect Index: the ratio of positive sentiment posts to the number of monthly active Weibo users	Weibo platform, Weibo financial report
	Neg_EA	Negative Environmental Affect Index: the ratio of negative sentiment posts to the number of monthly active Weibo users
	Neu_EA	Neutral Environmental Affect Index: the ratio of neutral sentiment posts to the number of monthly active Weibo users
Dependent variable	Share	The sales volume of electric passenger vehicles purchased privately as a proportion of total passenger vehicle sales in China	National Passenger Vehicle Market Information Joint Meeting
Control variables	Price	The price of electric vehicles, using the manufacturer’s suggested retail price (MSRP) for BYD’s Qin series (Shenzhen, China), in ten thousand yuan	BYD Auto Official Website
	Disparity	The price difference between gasoline and residential electricity, in yuan	East Money Information, NDRC Official Website
	Patent	The monthly number of patent disclosures related to electric vehicles, in units	Baiten Patent
	Charger	The monthly increase in the number of public charging stations nationwide, in units	Charging Alliance
	Subsidy	Defined based on the subsidy policy, with four dummy variables used for classification	Official Website of MOF\MOST\MIIT\NDRC

summarizes the variables and data sources used in the empirical analysis.

3.1.1. Independent Variables

• Independent variable definitions

Environmental affect refers to an individual’s attitudinal response to whether environmental issues or behaviors meet personal needs [13] and serves as the independent variable in this study. Using deep learning techniques, the texts of Weibo posts were decomposed into parts with positive, negative, and neutral sentiments. Three indices were constructed based on the number of various types of posts and active Weibo users for each month: Positive Environmental Affect Index (Pos_EA), Negative Environmental Affect Index (Neg_EA), and Neutral Environmental Affect Index (Neu_EA). These indices represent the ratio of posts with the three emotional tendencies toward the environment to the monthly active users on Weibo.

Pos_EA reflects the extent to which users express positive sentiments toward environmental issues, including support, praise, and encouragement for topics such as environmental action, sustainable development, and nature conservation. In contrast, Neg_EA indicates the degree of negative sentiments expressed by users on environmental issues, such as criticism, concern, and dissatisfaction regarding environmental degradation, pollution, and harmful practices. Neu_EA reflects the extent of neutral sentiments, including objective reporting, neutral viewpoints, or balanced evaluations of environmental issues.

Table 2. Examples of each cognitive theme.

Types of Posts	Description and Examples
Positive sentiment posts	“Planted 10 trees today to green the planet and slow down climate change! #Green Earth #World Earth Day”
Negative sentiment posts	“It’s heartbreaking to see news reports again about melting glaciers and rising sea levels due to global warming! Will the world ever be good again”?
Neutral sentiment posts	“The latest research shows that the impact of climate change on agricultural production is complex and requires more attention and research”.

provides concrete examples of Weibo posts classified under each sentiment category (positive, negative, and neutral).

2.

Independent Data collection

A common method to assess public emotional responses to a given issue is through surveys of individual emotions. However, surveys are not the only approach to capturing public affect. With the emergence of advanced methods like machine learning, researchers increasingly employ neural network models to analyze the sentiments expressed in public opinions to measure emotional responses. In recent years, social media platforms such as Twitter, Weibo, Reddit, Xiaohongshu, and TikTok have garnered significant academic attention for exploring public opinion. Social media platforms have become a “barometer” of public sentiment in China. Particularly during significant public events, official media can effectively reach and influence a broad audience through platforms such as Weibo. The emotional dissemination via these platforms is representative across demographic characteristics [27]. As one of the most representative Chinese social media platforms, over the past decade, Weibo has become a central public space for social interaction due to its high-quality original content, strong interactivity, and rapid information dissemination. Compared to traditional survey methods, Weibo offers a large dataset that can serve as a valuable supplement to conventional approaches. As a tool for measuring public opinion and sentiment, social media is gaining popularity in sustainability research.

This study employs a well-established and open-source Python program (3.9.1) (Weibo-search) to obtain a total of 723,864 publicly available Weibo posts from January 2014 to June 2023, leveraging the Weibo public Application Programming Interface (API) as the data source.

3.1.2. Dependent Variable

The dependent variable, Share, represents the market share of electric passenger vehicles (EPVs) in China, defined as the proportion of monthly EPV sales for private purchase relative to the total passenger vehicle sales during the same period. The monthly sales include only EPVs purchased privately, excluding those procured by government or public institutions. Market share, rather than sales volume, is chosen as the dependent variable because purchasing decisions for public procurement are largely policy-driven rather than influenced by individual factors, unlike private consumer behavior. Furthermore, prior studies frequently employ market share instead of sales volume, as using market share controls for external shocks—such as economic growth, rising household income, and automotive industry cycles—that can affect total vehicle sales (including fuel vehicles), thereby providing a more accurate measure of EV demand changes [26,28,29].

3.1.3. Control Variables

As the focal variable is EA, other factors influencing EV sales representing government regulation and market strategies are treated as control variables, which include the price of EVs (Price), battery cost (Battery), price disparity between gasoline and electricity (Disparity), monthly patent disclosures related to EVs (Patent), monthly increase in the number of charging stations nationwide (Charger), and the subsidy reduction policy (Subsidy).

Price represents the price of EVs, using the manufacturer’s suggested price for BYD’s Qin series as an indicator. These data cover the study period from January 2014 to June 2023, reflecting the overall price trends in the EV market. The Qin series serves as a standardized reference due to its consistent market positioning, helping control for price differences across vehicle models. Although actual transaction prices would be ideal as they incorporate promotional discounts and other adjustments, these data are challenging to obtain. Thus, this research follows previous studies in using manufacturer’s suggested price, which remains stable across markets [30,31,32].

Disparity captures the price gap between gasoline and electricity, calculated as the per-liter price of gasoline minus the per-kWh price of residential electricity. This measure reflects changes in the relative cost advantage of EVs in terms of operating expenses. Patent represents the monthly volume of patent disclosures related to EVs, serving as a proxy for the overall technological level of China’s EV industry. Charger measures the monthly increase in the number of charging stations nationwide, capturing the convenience of EV use.

Subsidy refers to the national subsidy reduction policy for EVs. This research period (January 2014 to June 2023) is divided into four phases based on the extent of subsidy reductions, represented by four dummy variables: Subsidy1, Subsidy2, Subsidy3, and Subsidy4. The correspondence between policy periods and dummy variables (Subsidy1–Subsidy4) is summarized in

Table 3. Policy dummy variables.

Time Period	Subsidy1	Subsidy2	Subsidy3	Subsidy4
January 2014–December 2016	1	0	0	0
January 2017–December 2018	0	1	0	0
January 2019–December 2021	0	0	1	0
January 2021–June 2023	0	0	0	1

. These variables are later used in the regression analysis to isolate the impact of subsidy reductions across phases.

Period I (January 2014–December 2016) is the subsidy refinement phase, during which the subsidy standards for EVs were first differentiated based on driving range and adjusted annually, with the scope of subsidies gradually expanded. The initial subsidy policy for private EV purchases was in place in March 2010, providing a one-time subsidy for private EV buyers in five pilot cities. Subsidy standards were determined based on battery capacity, with funds allocated to vehicle manufacturers rather than directly to consumers, allowing manufacturers to offer a discounted price depending on the subsidy.

Period II (January 2017–December 2018) is the initial dual-drive phase, characterized by further reductions in subsidy amounts and an increase in subsidy eligibility thresholds. In December 2016, four related government ministries (the National Development and Reform Commission (NDRC) of China, the Ministry of Industry and Information Technology (MIIT) of China, the Ministry of Science and Technology (MOST) of China, and the Ministry of Natural Resources (MNR) of China) jointly issued a notice on adjusting the financial subsidy policy for EV promotion, which raised subsidy thresholds, refined subsidy standards, improved fund allocation mechanisms, and set upper limits for central and local subsidies. Starting in 2017, the subsidy standards became more stringent, with additional criteria for battery energy density and electricity consumption per 100 kilometers, while the highest subsidy level was reduced by 20% from 2016.

Period III (January 2019–December 2021) is considered the mature dual-drive phase, during which the government further reduced subsidy levels and allowed foreign car manufacturers to enter the EV market, contributing to a more market-oriented landscape. In May 2018, the four government ministries (NDRC, MIIT, MOST, and MNR) jointly released a notice on adjusting and improving the EV subsidy policy, emphasizing the need to dismantle local protectionism, create a unified market, and eliminate local government purchase subsidies for EVs. Fiscal funds were redirected to support charging infrastructure development and associated services, with purchase subsidies for electric passenger vehicles discontinued.

Period IV (January 2021–June 2023) represents the market-driven phase, with continued subsidy reductions and an increasingly market-oriented environment. In December 2020, the four government ministries (NDRC, MIIT, MOST, and MNR) jointly issued a notice on improving the EV subsidy policy, stipulating that from 2021, subsidy standards would be reduced by 20% from 2020 levels. Unlike previous subsidy policies, this policy did not include a transitional period, as the final reduction phase was widely accepted within the EV industry.

3.2. Time-Series Cointegration Model

Given the longitudinal structure of our data, we carry out natural logarithmic processing on each time series and apply time-series cointegration model to evaluate the long-term effects of environmental emotions [28]. As Formula (1) shows, the monthly market share of the EVs is taken as the dependent variable, the three kinds of environmental emotions index (positive, negative, and neutral) as the independent variables, adding EV price, government subsidy, etc., as the control variables, where t represents the current period and t − 1 denotes the previous period in the time series. All independent, dependent, and control variables are elaborated in Table 1 of Section 3.1. The natural logarithm (ln) is applied to variables due to their varying magnitudes, aiming to simplify the scale. ε_t is the error term.

$$\begin{gathered} ln\_{Share}_t=\beta ₀+\beta ₁ln\_{Pos\_EA}_{t-1}+\beta ₂ln\_{Neg\_EA}_{t-1}+\beta ₃ln\_{Neu\_EA}_{t-1}+\beta ₄ln\_{Disparity}_t \\ +\beta ₅ln\_{Charger}_t+{\beta }_{6}ln\_{Patent}_t+{\beta }_{7}Subsidy+{\beta }_{8}Season+{\epsilon }_t \end{gathered}$$

(1)

Since car sales often exhibit seasonal fluctuations, seasonal dummy variables are used to control for cyclical patterns within the year, representing each quarter [26]. Additionally, to account for the lagged effects of social media information dissemination and emotional states on actual behavior, lagged values of the independent variables are included in the regression model [28]. This approach allows the regression model to more accurately capture these time-lagged effects, thereby improving its predictive capacity for the dependent variable.

A time-series cointegration model is favorable for time-series analysis. Cointegration occurs when two or more non-stationary time series are integrated together in such a way that there exists a long-run equilibrium relationship among them. Cointegration tests are used to determine whether there are long-term equilibrium relationships among these time series.

4. Sentiment Analysis of Public Environment Affect

We trained multiple models using a well-established public dataset along with an additional manually annotated dataset and selected the optimal model. This model enabled us to identify environmental affect from a total of 723,864 valid entries. Additionally, we filtered out 216,707 irrelevant entries to prevent any impact on subsequent time-series cointegration regression analysis. The detailed process information is provided next.

4.1. Data Preprocessing

The text preprocessing stage involved preparing the raw Weibo text data by cleaning irrelevant entries, removing duplicates, and conducting initial normalization. This step enhanced data quality, ensuring the accuracy and effectiveness of subsequent model training and application. The text preprocessing in this study was conducted through the following steps:

(1)

Data Cleaning: The Python regular expression module (re) was used to initially clean the text by removing irrelevant emojis (e.g., “\u200b”), special characters (e.g., “@” and “#”), URL links, and redundant punctuation. This initial cleaning step effectively reduced noise within the dataset, enhancing the efficiency of subsequent data processing.

(2)

Tokenization: The jieba segmentation tool was applied to segment the cleaned Weibo text. Known for its high accuracy and efficiency in Chinese text segmentation, jieba effectively handles the complexities of Chinese text.

(3)

Stopword Removal: A stopword list was constructed based on the Modern Chinese Word List by the Institute of Computational Linguistics at Peking University. This list includes common function words such as prepositions, conjunctions, and articles. Removing these frequently appearing but sentiment-neutral words reduces noise during model training and improves the precision of subsequent analysis.

(4)

Word Vector Conversion: Pre-trained Chinese Word Vectors were used to convert tokenized words into numerical vectors. Developed by the Natural Language Processing and Computational Social Science Lab at Tsinghua University, this model was trained on a large-scale Chinese corpus and captures semantic relationships between words effectively.

4.2. Corpus Construction

The quality of the corpus directly affects the performance of the model.

(1)

Base Corpus Selection: We selected a publicly available and academically validated Weibo sentiment corpus as the foundation, primarily used for training the model to identify positive and negative environmental affect. The initial corpus used was the Chinese Weibo Sentiment Analysis dataset (FudanNLP Weibo Sentiment Corpus) released by the Natural Language Processing Lab at Fudan University.

(2)

Supplementary Corpus Annotation: In response to the research needs, we specifically supplemented the corpus with neutral and irrelevant samples. A cross-validation annotation process was adopted, where each entry was independently annotated by two trained graduate students. In cases of inconsistent annotation results, a panel of three experts was consulted to ensure the consistency and reliability of the annotations.

(3)

Corpus Integration: The corpus extracted from the established dataset was combined with the manual coding corpus to form the final training dataset. The integrated dataset includes four categories: positive, negative, neutral, and irrelevant.

(4)

Data Balancing: To address the issue of class imbalance in the dataset and prevent the influence of label distribution disparities on model training, we applied data augmentation techniques, such as synonym replacement, back-translation, and random insertion, to balance the dataset’s label distribution.

4.3. Model Training and Application

During the model training phase, 70% of the corpus was used as the training set, while the remaining data were split into validation and test sets to ensure sufficient samples for training and adequate data for validation and final testing. The initial model was trained, and its performance was optimized by adjusting the model architecture and hyperparameters (e.g., learning rate, batch size, and dropout rate). To prevent overfitting, we also employed Early Stopping for the deep learning models. The results were as follows: the accuracy of Logistic Regression (LR) was 69.92%, Support Vector Machine (SVM) achieved 70.23%, Convolutional Neural Network (CNN) reached 75.31%, and Recurrent Neural Network (RNN) attained 72.07%. The model with the highest accuracy, CNN, was ultimately chosen for this study. As shown in Figure 1, the CNN model was selected for this study not only because it achieved the highest accuracy but also due to its inherent advantages in processing short-text sentiment analysis. CNNs excel at capturing local patterns and hierarchical features through convolutional layers, which is particularly effective for identifying sentiment-bearing phrases or keywords in short texts (e.g., Weibo posts). Additionally, CNNs are less prone to the vanishing gradient problem compared to RNNs and can efficiently handle spatial invariance in text data through weight sharing and pooling operations. These characteristics make CNNs a robust choice for sentiment analysis tasks where contextual nuances and localized sentiment cues are critical.

In conclusion, this study overcame the challenges posed by the diversity of internet short-text contexts in traditional sentiment analysis by employing supervised deep learning neural networks. The model successfully identified the sentiment attributes of 723,864 Weibo posts, filtered out 216,707 irrelevant entries (e.g., “Before and after the beginning of summer, the climate changes significantly, with large temperature differences between morning and evening, which makes people prone to illness…”), and then calculated the proportion of each sentiment category’s posts relative to the number of active Weibo users in the corresponding month. This led to the construction of the environmental affect indices for this study.

To ensure the scientific rigor and robustness of the constructed indices, both the validity and reliability of the sentiment classification process were carefully considered. The model was trained using a publicly validated corpus—the Weibo Sentiment Corpus from Fudan University—and was further supplemented with manually annotated neutral and irrelevant texts to align with the specific research context. Dual annotations with expert arbitration were adopted to enhance labeling consistency. After comparing multiple model performances, Convolutional Neural Network (CNN) was selected for its highest classification accuracy (75.31%). In addition, robustness tests using an alternative dependent variable produced consistent results, further confirming the internal consistency and empirical reliability of the environmental affect indices.

5. Empirical Results and Analysis

5.1. Descriptive Statistics

As shown in

Table 4. Descriptive statistics of major variables.

Variable Type	Variable Name	Observations	Mean	Standard Deviation	Minimum	Maximum
Independent	Pos_EA	114	468.332	399.752	27.019	2312.284
	Neg_EA	114	289.690	272.777	25.648	1109.075
	Neu_EA	114	150.399	125.168	9.408	695.462
Dependent	Share	114	0.077	0.098	0.001	0.332
Control	Price	114	16.564	5.260	9.980	25.980
	Disparity	114	6.091	0.914	4.240	8.532
	Patent	114	62.105	34.609	14.000	228.000
	Charger	114	53.476	57.693	2.400	214.900

, the independent variables Pos_EA (468.332), Neg_EA (289.690), and Neu_EA (150.399) exhibit substantial variability, as reflected by their wide ranges and large standard deviations. This variability highlights the heterogeneity of the sample data and underlines the need for robust statistical models.

For the dependent variable, Share, the mean is 0.077, with a standard deviation of 0.098. The minimum value is 0.001, and the maximum value is 0.332, reflecting a broad range of market share values within the sample, along with a certain degree of variability.

The mean values for the control variables Price, Disparity, Patent, and Charger are 16.564, 6.091, 62.105, and 53.476, respectively. The ranges and standard deviations for these control variables also vary, indicating notable differences in the observations across these variables within the sample.

5.2. Time-Series Cointegration Regression Results

We applied a time-series cointegration model to evaluate the long-term effects of environmental emotions on new EV sales. First, we used a unit root test to determine whether there were stable cointegration relationships among these time series and then performed OLS regression with stationary time-series data.

5.2.1. Unit Root Test

Non-stationary time-series data can result in spurious regression, where apparent statistical significance masks the absence of genuine relationships among variables. To address this, unit root tests were performed, ensuring that the series met stationarity requirements prior to further analysis. First, unit root tests were conducted for variables in Formula (1), i.e., original time series.

Based on the results indicated in

Table 5. Unit root test results for original series.

Variable	Z(t) Statistic	1% Critical Value	p-Value
ln_Share	−2.156	−3.506	0.223
ln_Pos_EA	−4.227	−3.506	0.001
ln_Neg_EA	−3.089	−3.506	0.027
ln_Neu_EA	−3.594	−3.506	0.006
ln_Disparity	−2.377	−3.506	0.149
ln_Charger	−1.228	−3.506	0.662
ln_Patent	−3.709	−3.506	0.004
ln_Price	−0.419	−3.506	0.907

, the following variables are determined to be stationary series at the 1% significance level: ln_Pos_EA, ln_Neg_EA, ln_Neu_EA, and ln_Patent. Conversely, the Z(t) statistics for ln_Share, ln_Disparity, ln_Charger, and ln_Price were all greater than the 1% critical value of −3.506, and their p-values were significantly greater than the 0.01 significance level. Therefore, the null hypothesis that these series were non-stationary could not be rejected, indicating that they were likely non-stationary series. Non-stationary series need to further undergo differencing to achieve stationarity for further analysis.

From

Table 6. Unit root test results for first-differenced series.

Variable	Z(t) Statistic	1% Critical Value	p-Value
ln_Share_diff	−13.063	−3.506	0.000
ln_Pos_EA_diff	−14.232	−3.506	0.000
ln_Neg_EA_diff	−15.808	−3.506	0.000
ln_Neu_EA_diff	−16.114	−3.506	0.000
ln_Disparity_diff	−8.675	−3.506	0.000
ln_Charger_diff	−7.661	−3.506	0.000
ln_Patent_diff	−16.560	−3.506	0.000
ln_Price_diff	−10.607	−3.506	0.000

, it is evident that after taking the first differences of all variables, their Z(t) statistics are significantly less than the 1% critical value, with corresponding p-values of 0.000. This indicated that at the 1% significance level, all first-differenced series were stationary and did not have unit roots, confirming the absence of a unit root problem.

5.2.2. OLS Regression

A multicollinearity test was performed first, and the results indicated that there was no obvious problem of high multicollinearity among these given variables. This implied that the correlations among the variables were relatively low. Subsequently, OLS regression was conducted, and the residual series were examined using Stata 16 with $ln\_{Share}_t$ as the dependent variable. The results of cointegration regression are presented in

Table 7. Time-series cointegration regression results.

Variable	Coefficient	Standard Error	t-Value	p-Value
l1_ln_Pos_EA	0.196	0.096	2.040	0.044
l1_ln_Neg_EA	−0.114	0.102	−1.110	0.268
l1_ln_Neu_EA	−0.202	0.116	−1.740	0.085
ln_Disparity	1.174	0.411	2.860	0.005
ln_Charger	1.185	0.111	10.710	0.000
ln_Patent	0.367	0.106	3.480	0.001
ln_Price	−1.347	0.343	−3.930	0.000
q2	0.338	0.104	3.250	0.002
q3	0.317	0.109	2.900	0.005
q4	0.427	0.113	3.770	0.000
Subsidy2	−0.652	0.202	−3.240	0.002
Subsidy3	−1.311	0.352	−3.720	0.000
Subsidy4	−1.423	0.481	−2.960	0.004
R2	0.950		Adjusted R2	0.943

.

The model estimation results showed that the R² and Adjusted R² are 0.950 and 0.943, respectively, indicating a good fit between the regression equation and the actual trends of the variables. The unit root test was continued on residual series, and the results are shown in

Table 8. Unit root test results for residual series.

Variable		Z(t) Statistic	p-Value
Residuals		−7.534	0.000
Critical Value	1% level	−3.506
	5% level	−2.889
	10% level	−2.579
Residuals		−7.534	0.000

.

The test results show that the t-value of the cointegration regression model is smaller than the 1% critical value, indicating that the residuals do not have a unit root. This confirms that the residual series is stationary, and there is a long-term stable relationship between the dependent and independent variables in the cointegration regression model, i.e., the variables are cointegrated.

5.2.3. Analysis of Cointegration Regression Results

The resulting beta coefficients, Standard Error, t-statistics and p-Value are reported in Table 7.

Environmental Affect Indices. The beta coefficients for two of the sentiment variables, Negative and Neutral Index, were negative but not significant. However, the beta coefficient for the Positive Index, the ratio of positive sentiment posts to the number of monthly active Weibo users, was substantial (0.196) and statistically significant (p < 0.05). Hence, a direct association of environmental affect with EV sales was found for only one of the three sentiment variables, namely, the Positive Environmental Affect Index.

The control variables influencing EV sales originate from three types and are listed in Table 7. The first control variable is subsidy, which represents the initial driving force from the Chinese government. The beta coefficients for subsidy reduction were all substantial (−0.65, −1.31, and −1.42, respectively) and significant (p < 0.001), which means that in the following three periods, subsidy reduction slowed down EV sales more in comparison to period I.

The other control variables are four market factors, disparity, charger, patent, and price. The beta coefficients were all significant (p < 0.005) and substantial (1.174, 1.185, 0.367, and −1.347, respectively), indicating that market incentives promoted EV sales while government subsidy kept reducing. The final control variables were related to seasonal fluctuations, and the beta coefficients for q2 to q4 were substantial (0.338, 0.317, and 0.427, respectively) and significant (p < 0.01) compared with the first quarter.

5.2.4. Robustness Test

To validate the robustness of the regression results, this study employed an alternative dependent variable, electric passenger vehicle sales, following the approach of Li and Liu (2022) [33]. The results in

Table 9. Robustness test results.

Variable	Coefficient	Standard Error	t-Value	p-Value
l1_ln_Pos_EA	0.2644951	0.1106919	2.39	0.019
l1_ln_Neg_EA	−0.1344074	0.1178351	−1.14	0.257
l1_ln_Neu_EA	−0.257	0.1340404	−1.92	0.058
ln_Disparity	1.1653	0.4740098	2.46	0.016
ln_Charger	1.2905	0.1275645	10.12	0.000
ln_Patent	0.3913	0.1218374	3.21	0.002
ln_Price	−1.113	0.3953113	−2.81	0.006
q2	0.3690	0.1199907	3.08	0.003
q3	0.4407	0.1260913	3.49	0.001
q4	0.7326	0.1306614	5.61	0.000
Subsidy2	−0.622	0.2323227	−2.68	0.009
Subsidy3	−1.445	0.405784	−3.56	0.001
Subsidy4	1.579	0.5540184	−2.85	0.005
R2	0.9369		Adjusted R2	0.9286

remained consistent, further supporting the reliability of the initial findings.

6. Discussion and Conclusions

This study aims to empirically examine whether environmental affect, triggered by environmental events discussed on social media over the past decade, has influenced purchase behavior in the world-leading EV market.

Contrary to the findings of Austmann and Vigne (2021) in Europe [26], we found that environmental affect is positive and significant in predicting EV sales in China for only one of the three sentiment variables, namely, the Positive Environmental Affect Index. This provides evidence of changes in public awareness over the past decade. Public support for green and sustainable consumption plays an important role from a macroeconomic perspective, which reveals a distinct regional pattern. While previous short-term research reported relationships for both positive and negative emotions, the long-term positive environmental affect is proved to be more directly linked to sustainable behavior than the negative and neutral types, reconciling prior studies.

A second conclusion is that EV-specific factors were discovered to be significant when checking control variables, namely, subsidy, disparity, charger, patent, and price. China has gradually phased out national and local subsidy policies since 2016. Subsidy reduction was all negative and significant, confirming that the subsequent three stages of subsidy reduction hindered EV sales and were conducive to bringing EV development back to market-driven mechanisms. Four market factors (disparity, charger, patent, and price) were all significant and substantial, corresponding with the findings of prior studies about determinants of EVs, such as Austmann and Vigne (2021) [26]. Our results further indicate that market forces promoted EV sales while government subsidies were being phased out. This also shows that the Chinese EV market itself was upgraded with measurable improvements during the transition from government-led to market-driven growth.

To accelerate this transition while sustaining growth, policymakers should strengthen positive environmental messaging through targeted social media campaigns that showcase EV benefits and real-world impact. The reallocation of subsidy savings toward critical infrastructure development, particularly the strategic expansion of fast-charging networks in urban hubs and along major transportation corridors, would address key adoption barriers. Simultaneously, establishing regional innovation hubs with tax incentives for technology commercialization could bridge the gap between patent development and market implementation. These measures would collectively reinforce the market-driven evolution of China’s EV sector while maintaining its global leadership position. The implementation of tiered electricity pricing for off-peak charging, coupled with public awareness initiatives demonstrating cost savings, would further incentivize consumer adoption. By aligning these policy interventions with the empirical findings on environmental affect and market forces, China can ensure a smooth transition from policy-led support to sustainable, market-driven growth in its EV industry.

The present study exhibits several limitations that warrant acknowledgment. First, certain constraints exist in data investigation. Although this research employs a decade-long longitudinal dataset, the measurement of environmental affect predominantly relies on social media text analysis, which may introduce sample bias. Specifically, Weibo users skew toward younger (aged 18–35), urban, and educated demographics compared to the general Chinese population, potentially underrepresenting older and rural consumers’ environmental attitudes. The generalizability of findings could be compromised by the representativeness of social media user demographics and the authenticity of posting behaviors. Furthermore, the environmental affect indices utilized in this study may not fully capture consumers’ latent emotional fluctuations, particularly manifesting limitations in measuring emotional intensity and temporal persistence.

Second, inadequacies emerge in variable control methodology. When analyzing the impact of environmental affect on purchasing behavior, this study potentially fails to comprehensively account for confounding effects from other influential factors. These include macroeconomic conditions, competitive brand strategies, and individual consumption patterns, among other variables. The presence of these unmeasured covariates may undermine the robustness of the research conclusions.

Building upon the identified limitations, future studies could address these gaps through methodological and theoretical advancements. First, researchers could enhance data robustness by integrating multi-source data collection approaches. Combining social media text analysis with survey-based measures, behavioral tracking (e.g., sensor data or purchase records), and qualitative interviews may mitigate sample bias and improve the granularity of environmental affect measurement. Advanced text mining techniques, such as sentiment-aware deep learning models, could further refine the assessment of emotional intensity and temporal dynamics. Longitudinal designs with higher-frequency sampling intervals or panel data tracking individual-level changes over time could also elucidate the persistence and evolution of environmental emotions.

Second, future work should prioritize rigorous variable control and causal inference frameworks. Employing quasi-experimental designs (e.g., natural experiments or instrumental variable approaches) could help disentangle the causal relationship between environmental affect and consumer behavior while accounting for confounding factors like macroeconomic shifts and competitive market dynamics. Multi-level modeling or machine learning methods that incorporate heterogeneous consumer characteristics (e.g., demographics and cultural values) and contextual variables (e.g., policy interventions and brand campaigns) would strengthen the generalizability and specificity of findings.

Critically, the current findings on the link between positive environmental affect and electric vehicle sales in China may not be generalized across diverse cultural and market contexts. To address this, cross-cultural comparative studies should be systematically integrated into future research. For example, parallel analyses in markets with distinct regulatory frameworks (e.g., the EU’s stringent emissions standards), cultural values (e.g., individualist vs. collectivist societies), or economic structures (e.g., mature vs. emerging EV markets) could reveal how environmental affect interacts with local factors to shape consumer behavior. Such comparisons would not only test the universality of the observed effects but also identify context-specific mechanisms—offering actionable insights for policymakers and industry stakeholders aiming to promote sustainable consumption globally.