Digitalization and Household Consumption-Based Carbon Emissions: Evidence from China’s CFPS Panel

Abstract

The rapid expansion of the digital economy is reshaping consumption patterns and carries significant implications for environmental sustainability and low-carbon transition. Against the backdrop of global carbon neutrality goals and the United Nations Sustainable Development Goals (SDGs), understanding how digitalization influences household consumption-based carbon emissions is considered critical for promoting sustainable consumption systems. Using five waves of balanced panel data from the China Family Panel Studies (CFPS) from 2014 to 2022, this study measures household carbon emissions based on the Consumer Lifestyle Approach (CLA) and constructs a multidimensional digitalization index covering digital access, usage, and cognition. Employing two-way fixed effects and chain mediation models, this study examines both the direct effects and underlying mechanisms linking digitalization and household carbon emissions. The results show that digitalization is significantly and positively associated with household carbon emissions, with notable heterogeneity across individuals, households, and regions. Mechanism analysis reveals that digitalization alleviates liquidity constraints and promotes consumption upgrading, forming a chain mediation pathway that drives emission growth. This study contributes to the literature by providing micro-level evidence on the environmental consequences of digitalization and by integrating behavioral mechanisms into the analysis of digital economy-environment interactions. It further offers policy insights for aligning digital transformation with low-carbon development and advancing sustainable consumption in the digital era.

1. Introduction

Global climate change has intensified the urgency of achieving sustainable development, making the transition toward low-carbon economies a central global agenda. In this context, the United Nations Sustainable Development Goals (SDGs), particularly SDG 11 (Sustainable Cities and Communities) and SDG 13 (Climate Action), highlight the importance of transforming consumption patterns toward sustainability. While previous efforts have primarily focused on production-side emission reductions, growing evidence suggests that household consumption has become a critical and increasingly dominant source of carbon emissions [1,2]. According to the China Energy Statistical Yearbook, carbon emissions generated by residential consumption account for more than 40 percent of the national total, and encompass various aspects such as clothing, food, housing, and transportation. As consumption-based emissions are decentralized, behavior-driven, and difficult to regulate directly, promoting sustainable consumption has emerged as a key pathway for achieving long-term carbon mitigation [3]. Scholars generally agree that understanding the formation mechanism of household carbon emissions and identifying their determinants are essential for promoting green transformation and establishing sustainable consumption systems [4,5]. However, existing studies have mainly focused on traditional economic factors such as income level, energy prices, and consumption preferences [6,7]. Fewer studies have explored how digitalization influences household carbon emission behaviors in the context of the digital economy. Digitalization, as a new social force shaping both production and consumption patterns, is likely to exert profound structural effects on household carbon emissions through its influence on information access, consumption modes, and innovation capacity.

In recent years, the rapid expansion of the digital economy has fundamentally transformed individuals’ production modes and lifestyles [8,9,10]. The widespread adoption of digital technologies such as the Internet, big data, and artificial intelligence has promoted economic growth and social efficiency and has also generated complex and sometimes conflicting impacts on household carbon emissions. On the one hand, digitalization can reduce carbon emissions by optimizing energy efficiency and enhancing low-carbon awareness [11]. For instance, the use of smart home technologies enables precise control of household energy consumption and minimizes unnecessary waste. Remote work and online services supported by digital platforms reduce the need for commuting and thus lower transport-related emissions. On the other hand, digitalization may also be associated with higher household carbon emissions by expanding digital consumption and increasing the energy demand associated with electronic devices [12]. The growing popularity of smartphones, computers, and smart appliances has intensified household energy consumption. The rapid growth of e-commerce has added to logistics and packaging emissions, while the widespread use of online entertainment services such as video streaming and gaming has extended device usage time and further raised energy consumption. A review of the relevant literature reveals that existing research on digitalization and carbon emissions remains limited and inconclusive. First, many studies, including Kojić et al. [13], Gao et al. [14] and Qin et al. [15], rely on single indicators, such as Internet usage or regional digital finance development, as proxies for digitalization, which fail to provide a comprehensive and multidimensional measure. This limitation not only reduces measurement accuracy but also obscures the heterogeneous behavioral effects associated with different dimensions of digitalization. Second, most studies focus on macro-level analysis of digital technologies and regional carbon emissions, while fewer have conducted analyses at the micro-level, which obscures the significant heterogeneity among individuals and restricts the identification of micro-level behavioral mediation mechanisms [16,17]. In contrast, this study conducts an empirical analysis using panel data from micro-level survey projects, providing new micro-level evidence on the relationship between digitalization and carbon emissions. Third, most studies typically rely on isolated mediation frameworks and overlook the potential interdependence among different mechanisms. As a result, the chained processes through which digitalization influences carbon emissions remain insufficiently understood [18]. To address this gap, this study develops a chain mediation framework of digitalization, liquidity constraints, consumption structure and carbon emissions, thereby providing a more systematic understanding of how digitalization shapes household economic behavior and ultimately affects environmental outcomes.

Building on this background, this study conducts a micro-level analysis of the relationship between digitalization and household carbon emissions from a household behavior perspective. It proposes two mediation mechanisms, i.e., liquidity constraints and consumption structure, and through systematic theoretical analysis and literature review, summarizes their respective mediation effects as well as the chain mediation effect between them. Within this theoretical framework, this paper conducts empirical validation using micro-level data from five waves of the China Family Panel Studies (CFPS) conducted in 2014, 2016, 2018, 2020, and 2022 (as shown in Figure 1). Specifically, this study measures household carbon emissions based on the Consumer Lifestyle Approach (CLA). It constructs a composite index of digitalization that consists of three dimensions: digital access, digital usage, and digital cognition. Liquidity constraints and household consumption structure are incorporated as chain mediators to examine the indirect transmission channels linking digitalization and household carbon emissions through a chained mediation effect model.

This study makes three main contributions. First, methodologically, it develops a three-dimensional evaluation framework for digitalization that captures digital access, usage, and cognition. Unlike existing studies that rely on single indicators [19,20,21], this multidimensional index reflects its multifaceted characteristics and enhances the comprehensiveness and accuracy of measurement. Second, it proposes a chain mediation theoretical framework that connects digitalization, liquidity constraints, consumption structure, and carbon emissions. While prior studies typically examine these mechanisms in isolation [22,23,24], this paper explicitly models their sequential interdependence, elucidating the intrinsic transmission logic through which digitalization impacts household carbon emissions, and thereby enriching interdisciplinary research at the intersection of digital economics and environmental economics. Third, in contrast to current research that primarily relies on macro-level data for analysis [25,26,27], it systematically investigates the impact of digitalization on carbon emissions, thereby complementing existing research in this area and providing micro-level evidence for the coordinated development of digitalization and low-carbon initiatives. Furthermore, by incorporating results from heterogeneity tests, it offers policy insights for designing differentiated green policies in the digital context, thus supporting the synergistic advancement of digital economic transformation and carbon emission reduction.

2. Theoretical Analysis and Hypotheses

2.1. Digitalization and Household Carbon Emissions

With the rapid development of digital technologies and the widespread availability of Internet infrastructure, digitalization has progressively penetrated many aspects of residents’ daily lives. The level of digitalization not only shapes consumption patterns and production behaviors but also subtly influences household energy use and the structure of carbon emissions. From a theoretical perspective, the impact of digitalization on household carbon emissions may operate through two distinct channels. The first is the promotion effect, through which digitalization expands the scale of consumption and enhances convenience in daily life, thereby increasing energy demand and carbon emissions [28,29]. The second is the substitution effect, in which digitalization improves information efficiency and strengthens environmental awareness, leading to reductions in unnecessary energy consumption and high-carbon activities [15]. The relative strength of these two effects may vary across different stages of digitalization, potentially leading to an inverted U-shaped relationship as described by Chen et al. [12]. In the early stages of digitalization, the promotion effect tends to dominate, whereas in later stages, the substitution effect may gradually become stronger. In the context of the rapid expansion of digital lifestyles in China, where the digital economy is still in a phase of accelerated growth and supporting low-carbon consumption norms and regulatory frameworks are not yet fully developed, most households remain in the early to middle stages of digital transformation. In addition, rising income levels and a strong preference for convenience further reinforce the expansion of consumption associated with digitalization. Therefore, this study posits that the promotion effect is likely to dominate, leading to an increase in household carbon emissions as the level of digitalization rises.

Specifically, the widespread adoption of e-commerce and online payment systems lowers transaction costs. This facilitates instant and diversified consumption, particularly for energy-intensive durable goods such as home appliances and automobiles [30]. Moreover, as digitalization advances, residents increasingly engage in online entertainment, e-shopping, and virtual education. Although these are forms of “virtual consumption”, they rely heavily on data centers, cloud computing, and logistics systems, all of which indirectly raise energy consumption and carbon emissions. According to Festinger’s social comparison theory, digitalization may also exacerbate carbon emissions by altering consumer culture and increasing upward social comparison. Digital platforms tend to reinforce conspicuous consumption, encouraging individuals to purchase high-energy or high-value products to maintain social status. Therefore, under the current stage of China’s economic development, improvements in digitalization may not yet yield the expected energy-saving and emission-reduction effects. Instead, it may increase household carbon emissions by reshaping consumption structures and lifestyles.

Although digitalization generally exerts a positive effect on household carbon emissions, this effect is not uniform across all groups. Differences among individuals and households in terms of Internet infrastructure and digital services acquisition lead to variations in how deeply digitalization permeates daily life [31]. Groups with higher levels of digital access often have more stable Internet connections, higher ownership of digital devices, and richer experience in digital utilization. These factors increase their consumption frequency in areas such as transportation, entertainment, and household services, thereby resulting in higher indirect carbon emissions. In addition, regional disparities in energy structure and technological capacity may amplify or moderate the carbon effects of digitalization [32]. In regions with a higher share of clean energy and stronger technological support, the carbon-emission increase caused by digitalization can be mitigated by policy interventions and energy efficiency.

Based on the above theoretical analysis, the following hypotheses are proposed:

Hypothesis 1. 

Digitalization has a positive effect on household carbon emissions.

Hypothesis 2. 

The effect of digitalization on carbon emissions is heterogeneous across different groups.

2.2. The Mechanisms of Digitalization and Household Carbon Emissions

2.2.1. The Mediating Role of Liquidity Constraints

In the context of the digital economy, improvements in digitalization have profoundly influenced financial behaviors and credit constraints. Traditional economic theory suggests that liquidity constraints serve as an important factor shaping household consumption behavior, as the degree of constraint directly determines both the level and the structure of consumption [33]. When credit markets are underdeveloped, household consumption is heavily dependent on cash flow conditions, often leading to underconsumption and structural distortions. Increasing digitalization helps to alleviate this constraint through several mechanisms. First, the widespread adoption of digital finance and mobile payment systems reduces transaction costs and credit barriers. This has made it easier for households to access financial services such as online credit and installment payments, thereby significantly enhancing their consumption capacity [34]. Second, improvements in information symmetry and the development of digital credit assessment systems reduce discriminatory lending practices in traditional financial institutions against low-income or informally employed groups. As a result, credit accessibility is significantly improved [35,36]. Third, the application of digital technology enables the entire process—from credit applications and document verification to contract signing—to be completed online. This significantly shortens loan approval cycles while eliminating various implicit costs associated with offline transactions. When households face sudden consumption needs, they can swiftly alleviate cash flow pressures through digital channels.

Liquidity constraints play a crucial role in shaping household consumption behavior, particularly by limiting the ability to engage in intertemporal consumption. When such constraints are binding, households tend to suppress current consumption due to restricted access to financial resources. However, as liquidity constraints are relaxed, partly due to the development of digital financial services and improved access to credit, households become more capable of expanding consumption and engaging in advance consumption. According to the life-cycle consumption theory, an increase in consumption induces growth in energy demand across various categories, especially in high-carbon areas such as transportation, housing, and entertainment [37]. Furthermore, according to the theory of mental accounting in behavioral economics, the convenience of digital platforms and the immediacy of credit-based payments lower the psychological cost of spending, thereby stimulating advance consumption and impulsive consumption that intensify energy use and embodied carbon emissions. However, liquidity constraints enhance households’ ability to allocate financial resources and may increase their willingness and capacity to engage in green consumption. This can reduce the carbon intensity of consumption to some extent, partially offsetting the rise in total carbon emissions.

Therefore, digitalization indirectly increases carbon emissions by alleviating liquidity constraints. Based on the above theoretical analysis, the following hypothesis is proposed:

Hypothesis 3. 

Liquidity constraints mediate the relationship between digitalization and household carbon emissions.

2.2.2. The Mediating Role of Consumption Structure

Digitalization affects not only the scale of household consumption but also the composition of that consumption. Changes in consumption structure reflect the evolution from survival-oriented to enjoyment-oriented consumption, a process often accompanied by variations in carbon intensity. The development of digital technologies influences consumption structure primarily by reducing information asymmetry and transaction costs, thereby reshaping households’ economic behavior and consumption choices. Specifically, digital platforms provide residents with abundant and real-time information, enabling them to access a wider range of products and services and facilitating structural upgrading [38]. Under traditional consumption modes, households could access product and service information only through physical stores or personal networks, which limited their awareness of emerging products and high-quality services. In contrast, e-commerce websites, social media, and short-video applications deliver personalized recommendations and marketing content that promote non-essential consumption such as transportation, tourism, and education-related services. Meanwhile, features such as online reviews and user feedback reduce decision-making risks and increase consumer trust in hedonic goods, leading to a greater share of spending on such activities. In addition, the advancement of digital payment and logistics systems lowers transaction costs and enables cross-regional and time-flexible consumption, thereby facilitating the upgrading of household consumption structures [39,40].

Upgrading the consumption structure influences total household carbon emissions by altering the carbon intensity of different expenditure categories. Compared to survival-oriented consumption, which has relatively stable and lower carbon intensity, durable goods and services consumption involves higher carbon emissions, including fuel consumption emissions, electricity consumption emissions, and transportation emissions, thereby driving up total household carbon emissions. For example, an increase in transportation and communication spending indicates more frequent use of private cars, air travel, or high-speed rail, all of which are more energy-intensive than public transport and directly contribute to higher emissions. Healthcare-related consumption also carries a carbon footprint due to the energy and materials used in medical equipment and pharmaceutical production. Although certain forms of non-essential consumption, such as online entertainment, have relatively low carbon intensity, the overall shift toward enjoyment-oriented consumption tends to increase total household carbon emissions.

In summary, digitalization affects household carbon emissions indirectly by promoting the upgrading of the consumption structure. Based on this reasoning, the following hypothesis is proposed:

Hypothesis 4. 

Consumption structure mediates the relationship between digitalization and household carbon emissions.

2.2.3. The Theoretical Mechanism of the Chain Mediation

Liquidity constraints and consumption structure are intrinsically connected, forming a chain transmission pathway. The core logic of this mechanism is that when digitalization alleviates liquidity constraints, households gain greater disposable resources that can be allocated to non-essential consumption. This process promotes the upgrading of the consumption structure and ultimately affects the level of household carbon emissions. The underlying mechanism can be interpreted from three perspectives: financial capacity, consumption preferences, and risk expectations.

First, the relaxation of liquidity constraints provides households with enhanced purchasing power and enables intertemporal consumption. When residents can access credit, installment payments, or other digital financial tools, their previously restricted expenditure capacity is released. Improved credit accessibility encourages greater spending on durable goods, services, and entertainment. This shift facilitates the transformation of household consumption from a survival-oriented pattern to an enjoyment-oriented one [41]. In contrast, residents facing severe liquidity constraints often have limited budgets that must be devoted to essential items such as food and housing, making it difficult to upgrade their consumption structure.

Second, liquidity constraints influence household consumption behavior by shaping psychological and behavioral preferences. Under constrained conditions, households have limited risk tolerance and tend to prioritize immediate and essential consumption. Once constraints are alleviated, greater access to credit and enhanced income flexibility alter consumption preferences, leading households to place more emphasis on quality of life, emotional satisfaction, and personalized experiences. The change in marginal propensity to consume fosters the upgrade in consumption structure. Particularly within the digital economy, new forms of spending such as online education, cultural entertainment, and shared mobility have become important avenues for households’ expenditure with improved liquidity.

Third, households facing tighter liquidity constraints are more likely to adopt precautionary saving behavior in response to future uncertainty, thereby suppressing current consumption and delaying structural optimization, which is consistent with the theory of precautionary savings [42,43]. As liquidity improves, households are able to rely on credit to manage unexpected expenditures rather than maintaining high savings rates, thereby releasing part of their accumulated wealth into the consumption domain. The process allows household consumption to become smoother and more rational, facilitating an evolution of the consumption structure toward long-term and developmental patterns. This mechanism reflects both the financial capacity effect of relaxed liquidity constraints and the adaptive adjustment in household risk perception.

Taken together, liquidity constraints influence the direction and degree of consumption structure upgrading by increasing disposable resources, reshaping consumption preferences, and reducing risk expectations. These three channels operate sequentially rather than independently: increased disposable resources provide the material foundation, reshaped consumption preferences determine the direction of consumption, and reduced risk expectations facilitate the realization of such consumption. Improvements in digitalization play a pivotal role in this process. On the one hand, digitalization alleviates liquidity constraints through channels such as digital payments and inclusive finance. On the other hand, improved liquidity fosters a transformation in consumption patterns, making them more diversified, service-oriented, and sophisticated, which subsequently induces changes in total carbon emissions. In other words, the mechanism by which digitalization affects household carbon emissions operates through a chain mediation pathway in which liquidity constraints influence consumption structure.

Based on the theoretical framework of chain mediation, this study proposes the following hypothesis:

Hypothesis 5. 

Liquidity constraints and consumption structure jointly serve as chain mediators in the relationship between digitalization and household carbon emissions.

In summary, from a theoretical perspective, the interrelationships between digitalization, liquidity constraints, consumption structure, and household carbon emissions are illustrated in Figure 2 below:

3. Materials and Methods

3.1. Data Sources

The individual- and household-level data used are drawn from the CFPS conducted by the Institute of Social Science Survey (ISSS) at Peking University. This database is selected for three main reasons. First, the CFPS provides comprehensive longitudinal information on multiple domains, including society, demographics, economy, education, and health. It contains extensive measurements at both the individual and household levels that are highly relevant to the variables used in this study. Second, the CFPS covers 25 provinces, autonomous regions, and municipalities across China, with a target size of approximately 16,000 households. The dataset is regarded as nationally representative and has been widely applied in prior research. Third, the CFPS survey has been conducted biennially since 2010 and currently provides seven waves of publicly available data for the years 2010, 2012, 2014, 2016, 2018, 2020, and 2022. In addition to CFPS, regional-level data is obtained from the China Energy Statistical Yearbook, the China Statistical Yearbook, and the China Environmental Statistical Yearbook.

This study identifies the household representative based on the CFPS question “Who is the decision-maker in your household?”. If this information is missing, this study alternatively selects the member who is most familiar with the household’s income and expenditure over the past 12 months. These two types of respondents are well-suited to represent household-level behaviors, particularly those related to consumption and digitalization. Considering the availability and consistency of the variables, this study selects data from five survey waves, namely 2014, 2016, 2018, 2020, and 2022. Using province codes and survey years provided in the CFPS, individual-, household-, and regional-level data are matched and integrated. After excluding observations with missing key variables, the final dataset is a balanced panel consisting of 13,605 observations from 2721 households over five periods.

3.2. Variable Descriptions

3.2.1. Explained Variable

The explained variable is the level of household carbon emissions (Carbon). Following the CLA [44], the carbon emissions associated with various categories of household consumption are calculated using the following formula:

$$Carbon={\displaystyle \sum _{i}Intensit{y}_i}\times Consumptio{n}_i$$

(1)

where $Carbon$ represents the carbon emissions of the household, $Intensit{y}_i$ denotes the carbon emission intensity corresponding to the consumption category $i$, and $Consumptio{n}_i$ refers to the total household expenditure on that category. Specifically, samples are divided into urban and rural groups according to the urban-rural classification provided in the CFPS questionnaire. The calculation of carbon emission intensity for different consumption categories is based on the estimates of Liu et al. [45], which distinguish between urban and rural residents. The relevant household expenditure data, including spending on food, clothing and footwear, housing, household equipment and daily goods, healthcare, transportation and communication, culture and education, and other consumption categories, are obtained from the CFPS household economic database. The total household carbon emissions are thus computed by aggregating the embodied carbon emissions across all consumption categories.

3.2.2. Core Explanatory Variable

The core explanatory variable is the level of digitalization (Digital). To measure residents’ level of digitalization in a more comprehensive and rigorous manner, this study constructs an evaluation system comprising three dimensions and eleven indicators, drawing on the Digital Inclusive Finance Index System developed by Peking University [46]. The measurement methods for each dimension and indicator are presented in

Table 1. Digitalization evaluation system.

Dimension		Indicator	Measurement Method	Weight
Prerequisite Dimension	Digital Access	Mobile Device Internet Access	If mobile devices are used for Internet access, assign a value of 1; otherwise, assign 0	-
		Fixed Device Internet Access	If computers are used for Internet access, assign a value of 1; otherwise, assign 0	-
Auxiliary Dimension	Digital Usage	Digitalization in the Financial Field	If the household holds financial products, assign a value of 1; otherwise, assign 0	1/8
		Digitalization in the Living Field	If the Internet is used for commercial activities, assign a value of 1; otherwise, assign 0	1/8
		Digitalization in the Entertainment Field	If the Internet is used for entertainment, assign a value of 1; otherwise, assign 0	1/8
		Digitalization in the Education Field	If the Internet is used for education, assign a value of 1; otherwise, assign 0	1/8
	Digital Cognition	The Importance of the Internet for Learning	Based on the rating of “How important is the Internet for learning”; assign a value of 1 if the value is greater than 3, otherwise assign 0	1/10
		The Importance of the Internet for Work	Based on the rating of “How important is the Internet for work”; assign a value of 1 if the value is greater than 3, otherwise assign 0	1/10
		The Importance of the Internet for Social Interaction	Based on the rating of “How important is the Internet for social interaction”; assign a value of 1 if the value is greater than 3, otherwise assign 0	1/10
		The Importance of the Internet for Entertainment	Based on the rating of “How important is the Internet for entertainment”; assign a value of 1 if the value is greater than 3, otherwise assign 0	1/10
		The Importance of the Internet for Commercial Activities	Based on the rating of “How important is the Internet for commercial activities”; assign a value of 1 if the value is greater than 3, otherwise assign 0	1/10

The rating scale for the five indicators in the Digital Cognition dimension is: 1 = very unimportant, 5 = very important. Financial products in the Digitalization in the Financial Field indicator include stocks, funds, national bonds, trust products, foreign exchange products, Yu’ebao and other Internet wealth management products.

.

The digitalization evaluation system developed in this study consists of three dimensions: (1) Digital access, which refers to the availability of digital hardware and network connectivity. (2) Digital usage, which measures the ability and frequency of individuals using digital technologies and services in daily life. (3) Digital cognition, which reflects individuals’ attitudes toward, trust in, and literacy regarding digital technologies. Among these, digital access is considered the foundational and prerequisite dimension, as it determines whether residents can participate in digital activities at all. Without digital access, even residents with a strong willingness or cognitive ability to use digital technologies cannot effectively engage in the digital process. Therefore, digital access is treated as a necessary condition with a “veto” property, while digital usage and digital cognition are regarded as auxiliary dimensions. Additionally, these two auxiliary dimensions are equally important and mutually reinforcing: digital usage represents the practical manifestation of digital engagement, directly reflecting residents’ ability to leverage digital technologies to solve real problems; digital cognition serves as the intrinsic guide for digital behavior, determining their willingness, direction, and depth of digital technology adoption. Together, they form the complete auxiliary layer of residents’ digital proficiency, comprising both practice and awareness, with neither being overlooked. Based on the functional symmetry and complementarity between digital usage and digital cognition, equal weights are assigned to these two dimensions (1/2 each). Within each auxiliary dimension, given that the individual indicators do not exhibit a clear hierarchy in representing the underlying construct, an equal-weighting scheme is further adopted. Specifically, the digital usage dimension consists of four indicators, each assigned a weight of 1/8, while the digital cognition dimension includes five indicators, each assigned a weight of 1/10. This hierarchical weighting approach ensures a balanced contribution across dimensions and indicators, thereby avoiding systematic underestimation or bias in measuring residents’ level of digitalization.

Following the approach of Alkire and Foster [47] and Chen et al. [48], this study calculates the comprehensive index of the auxiliary dimensions, and then measures the overall digitalization level by combining the digital access dimension and the composite auxiliary index, as follows: (1) If digital access is absent, or if digital access exists but the composite auxiliary index equals 0, the digitalization level is set to 0; (2) If digital access exists and 0 < composite auxiliary index ≤ 0.2, the level is set to 1; (3) If digital access exists and 0.2 < composite auxiliary index ≤ 0.4, the level is set to 2; (4) If digital access exists and 0.4 < composite auxiliary index ≤ 0.6, the level is set to 3; (5) If digital access exists and 0.6 < composite auxiliary index ≤ 0.8, the level is set to 4; (6) If digital access exists and the composite auxiliary index is greater than 0.8, the level is set to 5. Consequently, the core explanatory variable of digitalization level is defined as an ordinal variable that ranges from 0 to 5. This ordinal variable helps to characterize the phased and hierarchical differences in digital access. To ensure robustness, this study further employs alternative measures of digitalization, including an equal-weight index across dimensions, and a binary indicator reflecting whether digital transformation and a continuous variable of the auxiliary dimension index have been achieved.

3.2.3. Mechanism Variables

This study considers two mechanism variables: liquidity constraints and consumption structure. Following Kaplan et al. [49], the liquidity constraint is measured as the ratio of a household’s liquid assets to its net income. Based on the CFPS household economic database, liquid assets include household cash and deposits, financial assets, and funds lent to others, minus non-mortgage financial liabilities. Net income refers to the household’s disposable income after accounting for taxes and transfers. It is important to clarify that a higher ratio indicates greater availability of liquid resources relative to income and thus weaker liquidity constraints. Conversely, a lower value reflects tighter liquidity constraints. Following Dyer et al. [50], given that some observations of household liquid assets take zero values, this study applies the inverse hyperbolic sine (IHS) transformation to ensure the variable is well-defined.

The household consumption structure reflects the proportion of non-essential consumption within total household consumption. Non-essential consumption is a type of spending that contrasts with survival-oriented consumption. Its purpose is not to sustain basic individual survival, but rather to fulfill higher-level consumption demands such as spiritual enjoyment, personalized needs, and health and quality enhancement. It possesses core attributes of experiential nature and quality orientation. Drawing on Hu et al. [51], Yan [52] and Wei et al. [53], this study defines household expenditures on appliances and daily necessities, health care, transportation and communication, culture and entertainment, and other goods and services as non-essential consumption. The ratio of these expenditures to total household consumption is used to measure the consumption structure.

3.2.4. Control Variables

To ensure the accuracy of model estimation, this study includes a set of control variables at both the individual and household levels. At the individual level, the following variables are controlled for: gender (gender), health status (health), marital status (marriage), and educational attainment (edu). At the household level, the control variables include the age of the household head (headage), educational attainment of the household head (headedu), household size (familysize), per capita household income (income), and residential location (urban).

The specific measurement methods and descriptive statistics for all variables are presented in

Table 2. Variable descriptive statistics.

Category	Symbol	Variable	Measurement Method	Mean	S.D.	Min	Max
Explained Variable	Carbon	household carbon emissions	the sum of carbon emissions from all consumption categories (taking the logarithm)	7.8439	0.9560	5.6340	10.6540
Core Explanatory Variable	Digital	Residents’ level of digitalization	the digitalization evaluation system	0.6043	1.3524	0	5
Mechanism Variables	liquidity	liquidity constraint	the ratio of household liquid assets to net income (taking the logarithm)	0.4474	0.9972	−6.1704	8.2940
	structure	household consumption structure	the ratio of non-essential consumption in total household consumption	0.4286	0.2044	0	1
Control Variables	gender	gender	Assign 1 if the gender is male, otherwise 0	0.4786	0.4996	0	1
	health	health status	Based on the question “How do you perceive your physical health status” in the CFPS questionnaire: 1 = unhealthy, 2 = average, 3 = relatively healthy, 4 = healthy, 5 = very healthy	2.8515	1.2800	1	5
	marriage	marital status	Assign 1 if the marital status is married, otherwise 0	0.8915	0.3110	0	1
	edu	educational attainment	the years of education completed	6.5360	4.6147	0	19
	headage	the age of the household head	age of the financial respondent	53.9441	12.4460	18	90
	headedu	educational attainment of the household head	the years of education completed by the financial respondent	7.4361	4.3807	0	19
	familysize	household size	total number of family members	3.8953	1.9041	1	16
	income	per capita household income	per capita household income (taking the logarithm)	9.3666	1.0266	6.2366	11.7464
	urban	residential location	Assign 1 if the household residence is in an urban area, otherwise 0	0.3896	0.4877	0	1

All variables have 13,605 observations in the sample.

.

The descriptive statistics show that the explained variable Carbon exhibits a relatively stable distribution after logarithmic transformation, with a moderate degree of dispersion. This suggests that while substantial variation still exists across households, the influence of extreme values has been effectively mitigated. The core explanatory variable Digital displays a right-skewed distribution. Although most observations are concentrated at lower levels, a non-negligible proportion is observed at medium and higher levels, and no single category is excessively dominant. This distribution pattern not only reflects the stage-specific characteristics of digital development in the background of the digital divide, but also indicates that the constructed index possesses adequate discriminatory power. Therefore, from the perspective of variable distribution, the selected variables are appropriate for subsequent econometric analysis.

3.3. Model Specification

3.3.1. Baseline Regression Model

To examine the impact of digitalization on household carbon emissions, the baseline regression model is specified as follows:

$$Carbo{n}_{it}={\alpha }_0+{\alpha }_{1}Digita{l}_{it}+{\alpha }_2{X}_{it}+{\delta }_j+{\lambda }_t+{\epsilon }_{it}$$

(2)

where $Carbo{n}_{it}$ denotes the carbon emission level of individual $i$’s household in year $t$; $Digita{l}_{it}$ represents the digitalization level of individual $i$ in year $t$; and $X_{it}$ is a vector of individual- and household-level control variables. ${\alpha }_0$ and ${\epsilon }_{it}$ denote the constant term and random error, respectively. ${\alpha }_1$ captures the effect of the core explanatory variable Digital, and ${\alpha }_2$ represents a vector of coefficients associated with the control variables. Given that unobserved characteristics may simultaneously affect both the explained and explanatory variables, thereby giving rise to endogeneity due to omitted variables, failing to control for them may lead to biased estimates. Therefore, the model includes county fixed effects (${\delta }_j$) that capture the shared characteristics of individuals within the same county, as well as time fixed effects (${\lambda }_t$) that account for temporal variations in macroeconomic or policy factors. Given that individuals within the same household may exhibit correlation across years, the standard errors are clustered at the household level.

To further identify the heterogeneous effects across different dimensions, this study decomposes the digitalization index into three components (e.g., digital access, digital usage, and digital cognition) and incorporates them separately into the regression model for estimation. The regression models are specified as follows:

$$Carbo{n}_{it}={\gamma }_0+{\gamma }_{1}Acces{s}_{it}+{\gamma }_2{X}_{it}+{\delta }_j+{\lambda }_t+{\epsilon }_{it}$$

(3)

$$Carbo{n}_{it}={\theta }_0+{\theta }_{1}Usag{e}_{it}+{\theta }_2{X}_{it}+{\delta }_j+{\lambda }_t+{\epsilon }_{it}$$

(4)

$$Carbo{n}_{it}={\rho }_0+{\rho }_{1}Cognitio{n}_{it}+{\rho }_2{X}_{it}+{\delta }_j+{\lambda }_t+{\epsilon }_{it}$$

(5)

where $Acces{s}_{it}$, $Usag{e}_{it}$ and $Cognitio{n}_{it}$ represent individual $i$’s level of digital access, digital usage, and digital cognition in year $t$, respectively, and ${\gamma }_1$, ${\theta }_1$ and ${\rho }_1$ denote the corresponding regression coefficients. Moreover, ${\gamma }_2$, ${\theta }_2$ and ${\rho }_2$ represent a vector of coefficients associated with the control variables. ${\gamma }_0$, ${\theta }_0$ and ${\rho }_0$ are constant terms. ${\epsilon }_{it}$, ${\delta }_j$ and ${\lambda }_t$ are defined in the same way as in the previous equations.

Based on Formulas (2)–(5), the structural diagram of the baseline regression model is shown in Figure 3 below:

3.3.2. Mechanism Analysis Model

This study posits that liquidity constraints and consumption structure serve as chain mediators in the relationship between digitalization and household carbon emissions. Following the analytical procedure proposed by Preacher and Hayes [54], the mediation model is specified as follows:

$$liquidit{y}_{it}={\beta }_0+{\beta }_{1}Digita{l}_{it}+{\beta }_2{X}_{it}+{\delta }_j+{\lambda }_t+{\epsilon }_{it}$$

(6)

$$structur{e}_{it}={\phi }_0+{\phi }_{1}liquidit{y}_{it}+{\phi }_{2}Digita{l}_{it}+{\phi }_3{X}_{it}+{\delta }_j+{\lambda }_t+{\epsilon }_{it}$$

(7)

$$Carbo{n}_{it}={\eta }_0+{\eta }_{1}liquidit{y}_{it}+{\eta }_{2}structur{e}_{it}+{\eta }_{3}Digita{l}_{it}+{\eta }_4{X}_{it}+{\delta }_j+{\lambda }_t+{\epsilon }_{it}$$

(8)

In these equations, $liquidit{y}_{it}$ represents the liquidity constraints of individual $i$’s household in year $t$, $structur{e}_{it}$ denotes the household consumption structure of individual $i$’s household in year $t$, and other variables and parameters are the same as in Formula (2). Coefficient ${\eta }_3$ is used to test the direct effect of the variable Digital, i.e., to examine Hypothesis 1. Coefficients ${\beta }_1$ and ${\eta }_1$ are used to test the mediation effect of the variable liquidity, i.e., to examine Hypothesis 3. Coefficients ${\phi }_2$ and ${\eta }_2$ are used to test the mediation effect of the variable structure, i.e., to examine Hypothesis 4. Coefficients ${\beta }_1$, ${\phi }_1$ and ${\eta }_2$ are used to test the chain mediation effect of the variables liquidity and structure, i.e., to examine Hypothesis 5. Furthermore, ${\beta }_0$, ${\phi }_0$ and ${\eta }_0$ are constant terms. ${\beta }_2$, ${\phi }_3$ and ${\eta }_4$ represent a vector of coefficients associated with the control variables. ${\epsilon }_{it}$, ${\delta }_j$ and ${\lambda }_t$ are defined in the same way as in the previous equations.

Based on Formulas (6)–(8), the structural diagram of the mechanism analysis model is shown in Figure 4 below:

In addition, sampling weights provided by the CFPS are applied in all model estimations to enhance the national representativeness of the results.

4. Results

4.1. The Impact of Digitalization on Household Carbon Emissions

4.1.1. Baseline Regression Results

Following the model specification, this study employs a stepwise regression approach to examine the effect of digitalization on household carbon emissions. Specifically, in

Table 3. Baseline regression results.

Variable	(1)	(2)	(3)	(4)
Digital	0.1915 ***	0.0682 ***	0.0342 ***	0.0342 ***
	(0.0058)	(0.0059)	(0.0064)	(0.0072)
gender		−0.0296 **	−0.0218	−0.0218
		(0.0143)	(0.0143)	(0.0150)
health		−0.0254 ***	−0.0169 ***	−0.0169 ***
		(0.0054)	(0.0054)	(0.0065)
marriage		0.0987 ***	0.1190 ***	0.1190 ***
		(0.0223)	(0.0228)	(0.0311)
edu		0.0050 **	0.0052 **	0.0052 **
		(0.0022)	(0.0023)	(0.0026)
headage		−0.0056 ***	−0.0063 ***	−0.0063 ***
		(0.0006)	(0.0007)	(0.0010)
headedu		0.0117 ***	0.0114 ***	0.0114 ***
		(0.0022)	(0.0023)	(0.0031)
familysize		0.1453 ***	0.1411 ***	0.1411 ***
		(0.0039)	(0.0042)	(0.0066)
income		0.3257 ***	0.2646 ***	0.2646 ***
		(0.0075)	(0.0082)	(0.0121)
urban		0.2218 ***	0.1961 ***	0.1961 ***
		(0.0149)	(0.0179)	(0.0281)
Constant term	7.7282 ***	4.2813 ***	4.8939 ***	4.8939 ***
	(0.0086)	(0.0833)	(0.0916)	(0.1357)
County Fixed Effects	No	No	Yes	Yes
Time Fixed Effects	No	No	Yes	Yes
Cluster-Robust Standard Errors	No	No	No	Yes
Observations	13,605	13,605	13,605	13,605
Within-R2	0.0733	0.3127	0.3593	0.3593

Clustered robust standard errors are reported in parentheses. Statistical significance is indicated by ** p < 0.05, and *** p < 0.01.

, column (1) includes only the explained variable and the core explanatory variable. Column (2) adds individual- and household-level control variables. Columns (3) and (4) further incorporate two-way fixed effects and cluster-robust standard errors, respectively. The regression results show that the coefficient of the core explanatory variable Digital remains significantly positive at the 1% level across all specifications. This indicates a statistically robust positive association between digitalization and household carbon emissions: Residents with higher levels of digitalization tend to generate more carbon emissions, thereby confirming Hypothesis 1. From Column (1) to Column (4), the Within-R² values increase gradually (from 0.0733 to 0.3593), indicating that the models fit the data increasingly well.

Taking Column (4) as an example, the regression coefficient of the variable Digital is 0.0342, suggesting that a one-level increase in residents’ digitalization level is associated with an increase of 0.0342 units in household carbon emissions. This finding suggests that the convenience-enhancing effects of digitalization outweigh its potential efficiency gains, reflecting a typical “rebound effect,” whereby technological progress leads to increased overall energy consumption through behavioral responses. From a behavioral economics perspective, this result highlights the trade-off between digital convenience and carbon costs. Individuals often prioritize behavioral convenience and immediate utility in consumption decisions while tending to overlook the implicit environmental costs of consumption. Digital technologies, by restructuring consumption scenarios, significantly lower consumption barriers and transaction costs, markedly enhancing convenience and satisfaction. This, in turn, stimulates individuals to expand consumption scale, increase consumption frequency, and alter consumption patterns. This process is frequently accompanied by heightened energy demands and material consumption intensity. Examples include the logistics energy consumption from food delivery services, the ongoing electricity usage of smart home devices, and the packaging waste and transportation carbon emissions generated by online shopping. These factors collectively form the core logic behind the significant positive correlation observed between the two variables. These results provide theoretical support for the chain mechanism analysis, which investigates how digitalization affects household carbon emissions through channels such as liquidity constraints alleviation and consumption structure upgrading.

Regarding the control variables, the results reveal several noteworthy patterns. The variable gender is statistically insignificant, possibly because gender-based differences in consumption behavior have narrowed in modern society, and the degree of equality in consumption opportunities has increased in the digital age. Health status is significantly negative, suggesting that individuals in better health tend to adopt low-carbon, environmentally friendly, and sustainable lifestyles, thereby actively reducing their carbon footprint. Marital status is significantly positive, implying that marriage typically leads to household consolidation, increased income, and expanded consumption scale, all of which contribute to higher emissions. Both individual education level (edu) and household head education level (headedu) are significantly positive. Although education may enhance environmental awareness, it can simultaneously promote more open and diversified consumption preferences, driving demand for convenience and comfort, which in turn increases carbon emissions. The coefficient of household head age (headage) is significantly negative, reflecting the consumption contraction characteristic described in life-cycle theory. This suggests that older household heads may have more frugal consumption habits and use energy more cautiously, thereby reducing household emissions. Household size (family size) exhibits a significantly positive effect, as larger households naturally demand more electricity, gas, water, transportation, and food, leading to higher total carbon emissions, a result that represents the scale effect. Per capita household income (income) also shows a significantly positive coefficient, indicating that higher-income households can afford more energy-intensive goods and services and tend to have more diverse and frequent consumption activities. Residential location (urban) is significantly positive, possibly because urban residents generally enjoy higher income levels and living standards, and have greater opportunities for carbon-intensive consumption compared with their rural counterparts.

In addition,

Table 4. Regression results using decomposed digitalization components.

Variable	(1)	(2)	(3)
Access	0.0859 ***
	(0.0235)
Usage		0.0379 *
		(0.0207)
Cognition			0.0235 ***
			(0.0087)
gender	−0.0231	−0.0109	0.0048
	(0.0150)	(0.0330)	(0.0286)
health	−0.0166 **	−0.0161	−0.0174
	(0.0065)	(0.0150)	(0.0136)
marriage	0.1175 ***	0.1189 **	0.1088 **
	(0.0311)	(0.0579)	(0.0517)
edu	0.0065 **	−0.0016	−0.0008
	(0.0026)	(0.0055)	(0.0047)
headage	−0.0065 ***	−0.0036 **	−0.0039 **
	(0.0010)	(0.0017)	(0.0016)
headedu	0.0112 ***	0.0237 ***	0.0226 ***
	(0.0031)	(0.0062)	(0.0057)
familysize	0.1409 ***	0.1319 ***	0.1356 ***
	(0.0066)	(0.0124)	(0.0107)
income	0.2652 ***	0.3532 ***	0.3368 ***
	(0.0121)	(0.0264)	(0.0244)
urban	0.1984 ***	0.1528 ***	0.1722 ***
	(0.0282)	(0.0483)	(0.0436)
Constant term	4.8983 ***	4.0300 ***	4.1822 ***
	(0.1359)	(0.2957)	(0.2729)
County Fixed Effects	Yes	Yes	Yes
Time Fixed Effects	Yes	Yes	Yes
Cluster-Robust Standard Errors	Yes	Yes	Yes
Observations	13,605	5670	5670
Within-R2	0.3587	0.3848	0.3790

Clustered robust standard errors are reported in parentheses. Statistical significance is indicated by * p < 0.1, ** p < 0.05, and *** p < 0.01.

presents the results of separate regression analyses using the three dimensions of digitalization as the core explanatory variables.

In Table 4, the results show that all three dimensions exert a positive and statistically significant impact on household carbon emissions, with coefficients of 0.0859, 0.0379, and 0.0235, respectively. Notably, digital access exhibits the largest effect, followed by digital usage and digital cognition. This pattern is consistent with the conceptual structure of the digitalization index. Digital access represents the foundational condition for participation in the digital economy. Without access to digital infrastructure, individuals cannot effectively utilize digital technologies regardless of their skills or awareness. Therefore, its relatively larger coefficient provides empirical support for its “veto” role in the index construction. In contrast, digital usage and digital cognition function as complementary dimensions. The smaller yet significant and similar coefficients of these two dimensions suggest that they play supportive but non-negligible roles, which is consistent with the equal weighting scheme adopted in the index.

4.1.2. Endogeneity Tests

The CFPS database provides large-scale, randomly sampled, and independently observed microdata, which substantially enhances the reliability of the research conclusions. However, it should be noted that both digitalization and household carbon emissions may be jointly influenced by a variety of unobservable factors, such as personality traits, social environment, and cultural values. Although this study controls for individual characteristics, household attributes, and county fixed effects, data limitations and research design constraints make it impossible to fully account for all potential influencing variables (e.g., individual preferences or social norms), thereby raising concerns of possible omitted variable bias. Moreover, there may exist a bidirectional causal relationship between digitalization and household carbon emissions. Digitalization may affect household carbon emissions, while changes in household carbon emissions may in turn influence digitalization behavior, which could lead to endogeneity bias in estimation. To address this issue, two strategies are employed: (1) lagging the core explanatory variable by one period, and (2) applying the instrumental variable (IV) approach.

Since the CFPS conducts national panel surveys biannually, the first strategy involves matching the digitalization level in 2014 with household carbon emissions in 2016, the digitalization level in 2016 with emissions in 2018, the digitalization level in 2018 with emissions in 2020, and the digitalization level in 2020 with emissions in 2022. This procedure generates a balanced panel of four survey periods. For the instrumental variable strategy, following Du et al. [55], Yang et al. [56], Ivus and Boland [57], two external instruments are adopted: the spherical distance from each household’s region to Hangzhou (distance), and the topographic relief of the household’s residential area (TRI). Hangzhou is recognized as the origin of China’s digital finance, represented by Alipay, and the diffusion of digital finance exhibits strong spatial dependence [58]. Therefore, the geographic distance to Hangzhou is highly correlated with household digitalization. In addition, regional topographic relief influences the deployment of digital infrastructure and the quality of signal transmission, thereby affecting residents’ access to and use of digital technology. Both instruments meet the relevance condition. With respect to the exclusion restriction, it is acknowledged that spherical distance and topographic relief may be correlated with regional development, transportation conditions, and other factors that could potentially affect household carbon emissions. However, this study controls for key confounding factors, including household income, urban-rural status, and county fixed effects, which helps mitigate these alternative channels. Conditional on these controls, the instruments are more likely to affect household carbon emissions primarily through their impact on digitalization. However, because both instrumental variables are time-invariant, their explanatory power may be weakened when two-way fixed effects are included in the estimation. To address this, this study introduces the interaction term between each instrument and the average digitalization level of other residents within the same community (excluding the respondent) [59]. Since the community-level average digitalization varies over time, this interaction term becomes time-varying, thereby allowing them to be identified in the presence of time fixed effects. This strategy not only preserves the relevance of the instruments but also ensures sufficient variation for identification. The results are shown in

Table 5. Results of endogeneity tests.

Variable	One-Period Lag	IV 1		IV 2
		Stage 1	Stage 2	Stage 1	Stage 2
Distance × average		−0.0793 ***
		(0.0042)
TRI × average				−0.1393 ***
				(0.0092)
Digital	0.0471 ***		0.1124 ***		0.0939 ***
	(0.0086)		(0.0289)		(0.0313)
gender	−0.0319 *	−0.1123 ***	−0.0103	−0.1336 ***	−0.0234
	(0.0166)	(0.0264)	(0.0158)	(0.0282)	(0.0160)
health	0.0016	0.0303 ***	−0.0196 ***	0.0321 ***	−0.0166 **
	(0.0073)	(0.0081)	(0.0066)	(0.0086)	(0.0066)
marriage	0.1208 ***	−0.0538	0.1255 ***	−0.0797 *	0.1181 ***
	(0.0332)	(0.0430)	(0.0313)	(0.0470)	(0.0313)
edu	0.0114 ***	0.0924 ***	−0.0034	0.1012 ***	0.0064
	(0.0030)	(0.0052)	(0.0041)	(0.0055)	(0.0047)
headage	−0.0052 ***	−0.0143 ***	−0.0049 ***	−0.0162 ***	−0.0065 ***
	(0.0011)	(0.0012)	(0.0011)	(0.0013)	(0.0012)
headedu	0.0074 **	−0.0273 ***	0.0134 ***	−0.0242 ***	0.0111 ***
	(0.0036)	(0.0037)	(0.0032)	(0.0040)	(0.0032)
familysize	0.1234 ***	−0.0066	0.1420 ***	−0.0117	0.1409 ***
	(0.0074)	(0.0068)	(0.0067)	(0.0076)	(0.0067)
income	0.1826 ***	0.0745 ***	0.2572 ***	0.0986 ***	0.2656 ***
	(0.0134)	(0.0111)	(0.0124)	(0.0126)	(0.0128)
urban	0.2034 ***	0.0049	0.1825 ***	0.1453 ***	0.1979 ***
	(0.0304)	(0.0309)	(0.0284)	(0.0361)	(0.0290)
Constant term	5.6595 ***
	(0.1474)
County Fixed Effects	Yes	Yes	Yes	Yes	Yes
Time Fixed Effects	Yes	Yes	Yes	Yes	Yes
Cluster-Robust Standard Errors	Yes	Yes	Yes	Yes	Yes
KP-LM		122.80 ***		53.84 ***
KP-F		60.81		54.02
CD-F		74.62		84.09
Endogeneity test p value		0.0028		0.0026
Observations	10,884	13,605	13,605	13,605	13,605
Within-R2	0.2981

Clustered robust standard errors are reported in parentheses. Statistical significance is indicated by * p < 0.1, ** p < 0.05, and *** p < 0.01.

below.

According to the regression results in Table 5, the coefficient of the lagged Digital variable remains significantly positive. The instrumental variable estimation further provides supportive evidence for the baseline findings. The KP-LM statistics (122.80 and 53.84) are statistically significant, indicating that the null hypothesis of underidentification can be rejected. In other words, the instruments are sufficiently correlated with the endogenous explanatory variable, satisfying the identification condition. The KP-F and CD-F statistics are all well above the critical value of 16.38 at the 10% significance level. Specifically, the KP-F statistics are 60.81 and 54.02, and the CD-F statistics are 74.62 and 84.09. This confirms that the instruments are strong and that the weak instrument concern is unlikely. In addition, when multiple instruments are used, an overidentification test is conducted. The Hansen J test yields a p-value of 0.2475, failing to reject the null hypothesis that the instruments are jointly exogenous. This indicates that the overidentifying restrictions are valid. The first-stage coefficients of the instruments are all statistically significant, and the second-stage coefficients of the core explanatory variable remain consistent with the baseline regression results. Overall, these findings suggest that the positive association between digitalization and household carbon emissions is robust to alternative specifications that address potential endogeneity concerns.

4.1.3. Robustness Tests

To ensure the reliability of the empirical results, this study conducts a series of robustness tests using three alternative approaches. (1) Replacing the explained variable. To eliminate the potential influence of scale effects, the explained variable is replaced with household per capita carbon emissions. In addition, this study multiplies household carbon emissions by the ratio of regional carbon intensity to the national average level. The resulting interaction term serves as a new explained variable. This approach effectively controls for heterogeneity in regional energy structures and technological efficiency, thereby reducing potential interference in the estimation results. (2) Replacing the core explanatory variable. Regarding the weighting scheme, the three dimensions of digitalization are treated as equally important and assigned equal weights to recalculate the digitalization index, which is then used to replace the original core explanatory variable. Moreover, the study introduces a binary variable representing whether a household has achieved digital transformation. A household is considered digitally transformed when it has nonzero indicator values in all three dimensions of digitalization, and it is assigned a value of 1; otherwise, it is assigned a value of 0. Additionally, this study focuses on the group that has completed digital access and uses the auxiliary dimension index as a continuous variable to replace the core explanatory variable in the regression analysis. (3) Excluding special samples. Given the distinctive characteristics of municipalities directly under the municipal samples in terms of policy resources, economic development, and digital infrastructure, these samples are excluded from the analysis. In addition, since schooling and retirement may significantly influence the household head’s digital technology usage and consumption decisions, households with heads younger than 22 years or older than 65 years are also excluded. (4) Replacing fixed effects. Although the baseline model controls for individual characteristics, household characteristics, as well as county and year fixed effects, there may still exist unobserved and time-invariant household-level factors (e.g., consumption preferences, environmental attitudes, and long-term socioeconomic status) that could bias the estimates. To further address this concern, this study incorporates a household and year fixed effects model. (5) Replacing the econometric model. As discussed earlier, Chen et al. [12] identify an inverted U-shaped relationship between digitalization and carbon emissions. Although differences in the sample period and the construction of digitalization indicators may lead to divergent findings, this study further examines the validity and robustness of the model specification by incorporating the squared term of the core explanatory variable into the baseline regression, in order to test whether a significant nonlinear relationship exists.

According to the results presented in

Table 6. Results of robustness tests.

Variable	Robustness Test 1		Robustness Test 2			Robustness Test 3		Robustness Test 4	Robustness Test 5
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)
Digital	0.0297 ***	0.0365 ***	0.2152 ***	0.0898 ***	0.2381 ***	0.0382 ***	0.0395 ***	0.0238 ***	0.0632 ***
	(0.0072)	(0.0073)	(0.0429)	(0.0237)	(0.0746)	(0.0075)	(0.0076)	(0.0073)	(0.0219)
Digital2									0.0028
									(0.0052)
gender	−0.0125	−0.0221	−0.0212	−0.0234	−0.0137	−0.0245	−0.0160	0.0040	−0.0219
	(0.0150)	(0.0150)	(0.0150)	(0.0150)	(0.0329)	(0.0156)	(0.0159)	(0.0039)	(0.0150)
health	−0.0167 **	−0.0163 **	−0.0172 ***	−0.0165 **	−0.0186	−0.0173 **	−0.0135 *	−0.0144 **	−0.0169 ***
	(0.0065)	(0.0065)	(0.0065)	(0.0065)	(0.0151)	(0.0067)	(0.0070)	(0.0062)	(0.0065)
marriage	−0.0191	0.1134 ***	0.1186 ***	0.1176 ***	0.1146 *	0.1284 ***	0.1169 ***	0.0985 **	0.1191 ***
	(0.0305)	(0.0314)	(0.0311)	(0.0312)	(0.0584)	(0.0328)	(0.0329)	(0.0452)	(0.0311)
edu	0.0039	0.0051 *	0.0048 *	0.0067 ***	−0.0032	0.0046 *	0.0013	−0.0001	0.0052 **
	(0.0026)	(0.0026)	(0.0027)	(0.0026)	(0.0055)	(0.0027)	(0.0028)	(0.0014)	(0.0026)
headage	−0.0034 ***	−0.0062 ***	−0.0063 ***	−0.0066 ***	−0.0032 *	−0.0063 ***	−0.0082 ***	−0.0054 ***	−0.0063 ***
	(0.0010)	(0.0010)	(0.0010)	(0.0010)	(0.0017)	(0.0010)	(0.0012)	(0.0015)	(0.0010)
headedu	0.0127 ***	0.0118 ***	0.0115 ***	0.0112 ***	0.0247 ***	0.0118 ***	0.0130 ***	0.0081 **	0.0114 ***
	(0.0031)	(0.0031)	(0.0031)	(0.0031)	(0.0062)	(0.0032)	(0.0034)	(0.0041)	(0.0031)
familysize	−0.1036 ***	0.1410 ***	0.1409 ***	0.1409 ***	0.1321 ***	0.1396 ***	0.1380 ***	0.1064 ***	0.1411 ***
	(0.0067)	(0.0067)	(0.0066)	(0.0066)	(0.0124)	(0.0068)	(0.0071)	(0.0110)	(0.0066)
income	0.2639 ***	0.2606 ***	0.2635 ***	0.2656 ***	0.3521 ***	0.2629 ***	0.2669 ***	0.1615 ***	0.2647 ***
	(0.0121)	(0.0122)	(0.0121)	(0.0121)	(0.0264)	(0.0124)	(0.0133)	(0.0143)	(0.0121)
urban	0.1889 ***	0.1960 ***	0.1950 ***	0.1983 ***	0.1493 ***	0.1938 ***	0.2010 ***	0.1162 *	0.1958 ***
	(0.0281)	(0.0281)	(0.0281)	(0.0281)	(0.0479)	(0.0287)	(0.0309)	(0.0641)	(0.0281)
Constant term	4.5815 ***	5.3121 ***	4.9003 ***	4.8970 ***	3.9894 ***	4.8928 ***	4.9638 ***	6.0527 ***	4.8935 ***
	(0.1371)	(0.1362)	(0.1357)	(0.1358)	(0.2939)	(0.1387)	(0.1500)	(0.1801)	(0.1357)
County Fixed Effects	Yes	Yes	Yes	Yes	Yes	Yes	Yes	No	Yes
Household Fixed Effects	No	No	No	No	No	No	No	Yes	No
Time Fixed Effects	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Cluster-Robust Standard Errors	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Observations	13,605	13,605	13,605	13,605	5670	12,823	10,923	13,605	13,605
Within-R2	0.3499	0.3900	0.3595	0.3587	0.3865	0.3383	0.3493	0.5071	0.3593

Clustered robust standard errors are reported in parentheses. Statistical significance is indicated by * p < 0.1, ** p < 0.05, and *** p < 0.01.

, the coefficient of the core explanatory variable remains significantly positive across all alternative specifications, including replacing the explained variable, replacing the core explanatory variable, excluding special samples and replacing fixed effects. In addition, the coefficient of the squared term of digitalization (Digital²) is statistically insignificant, suggesting that no significant nonlinear relationship is detected within the sample. These findings demonstrate that the conclusions of this study are robust and reliable.

4.2. Further Analysis of the Impact of Digitalization on Household Carbon Emissions

4.2.1. Heterogeneity Analysis

To further test Hypothesis 2, which posits that the impact of digitalization on household carbon emissions is heterogeneous, this study conducts heterogeneity tests and analyses from three perspectives: individual characteristics, household characteristics, and regional characteristics.

(1) Individual Characteristics Heterogeneity. Social status reflects a comprehensive measure of income level, educational attainment, and occupation. It influences not only individuals’ ability and motivation to access and use digital technologies but also their consumption scale and structure. Environmental awareness, on the other hand, represents an individual’s perception of environmental issues, degree of concern, and willingness to adopt pro-environmental behaviors. It determines how strongly individuals recognize and respond to the low-carbon potential of digitalization. Therefore, this paper argues that the impact of digitalization on household carbon emissions exhibits heterogeneity among individuals of different social status and environmental awareness.

According to the CFPS questionnaire, respondents were asked, “What score would you give yourself for your social status in your local area?” Individuals who rated themselves as 1 were categorized as low social status, while those who rated themselves as 5 were categorized as high social status (N = 1007), those with scores of 2–4 as middle social status (N = 10,494), and those with a score of 5 as high social status (N = 2104). Similarly, based on the question “How serious do you think environmental problems are in China?”, individuals who gave a score of 0 were classified as low environmental awareness (N = 779), those with scores of 2–9 as middle environmental awareness (N = 10,658), and those with a score of 10 as high environmental awareness (N = 2168). Given that this study focuses on differences across extreme groups, the subsequent analysis primarily reports the regression results for the low and high social status groups, as well as the low and high environmental awareness groups.

The results of the subgroup regressions are presented in

Table 7. Heterogeneity tests based on individual characteristics.

Variable	Low Social Status	High Social Status	Low Environmental Awareness	High Environmental Awareness
Digital	0.0547	0.0462 **	0.0512 ***	0.0620
	(0.0338)	(0.0211)	(0.0160)	(0.0499)
Control variables	Controlled	Controlled	Controlled	Controlled
Constant term	4.4196 ***	5.3298 ***	5.2005 ***	5.4631 ***
	(0.3838)	(0.2703)	(0.3204)	(0.4199)
County Fixed Effects	Yes	Yes	Yes	Yes
Time Fixed Effects	Yes	Yes	Yes	Yes
Cluster-Robust Standard Errors	Yes	Yes	Yes	Yes
Observations	1007	2104	779	2168
Within-R2	0.3670	0.3293	0.3316	0.3291
Chow test p value	0.034		0.008

Clustered robust standard errors are reported in parentheses. Statistical significance is indicated by ** p < 0.05, and *** p < 0.01. The full sample includes a middle category; only extreme groups are reported for comparison.

. Columns (1) and (2) display the results for the heterogeneity test based on social status. The coefficient of the core explanatory variable is statistically insignificant for the low social status group, but significantly positive for the high social status group. Individuals with low social status tend to have poorer access to digital infrastructure, lower penetration of digital devices, and weaker digital literacy. Consequently, their overall level of digitalization is relatively low, and the actual impact of digitalization on their lifestyle, energy use, and consumption behavior remains limited. In contrast, the high social status group not only demonstrates a higher degree of digitalization but also possesses higher income and consumption capacity. While enjoying the convenience brought by digital technologies, they are also more likely to engage in high-carbon consumption activities, such as online shopping, food delivery, and commuting with private vehicles. This aligns with status consumption theory, which posits that high-status groups tend to signal their class identity through scarce, symbolic consumption to satisfy their display needs. Digital technologies expand high-end consumption channels, strengthen the social display of consumption, and further stimulate high-carbon consumption.

Columns (3) and (4) report the heterogeneity results based on environmental awareness. The coefficient of the core explanatory variable is significantly positive for the low environmental awareness group, but statistically insignificant for the high environmental awareness group. Environmental behavior theory indicates that environmental awareness directly determines willingness and intensity to engage in environmentally friendly actions. Individuals with stronger environmental consciousness tend to actively apply digital technologies in environmentally friendly and low-carbon behaviors, thereby mitigating the potential carbon-intensive effects of digitalization. Moreover, such individuals are more likely to use digital services in a rational and restrained manner, avoiding unnecessary or excessive consumption that could lead to additional environmental costs. In addition, Chow tests for coefficient differences across groups yield p-values of 0.034 and 0.008, respectively. Both are statistically significant, indicating that the differences in coefficients across groups are statistically meaningful, which further supports the heterogeneity findings discussed above.

(2) Household Characteristics Heterogeneity. According to Menz and Welsch [60] and Wang et al. [61], compared with younger individuals, elderly people generally have lower income, more conservative and rational consumption behaviors, and distinct consumption scales and structures. The elderly often face both physiological and psychological barriers in using digital devices, while limitations in social service provision and traditional cultural perceptions further contribute to the digital divide. Therefore, household aging is expected to influence both the level of digitalization and household carbon emissions. In addition, social capital typically reflects the density of a household’s social networks, the quality of interpersonal relationships, the frequency of reciprocal interactions, and the degree of community trust. It shapes consumption behavior, lifestyle, access to information, and value orientations, thereby affecting the relationship between digitalization and carbon emissions. Therefore, this paper argues that the impact of digitalization on household carbon emissions exhibits heterogeneity among households of different levels of aging and social capital.

This study measures household aging by the proportion of members aged 65 and above within a household, and measures household social capital by per capita gift-giving expenditure. For both variables, sample households are divided annually into terciles. Households in the top 30 percent are categorized as high aging households (N = 4309) or high social capital households (N = 3775), those in the bottom 30 percent are categorized as low aging households (N = 7973) or low social capital households (N = 4652), and the remaining households are categorized as medium aging households (N = 1323) or medium social capital households (N = 5178). The regression results for the low- and high-level groups are shown in

Table 8. Heterogeneity tests based on household characteristics.

Variable	Low Aging	High Aging	Low Social Capital	High social Capital
Digital	0.0403 ***	0.0242	0.0417 ***	0.0301 **
	(0.0085)	(0.0198)	(0.0122)	(0.0117)
Control variables	Controlled	Controlled	Controlled	Controlled
Constant term	4.9665 ***	5.1566 ***	5.3417 ***	4.7898 ***
	(0.1962)	(0.2816)	(0.2043)	(0.2511)
County Fixed Effects	Yes	Yes	Yes	Yes
Time Fixed Effects	Yes	Yes	Yes	Yes
Cluster-Robust Standard Errors	Yes	Yes	Yes	Yes
Observations	7973	4309	4652	3775
Within-R2	0.3344	0.3829	0.3501	0.3729
Chow test p value	0.018		0.010

Clustered robust standard errors are reported in parentheses. Statistical significance is indicated by ** p < 0.05, and *** p < 0.01. The full sample includes a middle category; only extreme groups are reported for comparison.

.

In Table 8, columns (1) and (2) show that the increase in digitalization level is positively associated with household carbon emissions, but this relationship becomes statistically insignificant among highly aged households. The life cycle theory shows that household consumption patterns adjust according to changes in the age structure of its members. Households with low aging levels are generally composed of middle-aged and young members who are more receptive to and frequent users of digital technologies. Their digital activities are closely linked to high-carbon consumption behaviors such as online shopping, streaming entertainment, and the use of smart home devices. In contrast, elderly households tend to maintain traditional consumption habits and rely less on digital platforms in daily life; therefore, changes in digitalization have a limited impact on their overall energy consumption and carbon emissions. Columns (3) and (4) indicate that the effect of digitalization on household carbon emissions is more pronounced among households with lower social capital. Such households are less likely to rely on informal networks, community assistance, or interpersonal exchanges to obtain resources and services. Instead, they depend more on market-based transactions and digital platforms to meet their daily needs, which increases their exposure to carbon-intensive consumption patterns. Conversely, households with higher levels of social capital are more strongly influenced by traditional norms and reciprocal ethics, emphasizing face-to-face interactions and offline social engagements. These behaviors, to some extent, reduce their dependence on high-carbon digital services. Furthermore, Chow tests produce p-values of 0.018 and 0.010, indicating that the estimated coefficients differ significantly across groups, providing additional evidence in support of the heterogeneity findings.

(3) Regional Characteristics Heterogeneity. The level of regional environmental governance reflects not only the local government’s commitment to ecological protection and the intensity of its regulatory enforcement, but also the strength of environmental regulations and the availability of environmental infrastructure that residents experience. These factors jointly influence household energy use, consumption patterns, and the ways in which digital technologies are applied in daily life. In addition, variations in regional energy intensity are associated with differences in economic structure, technological capability, resource endowment, and policy orientation. Therefore, both the level of environmental governance and the degree of energy intensity at the regional level are likely to shape the relationship between digitalization and household carbon emissions.

This study measures environmental governance by the ratio of investment in environmental pollution control to regional GDP, and measures energy intensity by the ratio of carbon emissions to GDP. Regions in the top 30 percent are defined as having high environmental governance (N = 3624) or high energy intensity (N = 4172), those in the bottom 30 percent are categorized as low environmental governance (N = 4683) or low energy intensity (N = 3624), and the remaining regions are categorized as medium environmental governance (N = 5298) or medium energy intensity (N = 3905). The results of the subgroup regressions for low and high environmental governance, as well as low and high energy intensity, are reported in

Table 9. Heterogeneity tests based on regional characteristics.

Variable	Low Environmental Governance	High Environmental Governance	Low Energy Intensity	High Energy Intensity
Digital	0.0629 ***	0.0221 *	0.0188 *	0.0535 ***
	(0.0146)	(0.0115)	(0.0108)	(0.0138)
Control variables	Controlled	Controlled	Controlled	Controlled
Constant term	5.4896 ***	4.5988 ***	4.4057 ***	5.4796 ***
	(0.2586)	(0.2082)	(0.2198)	(0.2507)
County Fixed Effects	Yes	Yes	Yes	Yes
Time Fixed Effects	Yes	Yes	Yes	Yes
Cluster-Robust Standard Errors	Yes	Yes	Yes	Yes
Observations	4683	3624	5528	4172
Within-R2	0.2889	0.4055	0.3898	0.3303
Chow test p value	0.055		0.005

Clustered robust standard errors are reported in parentheses. Statistical significance is indicated by * p < 0.1, and *** p < 0.01. The full sample includes a middle category; only extreme groups are reported for comparison.

below.

According to Table 9, columns (1) and (2) show that the impact of digitalization on household carbon emissions is stronger and more significant in regions with lower levels of environmental governance. Based on environmental regulation theory, stringent environmental regulations curb carbon emissions growth by constraining high-carbon activities and promoting the substitution of clean energy. In areas with weak environmental governance, insufficient regulatory oversight and a lack of technological support allow the positive carbon effects of digitalization to be more fully manifested. In contrast, regions with higher environmental governance tend to have more stringent policy guidance and cleaner energy structures, which can partially offset the carbon-increasing effect associated with digitalization. Columns (3) and (4) report the heterogeneity results with respect to energy intensity. The findings indicate that the effect of digitalization on household carbon emissions is greater and more significant in regions with higher energy intensity compared with regions characterized by lower energy intensity. As Du et al. [62] pointed out, in areas with lower energy intensity, even if residents do not alter their consumption scale or habits, the cleaner and more sustainable production of goods and services helps to reduce associated carbon emissions. Conversely, regions with carbon-intensive energy structures amplify the emission-increasing effects brought about by digitalization. In addition, Chow tests yield p-values of 0.055 and 0.005, respectively. Taken together, these results suggest that the estimated coefficients differ across groups, providing supplementary evidence for the heterogeneity analysis.

To enhance the rigor of the heterogeneity analysis, this study further estimates interaction models within a unified regression framework by interacting the core explanatory variable with subgroup indicators. The results are reported in

Table 10. Heterogeneity tests based on interaction terms.

Variable	Individual Characteristics Heterogeneity		Household Characteristics Heterogeneity		Regional Characteristics Heterogeneity
	Social Status	Environmental Awareness	Aging	Social Capital	Environmental Governance	Energy Intensity
Digital	0.0623 ***	0.0300 *	0.0270 *	0.0394 **	0.0453 ***	0.0336 **
	(0.0204)	(0.0159)	(0.0147)	(0.0162)	(0.0159)	(0.0157)
Group	0.0163 **	0.0060 *	0.0288 *	0.2080 ***	0.0406 **	−0.0505 *
	(0.0080)	(0.0031)	(0.0154)	(0.0136)	(0.0209)	(0.0264)
Interaction term	0.0205 ***	−0.0104 ***	−0.0238 **	−0.0136 *	−0.0248 ***	0.0160 **
	(0.0058)	(0.0021)	(0.0101)	(0.0070)	(0.0075)	(0.0075)
Control variables	Controlled	Controlled	Controlled	Controlled	Controlled	Controlled
Constant term	4.8879 ***	4.8599 ***	4.8862 ***	4.7838 ***	4.8363 ***	4.9713 ***
	(0.1369)	(0.1374)	(0.1362)	(0.1341)	(0.1414)	(0.1473)
County Fixed Effects	Yes	Yes	Yes	Yes	Yes	Yes
Time Fixed Effects	Yes	Yes	Yes	Yes	Yes	Yes
Cluster-Robust Standard Errors	Yes	Yes	Yes	Yes	Yes	Yes
Observations	13,605	13,605	13,605	13,605	13,605	13,605
Within-R2	0.3595	0.3595	0.3597	0.3746	0.3594	0.3594

Clustered robust standard errors are reported in parentheses. Statistical significance is indicated by * p < 0.1, ** p < 0.05, and *** p < 0.01.

. The interaction terms are statistically significant, and their signs are broadly consistent with those obtained from the subgroup regressions, suggesting that the effect of digitalization on household carbon emissions varies across individual, household, and regional groups.

4.2.2. Mechanism Analysis

Following Formulas (6)–(8), this study conducts a chain mediation effect test to explore the mechanisms through which digitalization affects household carbon emissions. If the core explanatory variable, Digital, and the mechanism variables, liquidity and structure, are significant, this would confirm the existence of a chain mediation effect. The regression results are presented in

Table 11. Chain mediation model results.

Variable	Formula (6)	Formula (7)	Formula (8)
Digital	0.0304 ***	0.0291 ***	0.0276 ***
	(0.0042)	(0.0018)	(0.0072)
liquidity		0.0206 ***	0.0378 *
		(0.0050)	(0.0198)
structure			0.3240 ***
			(0.0586)
gender	0.0019	−0.0037	−0.0211
	(0.0092)	(0.0037)	(0.0149)
health	−0.0175 ***	−0.0143 ***	−0.0130 **
	(0.0040)	(0.0016)	(0.0065)
marriage	−0.0402 **	0.0166 **	0.1169 ***
	(0.0185)	(0.0074)	(0.0310)
edu	0.0053 ***	0.0007	0.0049 *
	(0.0015)	(0.0007)	(0.0026)
headage	0.0006	−0.0012 ***	−0.0061 ***
	(0.0006)	(0.0002)	(0.0010)
headedu	−0.0146 ***	−0.0002	0.0120 ***
	(0.0018)	(0.0008)	(0.0031)
familysize	0.0041	0.0129 ***	0.1380 ***
	(0.0037)	(0.0016)	(0.0067)
income	−0.0559 ***	−0.0019	0.2673 ***
	(0.0070)	(0.0028)	(0.0122)
urban	−0.0023	−0.0395 ***	0.2050 ***
	(0.0169)	(0.0067)	(0.0281)
Constant term	1.0320 ***	0.4917 ***	4.7423 ***
	(0.0803)	(0.0320)	(0.1398)
County Fixed Effects	Yes	Yes	Yes
Time Fixed Effects	Yes	Yes	Yes
Cluster-Robust Standard Errors	Yes	Yes	Yes
Observations	13,605	13,605	13,605
Within-R2	0.1377	0.1083	0.3616

Clustered robust standard errors are reported in parentheses. Statistical significance is indicated by * p < 0.1, ** p < 0.05, and *** p < 0.01.

.

Specifically, in Table 11, the direct effect of the core explanatory variable Digital on the explained variable Carbon is significantly positive, with a coefficient of 0.0276. This indicates that higher levels of digitalization lead to increased household carbon emissions, further supporting Hypothesis 1. The decomposition of the indirect effects reveals several important pathways. First, the indirect effect of Digital on the mediating variable liquidity is significantly negative at 0.0304, and the effect of liquidity on Carbon is also significantly negative at 0.0378. These results suggest that digitalization alleviates household liquidity constraints, which in turn raises household carbon emissions. The estimated partial mediation effect is 0.0011, providing empirical support for Hypothesis 3. Second, the indirect effect of digital on the mediating variable structure is significantly positive at 0.0291, and structure has a significantly positive effect on carbon with a coefficient of 0.3240. This implies that improvements in digitalization can enhance or optimize household consumption structures, leading to higher carbon emissions. The corresponding partial mediation effect of 0.0094 confirms Hypothesis 4. Moreover, the mediating variable liquidity exerts a significantly negative effect on structure (0.0206), indicating that liquidity constraints hinder the upgrading of household consumption structures. Combining the effect of Digital on liquidity and that of structure on Carbon, the estimated value of the chain mediation effect is 0.0001, thereby supporting Hypothesis 5.

To further verify the robustness of these mediation effects, the Bootstrap method is applied. The specific effect values and 95% confidence intervals are obtained using a cluster-bootstrap with 1000 replications, resampling at the household level. The results, summarized in

Table 12. Results based on bootstrap method.

Effect		Value	Share of Total Effect	95% Confidence Intervals
				Lower Limit	Upper Limit
Direct Effect	Digital→Carbon	0.0276	80.70%	0.0203	0.0450
Indirect Effect	Digital→liquidity→Carbon	0.0011	3.22%	0.0003	0.0021
	Digital→structure→Carbon	0.0094	27.49%	0.0079	0.0110
	Digital→liquidity→structure→Carbon	0.0002	0.58%	0.0001	0.0003
Total Effect	Digital→Carbon	0.0342	31.29%	0.0196	0.0500

, show that for both the direct and indirect effects, the 95% confidence intervals do not include zero, confirming their statistical significance. Among the decomposed indirect effects, the effects through liquidity constraints and consumption structure are 0.0011 and 0.0094, respectively, indicating that a one-level increase in digitalization leads to increases of 0.0011 and 0.0094 units in household carbon emissions through the relaxation of liquidity constraints and the optimization of consumption structure. The chain mediation effect is significantly estimated at 0.0002, suggesting that digitalization not only affects household carbon emissions through individual mechanisms but also operates through a sequential transmission pathway, thereby confirming the internal linkage between these two mechanisms. The total indirect effect, calculated as the sum of all decomposed indirect effects, equals 0.0107. According to the baseline regression results in Table 3, the total effect of digitalization on household carbon emissions is 0.0342, suggesting that the mediation effects jointly account for approximately 31.29% of the total effect.

The above chain mediation test not only provides evidence on the complex transmission path through which digitalization affects household carbon emissions via liquidity constraints and consumption structure, but also addresses a limitation in previous studies that paid insufficient attention to the multiple mediating relationships between digitalization and its environmental effects. Moreover, it deepens the understanding of the intrinsic mechanisms through which digital technologies empower household economic behavior and influence environmental outcomes. These findings provide crucial empirical evidence for comprehending the internal trade-offs between digital economic development and carbon mitigation, and offer precise micro-level insights for policymakers to coordinate the advancement of digital transformation with the achievement of “dual carbon” goals.

In addition, considering that the explained variable Carbon and the mechanism variable structure are constructed using partially overlapping expenditure categories, there is a potential concern of mechanical correlation, which may affect the identification of the chain mediation effect. To address this issue, this study constructs an alternative measure of carbon emissions by excluding the expenditure categories used in the consumption-structure, thereby reducing potential overlap in variable construction. Based on the baseline mediation framework, this study re-estimates the model using this non-overlapping measure, and the results are reported in

Table 13. Robustness test for the chain mediation model.

Variable	Formula (1)	Formula (6)	Formula (7)	Formula (8)
Digital	0.0250 ***	0.0304 ***	0.0291 ***	0.0297 ***
	(0.0083)	(0.0042)	(0.0018)	(0.0078)
liquidity			0.0206 ***	0.0481 **
			(0.0050)	(0.0208)
structure				0.3131 ***
				(0.0595)
gender	−0.0195	0.0019	−0.0037	−0.0255
	(0.0166)	(0.0092)	(0.0037)	(0.0161)
health	0.0094	−0.0175 ***	−0.0143 ***	−0.0135 *
	(0.0073)	(0.0040)	(0.0016)	(0.0070)
marriage	0.1140 ***	−0.0402 **	0.0166 **	0.1412 ***
	(0.0349)	(0.0185)	(0.0074)	(0.0342)
edu	0.0048 *	0.0053 ***	0.0007	0.0059 **
	(0.0029)	(0.0015)	(0.0007)	(0.0028)
headage	−0.0056 ***	0.0006	−0.0012 ***	−0.0075 ***
	(0.0012)	(0.0006)	(0.0002)	(0.0011)
headedu	0.0107 ***	−0.0146 ***	−0.0002	0.0106 ***
	(0.0035)	(0.0018)	(0.0008)	(0.0033)
familysize	0.1268 ***	0.0041	0.0129 ***	0.1476 ***
	(0.0072)	(0.0037)	(0.0016)	(0.0071)
income	0.2647 ***	−0.0559 ***	−0.0019	0.2622 ***
	(0.0132)	(0.0070)	(0.0028)	(0.0129)
urban	0.2764 ***	−0.0023	−0.0395 ***	0.2127 ***
	(0.0311)	(0.0169)	(0.0067)	(0.0297)
Constant term	4.2951 ***	1.0320 ***	0.4917 ***	5.0769 ***
	(0.1494)	(0.0803)	(0.0320)	(0.1482)
County Fixed Effects	Yes	Yes	Yes	Yes
Time Fixed Effects	Yes	Yes	Yes	Yes
Cluster-Robust Standard Errors	Yes	Yes	Yes	Yes
Observations	13,605	13,605	13,605	13,605
Within-R2	0.3151	0.1377	0.1083	0.4029

Clustered robust standard errors are reported in parentheses. Statistical significance is indicated by * p < 0.1, ** p < 0.05, and *** p < 0.01.

. The findings show that the chain mediation effect through liquidity constraints and consumption structure remains statistically significant and retains the same direction, further strengthening the credibility of the chain mediation mechanism proposed in this paper.

5. Discussion

5.1. Summary of Key Findings

Existing theories on digitalization and sustainable development generally posit that digital technologies exert both promoting effects and substitution effects on carbon emissions, yet the relative strength of these effects remains controversial [63,64]. Based on micro-level evidence from individual and household perspectives, the empirical results of this study reveal that, under a two-way fixed effects model with strong explanatory power as reflected by the coefficients of determination, residents’ digitalization levels have a significant positive total effect on household carbon emissions. This finding provides key micro-level evidence for the dual-effect discussion. In the context of Chinese household consumption, the promoting effects, which enhance consumption convenience, stimulate consumption expansion, and improve consumption structure, currently dominate, while the substitution effects that optimize energy efficiency and promote green consumption have not yet been widely realized.

From the perspective of behavioral economics, this study confirms the crowding-out effect of immediate convenience over long-term environmental costs. Digitalization substantially reduces time costs and transaction barriers, making it easier to engage in impulsive, frequent, and high-carbon consumption behaviors. At the same time, consumers may lack sufficient awareness of the hidden carbon emissions behind such consumption. This phenomenon enriches the micro-behavioral theoretical framework of the environmental effects of digitalization and offers a new perspective for the theory of sustainable development. Digital technologies are not inherently low-carbon tools; their environmental impact depends on the guidance and constraints imposed on consumer behavior. The findings are broadly consistent with Rangel Luzuriaga [65], who emphasizes that digital competencies serve as a key driver of transformation, while generating differentiated outcomes depending on contextual conditions. Similarly, although digitalization enhances household capabilities, it is also associated with increased carbon emissions, driven by the expansion of consumption behaviors rather than efficiency gains. This comparison highlights that the environmental consequences of digitalization are not fixed, but are shaped by the interaction of contextual factors.

5.2. Comparison with Existing Literature

Several previous studies have reached results different from those of this study, which may be explained by differences in data, variable measurement, and research context. First, differences in data. This study uses micro-level household panel data from the CFPS and focuses on carbon emissions from the consumption side. In contrast, other studies often rely on provincial or national macro-level data and focus on the production side or total regional carbon emissions. For example, He et al. [66] used 283 Chinese cities from 2006 to 2021 as their research sample; Liu et al. [67] examined the relationship between the digital economy and carbon emissions in the agricultural production sector; Pu et al. [68] incorporated the impacts of residential and non-residential carbon emissions into their study. Moreover, variation in sample periods within the same CFPS dataset may also contribute to divergent findings. For instance, Chen et al. [12] analyze data from 2014 to 2018, while this study extends the period to 2022, capturing a later stage of digital expansion in China. As digitalization evolves, the balance between promotion and substitution effects may shift, resulting in different empirical outcomes. Second, differences in measurement. This study constructs a three-dimensional digitalization index that includes digital access, digital usage, and digital cognition, which comprehensively reflects residents’ multidimensional characteristics of digitalization. In contrast, other studies often use a single indicator to measure, which may overlook its structural heterogeneity. For instance, some studies measure digitalization by internet penetration rate, which only reflects access levels [69]. Furthermore, many previous studies use aggregate regional emission data, while this study adopts the CLA to measure household consumption-based carbon emissions, which captures the direct link between consumption behavior and emissions more precisely. These measurement differences help explain the discrepancies among research findings. Third, differences in research context. This study focuses on Chinese households, where the digital economy is rapidly expanding but low-carbon behavioral constraints have not yet matured. Other studies based on data from developed economies, where increased public environmental awareness, improved supporting environmental policies, widespread adoption of energy-efficient digital devices, and the development of low-carbon digital services may have already shifted digitalization into a stage of low-carbon transformation [70].

Despite these differences, the conclusions of this study are not in opposition to previous findings but rather reflect the multifaceted and context-dependent nature of the environmental effects of digitalization. The impact of digitalization on carbon emissions includes both suppressive effects on the production side and promoting effects on the consumption side. It is influenced not only by the characteristics of the technology itself but also by contextual factors such as data scale, measurement method, and development stage. The contribution lies in identifying the carbon-increasing effects and behavioral mechanisms of digitalization on the household consumption side. This complements existing research and contributes to a more comprehensive understanding of the environmental implications of digitalization.

5.3. Theoretical and Policy Implications

This study provides three main theoretical implications. First, by constructing a multidimensional digitalization index that encompasses digital access, usage, and cognition, this study enhances the measurement of digitalization. The results show that the effects of digitalization on carbon emissions vary significantly across dimensions, and neglecting this multidimensionality may lead to biased estimates and unclear mechanism identification. This study offers a more comprehensive and refined measurement method, contributing to the development of indicator systems in research on digitalization and environmental behavior.

Second, the mechanisms through which digitalization affects carbon emissions are not independent but exhibit a sequential linkage. The chain mediation model reveals an underlying transmission pathway in which digitalization relaxes economic constraints, reshapes consumption structure, and ultimately influences carbon emissions. This finding addresses the limitation of prior studies that largely overlook the interdependence among mediation mechanisms and enriches the interdisciplinary framework at the intersection of digital economics, behavioral economics, and environmental economics.

Third, this study extends the analytical focus from the macro-level perspective to the micro-level household perspective. In doing so, it deepens the understanding of how digitalization influences carbon emissions through consumption-related behavioral channels.

In terms of policy implications, the findings suggest that digitalization is not inherently low-carbon. Its environmental impact depends on how it shapes economic behavior. Therefore, policy efforts should guide digital development toward low-carbon transformation, strengthen environmental awareness, and promote sustainable consumption patterns. Moreover, given the significant heterogeneity in the effects of digitalization, differentiated policy approaches are needed to better align digital transformation with carbon reduction goals.

5.4. Limitations and Future Research

Although this study systematically examines the impact of digitalization on household carbon emissions and explores the chain mediation mechanism, several limitations remain. First, the proposed chain mediation framework does not incorporate potential variables such as social norms and psychological factors. Moreover, the possible interaction effects among variables have not been tested, which may result in an incomplete understanding of the underlying mechanism. Second, since the analysis is based on Chinese household data, the external validity of the conclusions may be limited by cultural and institutional contexts. For example, in developing economies, where credit constraints are more prevalent and digital finance is rapidly expanding, digitalization is more likely to generate a carbon-increasing effect. In contrast, in developed economies with more mature digital systems, stronger environmental regulations, and higher levels of environmental awareness, digitalization may primarily strengthen the substitution effect. Future research should test the generalizability of these findings using cross-country or cross-regional comparative data. Third, with regard to data, the use of a balanced panel, which retains only households observed in all survey waves, may introduce attrition bias. Regarding measurement, digitalization is specified as an ordinal variable, which limits the interpretation of marginal effects. Future research could address these limitations by adopting more diverse data and alternative measurement approaches.

6. Conclusions and Policy Implications

6.1. Conclusions

This paper systematically examines the impact of digitalization on household carbon emissions and its underlying mechanisms. The main findings are as follows: First, digitalization is significantly and positively associated with household carbon emissions. Second, the impact of digitalization on household carbon emissions exhibits significant heterogeneity. Specifically, the carbon-increasing effect of digitalization is more pronounced among individuals with higher social status and lower environmental awareness, within households characterized by low aging ratios and low social capital, and in regions with low environmental governance levels and high energy intensity. Third, digitalization affects carbon emissions through both liquidity constraints and consumption structure, with a significant chain mediation effect.

More broadly, this study indicates that digitalization is not inherently environmentally friendly. Its impact on carbon emissions essentially depends on the pathways for adjusting behavioral mechanisms and consumption patterns. Technological progress can be transformed into a driving force for sustainable development when accompanied by effective behavioral guidance and institutional constraints. The research findings not only provide micro-level policy evidence for coordinating the high-quality development of the digital economy with China’s “dual carbon” goals, but also offer important insights for other countries with similar cultural backgrounds, development stages, and institutional contexts.

6.2. Policy Implications

Based on the above findings and in view of the practical need to coordinate digital economic transformation with low-carbon development strategies, this study proposes the following targeted policy recommendations.

First, promote a low-carbon digital transition. Given the positive impact of digitalization on household carbon emissions, it is crucial to reduce the implicit carbon costs of digitalization through both digital infrastructure and digital service provision. In terms of infrastructure, greater investment should be directed toward the research and development of energy-efficient technologies and the green transformation of digital infrastructure. Efforts should be made to improve the coverage of digital access while prioritizing low-carbon construction schemes, thereby avoiding the “pollute first, mitigate later” development pattern. In terms of digital services, platform enterprises should be encouraged to develop low-carbon consumption scenarios, such as online office systems and remote conferencing. A carbon footprint accounting system should be established for major digital platforms, requiring e-commerce and social media platforms to regularly disclose carbon emissions from their operations and logistics. Platform enterprises should also be guided to reduce emissions through centralized delivery systems, recyclable packaging, and other sustainable practices.

Second, implement differentiated carbon reduction policies. In light of the heterogeneity of digitalization’s carbon effects, policy measures should be differentiated to avoid efficiency losses caused by uniform “one-size-fits-all” regulations. For instance, for low-socioeconomic groups, efforts should focus on enhancing digital access, advancing digital skills training, and promoting affordable energy-efficient products. For high-socioeconomic groups, a high-carbon consumption warning mechanism should be established to guide them in transforming their consumption advantages into drivers for low-carbon transition. Additionally, digital low-carbon awareness campaigns should be intensified for groups with low environmental consciousness, while green consumption subsidies should be provided to those with high environmental awareness. For low-aging households predominantly composed of young adults and middle-aged individuals, “scenario-based digital low-carbon living” initiatives should be implemented. Conversely, for high-aging households primarily consisting of elderly members, traditional low-carbon consumption channels should be preserved while developing age-friendly digital low-carbon services. In addition, governments should organize the establishment of community digital mutual aid platforms to create more diverse social networking opportunities for households with low social capital. Those with high social capital should consider transforming traditional social exchanges into low-carbon collaborations. At the regional level, areas with weak environmental governance and high energy intensity should further strengthen digital oversight and clean energy substitution. This includes publicly disclosing environmental governance information and exposing high-carbon pollution activities through digital platforms, while mandating energy-intensive enterprises to implement digital energy-saving upgrades.

Third, optimize the chain mediation transmission pathway. Given the chain mediation effect of liquidity constraints and consumption structure, policy efforts should focus on both financial and consumption guidance. In terms of financial guidance, the design of digital financial products should be standardized by incorporating low-carbon consumption preferences into digital credit evaluation systems. A special fund for digital low-carbon consumption could be established to provide fiscal subsidies, encouraging digital financial institutions to increase credit support for household low-carbon consumption and thereby ease liquidity constraints on the purchase of low-carbon products. In terms of consumption guidance, digital tools should be used to provide consumers with real-time information on product carbon emissions and low-carbon alternatives. Households that purchase low-carbon products or reduce high-carbon consumption could receive digital service vouchers or utility payment subsidies, which would further stimulate their willingness to engage in low-carbon consumption. In addition, the role of digital media in promoting low-carbon lifestyles should be strengthened by disseminating environmental knowledge through short videos, online discussions, and social media campaigns.