The impact of internet use on the subjective well-being of the elderly: New evidence from the China Family Panel Studies

With the widespread availability of 5G technology in China, the internet has deeply affected the lives of the elderly. This research analyzes the impact of internet use on subjective well-being (SWB) of the elderly using the latest 2020 China Family Panel Studies (CFPS) data through machine learning (ML) techniques and traditional regression models. (1) Machine learning �ndings indicate that the factors order of importance from strong to weak is the internet as an information channel, contact with family and friends, and work, leisure and entertainment, daily life and using WeChat, watching online short videos and learning. (2) Ordinary least squares (OLS) regression results show the internet for daily life and watching short-form videos have a signi�cant negative effect on the SWB of the elderly. (3) XGBoost with determination coe�cients (R 2 ) greater than 0.86 is signi�cantly outperformed the OLS with determination coe�cients of 0.35 in full sample sets. This work proves that the combination of machine learning and traditional regression methods can both take advantage of the interpretability of machine learning and reveal factors contributions through traditional regression models, having the ability to mine emerging and potential factors. Our research shows that continuously strengthening the digital construction of the elderly, promoting the digital literacy and skills cultivation of the elderly, and enhancing the sense of participation and happiness of the elderly can help improve the active digital aging of the elderly.


Introduction
In the context of digitalization and aging, the digital inclusion of the elderly is gradually becoming a hot topic of global concern.According to the Asia-Paci c Regional Congress of Gerontology, the old are those above the age of 60 in the Asia-Paci c area.The over-60s are also used as a criterion for the elderly in China.As a result, the over-60s are referred to as the old in this article.As of December 2021, the scale of China's elderly netizens aged 60 and above in China has reached 119 million, accounting for 11.5% of the total Internet users, and the Internet penetration rate of the elderly population aged 60 and over reached 43.2% (CNNIC, 2022).Technology offers tremendous opportunities to improve the quality of life of the elderly, enriching the ability to perform functional tasks in daily life (e.g., taking a taxi to get around, getting information from WeChat), enhancing the ability to learn and work through online information and so on, enabling the elderly to lead more independent lives (Berkowsky et al., 2018;Aggarwal et al., 2020).However, there is no systematic research on the effects of multiple aspects of internet use on the SWB of the elderly.
The interaction between internet use and life satisfaction among older adults is a complex network and has been explored in many researches (Zhang et al., 2020).Internet use can improve most aspects lives of the elderly, including improving their mental health and life satisfaction, decreasing depression, and keeping them physical health (Heo et Lee et al., 2021).Higher social technology use is associated with better self-assessed health, less chronic disease, higher SWB and fewer depressive symptoms, thus the elderly typically have positive attitudes towards technology (Chopik, 2016).
Research on the relationship between internet use and the SWB of the elderly has received increasing attention.Szabo et al. (2019) found that internet use can support the well-being of the elderly and affect well-being in multiple ways.Older participants assess the internet for three main purposes: social reasons (e.g., contact with friends and family), instrumental reasons (e.g., banking), and informational reasons (e.g., access to health-related information).The mechanisms by which internet use affects the elderly's SWB differ, with leisure time spent on social and recreational activities having a signi cant happiness motivating effect on the elderly, and learning activities having a signi cant happiness inhibiting effect (Peng et al., 2019).Jim and Li's (2021) research also showed that using the internet signi cantly improved the SWB of rural elderly and reduced their loneliness; the higher the frequency of internet access, the greater the improvement effect on SWB.
Although existing research have found a positive effect of internet use on SWB of the elderly, there are still certain de ciencies.Firstly, most research are limited to exploring individual characteristics factors that in uence the elderly's SWB, such as gender and education, and lack subjective research on the mechanisms and pathways between internet use and the elderly's SWB.Secondly, the factors used in previous relevant research were based on empirical selection, with relatively little data screening process.
And most of the research are mainly qualitative in nature, lacking quantitative research.Finally, the feature analysis used in previous research was linear in its approach to research, but there are intrinsic links and interactions between factors and non-linear associations which are often overlooked in traditional linear approaches, making it di cult to dig deeper into the non-linear effects of factors on the elderly's SWB, and the number of factors tested is relatively limited.These non-linear relationships are important for researching the complex determinants behind the subject.
Numerous previous studies have used traditional regression models that combine multiple factors driven by researchers' intuition to study how different internet use factors affect SWB of the elderly, which may result in in uencing factors being arti cially ignored.Although traditional regression models can reveal the contributions of variables, they are not the optimal choice for studies of nonlinear relationships among factors.In comparison, machine learning models are more effective for studies of nonlinear relationships among factors, but most of them cannot estimate the contribution of each variable (Zhang et al., 2022).This study aims to overcome this limitation by exploiting the interpretability of machine learning models in combination with traditional regression methods.Compared to traditional regression methods, machine learning methods have the following advantages: machine learning methods, especially non-linear machine learning methods, can establish nonlinear relationships between independent and dependent variables in a complex system environment, thus exploring the nonlinear effects between variables more comprehensively.And the computational and solving power of machine learning methods is signi cantly increased, allowing the impact of more variables to be examined (Chen et al., 2021), as shown in Fig. 1.
A comprehensive survey of internet use and skills of the elderly could provide more speci c and practical support, and it's also essential for a comprehensive understanding of how the elderly can bene t from technology (Hunsaker and Hargittai, 2018; Neves et al., 2018).To explore the impact of internet use on the elderly's SWB, we reidenti ed the factors in uencing internet use on the elderly's SWB using a nonlinear machine learning methods and traditional regression models based on the latest data from CFPS 2020, incorporating features used in previous research, and nally applying traditional methods for validation.
The remainder of the paper is structured as follows: Section 2 reviews the relevant literature; Section 3 describes the methodology, including a description of the data sources, factors, and research methods; Section 4 details the empirical results; Section 5 is the discussion section and practical implications; and Section 6 concludes the research.

China family panel studies (CFPS)
The CFPS is a nationally representative longitudinal survey administered by the Institute of Social Science Research at Peking University, covering 25 provinces in China, and it is a nationally representative survey of individuals, households, and communities in China.The CFPS survey re ects the social, economic, demographic, education and health changes in China and provides a data base for academic research and public policy analysis.
The survey provides a wealth of interesting information and has been used to examine a wide range of issues.For example, the impact of air pollution on healthcare use and healthcare cost in China (Liao et al.,2021); health effects of energy poverty (Zhang et al., 2019), or the relationship between Mandarin and health (Wang et al., 2019); the relationship between energy poverty and SWB (Nie et al., 2021), even between energy poverty and children's SWB (Zhang et al., 2021), or between income inequality and SWB (Zhang and Awaworyi Churchill, 2020);rural households' borrowing behavior (Zhang and Wang, 2022); poverty rates (Zhang et al., 2014); the impact of government spending on intergenerational mobility (Tang et al., 2021); the relationship between inequality of opportunity and family education expenditure (Song and Zhou, 2019) or risk asset investment (Song et al., 2020); there are also research on internet use and subjective well-being of the elderly (Li and Zhou, 2021).

Research on internet use and SWB of the elderly
Studies in China related to internet use and SWB of the elderly have used panel survey data, such as Chinese General Social Survey (CGSS), China Health and Retirement Longitudinal Research (CHARLS), and China Family Panel Studies (CFPS).Most studies have used traditional regression models, which have the advantage that they reveal the contribution of variables.Among them, ordinary least squares (OLS) regression models have been widely and successfully applied.In this paper, OLS models are also used for analysis and comparison.Increasing the frequency of internet use alone does not directly improve the SWB of the elderly, but a complex physical and mental health mechanisms and social are needed to effectively improve subjective wellbeing.

Liu and Xie CGSS2019 Regression analyses
There is a positive effect of the elderly's digital skills on their SWB, and the increase in the elderly's digital skills enhances their SWB.
Self-con dence was measured by the participant's con dence towards him/herself in the future.The answer ranges from 1 (very uncon dent) to 5 (very con dent).The percentage of the elderly with selfcon dence score of 3 and above in 2018 decreases from 91.8-69.5% in 2020.
Subjective well-being is calculated as: Mobile internet use was measured by the participants answering whether they used the internet through mobile devices (1 = yes, 0 = no).The percentage of the elderly with a life satisfaction score of 3 and above in 2018 increases from 11.7-16.1% in 2020.

Independent factors
Our research included 27 independent factors covering the factors used in previous research.ML combined with SHAP can provide a clear explanation of the different aspects of the mobile internet that affect the daily lives among the elderly, and enable researchers to intuitively understand the in uence of key features in the model.

Ordinary least squares (OLS) estimation
We used OLS model to explore the relationship between the digital divide and well-being in the elderly, which allows a more speci c analysis of the contribution between different factors.We apply OLS estimation based on the following model:

Evaluation criteria
To verify the forecasting accuracy of XGBoost, we adopt three main evaluation criteria: mean absolute error (MAE), Mean Square Error (MSE), and R-squared value (R 2 ).

Results
The overall framework of the article is shown in Fig. 2. We integrated the features utilized in previous studies and selected thematically relevant features.The features were ranked in importance based on interpretable machine learning models, and the selected features were analyzed and compared using machine learning models and traditional regression models.

Results of XGBoost-SHAP
The SHAP model was used to estimate additional Shapley values (SHAPVs) (i.e., importance scores) for the features and to explain the magnitude and direction of their effects.The left-hand side of Fig. 3. to Fig. 5. show the shape magnitudes for all variables based on the records, and illustrate the effect of the distribution of each variable on the explanatory variables.Red indicates high eigenvalues; blue indicates low eigenvalues.The absolute SHAPVs on the right-hand side of the gure explicitly rank the variables according to their importance.Using the results of the machine learning analysis, we selected keywords related to internet usage and ranked them by overall feature value, as shown in Fig. 6.
The selected correlation factors were predicted with the SWB of the elderly and found that the determination coe cients could reach 0.87, MAE of 0.5, and MSE of 0.528 (Table 5).In the regression results, the internet for work, the internet for learning, watching online short videos, internet for daily life, and internet for leisure and entertainment all had negative effects on SWB of the elderly, while the use of WeChat, contact with family and friends, and internet as an information channel were all positively correlated, and the positive correlation coe cient of internet as an information channel was over 0.5.

XGBoost-SHAP results
The method of selecting keywords based on feature importance was used in this paper, combining the use of XGBoost and SHAP to study the relationship between internet use and the SWB of the elderly, which has difference from previous studies.This research may contribute to an enhanced understanding of the manifold in uences of internet use on the SWB of the elderly.
Figure 7. illustrates the important factors based on XGBoost-SHAP.Results indicate that the factors order of importance from strong to weak is the internet as an information channel, contact with family and friends, and the Internet for work, the internet for leisure and entertainment, internet for daily life and using WeChat, watching online short videos and internet for learning.
The results of the online big data research showed that the focus of the digital divide among older adults was mainly on daily life (QR codes, Alipay, healthcare, insurance, etc.) and social media (WeChat) (Yuan and Jia, 2021).The results of panel data analysis showed that all these related factors have an impact on the SWB of the elderly.The empirical results of the panel data in this paper have some similarities with the results of previous studies on the digital divide among older adults using online big data.
The internet as an information channel as the most important factor has the greatest impact on the SWB of the elderly.Access to information is one of the main motivations for the elderly to use the mobile internet (Hur, 2016).According to Jin and Zhao's research (2019), 85% of the elderly read news via the internet, demonstrating the importance of the internet as an information channel.In the digital information age, the elderly need a wealth of information to help keep up with the news and events around them (Ijiekhuamhen et al., 2016).According to the latest survey data of China Internet Network Information Center (CNNIC), the utilization rate of internet news among senior internet users is 3.2 percentage points higher than that of internet users as a whole (CNNIC, 2022), which is the only application type used more by senior internet users, re ecting the circle characteristics of senior internet users in following current affairs and hot spots.
The utilization of internet for contact with family and friends also has a greater impact.For the elderly, the use of information and communication technology (ICT) is a low-cost, effective means of maximizing time spent with others and ultimately offsetting the considerable challenges to well-being encountered in the nal stages of life (Sims et al., 2017).The more the elderly interact with friends and relatives, etc., the more likely they are to use the internet to connect with each other due to its convenience, thereby increasing their use of the internet (Dong and Zhang, 2021).
The internet affects the extent to which the elderly are re-employed and thus affects their SWB.
Educational attainment has an impact on the re-employment pro le of the elderly, who may feel more empowered when they perform better in the job market (Hur,2016).The rapid popularity of online short videos also provides a platform for the re-employment of different groups of the elderly.In the important eld of online short video, the older population is placed in a new economic form of commercial capital and productive labor.They not only consume content in the online short video domain, but also present themselves through content production, producing digital products with great identities and distinctive style (Yuan, 2022).
The internet provides more leisure activities for the elderly.The research by Si (2021) showed that the internet has added new elements to the cultural life of the elderly, such as watching short online videos, playing online games, and watching TV dramas.Starting from the identity attributes of the elderly groups, they value the social function and entertainment function of the internet digital living space more.The current popular short videos and games meet the daily leisure and entertainment needs of the elderly groups.The application of the internet by the elderly has expanded from the past concentration on communication and information acquisition to the use of some convenient functions that seem exclusive to the young, such as watching videos, cell phone payment, cell phone navigation, and taxi services.These provide convenience for their daily life.
The rapid development of WeChat has made the social interaction of the elderly more convenient and has become the most important contact method for many people.As a social tool, Hu's (2020) research found that the elderly use WeChat to meet their needs for interpersonal interaction, information access and self-expression, effectively reducing their sense of isolation.At the same time, gender differences in the motivation for self-expression were evident, with women using WeChat more than men for selfexpression.And Zhang et al.'s (2021) research found that the willingness of the elderly to use WeChat to access health information is at a high level.
Internet use can facilitate learning for the elderly (Diniz et al.'s, 2020).As the elderly population in China is rising year by year, the elderly population is also gradually joining the wave of online learning.In the current situation, many elderly people's demand for education is gradually increasing.In the current social environment where various network terminals and platforms are inundated, the channels for elderly people to obtain information and the convenience of obtaining information have obviously increased (Wang, 2022).Access to information through short online videos is one of the main motivations for older adults to use the internet (Yang, 2021).
Elderly users have become an important source of increments for internet users, among which online short videos have become the main growth point of the elderly population's access to the network (Zhao et al., 2021).Online short-form videos provide a platform for the elderly to develop their hobbies and interests.As the elderly have more leisure time after retiring from their socio-professional roles, short videos meet the needs of the elderly to experience the world at large and learn about social dynamics.
Finally, in the context of the increasing demand for companionship, some elderly turn to short videos to nd supports due to a lack of emotion.On short video platforms, the elderly browse and post comments to make friends with like-minded people, relieving their feelings of loneliness and emptiness (Si, 2019; Cui et al., 2021).The widespread penetration of online short videos has changed the internet use behavior of the elderly.
The results of our analysis found that individual characteristic factors such as household registration type, gender and education have little effect on whether the elderly use the internet, while there are some effects on SWB of the elderly.But using the internet for learning has a greater effect on whether they use the internet, which deviates from the ndings of Peng et al. ( 2019) and Liu and Xie (2021).Our ndings support part of the ndings of Jiang and Chen (2021) that there is a correlation between the elderly's digital skills and SWB.However, their research shows that physical and mental health play a mediating effect between internet use and subjective well-being of older adults, but our results suggest that physical health status have little correlation with whether the elderly use the internet, but rather affects SWB of the elderly.Therefore, further research is needed to determine whether internet use contributes to improve physical and mental health among the elderly.The ndings are largely in line with those of Szabo et al. (2019).For the elderly who use the internet, the rst purpose is information, i.e., accessing information via the internet, along with social interaction, i.e., connecting with family and friends, and everyday use, such as scanning QR codes and online transactions, all have some in uences on their subjective well-being.

OLS regression results
The OLS regression results found that internet for work and learning, internet for daily life, watching shortform videos and internet for leisure and entertainment had a negative effect on the SWB of the elderly.
There is a more signi cant negative effect of watching short videos and internet for daily life.It could be attributed to the fact that since people's ability to work and learn gradually decreases with age, their willingness and motivation to learn gradually decreases as most elderly retire and leave the workforce, using the internet for learning activities may increase their burden and result in more psychological pressure, making them feel unhappy (Peng et al., 2019).The internet lowers the likelihood of reemployment among the elderly in urban China (Li et al., 2021).
As internet use becomes more in uential among the older population, they are acquiring more and more skills for daily life and recreation.Online shopping and online gaming also have an impact on their SWB, which is a point of concern.It is di cult for the elderly to fully grasp the information of products sold online, and their discrimination ability is limited, so they are more likely to encounter problems in online shopping, such as the physical object does not match the picture, the promotion method is too complicated, the e-commerce platform has not yet developed a mature and effective platform for the elderly groups, and some unscrupulous elements through online shopping to implement fraud on the elderly (Guan,2020).It seems that all the focus on the youth population who are addicted to online games.Outside of the mainstream, the phenomenon of older people being addicted to online games has been ignored intentionally or unintentionally.After retirement, the elderly have a lot of free time, and when their children are not around, they can waste the time through online games, which leads to the elderly being easily addicted to the internet (Wang, 2021).
According to CCNIC survey data, 84.8% of elderly internet users often use online short videos (CNNIC, 2022).Older people use short videos to upload their own life clips, share their daily life, get tra c, and earn income.However, "senior celebrities" is an emerging phenomenon in the communication eld of short video platform, and its development is still in the preliminary stage.The image of elderly people is often "scandalized" or "weakened" in the internet, bringing audiences a bad viewing experience.And older people are no longer restricted by working hours, and they have a lot of free time at their disposal (Yuan, 2022).Some reports show that elderly people spend a lot of time on short video platforms such as Tik Tok and Kwai, and have a high stickiness to these platforms.This can lead them to rely on watching short videos and become " digital addiction " (Chen,2021).

Comparing XGBoost-SHAP and OLS regression models
From Tables 5 and 6, it can be concluded that the R 2 values of determination coe cients for the complete data samples based on XGBoost and OLS regression are 0.866 and 0.353, respectively.The results indicate that the XGBoost model has higher R 2 values and better prediction results for the factors analysis compared to OLS regression.Probably, the reason is the inability of the OLS regression model cannot capture the nonlinearity between internet use factors and SWB of the elderly.
The SHAP model emphasizes the contribution of each feature to the corresponding prediction of the model and the global and local behavior gaining end-users trust in the ML approach.Furthermore, this study emphasizes that the use of interpretable machine learning does not inherently sacri ce accuracy or complexity, but rather enhances the predictions of the model by providing human-understandable explanations.
While traditional regression models are able to reveal the contributions of factors, they are not the best choice for factor analysis.In contrast, machine learning models are more effective for factor nonlinearity and importance analysis, but most of them cannot estimate the contribution of each factor.Combining the two and analyzing the factors together can uncover emerging and potential factors.

Practical implications
Based on our empirical results, we recommend that policymakers, researchers, and practitioners pay more attention to the impact of internet use on the elderly's well-being, and more attention should be paid to the increasingly important role of digital media in promoting healthy ageing (He et al., 2020).The following speci c insights are available to promote active aging.
Firstly, the government promotes the construction of information technology, improves and develops the supply of digital service facilities.Because of China's urban-rural dual structure, unbalanced regional economic development and the "digital divide" phenomenon, there are large differences in the extent and cost of internet infrastructure construction between regions, so internet resources should be tilted to less developed economic regions such as rural and western regions, and the construction of internet infrastructure in less developed regions should be accelerated.
Secondly, we should strengthen internet training for the elderly and customize courses according to their needs so that they can master the basic skills.to provide them with basic internet skills.Compared with young people, older people have less learning ability and energy.Compared with young people, older people have relatively weaker learning ability and energy, so Therefore, the process of using the internet is a bit tedious for them.Software can be improved to meet the special needs of the elderly and to simplify the navigation process to suit their needs.
Thirdly, the communities work together to promote active aging and provide more opportunities for digital participation, so that older people can play and show their self-worth and promote self-development on digital platforms.Public support can be effective in improving internet access for older persons (Tirado-Morueta et al., 2020).Governments and communities should improve digital infrastructure and provide greater internet access for the elderly (Sun and Zhou, 2021).The government should introduce corresponding policies to help the elderly broaden their digital participation channels, actively encourage them to learn new knowledge and skills through the internet, and to improve their human capital and promote the self-development of the elderly individuals.

Conclusion
This research uses the 2020 latest CFPS data to gain insight into the impact of internet use on the SWB of the elderly in China.we screened for factors of internet use that are not su ciently discussed in the relevant literature and combined ML and traditional regression methods to analyze.(1) Through a datadriven machine learning-based analysis approach, the empirical results indicate that the factors order of importance from strong to weak are the internet as an information channel, contact with family and friends, and the internet for work, the internet for leisure and entertainment, internet for daily life and using WeChat, watching online short videos and internet for learning.Several assumptions made in previous research regarding the impact of internet use on SWB in the elderly no longer apply.As technology advances and concerns about the digital divide among the elderly increase, individual characteristics factors such as household registration, education level, and gender are no longer the main reasons affecting the elderly' internet use.Instead, subjective factors, that is, the extent to which the elderly's bene t from internet use, such as the timeliness of access to information on the internet and the fact that the internet provides more entertainment activities for the elderly, etc., attract the elderly to better use the internet and thus increase their subjective well-being.The use of keywords in CFPS also needs to be dynamically adjusted as the dynamics of the problem change.
The limitation of this paper is the lack of options in the dataset related to whether the COVID-19 has affected subjective well-being.And the absence of network big data makes it di cult to dynamically issue this topic.
The introduction of interpretable methods in the machine learning black box when we use machine learning methods to analyze in uencing factors, which helps to further investigate the impact of different factors and the combinatorial mechanisms behind them.This sets an essential precedent for future research and selection of factors with more complex variables.Because of China's large population and uneven development, a survey of internet usage among the elderly on this basis is instructive and can provide references for overall telecommunications internet access.

Declarations Figures
Article     Features sorted by SHAP (Self-con dence).
Selection of mobile internet related features.
prediction accuracy of XGBoost is signi cantly better than other ML methods(Hamilton et al., 2019).The basic concept of XGBoost is to perform a second-order Taylor expansion on the objective function, train a tree model using the second-order derivative information of the function, and use a decision tree as a weak classi er to iteratively modify the residuals of the previous model.The complexity of the tree model is added as a regular term to the optimization objective to control the complexity of the tree in order to avoid over tting and simplify the model.According to Mo et al. (2019), its output function computes as follows: Traditional feature importance in XGBoost shows only which features are important, which does not show the impact of each feature on the prediction results.In other words, traditional feature importance primarily indicates the contribution of global features without any local interpretability.Local interpretability helps visualize the prediction trajectory as part of the prediction model process.The SHAP algorithm bridges the gap between global and local interpretability.Shapley additive explanations (SHAPs) is based on game theory and was originally proposed by Shapley in 1953(Shapley,1953).As a feature imputation method, this technique can explain the contribution of features to the model output of complex trained model.In addition, for each test sample, the model generates a predicted value so that the framework can provide interpretable analysis.It can explain the contribution of a speci c input (X) by calculating the impact of each feature on the results.The estimated Shapley value is calculated as follows:SHAPs are used to improve the interpretability of the results.the goal of SHAP is to explain the importance of features for the measured values by calculating the contribution of each feature to the measured values(Molnar, 2020).Lundberg et al. (2018) proposed a TreeSHAP for gradient enhancement models that included XGBoost.The TreeSHAP framework assigns consistent feature importance locally to each individual prediction of a single unit and provides a rich visualization of each feature attribute, which is an improvement over the classic feature importance and partial dependency graphs.According to Lundberg and Lee (2017), the TreeSHAP interaction values can be estimated as follows:

( 2 )
The OLS regression results found that the factors have a negative impact on the SWB in the elderly from strong to weak are watching short-form videos, internet for daily life, internet for learning, internet for leisure and entertainment and internet for work.(3) The interpretability of machine learning models can analyze the importance of research factors, having higher determination coe cients (R 2 ) and better prediction results.Traditional regression models can analyze the contribution of research factors.The combination of machine learning and traditional regression methods have the ability to tap emerging and potential factors. Framework.

Figure 2 Factors
Figure 2

Figure 3 Features
Figure 3

Figure 4 Features
Figure 4

Table 1
Research on internet use and subjective well-being of the elderly

Table 2
De nition of dependent variables.

Table 3
De nition of independent variables.XGBoost) is an extensible tree boosting system, which is a novel implementation of the gradient boosting decision tree (GBDT) ensemble algorithm, developed by Chen and Guestrin (2016).The algorithm has fast parallelism, controlled complexity, fault tolerance, and generalization capability; the predicted values are close to the actual values.It is veri ed that the data

Table 5
The metric values of the XGBoost.

Table 6
summarizes the OLS regression results for mobile internet use in the elderly, with determination coe cients of 0.353, and the model ultimately explains 35.3% of the total variance.

Table 6
Regression results with OLS.