Information Searching from New Media and Households’ Investment in Risky Assets: New Evidence from a Quasi-Natural Experiment

Abstract

In 2010, Google withdrew from mainland China unexpectedly, which is an important issue that significantly changed the information acquisition environment in China. After that, Baidu has dominated a search engine in China, which provides less informative results. We use Google’s withdrawal from mainland China as a quasi-experiment and the data from Chinese General Social Survey (CGSS) to test the relationship between the information searching in new media and household investment in risky assets. By using the difference-in-difference method, we find that Google’s withdrawal from mainland China significantly decreased households’ willingness to invest in risky assets. The results are robust after parallel trend test, PSM–DID, entropy balancing, placebo test, as well as changing the control and treatment group, using a Logit model and excluding other factors. As for the heterogeneity, the effects are different among females and males, rural and urban residents, and residents with different incomes. As for the plausible channels, we find that Google’s withdrawal from mainland China significantly affected firms’ information disclosure quality, the convenience of getting information and the risk preference, by which their investment behaviors are affected.

1. Introduction

1.1. Research Motivation

There is a puzzle that the household participation rate in the financial market is significantly less than theoretical model predictions, which is known as “Limited Participation Puzzles”. The existing literature explains the puzzles from multiple perspectives, including personal characteristics, cost, human capital, social capital and so on [1,2,3,4,5,6,7]. Among these perspectives, cost is one of the most important factors. Furthermore, prior research has found that the high information cost has become a main obstacle that decreases households’ willingness to invest in the stock market [8,9,10]. Therefore, reducing the cost of acquiring information can significantly encourage households to participate in the financial market.

The ability to access information at lower cost is one of the key features for new media. The existing literature has found that the obtaining information via the internet has significantly affected investors and financial markets [11]. Prior research has found that obtaining information from the internet can affect investors’ behavioral bias and ability to process information with more effectiveness [12,13,14]. The internet, as one form of new media, is an effective tool to get information. By using new media, especially using the internet to search for information, people can get larger amounts of information with lower costs and higher efficiency. Information obtained from new media can be divided into two categories. On the one hand, people can use new media to get information willingly. For example, people can search for the firms they are interested in to acquire knowledge of news and announcements, anywhere and at anytime. On the other hand, firms can use new media to disclose more effective information to the public. As The White Paper on Information Quality of China’s Capital Market and Information Transparency Index of Listed Companies in 2021 shows, besides disclosing information through announcements, listed firms can also disclose information through performance briefings, official websites, social media and news media. In 2021, the proportion of information disclosed through performance briefings and news media decreased significantly, but the proportion of information disclosed through social media increased, which reflects that listed firms are paying more attention to using social media, one of the new media, as an information disclosure channel. In addition, there are an increasing number of websites or applications for listed firms to disclose information, such as WIND, EastMoney and so on. The Shanghai and Shenzhen exchanges in China also use their websites to help firms to disclose information. Obtaining more information easily and effectively reduces the cost of getting information, which may affect households’ investment behaviors and decisions. Furthermore, relying on the new media, especially the internet, the goals of sustainable development can be achieved, because new media can help people get information more equally and with the new media. Additionally, information about sustainable development can be spread more rapidly and broadly by using new media, which is helpful in realizing sustainable development.

The role of new media, taking the internet as an example, can be understood from two aspects. On the one hand, the internet is one of the most important new media. By using the internet, we can get information at lower cost and with more efficiency. On the other hand, the new media relies on the technology of the internet. The internet provides technology support for new media. New media need the internet to function well. To be specific, information searching and information dissemination from new media need the internet. Therefore, using the internet is closely connected with new media.

1.2. Research Design

However, there exists a challenge in identifying the relationship between searching for information from new media and households’ investment in risky assets. The decision to participate in investing in financial markets and using new media to get information is made by households or investors themselves. Moreover, households using the internet may have higher incomes and may have more willingness to invest in risky assets. This phenomenon may cause endogeneity, resulting in inconsistent estimation.

In 2010, Google unexpectedly withdrew their searching business from mainland China due to the failure of negotiations with the Chinese government regarding its censorship of Google’s search results [15]. After that, Baidu has become a dominant search engine in mainland China. However, Baidu is notorious for manipulating its search results [16]. Xu et al. also find that, compared with Google, the information provided by Baidu is less informative [15]. Therefore, following Xu et al. [15], we use Google’s unexpected withdrawal from mainland China as an exogenous shock to test the relationship between information acquisition and households’ investment in risky assets, which can alleviate the influence of potential endogeneity.

To be specific, we use the sample of the Chinese General Social Survey (CGSS) from 2010 to 2017 and Google’s unexpected withdrawal from mainland China as an exogenous shock to test the relationship between information acquisition and households’ investment in risky assets. By treating Google’s unexpected withdrawal from mainland China as a quasi-experiment and using the difference-in-difference method, we test the relationship between information acquisition and households’ investment in risky assets, which provide the empirical evidence for the “Limited Participation Puzzles”.

1.3. Conclusions and Contributions

We find that Google’s unexpected withdrawal from mainland China makes households less willing to participate in financial markets. The effects are different among males and females, high-income and low-income residents, and rural and urban residents. The potential channels show that Google’s withdrawal affects firms’ information disclosure quality, households’ willingness to get information and their risk preference, which leads to a reduced willingness to participate in financial markets.

Compared with the existing literature, the potential contributions may include: (1) we use Google’s unexpected withdrawal from mainland China as an exogenous shock to test the relationship between information acquisition and households’ investment in risky assets, which alleviates the potential endogeneity to some extent. Therefore, we can get a causal relationship of these two relations. The existing literature mainly focuses on the relationship between internet use and investing behavior. In addition, studies always analyses the question from a forward perspective; that is to say, they always check whether investing behavior would be affected when households use the internet. We use Google’s unexpected withdrawal from mainland China as a quasi-experiment—that is to say, that households are negatively affected in acquiring information—to test the relationship. In this way, we can find the importance of information acquisition brought by new media; (2) we test the causal relations that affect households’ investment behaviors and check the importance of getting information for households, which may be helpful in realizing the goal of sustainable development.

We organize the remainder of the paper as follows: Section 2 provides a literature review and develops hypotheses; Section 3 introduces data, variables and the empirical strategy; Section 4 presents results and robustness checks; Section 5 provides further discussion on some plausible channels; and Section 6 concludes the paper.

2. Theory and Hypothesis

2.1. New Media and Information Acquisition

New media are forms of communication which apply digital technology to provide users with information and services through computer networks, wireless communication networks, satellites and other channels, as well as computers, mobile phones, digital television and other terminals [17]. From the perspective of location and time, the definition of new media is compared with the traditional media. Supported by digital compression and wireless network technology, new media can move beyond geographical boundaries and time limitations, and can cover global information by taking advantage of its large capacity, real time and interactivity.

There are several significant features of new media compared with traditional media. The communication channels are diversified [18]. Due to technical limitations, the communication channels of traditional media are very simple. New media, based on the internet, can rely on a variety of information terminals for information dissemination. The speed of information dissemination is much faster. Nowadays, with the continuous development of technology, the internet has gradually become used by every household. The information that relies on new media is spread at an amazing rate [19]. As long as households can use the internet, once the information is released on new media, households can immediately obtain it through computers, mobile phones, digital TVs and other digital equipment. Moreover, information in new media is highly interactive. Compared with traditional media, with which information only can be spread in one direction, new media has its own unique advantages in interaction. In the process of information spread in new media, households can communicate with information owners and can use the information they are interested in to recreate, which is helpful in further spread. For example, households can obtain information through popular ways, such as Weibo and WeChat official accounts. In this way, households can publicly express their opinions anytime and anywhere, as well as communicate with others from all over the world, and forward them to friends and relatives. These behaviors have changed the geographical and cultural restrictions caused by traditional media and greatly enhanced the initiative and participation in obtaining information. In addition, some new media with stronger interactivity, such as interactive movies, interactive animation and even game media, have added interactive design to media, which enables households to feel more immersed when using new media. The information that owners want to spread would be become more profound, which can be accepted and understood by households more easily.

According to the features of new media, the advantages of information spread by using media include timeliness, interactiveness and more freedom [20]. For the timeliness, information in new media can be updated immediately once it is released. Relying on internet technology, the speed of information spread is faster and faster. For example, through online new media platforms, we can listen to and watch the news broadcast live to understand the latest policies and arrangements, so that households who are not on the scene can get the latest information the first time. For the interactiveness, information in new media is always not in a single direction. Households can give comments and interact with others when obtaining information. More information is produced during the interactive process and the information can be understood more profoundly. For freedom, based on the support of internet technology, new media provides the households with more freedom. On the one hand, information spread through new media has the feature of large capacity, being updated at a fast speed and in rich presentation forms. These features enable households to choose more information with more controllable ability when obtaining information. On the other hand, by using new media, information spread has a wider scope and deeper penetration. Under the circumstance of the high pressure of people in contemporary society, as well as less available time, households can use new media to obtain the required information anytime and anywhere by using fragmented time, which helps them get information with more flexibility and freedom.

2.2. Factors That Affect Households’ Investment in Risky Assets

The theoretical foundation of households’ investment in risky assets is the portfolio theory proposed by Markowitz [21]. The idea of the portfolio theory is that households aim to maximize their expected return with minimum risk. In order to realize their objective, they adjust different asset portfolios, including risky assets and risk-free assets [22,23]. However, there is a puzzle that the reality of households’ investment in risky assets is different from the predictions when using the theoretical model. One of the issues is that, in reality, the participation rate in financial markets is significantly less than the model predictions, which is called “Limited Participation Puzzles” [24]. Prior research has discussed the reasons from different factors. Generally speaking, there are multiple factors that may affect households’ investment behaviors, including personal factors, psychological factors, human capital factors, social capital factors and cost factors. For the personal factors, existing literature finds gender, age, income, marital status, health conditions and so on [2,3,25]. For the psychological factors, some researchers find that happiness and risk preference would affect the determination of household investment in risky assets [26]. For the human capital factors, the existing literature always focuses on the effects of generalized and specific human capital perspective. The generalized human capital perspective includes intelligence, education background, cognitive ability and non-cognitive ability [4,5,6]. Specified human capital mainly focuses on the experience or knowledge of finance, such as financial education and financial literacy [27,28,29]. The social capital factors focus on the social ties or interactions with others that may have influence. As Ellison and Fudenberg [30] state, social interactions are behaviors that the subjects learn from others by observing their behaviors. When they cannot make a decision on choice of financial assets, they may observe what others choose to determine for their own portfolio. The role of social capital or social interaction can be seen as the information acquisition and bandwagon effect, indicating that the person with specific social ties may intend to have similar topics, similar behaviors and similar portfolios [31,32,33]. Liang and Guo [34] find a positive relationship between the frequency of visiting friends and participating in investing in stock markets. As for the cost factors, the main idea is that the cost of acquiring information would significantly affect households’ investment decisions. To be specific, the costs in financial markets mainly include the information acquisition cost and trading cost [35,36]. For the information acquisition cost, the likelihood of participating in financial markets is highly related to the information that can be obtained by the households [24]. Haliassos and Bertaut [37] find that one of the main reasons for households not choosing risky assets rather than risk-free assets is that there exists a high cost in acquiring information. Bertaut [7] also believes that the reason why households do not invest in stocks is that the information cost for households participating in stock markets is much higher than the expected returns that can be obtained from the stock market. The high information cost has become a main obstacle that decreases the households’ willingness to invest in stock markets [8,9,10]. The trading costs not only include different fees and taxes when trading risky assets, such as in China, but when trading stocks, investors need to pay stamp duties and commissions. Other trading costs are not directly related to trading, such as the transportation fee to the stock exchange and the time cost that is spent in selecting stocks.

2.3. New Media Use and Households’ Investment in Risky Assets

Typically, there are two effects of new media use on households’ investment in risky assets: information acquisition and social interaction.

2.3.1. Information Acquisition

From the information acquisition perspective, the functions of new media can be analyzed from information demand and information supply.

For the information demand, new media help the investors to get information with lower costs. As Barber and Odean [11] point out, the traditional way to get information on financial markets is to consult to the professionals and experts. Households need to spend time and money obtaining this information. However, when media appears, it becomes much easier to get information. The households can use newspaper, radio and TV. Although traditional media make it easier to get information, traditional media such as newspaper and TV are always limited within a specific area and period. Households can only get information for a specific site within a specific time. For example, when we read newspapers, the information is always out of date because the affairs happened before the newspaper was published. When we watch TV, we need to turn on the TV at the exact time, or we may miss the information disclosed by TV. Also constrained by the entities, the information disclosed on traditional media is limited. When new media appears, one of the main roles is obtaining information. Compared with traditional media, information on new media is richer and more timely. Almost all the newest information can be obtained by the public from new media when disclosed. In addition, information on the new media is beyond location and time limitations. For example, when we enter a word on the internet, one of the new media, a huge amount of information can be seen by us anytime and anywhere. The media provide information to financial markets [38,39,40,41]; therefore, the lack of media coverage limits the information availability for individual investors. The quantity and timeliness of the information from new media reduce the cost of getting information for households [24]. A household can get the related information in a timely and easy fashion, such as the latest news for the listed firms, announcements, economic data and new policies, which may affect the price or volatility of the risky assets. Moreover, with the development of technology, some new technologies are applied in new media, such as interaction technology, AI and VR. All of these new technologies help households have better awareness of the information. Not only information acquisition, but also information understanding may make the households have more willingness to invest in financially risky assets. Furthermore, because there exist many financial analysts and financial institutions in new media, households can get these reports and opinions more conveniently. The professional information alleviates the information asymmetry to some extent and reduces the cost of acquiring information. Besides that, financial analysts intend to provide optimistic opinions [42,43], especially in China [44,45]. If households can use new media to get more information, it can help them mitigate additional biases introduced by analysts [15]. Hence, new media use can provide households with more various information at a lower cost.

For the information supply, the listed firms can use new media to disclose information in a more timely fashion. For example, without new media, if the listed firms would like to disclose some information, they have to rely on the newspaper. By using new media, the listed firms can release information, including news, financial reports and important affairs, on their websites, official accounts and some financial websites. The public can have a better knowledge of this information quickly and timely. In addition, the cost of the listed firms using new media to release information is lower. Some researchers have found that internet searching makes the firms have better information disclosure quality [15]. Better information disclosure quality results in households getting information at lower cost. Stocks are one of the main risky assets in China, which are sensitive to the information. More information at a higher quality may make households invest in stock market [46]. Overall, we suppose that new media use can reduce information acquisition cost and make firms have better information disclosure quality, which may increase willingness to invest in risky assets.

To be specific, as mentioned above, Google’s withdrawal from China makes it harder for households to obtain information [47]. Baidu is notorious for manipulating its search results [16]. Xu et al. also find that, compared with Google, the information provided by Baidu is less informative [15]. As Xu et al. point out [15], when they searched all the 3694 A-share listed firms in 2019 from Google and Baidu, respectively, by analyzing the first three pages, they find there exists 133,057 and 108,496 results from Baidu and Google, respectively. Although the volume in Baidu is higher than the volume in Google, Baidu shows results for advertisements with a larger proportion. To be specific, advertisements account for 16.71% of search results from Baidu, while the proportion from Google is only 4.81%, because ads inserted into the search results can benefit Baidu [48]. Therefore, the information from Baidu is less informative. Moreover, for the information diversity, Xu et al. [15] find that the results of the top five websites from Baidu account for 55.72% of the total results, while the value for Google is only 18.77%. Google provides a broad coverage, with more meaningful information, so the high concentration of the results from Baidu on a few websites constrains the effectiveness of information obtained by investors. Moreover, Google is a global search engine; people may interact with other investors around the world, which may influence their opinions or risk preference. Overall, Google is not a complement but a substitute for Baidu [15]. Without Google, the cost of getting information from new media or the internet may increase. In addition, due to the difficulty for households to acquire information, firms may not intend to disclose timely and accurate information to the public. Worse information disclosure may reduce the likelihood of households to participate in financial markets.

2.3.2. Social Interaction

From the interaction perspective, traditional interaction mainly is based on face-to-face interaction, which is constrained by time and space. By using the new media, households can interact with each other in some forums or on internet platforms. The interaction with each other has two functions. On the hand, it may make information dissemination much faster. On the other hand, because of the convenience of interaction via new media, the bandwagon effects may affect investment decisions. When households find that others invest in stocks when they are chatting via the internet, they may also tend to buy the same stocks. Moreover, interaction via new media may affect households’ risk preference in changing investment decisions. Wasiuzzaman and Edalat [49] find that the frequency of using social networks is positively related to the individuals’ risk preference. Higher risk preference promotes households investing in risky assets. Overall, we consider that new media use may increase households’ interactions and then affect their investment behavior.

To be specific, by using Google, households can more easily to get in touch with others from outside of mainland China, which indicates that people can invest in broader networks of friends. When Google withdrew from mainland China, households may have found it harder to broaden their networks of friends. Moreover, Google is one of the new media, and Google’s withdrawal makes residents lose one channel of social interaction. Therefore, social interaction may be negatively affected after Google’s withdrawal.

Therefore, from the perspective of information acquisition and social interaction, we propose Hypothesis 1, as follows:

Hypothesis 1.

Google’s withdrawal from China makes households participate less in financial markets.

3. Methods

3.1. Sample

We use samples from the CGSS conducted in the years 2010, 2012, 2013, 2015 and 2017. The CGSS, as one of the national, comprehensive and continuous Chinese survey projects initiated in 2003, conducts a cross-sectional survey of more than 10,000 households among provinces in mainland China each year. However, only the surveys during the years 2010–2017 have detailed information about the respondents’ investment behaviors, as well as their internet use. Although the data of the CGSS is cross-sectional, we can append them together due to the random sampling around China.

3.2. Variables

The dependent variable is a dummy variable indicating whether the respondents invest in risky assets, including stocks, funds, bonds, futures, options, real estate, foreign exchange and others. If the respondents invest in risky assets, the variable Invest is coded as 1; otherwise, it is 0.

The independent variable is Treat_it × Post_t due to the difference-in-difference strategy applied. The variable Treat is used to identify the treatment group and control group. According to the surveys, the treatment group is defined as households who use the internet sometimes, often and always, and the control group is defined as households who use the internet rarely and never. For the treatment group, the variable Treat_it is defined as 1; otherwise, it is 0. As Xu et al. [15] state, the affairs that Google unexpectedly withdrew its searching business from China in 2010 can be treated as a quasi-natural experiment. For the CGSS (2010), the variable Post_t is defined as 0, indicating the pre-treatment period; otherwise, it is coded as 1, indicating the after-treatment period. In fact, the CGSS (2010) was mainly conducted in 2009, during which the residents in China could use Google to search for information; therefore, it is reasonable to treat the CGSS (2010) as a pre-treatment period.

According to the existing literature [50,51,52], we mainly use individual-level control variables that may affect residents’ decisions on investment, including gender (Gender), education level (Edu), health conditions (Health), nationality (Nation), age (Age), square of age (Age²), the area of house (LnHouseArea), income (LnIncome), the feeling of happiness (Happiness), the feeling of fairness (Fair), the feeling of taking advantage of others (Advantage), social class (Social Class), religions (Religion), as well as marital status (Marriage fixed effect) and political status (Political fixed effect). The individual-level control variables are all from CGSS, and the response for rejecting, not knowing and not applicable are treated as missing values. The definitions of the variables are shown in

Table A1. The definition of the variables.

Variables	Definition
Invest	a dummy variable indicating whether the respondents invest in risky assets, including stocks, funds, bonds, futures, options, real estate, foreign exchange and others. If the respondents invest in risky assets, Invest is coded as 1; otherwise, it is 0
Treat	a dummy variable to identify treatment group and control group. If the respondents use the internet sometimes, often and always, it is coded as 1. If the respondents use the internet rarely and never, it is coded as 0
Post	for the CGSS (2010), it is coded as 0; otherwise, it is coded as 1
Gender	a dummy variable. Males equal 1 and Females equal 0
Age	it equals to the conducting year minus the birth year
Age2	the square of Age
LnArea	the area of the respondents’ house. It is winsorized at upper and lower 1% and taken logarithmically
LnIncome	the total income of the respondents. It is winsorized at upper and lower 1% and taken logarithmically
Edu	from illiteracy to graduate or above are coded from 1 to 14, respectively
Health	unhealthy, relatively healthy, general, relatively healthy and very healthy is coded from 1 to 5, respectively
Trust	from negative responses to positive responses are coded from 1 to 5, respectively
Advantage	from negative responses to positive responses are coded from 1 to 5, respectively
Fair	from negative responses to positive responses are coded from 1 to 5, respectively
Happiness	from negative responses to positive responses are coded from 1 to 5, respectively
Social Class	From the lowest to the highest are coded from 1 to 10, respectively
Nationality	the Han equals to 1 and others equal to 0
Religion	non-believers are coded as 0 and believers are coded as 1
Marriage	fixed effect dummies of unmarried, cohabitation, first marriage with spouse, remarriage with spouse, separation as divorce, divorce, and widowhood
Political Status	fixed effect dummies of the masses, the members of the Communist Youth League, the members of Democratic parties, the members of the CPC

.

The summary description of the variables we used are shown in

Table 1. Summary description.

Variables	Obs	Mean	S.D	Min	Median	Max
Invest	51,497	0.089	0.285	0	0	1
Treat	51,548	0.374	0.484	0	0	1
Post	51,642	0.808	0.394	0	1	1
Treat × Post	51,548	0.324	0.468	0	0	1
Gender	51,642	1.503	0.500	1	2	2
Age	51,642	49.560	16.100	19	49	86
Age2	51,642	2716	1643	361	2401	7396
LnArea	51,642	4.514	0.624	2.708	4.500	6.215
LnIncome	51,640	8.415	3.378	0	9.616	12.61
Edu	51,642	4.942	3.076	1	4	14
Health	51,642	3.586	1.093	1	4	5
Trust	51,642	3.453	1.029	1	4	5
Advantage	51,642	3.043	1.071	1	3	5
Fair	51,642	3.077	1.062	1	3	5
Happiness	51,642	3.820	0.845	1	4	5
Social Class	51,642	4.204	1.694	1	4	10
Nation	51,593	0.082	0.274	0	0	1
Religion	51,632	0.118	0.323	0	0	1

. The comparison of the variables for control groups and treatment variables are shown

Table 2. The comparison of control and treatment group.

	Control Group		Treatment Group
Variables	Obs	Mean	Obs	Mean	Diff
Invest	32,156	0.031	19,254	0.188	−0.157 ***
Gender	32,256	1.516	19,292	1.481	0.035 ***
Age	32,256	56.17	19,292	38.50	17.669 ***
Age2	32,256	3349	19,292	1655	1693.676 ***
LnArea	32,256	4.548	19,292	4.458	0.089 ***
LnIncome	32,254	7.988	19,292	9.125	−1.137 ***
Edu	32,256	3.579	19,292	7.225	−3.646 ***
Health	32,256	3.340	19,292	3.997	−0.657 ***
Trust	32,256	3.520	19,292	3.343	0.177 ***
Advantage	32,256	3.016	19,292	3.089	−0.073 ***
Fair	32,256	3.156	19,292	2.945	0.211 ***
Happiness	32,256	3.768	19,292	3.911	−0.143 ***
Social Class	32,256	4.019	19,292	4.513	−0.494 ***
Nation	32,222	0.092	19,279	0.063	0.029 ***
Religion	32,248	0.132	19,290	0.094	0.038 ***

Notes: (1) the significance of the difference is tested by t test with unequal variance; (2) *** indicates statistical significance at the 1% level.

. The correlations of the variables are shown in

Table 3. The correlation of the variables.

	Invest	Treat	Post	Treat × Post	Gender	Age	LnArea	LnIncome	Edu	Health	Trust	Advantage	Fair	Happiness	Class	Nation	Religion
Invest	1
Treat	0.267 ***	1
Post	0.013 ***	0.115 ***	1
Treat × Post	−0.224 ***	0.896 ***	0.338 ***	1
Gender	−0.007	−0.034 ***	0.002	−0.025 ***	1
Age	−0.082 ***	−0.531 ***	0.052 ***	−0.463 ***	−0.022 ***	1
LnArea	−0.054 ***	−0.069 ***	0.035 ***	−0.059 ***	−0.005	−0.018 ***	1
LnIncome	0.133 ***	0.163 ***	0.028 ***	0.147 ***	−0.221 ***	−0.028 ***	−0.084 ***	1
Edu	0.322 ***	0.574 ***	0.024 ***	0.497 ***	−0.103 ***	−0.410 ***	−0.090 ***	0.242 ***	1
Health	0.066 ***	0.291 ***	−0.010 **	0.251 ***	−0.071 ***	−0.389 ***	0.040 ***	0.138 ***	0.269 ***	1
Trust	−0.014 ***	−0.083 ***	−0.034 ***	−0.080 ***	−0.018 ***	0.131 ***	0.039 ***	0.003	−0.031 ***	−0.003	1
Advantage	−0.008 *	0.033 ***	0.025 ***	0.041 ***	−0.044 ***	−0.038 ***	−0.016 ***	0.025 ***	0.004	−0.013 ***	−0.213 ***	1
Fair	−0.050 ***	−0.096 ***	0.038 ***	−0.073 ***	−0.007	0.131 ***	0.069 ***	−0.010 **	−0.074 ***	0.016 ***	0.314 ***	−0.177 ***	1
Happiness	0.051 ***	0.082 ***	0.029 ***	0.071 ***	0.019 ***	0.006	0.077 ***	0.061 ***	0.105 ***	0.231 ***	0.174 ***	−0.097 ***	0.283 ***	1
Class	0.130 ***	0.141 ***	0.044 ***	0.127 ***	0.030 ***	−0.039 ***	0.101 ***	0.121 ***	0.196 ***	0.199 ***	0.068 ***	−0.066 ***	0.148 ***	0.296 ***	1
Nation	−0.045 ***	−0.051 ***	−0.016 ***	−0.045 ***	0.008 *	−0.036 ***	0.071 ***	−0.025 ***	−0.052 ***	−0.011 **	0.017 ***	−0.010 **	0.038 ***	0.006	−0.004	1
Religion	−0.006	−0.057 ***	−0.011 **	−0.048 ***	0.074 ***	0.037 ***	0.031 ***	−0.015 ***	−0.087 ***	−0.044***	−0.014 ***	0.000	−0.007	0.015 ***	0.014 ***	0.232 ***	1

Notes: ***, **, * indicate statistical significance at the 1%, 5% and 10% levels.

. From Table 1, there are around 8.9% of respondents who invest in risky assets. From Table 2, we can find that the control group invests significantly less than the treatment group. From Table 3, the correlation between Invest and Treat_it × Post_t is significantly negative, which supports Hypothesis 1 to some extent.

3.3. Regression Model

As mentioned above, we treat the unexpected withdrawal of Google from mainland China in 2010 as a quasi-experiment. Therefore, the difference in difference (DID) can be used to test the relationship between the information search from new media and the willingness to invest in risky assets. Specifically, the baseline model is as follows:

$$Y_{it}={\alpha }_0+{\alpha }_1{Treat}_{it}+{\alpha }_2{Post}_t+{\alpha }_3{{Treat}_{it}\times Post}_t+{\alpha }_4{X}_{it}+{\omega }_j+{\mu }_t+{\epsilon }_{it}$$

where Y_it is the individual i’s investment decisions on risky assets at year t. Treat_it × Post_t is the variable we are interested in. X_it is a vector of control variables. ω_j and μ_t are the province fixed effect and year fixed effect, respectively. ε_it is the residual. Actually, because the dependent variable is a dummy variable, the Logit or Probit model could be used in the empirical analysis. However, we apply DID in the empirical analysis and the consistent estimation can only be obtained by using linear regression. Therefore, linear regression is used in the empirical analysis.

3.4. The Empirical Methods

The main methods we use include DID, PSM–DID, entropy balancing and a placebo test. The aim of these methods is shown in

Table 4. The main empirical methods.

Methods	Description
DID	A two-way fixed-effect model with the province fixed effect and year fixed effect
PSM–DID	Considering that the results may be affected by observable differences between treatment group and control group, DID with propensity score matching (PSM–DID) approach is applied
Entropy Balancing	Entropy balancing is a quasi-matching approach that weights each observation such that post-weighting distributional properties of treatment and control observations are virtually identical to ensure covariates balance
Placebo Test	To further address the concern that our DID results may be spurious instead of reflecting an actual effect of Google withdrawal, a placebo test is used.

.

4. Results

4.1. Baseline Results

Table 5. Information search from new media and risky assets investment.

	(1)	(2)
	Invest	Invest
Treat × Post	−0.041 ***	−0.041 ***
	(0.008)	(0.009)
Treat	0.153 ***	0.117 ***
	(0.007)	(0.008)
Post	−0.002	0.004
	(0.003)	(0.005)
Gender		0.008 ***
		(0.003)
Age		0.008 ***
		(0.001)
Age2		−0.000 ***
		(0.000)
LnHouseArea		0.011 ***
		(0.002)
LnIncome		0.001 **
		(0.000)
Edu		0.017 ***
		(0.001)
Health		−0.005 ***
		(0.001)
Trust		0.000
		(0.001)
Advantage		−0.004 ***
		(0.001)
Fair		−0.006 ***
		(0.001)
Happiness		0.002 *
		(0.001)
Social Class		0.008 ***
		(0.001)
Nation		−0.004
		(0.005)
Religion		0.009 **
		(0.004)
Constant	0.305 ***	−0.109 ***
	(0.009)	(0.021)
Marriage fixed effect	NO	YES
Political fixed effect	NO	YES
Province fixed effect	YES	YES
Year fixed effect	YES	YES
Obs.	58234	51,337
Adj_R2	0.140	0.184

Notes: (1) Robust standard errors in parentheses; (2) ***, **, * indicate statistical significance at the 1%, 5% and 10% levels.

reports the empirical result of the relationship between information search from new media and risky assets investment. Column 1 shows that, without any control variables, the coefficient of Treat_it × Post_t is significantly negative at 1% significant level. Column 2 shows that, with the whole control variables, the coefficient of Treat_it × Post_t is significantly negative at 1% significant level. All the results control the year and province fixed effect. The results indicate that Google’s withdrawal decreases the households’ willingness to invest in risky assets. The empirical results support the argument that information search from new media does affect the households’ investment behavior. When the information obtained from new media becomes less and harder, households would likely reduce the likelihood of investing in risky assets. Therefore, Hypothesis 1 is supported.

4.2. Robustness Checks

4.2.1. Parallel Trend Test

For DID analysis, in order to get a consistent estimation, the parallel trend should be satisfied. Because we only have one year of data before the treatment, the parallel trend cannot be tested by using traditional ways. In order to test whether the parallel trend is satisfied or not, we compare the difference of Invest between the treatment group and control group before and after the treatment. The results are shown in

Table 6. Parallel trend test.

	Before 2010		After 2010
Invest	Obs	Mean	Obs	Mean	Diff
Treatment Group	3256	0.223	18,779	0.180	−0.044 ***
Control Group	8459	0.030	27,918	0.031	0.001
Diff		0.194		0.149 ***

Notes: (1) the significance of the difference is tested by t-test with unequal variance; (2) *** indicates statistical significance at the 1% level.

. From Table 6, we can find that before 2010—that is to say, before the treatment—the difference between the treatment group and control group is not significant, indicating that without the treatment, the treatment group and control group are not significantly different. After 2010, with the treatment, the treatment group and control group are significantly different, at 1% significant level. Moreover, for the treatment group, the average value of Invest changes significantly before and after the treatment, while the average value of Invest is almost the same before and after the treatment for the control group. Therefore, we can conclude that the parallel trend is satisfied and using DID can get a consistent estimation.

4.2.2. Difference in Difference with Propensity Score Matching (PSM–DID)

Considering that the results may be affected by observable differences between treatment group and control group, the DID with propensity score matching (PSM–DID) approach is applied. Because to the data used are stacked cross-sectional data from 2010, 2012, 2013, 2015 and 2017, the PSM method would be used year by year in order to match the treatment group with the control group more accurately, to some extent. Taking the samples in 2010 as an example, at the beginning, a logit model is estimated in which the dependent variable is the treatment status indicator (Treat_it = 1), and the co-variate consists of all the control variables, including the province fixed effects, marriage fixed effect and political fixed effect. We perform the nearest-neighbor propensity score matching without replacement and the caliper of the matching is set at 0.01, indicating that the propensity score distance between two individuals can be no more than 0.01 if they are to be matched observations, to further assure that matched observations are similar in all dimensions. After that, we can get treatment group and matched control group for the year 2010. The same method is applied for the CGSS for the years 2012, 2013, 2015 and 2017. In this way, we can get treatment groups and control groups from 2010 to 2017. The variable Post_t still equals to 1 after 2010. The DID, the same as the baseline model, is estimated within the matched sample.

Table 7. The results of PSM–DID.

(A) Balance Test
Variables	Samples	Treat	Control	Bias	t-Test
Dependent Variables
Invest	Unmatched	0.188	0.031	52.100	62.640
	Matched	0.136	0.061	24.800	15.030
Covariants
Gender	Unmatched	1.481	1.516	−7.000	−7.720
	Matched	1.490	1.484	1.200	0.690
Age	Unmatched	38.503	56.170	−130.400	−142.140
	Matched	46.684	46.703	−0.100	−0.090
Age2	Unmatched	1655.500	3349.300	−123.400	−130.600
	Matched	2351.500	2352.400	−0.100	−0.040
LnHouseArea	Unmatched	4.458	4.548	−14.500	−15.790
	Matched	4.486	4.479	1.000	0.600
LnIncome	Unmatched	9.126	7.989	33.800	37.430
	Matched	8.929	8.921	0.200	0.150
EDU	Unmatched	7.224	3.579	136.700	158.830
	Matched	5.097	5.130	−1.200	−0.850
Health	Unmatched	3.997	3.340	64.400	68.910
	Matched	3.753	3.772	−1.800	−1.120
Trust	Unmatched	3.343	3.520	−17.200	−18.950
	Matched	3.380	3.370	0.900	0.530
Advantage	Unmatched	3.089	3.016	6.900	7.540
	Matched	3.117	3.121	−0.400	−0.230
Fair	Unmatched	2.944	3.156	−20.100	−21.950
	Matched	2.951	2.946	0.500	0.280
Happiness	Unmatched	3.911	3.768	17.300	18.610
	Matched	3.841	3.839	0.200	0.140
Social Class	Unmatched	4.513	4.019	29.700	32.300
	Matched	4.260	4.285	−1.500	−0.870
Nation	Unmatched	0.063	0.092	−10.800	−11.660
	Matched	0.074	0.072	0.800	0.480
Religion	Unmatched	0.094	0.132	−12.000	−12.930
	Matched	0.112	0.108	1.100	0.640
(B) DID with PSM Sample
		(1)		(2)
		Invest		Invest
Treat × Post		−0.038 ***		−0.037 ***
		(0.013)		(0.013)
Treat		0.108 ***		0.107 ***
		(0.013)		(0.012)
Post		−0.026 ***		−0.001
		(0.008)		(0.012)
Control Variables		NO		YES
Marriage fixed effect		NO		YES
Political fixed effect		NO		YES
Province fixed effect		YES		YES
Year fixed effect		YES		YES
Obs.		14192		14,182
Adj_R2		0.112		0.139

Notes: (1) Robust standard errors in parentheses; (2) *** indicates statistical significance at the 1% level.

shows the PSM–DID results. Panel A illustrates the results of the balance test. From Panel A, we can find that the t-values for all covariates are quite small after the PSM matching process, indicating that matching process is reasonable. Panel B delivers the DID estimates and we can find that the coefficient of Treat_it × Post_t is significantly negative for the model, both with and without control variables, which is similar with the baseline model, as shown in Table 5. Therefore, we can conclude that the results are robust.

4.2.3. Entropy Balancing

Following Hainmueller [53] and Chapman et al. [54], we use entropy balancing to do a robustness check, which is a quasi-matching approach that weights each observation such that post-weighting distributional properties of treatment and control observations are virtually identical to ensure covariates balance. All the control variables, the quadratic interaction term of the control variables and the cubic interaction term of control variables are used for matching. After that, DID is used among the entropy balanced sample.

Table 8. DID with entropy balancing.

(A) Balancing Test
Variable	Sample	Treatment Group			Control Group
		Mean	Variance	Skewness	Mean	Variance	Skewness
Gender	Before	38.790	176.000	0.650	54.720	217.100	0.117
	After	38.790	176.000	0.650	38.790	176.100	0.651
Age	Before	1.681	1.353	1.366	3.211	2.586	0.440
	After	1.681	1.353	1.366	1.681	1.354	1.369
Age2	Before	4.461	0.371	−0.163	4.540	0.396	−0.153
	After	4.461	0.371	−0.163	4.461	0.371	−0.163
LnArea	Before	9.132	12.240	−2.058	8.070	10.660	−1.768
	After	9.132	12.240	−2.058	9.132	12.250	−2.057
LnIncome	Before	7.150	10.290	0.310	3.885	5.605	1.445
	After	7.150	10.290	0.310	7.149	10.290	0.281
Edu	Before	3.982	0.812	−0.718	3.396	1.266	−0.285
	After	3.982	0.812	−0.718	3.982	0.812	−0.677
Health	Before	3.335	1.049	−0.626	3.511	1.054	−0.756
	After	3.335	1.049	−0.626	3.335	1.049	−0.575
Trust	Before	3.107	1.056	−0.076	3.013	1.189	−0.043
	After	3.107	1.056	−0.076	3.107	1.056	−0.111
Advantage	Before	2.964	1.062	−0.274	3.130	1.150	−0.378
	After	2.964	1.062	−0.274	2.964	1.062	−0.236
Fair	Before	3.908	0.575	−0.973	3.780	0.775	−0.952
	After	3.908	0.575	−0.973	3.908	0.575	−0.978
Happiness	Before	4.514	2.589	−0.209	4.055	2.933	0.134
	After	4.514	2.589	−0.209	4.514	2.589	−0.159
Social Class	Before	0.096	0.087	2.749	0.129	0.112	2.218
	After	0.096	0.087	2.749	0.096	0.087	2.749
Nation	Before	2.454	2.248	0.061	2.534	2.249	−0.045
	After	2.454	2.248	0.061	2.454	2.248	0.061
Religion	Before	5.739	7.723	0.955	8.239	1.259	0.561
	After	5.739	7.723	0.955	5.740	7.727	0.959
(B) DID with Entropy-Balancing Sample
		(1)			(2)
		Invest			Invest
Treat × Post		−0.075 ***			−0.041 ***
		(0.016)			(0.009)
Treat		0.159 ***			0.117 ***
		(0.015)			(0.008)
Post		−0.009			0.004
		(0.011)			(0.005)
Control Variables		NO			YES
Marriage fixed effect		NO			YES
Political fixed effect		NO			YES
Province fixed effect		YES			YES
Year fixed effect		YES			YES
Obs.		51,354			51,337
Adj_R2		0.124			0.184

Notes: (1) Robust standard errors in parentheses; (2) *** indicates statistical significance at the 1% level.

shows the results of using the entropy balanced sample. Panel A presents a balanced test after the entropy balancing and we can see that the treatment groups and control groups are balanced. Panel B show the results of the DID. We can find that the Treat_it × Post_t is significantly negative, indicating the results are robust.

4.2.4. Placebo Test

To further address the concern that our DID results may be spurious instead of reflecting an actual effect of Google withdrawal, we follow prior literature [55,56] to conduct placebo tests by running simulations that artificially assign the internet users or not. For each simulation, we draw a random sample of the internet users from the pool of actual samples, and then treat the remaining firms as the hypothetical “control groups”. We then conduct the DID regressions with the same specification as in Table 5 based on this simulated sample. We then repeat this simulation process 500 times. As Bertrand et al. [57] point out, DID may grossly underestimate the standard errors due to correlations across time. In order to solve this problem, placebo tests can also be used to avoid underestimates of the standard errors.

Figure 1 shows the distribution of the estimated coefficient of the primary variable of interest, Treat_it × Post_t, from the simulations along with the benchmark estimates, −0.041 (without control variables) and −0.041 (with control variables) reported in Table 5. Panel A presents the results without control variables, which show that the distribution of estimated coefficients derived from the randomly assigned simulated samples are clearly centered around 0, and the benchmark estimate is located outside the entire distribution (the line in the far left in the figure). Similarly, the results with control variables, presented in Panel B, deliver the same information, that the distribution of estimated coefficients derived from the randomly assigned simulated samples are centered around 0, and the benchmark estimate is located outside the entire distribution (the line in the far left in the figure). Combined, these simulation tests suggest that the negative and significant effect is not driven by chance.

4.2.5. Changing the Treatment Group and Control Group

In the baseline model, the residents who use the internet sometimes, often and always are defined as the treatment group, while the others are treated as the control group. For robustness checks, we change the definition of control group and treatment group. Firstly, the residents who use the internet in their spare time are used to identify the treatment group and control group. To be specific, the residents who use the internet in their spare time, multiple times a day, week, and month, are treated as the treatment group, and others are treated as the control group. The results are shown in columns (1) and (2) of

Table 9. Change in the treatment and control groups.

	(1)	(2)	(3)	(4)	(5)	(6)
	Using the Internet in Spare Time		Never Using the Internet as the Control Group		the Internet as a Main Information Sources as Treatment Group
	Invest	Invest	Invest	Invest	Invest	Invest
Treat × Post	−0.036 ***	−0.036 ***	−0.028 ***	−0.028 ***	−0.034 ***	−0.033 ***
	(0.008)	(0.008)	(0.007)	(0.007)	(0.011)	(0.013)
Treat	0.147 ***	0.111 ***	0.131 ***	0.088 ***	0.148 ***	0.105 ***
	(0.007)	(0.008)	(0.006)	(0.007)	(0.011)	(0.012)
Post	−0.003	0.003	0.00002	0.006	−0.012 ***	−0.002
	(0.003)	(0.005)	(0.003)	(0.005)	(0.004)	(0.006)
Control Variables	NO	YES	NO	YES	NO	YES
Marriage fixed effect	NO	YES	NO	YES	NO	YES
Political fixed effect	NO	YES	NO	YES	NO	YES
Province fixed effect	YES	YES	YES	YES	YES	YES
Year fixed effect	YES	YES	YES	YES	YES	YES
Obs.	58,209	51,324	58,234	51,337	57,102	50,447
Adj_R2	0.140	0.184	0.135	0.181	0.132	0.181

Notes: (1) Robust standard errors in parentheses; (2) *** indicates statistical significance at the 1% level.

. Secondly, the residents that never use the internet are defined as the control group. The results are shown in columns (3) and (4) of Table 8. Lastly, the residents that use the internet as a main information source are treated as the treatment group and the results are shown in Columns (5) and (6). The results in Table 9 show that the results are robust.

4.2.6. Using Logit Model

Considering the dependent variable as the dummy variable, the logit model can be used. We still use the control variables as mentioned above. The results are shown in column (1) of

Table 10. Logit Model and Excluding the Affects of Economic Development.

	(1)	(2)	(3)	(4)	(5)	(6)
	Logit	Linear	Linear	Linear	Linear	Linear
	Invest	Invest	Invest	Invest	Invest	Invest
Treat × Post	−0.341 ***	−0.041 ***	−0.044 ***	−0.044 ***	−0.050 ***	−0.044 ***
	(0.099)	(0.009)	(0.009)	(0.009)	(0.009)	(0.009)
Treat	1.447 ***	0.117 ***	0.119 ***	0.119 ***	0.124 ***	0.127 ***
	(0.095)	(0.008)	(0.009)	(0.009)	(0.009)	(0.009)
Post	0.025	0.005	0.047 **	0.065 ***	0.048 *	−0.036
	(0.092)	(0.008)	(0.020)	(0.022)	(0.027)	(0.031)
Fin		−0.054		−0.311
		(0.229)		(0.225)
Internet Development			−0.030 **	−0.036 ***
			(0.013)	(0.013)
Province × Year	NO	NO	NO	NO	YES	NO
City × Year	NO	NO	NO	NO	NO	YES
Control Variables	YES	YES	YES	YES	YES	YES
Marriage fixed effect	YES	YES	YES	YES	YES	YES
Political fixed effect	YES	YES	YES	YES	YES	YES
Province fixed effect	YES	YES	YES	YES	YES	NO
City fixed effect	NO	NO	NO	NO	NO	YES
Year fixed effect	YES	YES	YES	YES	YES	YES
Obs.	51,505	51,337	51,337	51,337	51,337	40,069
Pseudo_R2/Adj_R2	0.261	0.184	0.184	0.184	0.193	0.189

Notes: (1) Robust standard errors in parentheses; (2) ***, **, * indicate statistical significance at the 1%, 5% and 10% levels.

. We can find that the coefficient of Treat_it × Post_t, is significantly negative at the 1% significance level. The result is robust.

4.2.7. Excluding the Other Factors

Considering that economic development may affect residents’ investment behaviors, we try to exclude the effects of economic development. Firstly, we suppose that financial development may influence residents’ investment behaviors. We use the ratio of added value of the financial industry to GDP of each province (Fin_it) as the measurement of financial development. We add Fin as a control variable and the result is shown in column (2) of Table 10. We can find that the coefficient of Treat × Post is still significantly negative at the 1% significance level, indicating that the result is robust. In addition, we consider that internet development may also influence residents’ investment behaviors. We use the number of network ports in each province (Internet Development) to measure internet development and add Internet Development as a control variable. The result is shown in column (3) of Table 10. The coefficient of Treat_it × Post_t is still significantly negative at the 1% significance level. In column (4) of Table 10, we add the variable Fin and Internet Development together as control variables. The coefficient of Treat_it × Post_t is still significantly negative at the 1% significance level. Furthermore, we add the interaction fixed effects of province and year as well as city and year to exclude the influence of unobservable regional development, which may affect residents’ investment behaviors. The results are shown in columns (5) and (6) of Table 10. We can find that the coefficient of Treat×Post is still significantly negative at the 1% significance level. Therefore, we can conclude that the results are robust when excluding regional economic development.

We also consider that if the withdrawal of Google affects other ways to get information, the results may not be driven by the withdrawal of Google. In order to test that the results are driven by the withdrawal of Google—that is to say, the withdrawal of Google may only affect information acquisition from the internet—we still use the same way as mentioned before, where the the treatment group is defined as the households who use the newspapers, magazines, radios and TV sometimes, often and always, and the control group is defined as the households who use those media rarely and never. For the CGSS (2010), the variable Post is defined as 0, indicating the pre-treatment period; otherwise, it is coded as 1, indicating the after-treatment period. The results are shown in

Table 11. Excluding the effects of other media and not using DID.

	(1)	(2)	(3)	(4)	(5)	(6)
	Full Sample	Full Sample	Full Sample	Full Sample	2010	>2010
	Invest	Invest	Invest	Invest	Invest	Invest
Paper × Post	−0.002
	(0.009)
Mag × Post		0.017
		(0.013)
Radio × Post			0.009
			(0.011)
TV × Post				−0.008
				(0.007)
Paper	0.025 ***
	(0.008)
Mag		0.024 **
		(0.011)
Radio			0.006
			(0.009)
TV	0.025	0.005	0.047 **	−0.022 ***
	(0.092)	(0.008)	(0.020)	(0.006)
Post	0.017 ***	0.014 **	0.011 **	0.015 *
	(0.006)	(0.006)	(0.006)	(0.008)
Internet					0.046 ***	0.029 ***
					(0.004)	(0.001)
Control Variables	YES	YES	YES	YES	YES	YES
Marriage fixed effect	YES	YES	YES	YES	YES	YES
Political fixed effect	YES	YES	YES	YES	YES	YES
Province fixed effect	YES	YES	YES	YES	YES	YES
Year fixed effect	YES	YES	YES	YES	YES	YES
p-value of Chow Test					0.000 ***
Obs.	51,388	51,372	51,344	51,396	9803	41,534
Pseudo_R2	0.175	0.175	0.174	0.175	0.201	0.187

Notes: (1) Robust standard errors in parentheses; (2) ***, **, * indicate statistical significance at the 1%, 5% and 10% levels; (3) Chow test is used to test the difference between the coefficient of 2010 and after 2010.

. We can find that the coefficients of Paper_it × Post_t, Mag_it × Post_t, Radio_it × Post_t and TV_it × Post_t are not significant, which indicates that the withdrawal of Google only affects the residents who use the internet. Therefore, we exclude the potential influence on other ways to get information and the results are robust.

4.2.8. Not Using Difference in Difference

Considering that we only have one period before the event, the parallel trend may not satisfy. In the robustness check, we try to use another way. Specifically, the frequency of using the internet (Internet_it) is coded from 1–5 for never, rarely, sometimes, often and always. In order to test whether Google’s withdrawal influences the household investment in risky assets, we investigate the effects for the samples of 2010 and after 2010. The results are shown in columns (5) and (6) of Table 11. We can see that the coefficients of Internet for the samples of 2010 are larger than the samples after 2010. The Chow test shows the difference between the two groups is significant, indicating that Google’s withdrawal affects households who use the internet. Therefore, the result is robust.

5. Further Discussion

5.1. The Discussion of the Heterogeneity

In order to identify the heterogeneous effects, we further investigate the effects across groups of different income, gender and location. The results are shown in

Table 12. The Heterogeneity Effects on Residents.

	(1)	(2)	(3)	(4)	(5)	(6)	(7)
	Urban	Rural	Male	Female	Low-Income	Middle-Income	High-Income
	Invest	Invest	Invest	Invest	Invest	Invest	Invest
Treat × Post	−0.058 ***	0.047 ***	−0.039 ***	−0.045 ***	0.003	−0.035 **	−0.036 **
	(0.010)	(0.011)	(0.011)	(0.013)	(0.016)	(0.014)	(0.014)
Treat	0.132 ***	0.032 ***	0.115 ***	0.119 ***	0.028 *	0.091 ***	0.125 ***
	(0.010)	(0.011)	(0.011)	(0.013)	(0.016)	(0.014)	(0.014)
Post	0.026 ***	−0.040 ***	0.007	−0.005	0.005	−0.028 ***	−0.049 ***
	(0.008)	(0.007)	(0.007)	(0.010)	(0.008)	(0.009)	(0.013)
Control Variables	YES	YES	YES	YES	YES	YES	YES
Marriage fixed effect	YES	YES	YES	YES	YES	YES	YES
Political fixed effect	YES	YES	YES	YES	YES	YES	YES
Province fixed effect	YES	YES	YES	YES	YES	YES	YES
Year fixed effect	YES	YES	YES	YES	YES	YES	YES
Obs.	27,105	24,232	25,535	25,802	18,024	17,061	16,252
Adj_R2	0.191	0.177	0.184	0.185	0.130	0.097	0.176

Notes: (1) Robust standard errors in parentheses; (2) ***, **, * indicate statistical significance at the 1%, 5% and 10% levels.

. Columns (1) and (2) show the heterogeneous effects comparing urban and rural residents. We can find that the coefficient of Treat_it × Post_t for urban residents is larger than the rural residents, indicating that the effect is larger for the urban residents compared with the rural residents. Columns (3) and (4) show the results comparing female and male sub-samples. In general, internet use has significant influence on investment behavior for both gender groups. However, using the effect significantly decreases the probability of investing in risky assets by 4.5% for females and 3.9% for males. Columns (5) and (7) show the heterogeneous effects across income groups. We divide the whole sample by income level into three groups: low-income (the bottom one third of the sample), middle-income (the middle one third of the sample) and high-income (the top one third of the sample). Results show that the effects are only significant for middle-income and high-income groups, reducing the likelihood of investing in risky assets by 3.5% and 3.6%, respectively.

Furthermore, in CGSS, as mentioned above, the risky assets include stocks, funds, bonds, futures, options, real estate, foreign exchange and others. Therefore, the heterogeneous effects for different risky assets are discussed as shown in

Table 13. The heterogeneity effects on different risky assets.

	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
	Stocks	Funds	Bonds	Futures	Options	Real Estate	Foreign Exchange	Others
Treat × Post	−0.036 ***	−0.025 ***	0.002	0.001	0.001 *	−0.002	−0.002	0.003
	(0.008)	(0.006)	(0.003)	(0.001)	(0.001)	(0.001)	(0.001)	(0.002)
Treat	0.101 ***	0.052 ***	0.003	−0.000	−0.000	0.002 *	0.002 *	0.001
	(0.008)	(0.006)	(0.003)	(0.001)	(0.000)	(0.001)	(0.001)	(0.002)
Post	−0.003	−0.006	−0.001	−0.000	−0.000	0.000	−0.001	0.004 **
	(0.005)	(0.004)	(0.002)	(0.001)	(0.001)	(0.001)	(0.001)	(0.002)
Control Variables	YES	YES	YES	YES	YES	YES	YES	YES
Marriage fixed effect	YES	YES	YES	YES	YES	YES	YES	YES
Political fixed effect	YES	YES	YES	YES	YES	YES	YES	YES
Province fixed effect	YES	YES	YES	YES	YES	YES	YES	YES
Year fixed effect	YES	YES	YES	YES	YES	YES	YES	YES
Obs.	51,337	51,337	51,337	51,337	51,337	51,337	51,337	41,263
Adj_R2	0.159	0.088	0.015	0.004	0.001	0.003	0.005	0.005

Notes: (1) Robust standard errors in parentheses; (2) ***, **, * indicate statistical significance at the 1%, 5% and 10% levels.

. We can find that the effects are only significantly negative for stocks and funds. For options, the effect is significantly positive at the 10% significant level. For other risky assets, the effect is not significant. The reasons may be that stocks and funds are the more common risky assets for households compared with other risky assets. As the Investigation Report on the Current Situation of Investor Education in China conducted by China Securities Journal shows, the the proportion of investing in stocks and funds among individual investors rose from 28.46% to 33.21% from 2018 to 2020, while investing in real estate declined from 9.33% in 2018 to 9.02% in 2020. In addition, investing in stocks and funds is much easier than others; for example, people can use their mobile phone to buy and sell stocks and funds. Therefore, the effects are more significant for stocks and funds.

5.2. Plausible Channels

As discussed above, we suppose that the withdrawal of Google may have two effects. On the one hand, as mentioned by Xu et al. [15], the withdrawal of Google may affect firms’ information disclosure quality, which make it more difficult for investors to get information. Difficulty in obtaining information may make investors invest less in risky assets. In order to test plausible channels of firms’ information disclosure quality, following existing studies [15,58], we conduct treatment groups and control groups based on the Google index. To be specific, the firms whose stock tickers have a higher search volume index than the sample median in 2009 are treated as treatment groups. The KV is used to measure the information disclosure quality of listed firms in China. The specific calculation of KV is shown in Appendix A. The results are shown in columns (1) and (2) of

Table 14. The effects on information disclosure and investors’ attention.

	(1)	(2)	(3)	(4)
	F.KV	F.KV	F.IA	F.IA
Treat × Post	0.076 ***	0.026 ***	−0.002 *	−0.003 ***
	(0.006)	(0.006)	(0.001)	(0.001)
LnAsset		−0.053 ***		−0.003 ***
		(0.004)		(0.001)
SGR		0.001		0.001 **
		(0.002)		(0.000)
LnFirmAge		−0.125 ***		−0.015 ***
		(0.020)		(0.003)
Duality		−0.002		−0.000
		(0.004)		(0.001)
IndRate		0.000		−0.015 **
		(0.027)		(0.007)
LnBoard		−0.008		−0.004
		(0.010)		(0.003)
LEV		0.041 ***		0.002
		(0.012)		(0.003)
ROE		−0.020 **		0.009 ***
		(0.010)		(0.003)
Tobins’Q		−0.004 **		−0.003 ***
		(0.002)		(0.000)
PPE		−0.059 ***		−0.005 *
		(0.014)		(0.003)
Top10		0.002 ***		−0.000 ***
		(0.000)		(0.000)
Constant	0.122 ***	1.424 ***	0.056 ***	0.176 ***
	(0.002)	(0.094)	(0.001)	(0.020)
Firm Fixed effect	YES	YES	YES	YES
Year Fixed effect	YES	YES	YES	YES
Obs.	13,761	12,305	12,805	12,370
Adj_R2	0.201	0.225	0.390	0.411

Notes: (1) Robust standard errors clustered at firm level in parentheses; (2) ***, **, * indicate statistical significance at the 1%, 5% and 10% levels; (3) F. indicates forward one period.

. We find that the coefficients of Treat_it × Post_t is significantly positive, indicating that the withdrawal of Google reduces the firms’ information disclosure quality, and then the residents’ willingness to invest in risky assets decreases.

We also suppose that the withdrawal of Google may decrease investors’ attention in the listed firms due to the difficulty getting accurate and timely information from new media, and that they then may be not willing to invest in risky assets. Following prior literature [59,60,61], the measurement of investors’ attention (IA) is used as a dependent variable. The specific calculation of IA is shown in Appendix A. The results are shown in columns (3) and (4) of Table 14. We can find that the coefficients of Treat × Post is significantly negative, indicating that the withdrawal of Google reduces investors’ attention. Therefore, we suppose that the withdrawal of Google makes firms’ information disclosure quality worse, and that households find it difficult to get information and then decrease their willingness to participate in financial market.

On the other hand, we consider that the withdrawal of Google may make the residents’ getting information hard and the households may be not willing to obtain the information willingly in their spare time. In order to test whether this channel is satisfied or not, we use studying during spare time as a proxy of willingness to obtain information. We use the questions from the CGSS about the frequency of studying in their spare time (Information) and responses including Never, Rarely, Sometimes, Usually and Very often. We code them from 1–5 and the results are shown in Panel A of

Table 15. The effects on residents.

(A) The Effect on Obtaining Information
	(1)		(2)
	Information		Information
Treat × Post	−0.151 ***		−0.074 ***
	(0.023)		(0.024)
Treat	1.089 ***		0.479 ***
	(0.022)		(0.024)
Post	−0.089 ***		−0.037 **
	(0.013)		(0.018)
Control Variables	NO		YES
Marriage fixed effect	NO		YES
Political fixed effect	NO		YES
Province fixed effect	YES		YES
Year fixed effect	YES		YES
Obs.	58,318		51,404
Adj_R2	0.233		0.357
(B) The Effect on Risk Preference
	(1)	(2)	(3)	(4)
	Risks	High Risk, High Return	Entrepreneurship	Criticize
Infor_Internet	−0.338 **	−0.303 **	−0.455 ***
	(0.159)	(0.156)	(0.132)
Treat × Post				−0.121 ***
				(0.030)
Treat				0.170 ***
				(0.030)
Post				−0.108 ***
				(0.022)
Control Variables	YES	YES	YES	YES
Marriage fixed effect	YES	YES	YES	YES
Political fixed effect	YES	YES	YES	YES
Province fixed effect	YES	YES	YES	YES
Year fixed effect	NO	NO	NO	YES
N	1553	1531	1615	50,828
Pseudo R2/Adj_R2	0.053	0.046	0.068	0.050

Notes: (1) Robust standard errors in parentheses; (2) ***, ** indicate statistical significance at the 1% and 5% levels; (3) F. indicates forward one period.

. We can find that the coefficients of Treat × Post are significantly negative, indicating that the withdrawal of Google makes the residents less willing to get information in their spare time and that their risk asset investment may be negatively affected.

Moreover, we suppose that the withdrawal of Google may affect residents’ risk preference. In order to test the effects on risk preference, we firstly use the question from 2015, that compared with ordinary and stable life, I prefer a life with risks and opportunities, and the question, If I have extra money, I would like to invest in risky but high return projects. The answers are from Strongly Agree to Strongly Disagree. We code them from 1 to 5 and use the ordered probit model. We also use the entrepreneurship willingness to measure the risk preference. The results are shown in columns (1) to (3) of Panel B in Table 13. We also use the inclination of criticizing government to measure the risk preference. The results are shown in column (4) of Panel B in Table 15. We can find that the coefficient of Infor_Internet is significantly negative, indicating that using new media would affect the residents’ risk preference; that is to say, the withdrawal of Google may affect their risk preference, and then their willingness to invest in risky assets is negatively affected.

Overall, the discussion of the plausible channels shows that the withdrawal of Google makes firms’ information disclosure quality worse, and makes it difficult for households to get information and change their risk preference, and then their investment behavior may be negatively affected.

6. Conclusions

In 2010, Google withdrew from mainland China unexpectedly, which significantly changed the information acquisition environment in China. After that, obtaining information from new media searching comes with high cost and more difficulty. We use the withdrawal from mainland China as a quasi-experiment and the data from CGSS to test the relationship between information searching in new media and households’ investment in risky assets. By using the difference-in-difference method, we find that Google’s withdrawal from mainland China significantly decreased households’ willingness to invest in risky assets. The results are robust after a set of robustness checks. As for the heterogeneity, the effects are different among females and males, rural and urban, and high-income, middle-income and low-income. As for the plausible channels, we find that Google’s withdrawal from mainland China significantly affects the information disclosure quality, the convenience of getting information and the risk preference, by which their investment behaviors are affected.

The fairness of getting information is an important aspect in sustainable development, with a feature of new media making the obtaining of information more easy, with lower costs, letting households have a better awareness of information. In addition, information on sustainable development can rely on new media to let the whole world know, which is helpful for sustainable development. Therefore, in the future, we should explore how to reduce costs in information dissemination by using new media. How to make new media disseminate information more effectively is also an issue to be considered in the future.

Information is important for investors, which may affect their willingness to invest in financial markets. Therefore, we propose two implications. On the one hand, new media should provide more effective and timely information to investors, which can reduce the cost of acquiring information and encourage them to participate in financial markets. On the other hand, we should further popularize new media to overcome the limitations of financial market frictions and promote greater participation of households in financial markets. However, it is also necessary to pay attention to the legality of various investments to prevent some illegal fundraising products from using the cloak of “Internet finance” and “P2P loans”. Furthermore, new media can also promote investment in financial markets through social interaction. While investment in risky assets usually requires certain professional knowledge, investors should always control risks in investment activities to avoid blindness in investment.