Information accessibility and knowledge creation: the impact of Google’s withdrawal from China on scientific research

ABSTRACT

How important is Google for scientific research? This paper exploits the exogenous shock represented by Google’s sudden withdrawal of its services from mainland China to assess the importance of access to information for the knowledge production function of scientific scholars in the field of economics. For economists, a type of scholar with a simple knowledge production function, results from difference-in-difference analyses, which compare their scientific output to scholars located in the neighbouring regions, show that the scientific productivity declines by about 28% in volume and 30% in terms of citations. These results are consistent with the view that information accessibility is an important driver of scientific progress. Considering that the negative effect of the shock is stronger for top scholars located in China, Google’s sudden exit bears the risk that researchers lose touch with the research frontier and persistently lag behind their foreign peers.

KEYWORDS:

• Information accessibility
• academic publications
• knowledge production
• Google
• China

JEL CLASSIFICATION:

• D83
• I23

1. Introduction

Since Google entered mainland China in 2006, its share of the search engine market of mainland China¹ rapidly increased to 40.08% by the end of 2009.² Together with the Chinese firm Baidu, which offers a similar service portfolio and held a market share of 58.47%,³ Google effectively became part of a duopoly (Kong et al. 2022). Google was, hence, a main source of information in China, especially of information from foreign countries (Kong et al. 2022; Wang, Yu, and Zhang 2020). Like any search engine provider operating in China, Google was obliged to follow the strict censorship guidelines imposed by the Chinese government, but, in January 2010, Google decided to discontinue the censoring of search results on its China search page (Google.cn).⁴ This decision rapidly escalated in a sudden and unannounced withdrawal of some Google services from China, leaving millions of users without access to the world’s top search engine overnight. From the 30^th of June 2010 onwards, users in China could not access some of the main Google services anymore (Kong et al. 2022; The Official Google Search Blog 2012; Quinn 2012; Roberts 2014; Xu, Xuan, and Zheng 2021).⁵

In this paper, we investigate the effect of Google’s⁶ sudden exit from China on the scientific research output of scholars in the field of economics located in China. Access to information in the form of books and research material has been shown to be crucial for the generation of new knowledge (Berkes and Nencka 2019; Biasi and Moser 2021; European Commission 2012; Furman and Stern 2011; Furman, Jensen, and Murray 2012; McCabe and Snyder 2015; Mueller-Langer, Scheufen, and Waelbroeck 2020; Waldinger 2016). A lack of access or high accessibility costs can, hence, be a key barrier to new discoveries and knowledge creation. Not surprisingly, information and communication technologies have been shown to enhance science production by increasing the availability of information and, hence, reducing search costs (Agrawal and Goldfarb 2008; Ding et al. 2010; Kim, Morse, and Zingales 2009; Winkler, Levin, and Stephan 2010). While the withdrawal of Google’s services from China does not completely shut down access to information for academic scholars, it surely leads to an increase in their search costs.⁷ Affected scholars are, therefore, still able to access information, but the lengthier research process generates a slowdown of their knowledge production and, hence, their short-term publication outcome.⁸ Google’s sudden exit from China, therefore, bears the risk that researchers located in China lose touch with the research frontier and persistently lag behind their foreign peers.

Using Google’s exit from China to assess the effect of barriers to information accessibility on scientific research has several advantages that address common challenges for causal estimation. First, Google’s exit was exogenous to science production and unexpected as it was the result of a rapid escalation of political tensions between the Chinese leadership and Google (Kong et al. 2022; Xu, Xuan, and Zheng 2021; Zheng and Wang 2020). Second, Google was, at the time of the sudden withdrawal of its services, one of the main sources of knowledge for China (Kong et al. 2022; Wang, Yu, and Zhang 2020) and its scientists (Qiu 2010).

Our empirical analysis focuses on the field of economics following prior studies such as Kim et al. (2009), McCabe and Snyder (2015), Liang, Gu, and Nyland (2022), and Piracha et al. (2022). Economics is a research field with a simple knowledge production function as it does not rely on material and expensive specialised equipment (Stephan and Levin 1992). In addition, new insights are published almost exclusively in scientific journals rather than in books and conference proceedings which are often not well covered in bibliometric databases (e.g. Michels and Fu 2014). Hence, an estimated effect of the sudden decrease in information accessibility on scientific output is less likely to be confounded by other effects resulting from the knowledge generating process or the publication strategy of the field.

To derive causal results, we use a Difference-in-Difference (DiD) approach employing a control group of researchers located in Taiwan and Hong Kong following Zheng and Wang (2020) who argue for a control group that is culturally, economically, and geographically closely related to China. Our results show that researchers in the field of economics affiliated with Chinese institutions experience a significant decline in both their research output quantity and impact as measured by citations received by the future literature. The magnitude is about 28% for co-author weighted publications and 30% for co-author weighted citations.

We explore the proposed underlying mechanism of information accessibility further and show that the productivity and impact of those scholars located in China who work with foreign co-authors are less affected by Google’s exit. These scholars can use their interpersonal networks as a channel for knowledge access (Mohnen 2022; Singh 2005). The publication output and impact of these scholars decreases by smaller shares of 20% and 22%, respectively, supporting that the mechanism of knowledge accessibility is responsible for the decline in publication output after Google’s withdrawal.

In further analysis, we find that the effect in terms of quantity and impact is stronger for those scholars with the highest impact as measured by their citation stock over publication stock before Google’s withdrawal. The publication output and impact of the top 25% scholars decrease by 39.5% and 37.5%, respectively, while the publication output of the scholars at the bottom of the impact distribution decreases by 20%. There is no significant effect for the scholars at the bottom of the impact distribution in terms of impact. The large effects on the top scholars raise concerns about the ability of China to stay in touch with the research frontier in the medium and long run with potentially harmful implications for economic growth (Griliches 1992; Jaffe 1989). Also, we find no significant differences in the negative effect of the shock on treated scholars affiliated with both top and less renowned universities.

While our study is limited to the field of economic research, we make several contributions to the literature. First, our work adds to our understanding of the determinants of knowledge creation (Stephan and Levin 1992; Stephan 1996, for an overview) and more specifically of the role of information and communication technology in knowledge creation (Agrawal and Goldfarb 2008; Ding et al. 2010). Prior studies have shown that access to network technology (Agrawal and Goldfarb 2008; Ding et al. 2010, for the case of BITNET) eases information accessibility and facilitates the knowledge production of scientists. Here, we focus on Google as a general search engine and complement prior findings for different technologies. Second, we contribute to recent literature that focuses on positive information shocks such as the availability of access to libraries (Berkes and Nencka 2019; Biasi and Moser 2021; Furman, Jensen, and Murray 2012), of research resources (Furman and Stern 2011) and of online access to scientific journals (McCabe and Snyder 2015; Mueller-Langer, Scheufen, and Waelbroeck 2020), and their impact on knowledge creation. We differ from these studies in two ways. First, these studies focus on the access to prior scientific knowledge available in the form of books, journals, and research resources while we focus on the access to a search engine that covers a much broader scope of information. Second, we explore a negative shock to information availability to assess the effects on science production while prior studies focus on positive shocks to information availability. We cannot assume that positive and negative shocks have a symmetric effect since this is rarely the case in reality (see, for instance, the large literature on asymmetric investor reactions in financial markets, e.g. Kuhnen 2015; Kluger and Wyatt 2004, or, a very different example, the asymmetric responses of individuals to positive and negative feedback about their intelligence and beauty; Eil and Rao 2011).

Third, our finding that scholars can use their interpersonal networks as a channel for knowledge access (Mohnen 2022; Singh 2005) contributes to the large literature on academic networks (e.g. Beaver and Rosen 1978; Fanelli, Larivière, and Dorta-González 2016; Greene 2007; Wuchty, Jones, and Uzzi 2007) and, in particular, to the smaller literature on informal links between researchers (Brown 2005; Laband and Tollison 2000; Oettl 2012; Rose and George 2021). Prior studies define informal links between researchers as providing feedback visible in the acknowledgement of the paper (Laband and Tollison 2000; Rose and George 2021) or through presentations (Brown 2005; Laband and Tollison 2000; Oettl 2012) and show that these informal collaborations increase citations. We provide suggestive evidence for co-author networks facilitating access to knowledge beyond joined projects which leads to a lower drop in scientific productivity and citations in the presence of an information shock.

Lastly, we add to the developing literature that focuses on the implications of Google’s withdrawal from China. These include a higher stock crash risk for firms (Xu, Xuan, and Zheng 2021) and a decrease in corporate innovation (Kong et al. 2022; Zheng and Wang 2020). Differently from these prior studies, our focus is on the scientific rather than on the corporate sector.

2. Background

The well-known cumulative nature of science requires research to evolve along specific lines where scientists build on and advance prior insights (Azoulay et al. 2015; Merton 1973; Mokyr 2002). Having access to the most recent worldwide developments in the respective research field is, hence, crucial for the generation of new state-of-the-art knowledge (Berkes and Nencka 2019; European Commission 2012; Furman, Jensen, and Murray 2012). Further, the nature of competition in science is a winner-takes-all game that promises high reputation gains, lucrative jobs, and research opportunities for the winner, i.e. the first to make a discovery, while the second to finish the race often leaves empty-handed (Merton 1973). This implies that the distribution of publications and citations at the individual level is extremely skewed (Lotka 1926; Price 1963) and only a minority of scientists are able to contribute to the advancement of science (Cole and Cole 1972; Partha and David 1994). The nature of competition in science emphasises the crucial importance of speedy access to recent information.

As scientific research becomes increasingly complex and multidisciplinary over time (Jones 2009; Wuchty, Jones, and Uzzi 2007), scientists’ costs of staying up to date with the latest discoveries in their research field increased tremendously over the past decades. Information and communication technologies have been shown to be crucial factors in today’s knowledge production function as they increase the availability of information and reduce search costs (Agrawal and Goldfarb 2008; Ding et al. 2010; Kim, Morse, and Zingales 2009; Winkler, Levin, and Stephan 2010). Agrawal and Goldfarb (2008) and Ding et al. (2010), for instance, show how access to BITNET facilitated collaboration between scientists and enhanced knowledge production. McCabe and Snyder (2015) and Mueller-Langer et al. (2020) show how online access to scientific journals improves citation rates and the creation of new scientific results in both developed and developing countries.

Here, we focus on Google and its search engine as an alternative technology that facilitates information access and reduces search costs (Kong et al. 2022; Xu, Xuan, and Zheng 2021; Zheng and Wang 2020). Google’s services have been shown to be crucial for corporate China by facilitating the development of novel technologies (Kong et al. 2022; Zheng and Wang 2020) and preventing investor overreactions leading to stock crashes in Chinese businesses (Xu, Xuan, and Zheng 2021). The importance of Google was not less significant for the academic sector. After Google’s exit, visits to Wiley Online Library from Google dropped by around 30%, and from Google Scholar by around 15% (Eassom 2016). According to a survey of almost 800 Chinese researchers conducted by Nature just before Google’s withdrawal, more than 80% of the respondents used Google’s search engine to find academic papers, close to 60% to get information about scientific discoveries or other scientists’ research programmes, and one-third of the survey respondents made use of Google’s products to find science-policy and funding news (Qiu 2010). This evidence highlights how important information is at all stages of the research process, from searching for input and defining a research project to access to funding. Google’s withdrawal from China, hence, significantly increased the barriers to access information for scientists located in China and they could hardly find alternatives, such as Virtual Private Networks (VPNs) ⁹ or mirror platforms,¹⁰ to overcome the search hurdle (Lu et al. 2017).¹¹

In this article, we, therefore, ask the question of whether and to what extent the scientific output of scholars located in China is negatively affected by Google’s withdrawal of its services. We expect a negative effect since access to information and prior knowledge is one key ingredient of the knowledge production function (Ding et al. 2010; Stephan and Levin 1992). The limited accessibility of information is expected to affect both publication outcome and impact. The underlying mechanisms are different though. Regarding publications, treated scholars may publish less due to a more difficult access to information and higher search costs. Regarding citations, treated scholars may receive fewer citations per publication due to a lower ‘quality’ of their work. Not having readily access to the most recent scientific advances implies that their research is not as close to the knowledge frontier as other articles. The resulting restricted novelty of the publications leads to fewer citations. Overall, difficulties to stay in touch with the research frontier delay the scientists and make them less likely to win the race for priority. This should be directly reflected in a lower publication output. In addition, Google’s search engine was especially important in China for accessing foreign information (Kong et al. 2022). Baidu’s search engine, in contrast, ranks local search results, i.e. search results in Chinese language, with higher priority than foreign information (Yi 2014). Google has therefore a comparative advantage in nonlocal information search and its exit enforced a tendency to source more local information (Zheng and Wang 2020). While local search can be more efficient for some topics, it may lead to a ‘local search trap’ resulting in rather incremental improvements to the state of the art (Laursen 2012; Wagner, Hoisl, and Thoma 2014; Zheng and Wang 2020).¹² Distant search, in contrast, tends to be explorative in nature and stimulates the arrival of novel ideas so that it increases the chances of breakthrough inventions (Arts and Fleming 2018). The difficulties in engaging in distant searches and the resulting decrease in novelty should be reflected in a decline in the impact of the scientific publications that scholars located in China publish after Google’s withdrawal. Chinese publications after Google’s exit are expected to be used to a lesser extent as building blocks for future research. In summary, we expect that the publication volume and impact of scholars located in China drops after Google’s exit from China.

3. Method

To analyse the effect of the sudden withdrawal of Google from China on the publication volume and impact of scholars located in China, we employ DiD methods. Our treatment group consists of authors who were only affiliated with one or more universities in China before and after Google’s withdrawal.¹³ The control group consists of scholars that were affiliated with one or more universities in Hong Kong or Taiwan before and after 2010 following Zheng and Wang (2020) who recommend using a control group that is culturally, economically, and geographically closely related to China. We believe that scholars affiliated in Hong Kong and Taiwan are a suitable control group for the following reasons: (1) in China, Hong Kong, and Taiwan, scholars are influenced by Chinese culture, history, and politics and they all face similar political and cultural pressures; (2) scholars in China, Hong Kong, and Taiwan had limited access to research funding and resources as compared to scholars in the U.S. and Europe (back in the years around the shock); (3) scholars in China, Hong Kong, and Taiwan are more likely to engage between themselves due to geographical proximity which is likely to affect their research visions and approaches; and (4) researchers in China, Hong Kong, and Taiwan speak the same language.

We estimate an equation of the form:

(1) $\mathrm{Publication}{\mathrm{s}}_{\mathrm{it}}=\mathrm{f}\left({\beta }_1\mathrm{Trea}{\mathrm{t}}_{\mathrm{i}}\ast \mathrm{Pos}{\mathrm{t}}_{\mathrm{t}}+\mathit{\delta }{\mathrm{\Gamma }}_{\mathrm{i}}+{\mathit{\phi }}_{\mathrm{t}}\right)+{\mathit{\epsilon }}_{\mathrm{it}}$(1)

where $\mathit{Publication}{s}_{\mathit{it}}$represents different dependent variables that capture the publication output and impact of author $\mathit{i}$ in year $\mathit{t}$. Those are the publication count, the fractional publication count, the citation count, the fractional citation count, and citations divided by publications in year $\mathit{t}$. As the dependent variables tend to follow a count distribution, we estimate the model as a Poisson model.

The variable $\mathit{Trea}{t}_i$ is a binary variable that indicates whether a scholar belongs to the treatment group or the control group. Note that the affiliation with the treatment or control group ($\mathit{Trea}{t}_i$) is time-invariant and, hence, included in the author’s fixed effect (${\mathrm{\Gamma }}_i$). ${\mathrm{\Gamma }}_{\mathrm{i}}$ controls for inherent differences between researchers caused by unobservable factors such as talent or ability in the form of researcher fixed effects. To show the robustness of our results, we also present specifications without fixed effects where we estimate a potential effect for the systematic difference between treatment and control groups and allow for control variables in the Appendix.

The variable $\mathit{Pos}{t}_t$ is a binary variable that takes the value one from the year after Google’s withdrawal from China, 2011, onwards.$\mathit{\hspace{0.17em}}{\phi }_t$ captures common time trends through a set of year dummies.

The main result of the model is provided by the coefficient ${\beta }_1$, which captures the average difference in the change of publication output between treatment and control observations after Google’s withdrawal. If scientists in the treatment group experience a decline in publications after having a more restricted access to information, while scientists in the control group do not, ${\beta }_1$shows a negative and significant effect.

4. Data, variables, and descriptive statistics

4.1. Data

To investigate the effect of Google’s withdrawal of its services from China on scientific productivity, we retrieve scholarly publications from English language journals and citation data for researchers in China, Taiwan, and Hong Kong for the time period 1995–2019 from Scopus. Scopus has been found to outperform its competitor World of Science in terms of coverage (Aksnes and Sivertsen 2019), especially in the field of economics research (Bosman et al. 2006). With our choice of research field, we follow prior studies such as Kim et al. (2009), McCabe and Snyder (2015), Liang, Gu, and Nyland (2022), and Piracha et al. (2022) which also based their empirical analyses on data for scholars in the field of economics. Studying economics has two important advantages. First, the science production function is relatively simple as it does not rely on material and expensive equipment so that the input factors reduce to effort, skills, and knowledge (Ding et al. 2010; Stephan and Levin 1992). Second, insights in the field of economics are published almost exclusively in scientific journals rather than in books and conference proceedings which are typically not well covered by publication databases. Hence, an estimated effect of the sudden decrease of information accessibility on scientific output in economics is less likely to be confounded by other effects resulting from the knowledge generating process or the publication strategy of the field.

To arrive at the author level, we aggregate our publication data at the author-year level relying on the Scopus author identifiers (Kawashima and Tomizawa 2015).¹⁴ This leads to an unbalanced panel that includes the complete publication record of each author from 1995 onwards. Each author enters the database with her first publication and leaves the dataset with the last publication observed in Scopus. Years in which an author has not published are treated as years of zero publications. After some data inspection at the author level, we exclude the earliest and latest years leading to a time window for the analysis of the years 2007–2017.

After excluding authors with missing country information¹⁵ and some further data cleaning¹⁶, we drop authors with double affiliations in China and elsewhere as we cannot be sure about their country of residency. This is crucial because an author with an affiliation in China and the U.S. could be working in the U.S. and, hence, not be affected by the exit of Google from China. In total, we excluded 1,644 authors with affiliations both in China and the rest of the world (7,680 observations) and 725 authors with affiliations both in China and Hong Kong or Taiwan (5,509 observations).

Further, we include only authors who did not change their country of affiliation during our time period of interest. This reflects a conservative approach because we only focus on authors with one affiliation region, i.e. China, HK, and Taiwan, or rest of the world, and, implicitly, we also account for authors that did not change their affiliation before and after 2010, potentially motivated by the restrictions faced in China. At this point, 1,249 authors are excluded. We also only keep authors that have at least one publication before and after 2010 to assure that the treated scholar is part of the same regime in the pre-and post-treatment period so that the performance before and after the shock can be meaningfully compared. In other words, we exclude scholars who might have left or joined Chinese academia for reasons potentially related to the treatment. Here 39,508 authors are removed.

After cutting some outliers, i.e. the top 1% of the distribution of each dependent variable, our final dataset comprises 16,750 observations at the author-year level which corresponds to 8,653 observations for 1,141 treated authors and 8,097 observations for 1,004 control authors. For a later investigation of the proposed mechanism of information accessibility, we further distinguish between treated authors who collaborated with foreign authors over our sample period and those who did not. 6,188 observations on 769 authors in the treatment group collaborated with foreign scholars.

4.2. Variables

We use five dependent variables to measure the scientific output of our scholars in terms of quantity and impact. Specifically, we use the number of publications and the co-author weighted number of publications, i.e. fractional publications, as simple output indicators. To account for publication impact, we weigh the publications by the citations they receive by future research. We therefore use the number of citations, the fractional citations, and the citations divided by publications per year. The citations are counted as aggregate citations in the publication year (see also Hussinger and Pellens 2019). Citations are a widely used indicator of the importance of scientists and their scientific findings, showing the extent to which results and insights are used as building blocks for future research.

While our main results employ a lean specification without control variables, we show robustness checks that control for scientists’ career age and funding in a regression setting without fixed effects. Career age accounts for the fact that scholars change their level of commitment to publishing as their career progresses (e.g. Stephan and Levin 1992). The age effect is found to be non-linear showing that scientists’ productivity peaks at a certain point in time. Our regressions account for this. Career age is measured as the number of years since the first publication. Access to funding is a means to facilitate productivity (e.g. Hottenrott and Lawson 2017; Hussinger and Carvalho 2022; Salter and Martin 2001). We use information provided by Scopus on whether a publication received any type of funding to generate a dummy that depicts whether a researcher received funding in the year of publication. As funding may influence productivity beyond the funding period (Hussinger and Carvalho 2022), we use the funding stock as our control variable:

(2) $\mathrm{Funding}\mathrm{stock}{ }_{\mathrm{it}}={\mathrm{Number}\mathrm{of}\mathrm{articles}\mathrm{that}\mathrm{received}\mathrm{funding}}_{\mathrm{it}}+(1-\mathit{\delta }){\mathrm{Funding}\mathrm{stock}}_{\mathrm{it}-1}$(2)

where δ is a depreciation rate of 15%. In order to have a meaningful stock measure, we use information from our initial dataset going back until 1995 and assume that funding stock₁₉₉₄ is equal to 0.

Furthermore, we show the robustness of our results on a sample in which we match treated and control scholars on their citations stock, co-author weighted citations stock, publications stock, co-author weighted publications stock, and funding stock as of 2010. As for the funding stock above, these stocks are a function of the scholars current and previous performances and are calculated according to Eq. 2. Some of these stocks as well as a measure of citations over publications stock (again calculated as in Eq. 2) are used in further robustness checks where we use a synthetic DiD method (Arkhangelsky et al. 2021) and an entropy balancing approach (Hainmueller 2012).

4.3. Descriptive evidence and statistics

To allow a first graphical inspection of the sample, Figure 1 shows the effect of Google’s exit on our five dependent variables for the treated and control groups. We observe a reduction for all our output measures after Google’s withdrawal for the treatment group, while the timeline for the control group seems to be unaffected. This simple descriptive evidence suggests a strong and immediate effect of Google’s exit on the scientific output of scholars located in China which then fades out over time.

Figure 1. Descriptive evidence.

Table 1 shows some before-after comparisons of the means of the dependent variables for the treated and control groups. We observe that the treated scholars experience a drop in their co-author weighted publications of 27% and their co-author weighted citation impact of 44%. The control group, in contrast, sees no change regarding their co-author weighted publications before and after Google’s exit from China and a significantly smaller decline of their co-author weighted citation impact of 25%. The changes are statistically significant at the 1% level as t-tests show.

Table 1. Descriptive statistics.

			Treated scholars						Control scholars
	Time	Variable	n	mean	sd	median	min	max	n	mean	sd	median	min	max
1	Pre-2010	number of publications	2789	0.60	0.72	0.00	0.00	5.00	3069	0.48	0.67	0.00	0.00	4.00
2	Pre-2010	fractional publications	2789	0.26	0.35	0.00	0.00	2.00	3069	0.23	0.36	0.00	0.00	2.00
3	Pre-2010	number of citations	2789	16.42	33.34	0.00	0.00	274.00	3069	8.42	21.57	0.00	0.00	203.00
4	Pre-2010	fractional citations	2789	6.44	13.49	0.00	0.00	106.00	3069	3.68	9.50	0.00	0.00	103.00
5	Pre-2010	citations per publication	2789	13.47	26.30	0.00	0.00	177.00	3069	7.29	19.00	0.00	0.00	177.00
6	Post-2010	number of publications	5864	0.48	0.76	0.00	0.00	5.00	5028	0.50	0.72	0.00	0.00	5.00
7	Post-2010	fractional publications	5864	0.19	0.32	0.00	0.00	2.00	5028	0.23	0.36	0.00	0.00	2.00
8	Post-2010	number of citations	5864	10.30	26.00	0.00	0.00	265.00	5028	6.46	18.02	0.00	0.00	213.00
9	Post-2010	fractional citations	5864	3.61	9.29	0.00	0.00	97.00	5028	2.75	7.89	0.00	0.00	97.00
10	Post-2010	citations per publication	5864	7.72	18.66	0.00	0.00	170.00	5028	5.43	15.37	0.00	0.00	177.00

Notes: Number of treated scholars: 1,141; number of control scholars: 1,004. Treated scholars refers to scholars only affiliated in China.

Also the control group suffers a decline in impact of about 25%. This is perfectly in line with the literature which documents a strong increase in the competition among scholar to publish due to significantly lower acceptance rates and number of published articles as well as higher submission rates (see, for example, Larivière, Gingras, and Archambault 2009 or Card and DellaVigna (2013) for top journals in the field of economics). This provides some evidence that even though our sample had to be significantly restricted to a small portion of the population of treated scholars, we still find that the general population trends hold for our sample.

5. Results

5.1. Parallel trends

One of the requirements for deriving causal effects from a DiD analysis is a parallel movement of the dependent variable in the pre-treatment period. Table 2 reports our results from a regression investigating the existence of pre-treatment parallel trends. The specification extends Equation (1)(1) $\mathrm{Publication}{\mathrm{s}}_{\mathrm{it}}=\mathrm{f}\left({\beta }_1\mathrm{Trea}{\mathrm{t}}_{\mathrm{i}}\ast \mathrm{Pos}{\mathrm{t}}_{\mathrm{t}}+\mathit{\delta }{\mathrm{\Gamma }}_{\mathrm{i}}+{\mathit{\phi }}_{\mathrm{t}}\right)+{\mathit{\epsilon }}_{\mathrm{it}}$(1) in that we interact the year dummies (${\phi }_t$) with the treatment indicator ($\mathit{Trea}{t}_i$).

Table 2. Parallel trends.

Dependent Variables:	number of publications	fractional publications	number of citations	fractional citations	citations per publication
Model:	(1)	(2)	(3)	(4)	(5)
Variables
treat × 2007	0.0817	0.1026	−0.1085	−0.1186	−0.1012
	(0.1074)	(0.1178)	(0.1818)	(0.1902)	(0.1785)
treat × 2008	0.0370	−0.0149	0.0199	−0.0093	0.0587
	(0.0922)	(0.1016)	(0.1675)	(0.1684)	(0.1685)
treat × 2009	−0.0784	−0.0398	−0.0732	−0.0373	−0.0508
	(0.0820)	(0.0928)	(0.1505)	(0.1506)	(0.1475)
treat × 2011	−0.3425^***	−0.3113^***	−0.4618^***	−0.4344^***	−0.4680^***
	(0.0817)	(0.0907)	(0.1429)	(0.1476)	(0.1445)
treat × 2012	−0.3858^***	−0.4217^***	−0.3192^**	−0.4343^***	−0.3692^**
	(0.0867)	(0.0948)	(0.1505)	(0.1500)	(0.1511)
treat × 2013	−0.2481^***	−0.2895^***	−0.3265^**	−0.3918^**	−0.3535^**
	(0.0907)	(0.0988)	(0.1541)	(0.1586)	(0.1550)
treat × 2014	−0.4082^***	−0.5284^***	−0.2537	−0.4975^***	−0.3062^*
	(0.0970)	(0.1050)	(0.1823)	(0.1844)	(0.1822)
treat × 2015	−0.2086^**	−0.2041^*	−0.3531^**	−0.4597^**	−0.4845^***
	(0.0961)	(0.1058)	(0.1730)	(0.1795)	(0.1733)
treat × 2016	−0.2090^**	−0.3234^***	−0.0135	−0.0693	−0.0848
	(0.1002)	(0.1085)	(0.1876)	(0.1843)	(0.1799)
treat × 2017	−0.1083	−0.1467	0.0249	−0.1194	−0.1766
	(0.1049)	(0.1142)	(0.2147)	(0.2081)	(0.1969)
Fixed-effects
Author	Yes	Yes	Yes	Yes	Yes
Year	Yes	Yes	Yes	Yes	Yes
Fit statistics
Observations	16,444	16,444	16,159	16,159	16,159
Squared Correlation	0.25931	0.26549	0.27743	0.26888	0.27172
Pseudo R^2	0.12860	0.11977	0.35738	0.33675	0.33729
BIC	47,842.0	36,031.8	373,062.6	170,670.2	309,882.0

Notes: Clustered standard errors at the author level in parentheses. Significance levels: ***: 0.01, **: 0.05, *: 0.1.

Poisson regression with standard errors clustered at the author level. Treated scholars refers to scholars only affiliated in China.

Table 2 shows that all our dependent variables moved in parallel for the treated and control groups before 2010. Only after Google’s exit from China, the interaction terms of the year dummies and the treatment group become significant showing different trajectories for the treatment and control group.

At first glance, it may seem surprising that the effect of the restricted access to information is visible immediately, i.e. in the first year after Google’s withdrawal of its services. For a discipline like economics with a lengthy peer-review process, one might expect that the effect would be visible only in later years after the treatment. Recent evidence, however, shows that when considering the globality of the available journals in this field, i.e. top as well as less-renowned ones, the average length of the review process in economics can be less than a year. Huisman and Smits (2017) report an average of 25 weeks from the submission to the acceptance and Björk and Solomon (2013) report an average of 18 months. In addition, it needs to be considered that the peer review process in economics typically takes several rounds (2.16 according to Huisman and Smits 2017). The nature of the reviewer comments changes over the different stages of the peer review process though. The first round of comments is the most important one for authors, as it defines how much time is lost in case of a rejection, as well as for academic journals for which the duration of the first-round review stage is an important indicator for journal management quality (Azar 2007, Solomon and Bjoerk, 2012; Huisman and Smits 2017). The most important comments are, hence, typically made in the first round. They relate to the core literature of the paper, raise technical issues or suggestions about the data and methods, and request further robustness checks (Allen et al. 2019). Once these key comments are addressed, in later rounds, reviewers tend to make more minor and more general comments which often span beyond the scope of the paper and target issues such as the broader implications of the study (Allen et al. 2019). Therefore, while robustness checks and comments targeting the core literature that authors receive during the first round of the peer review process can be addressed with limited new information, the papers that are close to publication tend to need more additional information dealing with the comments of a second or third round of the review process. This is reflected in the observed sudden drop in publications right after the shock which to a large extent may present a delay in the revision of second or third-round papers.¹⁷

5.2. Main results

Table 3 reports the results of the Poisson regressions for Equation (1)(1) $\mathrm{Publication}{\mathrm{s}}_{\mathrm{it}}=\mathrm{f}\left({\beta }_1\mathrm{Trea}{\mathrm{t}}_{\mathrm{i}}\ast \mathrm{Pos}{\mathrm{t}}_{\mathrm{t}}+\mathit{\delta }{\mathrm{\Gamma }}_{\mathrm{i}}+{\mathit{\phi }}_{\mathrm{t}}\right)+{\mathit{\epsilon }}_{\mathrm{it}}$(1) . It appears that, in line with the descriptive evidence, all our dependent variables are affected by the shock significantly at the 1% significance level. The average treated researcher loses 28% of her co-author weighted publications and 30% of her co-author weighted citations which corresponds to an average decrease of about 0.08 and 2 in fractional publications and citations, respectively. These results are robust if a synthetic difference-in-difference approach (Arkhangelsky et al. 2021) or an entropy balancing approach (Hainmueller 2012) is used. The results are presented in Tables A2 and A3 in the Appendix.

Table 3. Main results.

Dependent Variables:	number of publications	fractional publications	number of citations	fractional citations	citations per publication
Model:	(1)	(2)	(3)	(4)	(5)
Variables treat × post	−0.2868^***	−0.3306^***	−0.2670^***	−0.3570^***	−0.3422^***
	(0.0494)	(0.0536)	(0.0880)	(0.0885)	(0.0892)
Fixed-effects
Author	Yes	Yes	Yes	Yes	Yes
Year	Yes	Yes	Yes	Yes	Yes
Fit statistics
Observations	16,444	16,444	16,159	16,159	16,159
Squared Correlation	0.25799	0.26416	0.27542	0.26762	0.27007
Pseudo R^2	0.12815	0.11933	0.35654	0.33611	0.33669
BIC	47,768.6	35,952.2	373,439.7	170,730.2	310,057.0

Notes: Clustered standard errors at the author level in parentheses. Significance levels: ***: 0.01, **: 0.05, *: 0.1.

Poisson regression with standard errors clustered at the author level. Treated scholars refers to scholars only affiliated in China.

5.3. Foreign co-authors as a channel to access information

Above, we report that the publication volume and impact of scholars located in China dropped after Google’s withdrawal. The proposed mechanism is a decline in information accessibility. To investigate further whether this mechanism is at work, we distinguish between scholars located in China with and without foreign co-authors during our sample period. Interpersonal networks have been shown to be an important channel for knowledge diffusion (Mohnen 2022; Singh 2005) so that we expect that scholars located in China leverage their foreign co-author network to access information in the aftermath of the Google exit.

The regressions presented in Table 4 show the effect for the subsample of the treatment group which consists of scholars located in China with foreign co-authors versus the control group. We find that those scholars who still can access information through their co-authors’ network are affected less by Google’s exit supporting our proposed mechanism of information accessibility. The average treated researcher with foreign co-authors loses 20% of her co-author weighted publications and 22% of her co-author weighted citations.¹⁸

Table 4. Scholars located in China with foreign co-authors versus the control group.

Dependent Variables:	number of publications	fractional publications	number of citations	fractional citations	citations per publication
Model:	(1)	(2)	(3)	(4)	(5)
Variables treat × post	−0.1796^***	−0.2244^***	−0.1671^*	−0.2503^***	−0.2653^***
	(0.0528)	(0.0581)	(0.0931)	(0.0945)	(0.0934)
Fixed-effects
Author	Yes	Yes	Yes	Yes	Yes
Year	Yes	Yes	Yes	Yes	Yes
Fit statistics
Observations	14,057	14,057	13,834	13,834	13,834
Squared Correlation	0.25920	0.26454	0.26883	0.25192	0.26434
Pseudo R^2	0.12726	0.11747	0.35050	0.32541	0.33081
BIC	40,437.8	29,979.5	326,050.9	145,757.4	268,703.7

Notes: Clustered standard errors at the author level in parentheses. Significance levels: ***: 0.01, **: 0.05, *: 0.1.

Poisson regression with standard errors clustered at the author level. Treated scholars refers to scholars only affiliated in China with foreign co-authors.

5.4. A look at the top scholars and top institutions

The decline in publications would be more worrisome if top scholars were affected because only a small fraction of scientists is able to contribute to the advancement of science and because the top scholars are those who are likely to repeat their top performances (Cole and Cole 1972; Partha and David 1994; Merton, 1973). Should the top scholars’ performance, hence, decline, the threat for China to lose touch with the research frontier would be more severe.

Table 5 shows regressions for the subsamples of scholars with the highest and lowest impact in our sample, i.e. in the treated and control groups. The distinction is based on the citation stock divided by the publication stock in the year 2010. We chose the highest and lowest 25% percentile to have enough observations in each subsample for a credible regression analysis. The subsamples of the top and less impactful scholars contain an almost equal number of observations for scholars in the treatment and control groups which makes us confident that we are showing a meaningful comparison.¹⁹

Table 5. Top scholars and less impactful scholars.

	top 25% scholarsless impactful 25% scholars
Dependent Variables: Model:	number of publications (1)	fractional publications (2)	number of citations (3)	fractional citations (4)	citations per publication (5)	number of publications (6)	fractional publications (7)	number of citations (8)	fractional citations (9)	citations per publication (10)
Variables
treat × post	−0.4205^***	−0.5033^***	−0.3273^**	−0.4677^***	−0.3399^**	−0.1121	−0.2225^**	0.3489^*	0.2346	0.0902
	(0.1131)	(0.1279)	(0.1535)	(0.1571)	(0.1557)	(0.0953)	(0.1051)	(0.1886)	(0.1866)	(0.1872)
Fixed-effects
Author	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Year	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Fit statistics
Observations	3,439	3,439	3,439	3,439	3,439	4,599	4,599	4,430	4,430	4,430
Squared Correlation	0.28520	0.27909	0.25743	0.26856	0.25702	0.23239	0.23708	0.30331	0.25639	0.28599
Pseudo R^2	0.13515	0.11943	0.28715	0.29228	0.27304	0.11335	0.10786	0.37725	0.31569	0.33782
BIC	9,199.2	6,579.4	142,229	58,733.1	121,426.3	12,442.2	9,462.1	47,370.0	24,679.0	40,250.8

Notes: Clustered standard errors at the author level in parentheses. Significance levels: ***: 0.01, **: 0.05, *: 0.1.

Poisson regression with standard errors clustered at the author level. Treated scholars refers to scholars only affiliated in China.

Number of top 25% treated scholars: 281. Number of top 25% control scholars: 180. Number of less impactful 25% treated scholars: 293.

Number of less impactful 25% control scholars: 303.

The results show that the top scholars are more affected by the restricted access to information than the average researchers (compared to results in Table 4). When focusing on the bottom of the impact distribution, we find a smaller decline in co-author weighted publication outcome and no significant effect on the (already small) impact. More specifically we find that the average treated top researcher loses 39.5% of her co-author weighted publications and 37.5% of her co-author weighted citations. On the other hand, the average treated scholar in the lowest impact group only loses 20% of her co-author weighted publications and experiences no significant decline in her co-author weighted citations.²⁰

University’ s status may also play a role. It is expectable that top institutions might benefit from legalised access to Google services. We, therefore, identify treated scholars affiliated with the 39 universities within ‘Project 985’ and the 112 universities within ‘Project 211’, which are both nationwide projects aiming at creating elite universities, especially in terms of research output (Zhang, Patton, and Kenney 2013). We study whether these scholars, who might be expected to enjoy superior access to Google services through VPNs with the approval of the central government, are still negatively affected by the shock. As reported in Figures A1 and A2 in the Appendix, we find that both groups of treated scholars are equally negatively affected.

5.5. Robustness checks

Our results are robust to an estimation without fixed effects and to an estimation that includes control variables (see Tables A1 and A2 which show robustness for the full sample (Table A1) and the subsample of those that have foreign co-authors (Table A2)). Furthermore, we present matched sample regressions that account for differences in scientists’ productivity before Google’s exit from China (see Table A3). The results are presented in the Appendix.

We, further, investigated whether the additional local internet restrictions in the form of local web filters at the level of the province matter (Kong et al. 2022; Xu, Xuan, and Zheng, 2021). One could imagine that scientists in provinces with local filtering devices have more difficulty circumventing the GFC to access Google to search for information after 2010. We did not find a different impact of the withdrawal of Google’s services in provinces with local filtering, which affects 22% of our observations, and without local filtering. This result may be explained by the fact that only a minority of scientists try to circumvent the GFC (Lu et al. 2017) perhaps also due to the high risk of penalties.²¹ The results are available upon request.

6. Discussion

6.1. Summary

Knowledge and information are indispensable inputs for the knowledge production function (Ding et al. 2010; Stephan and Levin 1992). With the increasing complexity and multidisciplinarity of science over time (Jones 2009; Wuchty, Jones, and Uzzi 2007), access to information is of utmost importance for scientists aiming to contribute to the research frontier (Berkes and Nencka 2019; Biasi and Moser 2021; European Commission 2012; Furman and Stern 2011; Furman, Jensen, and Murray 2012; McCabe and Snyder 2015; Mueller-Langer, Scheufen, and Waelbroeck 2020; Waldinger 2016).

In this paper, we employ Google’s withdrawal from China in 2010 as an event that allows testing the impact of increased barriers to access information on scientists’ publication output and impact. Google’s search engine was a major channel for scholars located in China to obtain foreign information so that its sudden withdrawal severely hampered the ability of scientists to access the knowledge frontier (Qiu 2010). Our results from DiD analyses that compare scholars in the field of economics located in China to a culturally, economically, and geographically close control group show that publication output dropped by 25% and by 28% if weighted by co-authors. In addition, publication impact measured through co-author weighted citations dropped by 30% and citations per publication dropped by 29%. These results contribute to our understanding of the determinants of knowledge creation (Stephan and Levin 1992; Stephan 1996, for an overview) and more specifically of the role of information and communication technology in knowledge creation (Agrawal and Goldfarb 2008; Ding et al. 2010). While prior studies have shown that access to network technology eases information accessibility and facilitates the knowledge production of scientists (Agrawal and Goldfarb 2008; Ding et al. 2010), we complement prior findings and provide evidence for the importance of Google’s services for science production in economics.

By providing evidence for the effect of a negative shock of information availability to assess the effects on science production in economics, we contribute to recent literature that focuses on the knowledge creation effect of positive information shocks such as the availability of access to libraries (Berkes and Nencka 2019; Biasi and Moser 2021; Furman, Jensen, and Murray 2012), of research resources (Furman and Stern 2011) and of online access to scientific journals (McCabe and Snyder 2015; Mueller-Langer, Scheufen, and Waelbroeck 2020). Our findings confirm that reactions to a positive and negative shock are not symmetric. For a positive shock, it takes time for the knowledge production function of the majority of the scientists to adjust (e.g. Panel A and B of Figure 8 in Furman, Nagler, and Watzinger 2021, for the effect of increased knowledge access through the United States Patent and Trademark Library systems on local patenting: it takes some time until the full benefits for local patenting realise). The effect of a negative shock, on the contrary, is expected to occur immediately as an existing knowledge production process is suddenly interrupted as observed in our analysis: the effect of the negative shock kicks in immediately, and fades out over time, in stark contrast to the delayed reactions to a positive shock.

We propose that the underlying mechanism of this decline is information accessibility and test this hypothesis further by showing that the productivity decline is smaller for scientists who can leverage their foreign network to access information. Here, the number of publications decreases by 16.5% and co-author weighted publications by 20%. In terms of quality, co-author weighted citations dropped by 22% and citations per publication dropped by 23%. The finding that scholars can use their interpersonal networks as a channel for knowledge access (Mohnen 2022; Singh 2005) contributes to the large literature on the positive effects of academic networks (e.g. Beaver and Rosen 1978; Fanelli, Larivière, and Dorta-González 2016; Greene 2007; Wuchty, Jones, and Uzzi 2007) and, in particular, on informal collaboration between researchers (Brown 2005; Laband and Tollison 2000; Rose and George 2021). We add by providing suggestive evidence that co-author networks facilitate access to knowledge beyond joined projects which leads to a lower drop in scientific productivity and citations in the presence of an information shock. Further results show that especially the top scholars located in China are affected in terms of their output and impact. The decline in top scholars’ citations likely reflects a lack of novelty caused by access barriers to the novel frontier.^22,23 Losing touch with the research frontier can lead to a persistent lag behind foreign peers with potentially harmful implications for economic growth (Griliches 1992; Jaffe 1989) because the more novel discoveries have a higher chance to have an impact on technology development (Veugelers and Wang 2019). It is worth mentioning the extreme relevance that citations have with respect to labour market outcomes (Ellison 2013; Hamermesh and Pfann 2012), especially for low-ranked departments (Gibson, Anderson, and Tressler 2017). Leveraging the recent finding by Koffi (2021), who shows that fewer citations have a negative effect on authors’ future work, our results, especially for top researchers, are even more relevant for the future of scientists in China.

6.2. Broader implications

Our results also draw attention to the current influence of big tech firms, way beyond their commercial power (Igna and Venturini 2023; Khan 2018; Petit and Teece 2021). While large companies existed before and despite the fact that societies developed regulations to protect consumers, competition, and the environment, big tech firms are powerful in new ways that derive from their control over digital technology that can grant or limit access to information, the most crucial resource of our fast-moving world (e.g. Igna and Venturini 2023; Rikap and Lundvall 2022; Yu, Liang, and Wu 2021). In this position, big tech firms can also have a not neglectable influence on the creation of science, as we show using the example of economics.

The fact that Google’s withdrawal from China affected researchers affiliated to both top and less prestigious universities alike (see Figures A1 and A2 in the Appendix) demands a deeper reflection on the impact of big tech firms on a much wider variety of elements of our society including science and research.

6.3. Limitations and future research

A limitation of the paper is that our results might not be generalisable to other fields of science. Economics is a scholarly discipline with a simple production function. For hard sciences, where next to information about the state of the art in the field, also specialised equipment is often essential, we could imagine that the effect of limited access to information has even larger effects on science production since also information about the latest advances in specialised equipment can be missing. In addition, due to a lack of information on the location of authors with multiple affiliations, we needed to drop these scientists. Similarly, due to a lack of information about the reasons for mobility of individual scientists, we also needed to drop those scientists who changed their affiliation from one country to another from our sample. Ideally, we would have access to this information as our results might be biased towards the less successful scientists.

We also acknowledge the limitations related to our measure of funding. Bibliometric data on funding relies heavily on the authors’ funding acknowledgements (Sugimoto and Larivière 2023) and databases like Scopus and Web of Science are not able to present complete information (Liu 2020). Nonetheless, we are encouraged by the fact that policies regarding grant acknowledgement vary significantly by country, and China, for example, has extremely strict policies in this regard (Sugimoto and Larivière 2023, 96), alleviating, therefore, some of our concerns. This is also one of the main reasons we decide to study the field of economics.

We also acknowledge that our results may be specific to the context of China. Access to information is only one factor in the scientific production function and the entire academic and political environments are likely to matter as well. We expect that in an advanced academic environment, which is rich in material resources, the effects of limited access to knowledge may be somewhat smaller, while it matters more in less developed environments (Mueller-Langer, Scheufen, and Waelbroeck 2020).

Furthermore, we acknowledge that Google’s withdrawal from China cannot be entirely separated from coincident incremental policy changes that may have had an additional effect on publications in economics. To the best of our knowledge, no major policy change with implications for publications has taken place during our time frame of study such as the new evaluation policy of 2020 that requires Chinese scholars to publish at least one-third of their research in Chinese journals (Liang, Gu, and Nyland 2022). Nevertheless, we acknowledge that the Chinese research landscape is in constant development (Piracha et al. 2022) with potential implications for our estimated effects.

Lastly, we acknowledge that our difference-in-difference setting with a country-wide shock can only rely on a ‘second best’ type of control group. While in an ideal setting, we would observe treated and control scholars in the exact same conditions, this is by definition not possible for a country-wide shock. This is why we use a ‘second best’ control group following Zheng and Wang (2020).

The extension of our analysis to other science fields and other political and academic contexts is hence a straightforward avenue for future research. In addition, future research could explore the effect of governmental initiatives to censor access to information for scientific productivity (Ritchie, Driscoll, and Maron 2017).

Acknowledgements

We would like to thank Benjamin Balsmeier and Bettina Peters for helpful comments. In addition, we would like to thank participants of the KU Leuven Research Retreat 2022, the 2022 EPIP Conference, the 9^th ZEW/MaCCI Conference – INNOPAT, the 2022 RISE Workshop, the 16^th Workshop on Organisation, Economics and Policy of Scientific Research (WOEPSR), the PDW on Innovation, Technology and R&D Management at the University of Essex, the Munich Summer Institute 2023, and the participants to the internal seminar at ESADE and Copenhagen Business School.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

We cannot share our dataset since Scopus is a commercial provider.

Notes

¹ Hereafter, we refer to mainland China simply as ‘China’.

² https://gs.statcounter.com/search-engine-market-share/all/china/#quarterly-200901–201601

³ https://gs.statcounter.com/search-engine-market-share/all/china/#quarterly-200901–201601

⁴ In January 2010, following a major cyber-attack on Google, originating from China, it has been uncovered that accounts of dozens of human rights activists connected with China were being routinely hacked. This, ‘combined with attempts over the last year to further limit free speech on the web in China including the persistent blocking of websites such as Facebook, Twitter, YouTube, Google Docs and Blogger’ (Drummond 2010), led Google to discontinue its censoring activities on search results from Google.cn.

⁵ Towards the end of March 2010, frictions between the Chinese Communist Party (CCP) and Google’s executives due to censorship issues and hacking attempts led Google to withdraw its search engine from China, meaning that Google.cn was not working anymore. Google users’ search requests, after the 23^rd of March 2010, were redirected to Google’s Hong Kong servers, but, as Kong et al. (2022, 5) points out: ‘[t]he Chinese government criticized Google’s withdrawal as unfriendly and irresponsible and blocked Google’s Hong Kong search website and its search websites in all other languages on March 30, 2010. Google then stopped redirecting visits to its Chinese search website to its Hong Kong website starting from June 30, 2010 (Cheng 2010). From there on, accessing search results via Google has become excessively difficult from mainland China’. The initial redirection to the Hong Kong servers was applied only to Google Search, Google News, and Google Images, while other specialised services of Google, such as Google Maps, Google Music, and Google Shopping, remained available in China and were shut down from 2012 onward. While scholars like Zheng and Wang (2020) focus on 2014 as the year of the ‘shock’ since this is the date in which all of the services offered by Google were unavailable in China, we, in the spirit of Kong et al. (2022) and Xu et al. (2021), use the year 2010 as the relevant year of the ‘shock’, i.e. when Google’s search engine became unavailable in China. We do so because at that time Google search engine was the essential service to access information about science as Google Scholar was still underdeveloped and underfeatured as compared to its current version.

⁶ Note that we do not specifically refer to Google scholar.

⁷ In fact, for quite some time ‘there is growing evidence that both novice and experienced searchers are increasingly using simple single text box search interfaces such as those provided by search engines like Google (http://google.com)’ (Hemminger et al. 2007, 2214).

⁸ A scholar affected by the withdrawal of Google can still have access to specific websites. Without a centralised platform such as Google search, however, the researcher would either need to know the exact source of the piece of information she is looking for or would have to invest a significantly larger amount of time to look for it (compared to accessing it through Google search). This would lead to a delay of her publications.

⁹ As noted by Jennings (2010), ‘[t]he rise of VPNs comes as China defends its curbs on the internet after the world’s biggest search engine provider, Google Inc., threatened to shut down its Chinese Google.cn site over censorship and a severe hacking attack’. This means that VPNs in China back in 2010 were not a main instrument to circumvent the Great Firewall of China (GFC). In addition, there is evidence that the Chinese government was strongly against VPNs already back in 2011 with users reporting unstable connections when trying to access foreign websites. All the above evidence points to the fact that VPNs back in 2010 were not able to provide scholars with stable access to information.

¹⁰ Mirror platforms, such as ‘scholar.glgoo’, aim at mirroring existing but inaccessible websites due to the Chinese internet censorship. Such platforms were not available around the 2010 shock year (https://web.archive.org/web/20230000000000*/https://scholar.glgoo.org/).

¹¹ Lu et al. (2017) surveyed 371 faculty members and students at Tsinghua University, one of the top academic institutions in China, in 2015, on whether and how they can bypass the GFC. Even though 26% of the respondents claimed that they can regularly bypass the Great Firewall through VPNs, none of the commonly adopted solutions have provided satisfactory user experiences.

¹² In addition, (1) Baidu mixes and prioritizes a large proportion of advertisements in its search results (Yi 2014) while Google displays paid search results separately leading to a low overlap and little ranking similarity in the search results between the two search engines (Jiang 2014a, 2014b; (2) the quality of search results in Baidu was even poorer back in the years around 2010 (CNNIC 2011), in fact 44% of the respondents to the Annual national survey on search engines in China conducted by the China Internet Network Information Center (CNNIC) criticised Baidu for ‘garbage information and bad links’ (Kong et al. 2022, 5); and (3) Google appeared to continue providing uncensored search results from foreign websites despite the agreement with the Chinese government (Kong et al. 2022; Lau 2010; Thompson 2006; Wilson, Ramos, and Harvey 2007; Xu, Xuan, and Zheng 2021), hence providing higher-quality search results to users’ requests.

¹³ To obtain a clean setup for this study, we do not allow scholars in the treated group to be affiliated with institutions outside of China. That is, if a scholar is affiliated both in China and outside of China, she is not included in our treated group as we would not have information on her physical location. In fact, when analysing the impact of the shock on scholars with double affiliations, we do not find any significant effect.

¹⁴ According to Kawashima and Tomizawa (2015) Scopus authors’ identifiers are reliable, reaching about 98% in terms of recall and 99% in terms of precision (see also Baas et al. 2020).

¹⁵ This affected 1,189 observations of 237 authors.

¹⁶ For example, we found that the International Journal of Biological Macromolecules was wrongly assigned to the field of economics and dropped all the misclassified authors (15,942 authors with 24,844 observations) that published in the respective research discipline.

¹⁷ To investigate the immediateness of the effect further, we check the parallel trend only for the top journals in economics, i.e. authors that published in the top journals in economics before 2010. Those journals include: Econometrica, American Economic Review, Journal of Political Economy, Quarterly Journal of Economics, Journal of Finance, Review of Economic Studies, Journal of Financial Economics, Journal of Economic Literature, Review of Financial Studies, Journal of Economic Perspectives, American Economic Journal: Macroeconomics, Journal of Accounting and Economics, American Economic Journal: Applied Economics, Review of Economics and Statistics, American Economic Journal: Economic Policy, Journal of Marketing, Journal of Management, Review of Corporate Finance Studies, Journal of Consumer Research, Annual Review of Economics, NBER Macroeconomics Annual, Marketing Science, Journal of Accounting Research, American Economic Journal: Microeconomics. This leaves us with a sample of 309 or 299 observations respectively. Note that all of the authors that publish in top journals have foreign co-authors. We find no effect, which provides some validity to our explanation of the immediateness of the effect which is most likely due to the significantly shorter review process in less-renowned economics journals. The results are available upon request.

¹⁸ The parallel assumption holds for this subsample. Results are available upon request.

¹⁹ There are 281 treated scholars and 180 control scholars in the group of the top 25%. The group of the less impactful 25% includes 293 treated and 303 control scholars.

²⁰ The parallel assumption holds for this subsample. Results are available upon request.

²¹ Penalties for VPN usage start from as low as 100 CNY fines but can go up to 5 years in jail (see, for example, Hawkins 2023, or; Haas 2017).

²² The number of backward citations that are less than three years old drops for both treated and control groups from the pre- to the post-shock period, but the drop for the treated group is statistically larger than for the control group. Similar results are found when comparing the number of backward citations that are less than two, four, and five years old. In addition, when comparing the number of backward citations that are more than two, three, four, and five years old, we find that these increases are stronger and more significant for the treated group.

²³ The decline in citations to the treated scholars’ work might be also explained by a decrease in the articles’ visibility from the standpoint of other treated Chinese economists, i.e. the higher barriers to access information restrict the visibility of the treated articles such that the decline in citations might not be due to a decrease in quality, but, instead, to a decrease in visibility from a specific sample of scholars’ point of view. Nevertheless, we argue that if this is true, the decreased visibility is likely to be negatively affecting both control and treated scholars’ work alike.

²⁴ The parallel trend assumption holds for the matched sample. Results are available upon request.

Appendix

We present several robustness checks in this Appendix. The tables below show robustness checks for alternative specifications of Equation (1)(1) $\mathrm{Publication}{\mathrm{s}}_{\mathrm{it}}=\mathrm{f}\left({\beta }_1\mathrm{Trea}{\mathrm{t}}_{\mathrm{i}}\ast \mathrm{Pos}{\mathrm{t}}_{\mathrm{t}}+\mathit{\delta }{\mathrm{\Gamma }}_{\mathrm{i}}+{\mathit{\phi }}_{\mathrm{t}}\right)+{\mathit{\epsilon }}_{\mathrm{it}}$(1) . A1 and A2 show regression models without fixed effects and including control variables when the treated group of scholars does and does not have foreign co-authors, respectively. It appears that our main effect is robust against these different specifications of our model.

Table A1. Robustness checks I – full sample.

Dependent Variables:	number of publications		fractional publications		number of citations		fractional citations		citations per publication
Model:	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)	(10)
Variables (Intercept)	−0.7280^***	−0.3617^***	−1.465^***	−1.079^***	2.131^***	2.697^***	1.303^***	1.885^***	1.987^***	2.640^***
	(0.0260)	(0.0364)	(0.0376)	(0.0535)	(0.0062)	(0.0083)	(0.0094)	(0.0128)	(0.0067)	(0.0089)
treat	0.2223^***	−0.0016	0.1214^**	−0.1077^**	0.6676^***	0.3733^***	0.5590^***	0.2580^***	0.6133^***	0.3145^***
	(0.0356)	(0.0365)	(0.0528)	(0.0541)	(0.0078)	(0.0079)	(0.0120)	(0.0123)	(0.0084)	(0.0086)
post	0.0443	0.1344^***	−0.0110	0.1052^**	−0.2652^***	−0.0499^***	−0.2911^***	−0.0520^***	−0.2953^***	−0.0176^*
	(0.0327)	(0.0357)	(0.0478)	(0.0525)	(0.0083)	(0.0091)	(0.0127)	(0.0139)	(0.0090)	(0.0099)
treat × post	−0.2764^***	−0.3173^***	−0.3320^***	−0.3705^***	−0.2011^***	−0.2453^***	−0.2878^***	−0.3392^***	−0.2616^***	−0.2483^***
funding stock	(0.0449)	(0.0461) 0.2323^***	(0.0677)	(0.0693) 0.2275^***	(0.0104)	(0.0107) 0.2518^***	(0.0162)	(0.0167) 0.2522^***	(0.0114)	(0.0117) 0.2242^***
career age		(0.0038) −0.1465^***		(0.0062) −0.1499^***		(0.0008) −0.2136^***		(0.0013) −0.2162^***		(0.0010) −0.2407^***
career age square		(0.0089) 0.0068^***		(0.0135) 0.0067^***		(0.0021) 0.0093^***		(0.0033) 0.0091^***		(0.0023) 0.0103^***
		(0.0005)		(0.0007)		(0.0001)		(0.0002)		(0.0001)
Fit statistics
Observations	16,750	16,749	16,750	16,749	16,750	16,749	16,750	16,749	16,750	16,749
Squared Correlation	0.00366	0.04209	0.00625	0.03209	0.01801	0.02889	0.01547	0.02353	0.01856	0.02831
Pseudo R^2	0.00195	0.06820	0.00328	0.04751	0.03144	0.13730	0.02548	0.11974	0.03037	0.10373
BIC	31,420.1	29,365.3	17,516.2	16,769.3	543,145.3	483,797.9	225,449.7	203,668.6	432,936.5	400,198.8

Notes: Clustered standard errors at the author level in parentheses. Significance levels: ***: 0.01, **: 0.05, *: 0.1.

Poisson regression with standard errors clustered at the author level. Treated scholars refers to scholars only affiliated in China.

Table A2. Robustness checks II – treated scholars with foreign co-authors.

Dependent Variables:	number of publications		fractional publications		number of citations		fractional citations		citations per publication
Model:	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)	(10)
Variables (Intercept)	−0.7280^***	−0.3897^***	−1.465^***	−1.118^***	2.131^***	2.692^***	1.303^***	1.864^***	1.987^***	2.628^***
	(0.0260)	(0.0376)	(0.0376)	(0.0557)	(0.0062)	(0.0086)	(0.0094)	(0.0133)	(0.0067)	(0.0092)
treat	0.2001^***	0.0081	0.0295	−0.1676^***	0.7011^***	0.4446^***	0.5243^***	0.2639^***	0.6452^***	0.3873^***
	(0.0393)	(0.0399)	(0.0597)	(0.0607)	(0.0083)	(0.0084)	(0.0131)	(0.0133)	(0.0090)	(0.0092)
post	0.0443	0.1295^***	−0.0110	0.0923^*	−0.2652^***	−0.0408^***	−0.2911^***	−0.0551^***	−0.2953^***	−0.0132
	(0.0327)	(0.0359)	(0.0478)	(0.0529)	(0.0083)	(0.0091)	(0.0127)	(0.0140)	(0.0090)	(0.0100)
treat × post	−0.1304^***	−0.2283^***	−0.1810^**	−0.2848^***	−0.0724^***	−0.1647^***	−0.1421^***	−0.2477^***	−0.1632^***	−0.1931^***
funding stock	(0.0487)	(0.0499) 0.2225^***	(0.0751)	(0.0769) 0.2227^***	(0.0109)	(0.0112) 0.2373^***	(0.0174)	(0.0179) 0.2402^***	(0.0120)	(0.0123) 0.2099^***
career age		(0.0040) −0.1349^***		(0.0066) −0.1355^***		(0.0008) −0.2085^***		(0.0014) −0.2065^***		(0.0011) −0.2334^***
career age square		(0.0094) 0.0062^***		(0.0144) 0.0060^***		(0.0022) 0.0089^***		(0.0036) 0.0086^***		(0.0025) 0.0099^***
		(0.0005)		(0.0007)		(0.0001)		(0.0002)		(0.0001)
Fit statistics
Observations	14,285	14,285	14,285	14,285	14,285	14,285	14,285	14,285	14,285	14,285
Squared Correlation	0.00212	0.04159	0.00217	0.02904	0.02031	0.03209	0.01272	0.02310	0.01896	0.02995
Pseudo R^2	0.00116	0.06412	0.00114	0.04476	0.03631	0.13435	0.02152	0.11062	0.03167	0.09926
BIC	27,309.5	25,619.2	15,034.7	14,408.6	468,123.7	420,532.1	190,901.5	173,551.7	372,032.2	346,094.3

Notes: Clustered standard errors at the author level in parentheses. Significance levels: ***: 0.01, **: 0.05, *: 0.1.

Poisson regression with standard errors clustered at the author level. Treated scholars refers to scholars only affiliated in China with foreign co-authors.

Table A3. Matched sample.

Dependent Variables:	number of publications	fractional publications	number of citations	fractional citations	citations per publication
Model:	(1)	(2)	(3)	(4)	(5)
Variables treat × post	−0.3079^***	−0.3714^***	−0.3053^**	−0.4558^***	−0.3497^***
	(0.0685)	(0.0755)	(0.1307)	(0.1313)	(0.1318)
Fixed-effects
Author	Yes	Yes	Yes	Yes	Yes
Year	Yes	Yes	Yes	Yes	Yes
Fit statistics
Observations	10,131	10,131	9,912	9,912	9,912
Squared Correlation	0.24099	0.24957	0.23580	0.24758	0.24018
Pseudo R^2	0.11937	0.11694	0.33497	0.32342	0.32148
BIC	27,709.5	21,011.3	203,283.4	94,016.7	175,067.2

Notes: Clustered standard errors at the author level in parentheses. Significance levels: ***: 0.01, **: 0.05, *: 0.1.

Poisson regression with standard errors clustered at the author level. Treated scholars refers to scholars only affiliated in China. 665 pairs of matched scholars.

A surprising effect is that we find a U-shaped effect of career age, while typically an inverse U-shape is found. This is driven by the first publication year of our authors in which they have a high publication rate. If the first year of publication is excluded, we find the typical pattern of an increase in publications which at some age peaks and declines.

Table A3 shows robustness for a matched sample. We use nearest neighbour matching on authors’ citations stock, co-author weighted citations stock, publications stock, co-author weighted publications stock, and funding stock as of 2010. We have matched 665 treated scholars to their most comparable peers in the control group. The 665 treated scholars and their matched controls are observed on average in 7.62 years which leads to a total sample of 10,131 scientist-year observations. The matched sample is balanced in terms of the matching criteria and the results are in line with the main findings of Table 3. The average treated researcher, in fact, loses 31% and 36.5% of her co-author weighted publications and citations, respectively.²⁴

Tables A4 and A5 report the results when using a synthetic DiD (Arkhangelsky et al. 2021) and an entropy balancing approach (Hainmueller 2012), respectively. We use career age, funding stock, publication stock, and citations over publications stock to define the balancing.

Table A4. Synthetic DiD results.

Dependent variable	ATT	se	t	P > |t|	95% confidence interval
Number of publications	−0.10412	0.01536	−6.78	0.000	−0.13423	−0.07401
Fractional publications	−0.03706	0.00836	−4.44	0.000	−0.05343	−0.02068
Number of citations	−1.98438	0.65555	−3.03	0.002	−3.26923	−0.69953
Fractional citations	−0.77467	0.27070	−2.86	0.004	−1.30523	−0.24411
Citations per publication	−1.26016	0.57935	−2.18	0.030	−2.39567	−0.12465

Table A5. Entropy-balancing results.

Dependent Variables	number of publications	fractional publications	number of citations	fractional citations	citations per publication
Model:	(1)	(2)	(3)	(4)	(5)
Variables treat × post	−0.299*** (0.090)	−0.270** (0.132)	−0.426*** (0.020)	−0.440*** (0.030)	−0.283*** (0.022)
Fixed-effects Author Year Fit statistics Observations	Yes Yes 16174	Yes Yes 16174	Yes Yes 15890	Yes Yes 15890	Yes Yes 15890
Log likelihood	−3041.337	−1582.674	−49486.628	−21196.999	−40383.314
Prob>chi2	0.000	0.000	0.000	0.000	0.000

Notes: Clustered standard errors at the author level in parentheses. Significance levels: ***: 0.01, **: 0.05, *: 0.1. Poisson regression with standard errors clustered at the author level.

The Figures below show similarity in the effect of the shock on treated scholars affiliated with both top and less renowned universities.

Figure A1. Parallel trends for top universities’ scholars vs control group.

Figure A2. Parallel trend for less renowned universities’ scholars vs control group.