WHAT ROLE DO NEW DATABASES PLAY IN THE CODING PROCESS?

Many event datasets use the same basic coding process: (1) gather a corpus of potentially relevant documents, (2) determine if those documents meet the criteria for inclusion in coding, (3) assign coding determinations following predetermined rules, and (4) review the results for accuracy.

News databases enable researchers to quickly locate a corpus of relevant documents by searching for particular sources and keywords. Figure 1 is a conceptual diagram of the coding process.

Figure 1 Illustration of the Coding Process Using Newspaper Databases

The application programming interface (API) allows the search application to communicate with digital archives. APIs can be used directly through applications written by a researcher or through command-line entries. Access to the API often is restricted to specialized, paid licenses. The user interface (UI) is the front-end interface often found on a database’s proprietary programs. Information entered into the UI then is communicated through the API to the database corpus. These interfaces are managed by the databases themselves. For news databases, both the API and the UI serve as intermediaries between the search query and the articles that comprise the database.

The data-gathering process using news databases begins with the formation of the search query, which is defined as the combination of search terms, specified sources, and date ranges that a researcher wants to examine. Although researchers can type simple search terms into database UIs, most are compatible with Boolean search terms. Boolean search is a logic-based approach to search queries in which keywords are strung together with operators including the words AND, OR, and NOT as well as brackets and quotations to form detailed queries.

After specifying the terms, a researcher enters queries into the database UI. The UI interprets the query and passes it along to the API, which—in turn—sends the request to the database. The database then returns a list of all documents that meet the search criteria to the UI, where the researcher can view the results. When using the API directly, the query can specify that results are returned directly into various file types. Once the researcher has access to the documents, they are filtered through to find which are relevant. In some projects, researchers begin reading through the documents, coding as they proceed. In other projects, the numerous results from their initial query require narrowing down through machine, manual, or hybrid approaches.

There is significant variability in how coding teams using news databases proceed, including how the corpus of documents is determined and whether analysis of selected articles is accomplished manually or through machine processing. Table 1 illustrates the process of a few major event datasets. Although coding projects vary in how they distill the corpus into data, they share a reliance on databases to reliably and accurately provide requested documents.

Table 1 Comparison of Selected Conflict Datasets

Sources: National Consortium for the Study of Terrorism and Responses to Terrorism (2018); Schrodt, Palmer, and Hatipoglu (Reference Schrodt, Palmer and Hatipoglu2008)

WHAT ARE NEWS DATABASES AND WHAT IS CHANGING?

The most prominent databases for researchers are LexisNexis products (e.g., Nexis Uni), Factiva, and Google News. In the case of proprietary databases such as Nexis Uni and Factiva, news agencies are contracted to allow inclusion of the creator’s content within the database. When users search, they can view an article and, when permitted, download the complete text. These databases differ from open services such as Google News, in which the user is directed to the content creator’s website. Access to proprietary databases often is granted through university libraries due to the substantial cost of obtaining access.Footnote ³

There is a small but important body of literature that examines the databases themselves. This includes research that analyzes the similarities in news stories among databases (Weaver and Bimber Reference Weaver and Bimber2008); the overlap in stories from the same news sources between databases and physical print (Ridout, Fowler, and Searles Reference Ridout, Fowler and Searles2012); and the overlap in content availability between databases and microfilm (Youngblood, Bishop, and Worthington Reference Youngblood, Bishop and Worthington2013). Existing work, however, has not examined how databases change over time or how those changes affect the discipline.

It takes a significant disruption to alert researchers that anything may be amiss. Most news-database changes are gradual and invisible to the end user and, as such, can go unnoticed. We identify the main changes to these services and their consequences. For clarity, we divide them into two categories: changes to terms of use and changes in functionality.

Changes to Terms of Use

Changes to terms of use are alterations in what is allowed under the database licensing agreements. They are driven largely by contracts between databases and content creators. The most relevant changes that we identified are article-download restrictions, article-retention limits, cross-institutional collaboration, and restrictions on machine- or crowd-assisted coding and processing. Table 2 summarizes these policies for the major news databases. Although some of these activities have always been prohibited, it is only recently that database managers have started policing violations.

Table 2 Comparison of Database Terms of Use for 2020

Downloading source materials plays a central role in many coding efforts. Downloading articles allows coders to centrally store all source material and use alternative software to interact with the articles (e.g., Python). Although many databases allow downloading, functionality often is limited to one article at a time. This can be prohibitively slow if many articles are needed. When downloads are allowed, users may retain sources only for a finite amount of time, during which researchers must extract all needed information from texts before deleting them.Footnote ⁴ The ability to return to coding documentation to address questions regarding the data is crucial (Hensel and Mitchel Reference Hensel and Mitchell2015, 119). The inability to do so prevents replication and limits the ability to question or correct coding errors.

Current restrictions also prohibit sharing content with individuals outside of the user’s institution.Footnote ⁵ Many event-data creators collaborate with researchers at other institutions. Additionally, this restriction prohibits datasets from sharing the text of their source documents if it is requested. Many coding projects use machine document processing to expedite the coding process. This is explicitly prohibited by most databases, and violations risk the academic license of the user’s institution.Footnote ⁶ Users who want to use automated approaches must obtain alternative and often more-expensive licenses.

Changes in Functionality

Two changes to database functionality may cause researchers to miss events of interest: updated search-string interpretation and changing source lists. Search-string interpretation refers to situations in which the database UI interprets a researcher’s query in unintended ways. We observed this happening through changing abbreviations for common sources, interpretation of dates, and requirements for inclusion/exclusion terms.

For example, the search string for the MID4 project abbreviated the source Agence France-Presse as “AFP.” In coding the MID5 project, it was discovered that this shorthand no longer functions in Nexis Uni, resulting in “AFP” being excluded from the original batch of downloads for 2012. This exclusion was discovered in a subsequent audit of the search results. After correcting this and gathering all of the relevant AFP stories for 2012, 25 new incidents and three new disputes were discovered (Palmer et al. Reference Palmer, McManus, D’Orazio, Kenwick, Karstens, Bloch, Dietrich, Kahn, Ritter and Soules2022). Had the MID team not performed multiple audits of its search results, these disputes would not have been included in the final dataset. The audit also revealed that the initial batch of downloads excluded stories from Interfax/ITAR-TASS for 2011. Eight new disputes were coded from those retrieved stories.Footnote ⁷

Similarly, in Factiva, we encountered issues with dates in which specifying a range of dates (e.g., January 1–January 31) was interpreted in such a way that it excluded stories from the final day of the range (e.g., January 1, 12:01 a.m.–January 31, 12:01 a.m.). Updates such as these mean that researchers must perform periodic audits to ensure search strings are working as intended.

The sources available through the database also change over time. Contracts periodically expire or are renegotiated, resulting in sources dropping out of the database. This happens unannounced and intermittently. Additionally, the presence of the same source in two databases does not ensure access to the same publications. For example, Nexis Uni and Factiva both have Xinhua in their source lists but they contain different Xinhua publications.

Thus, search results may vary depending on factors hidden from the researcher. This is particularly disconcerting for projects that attempt to approximate the universe of cases. Whereas major events are unlikely to be omitted due to the substantial coverage they receive, smaller events that receive less media attention are more likely to be overlooked when a news story is lost in a database query.

To illustrate this issue, we tracked changes in search results on Factiva and Nexis Uni between February and October of 2019. Using search strings drawn from the MID Project (Palmer et al. Reference Palmer, McManus, D’Orazio, Kenwick, Karstens, Bloch, Dietrich, Kahn, Ritter and Soules2022) as well as SCAD (Salehyan et al. Reference Salehyan, Hendrix, Hamner, Case, Linebarger, Stull and Williams2012), we recorded the number of resulting articles for three periods: 2011–2018, January 2015, and January 2011 (figures 2 and 3). Both Factiva and Nexis Uni returned different numbers of articles for identical searches conducted over time within the same databases.Footnote ⁸ These week-to-week changes typically were small (i.e., <1% of total articles). We identified a small number of more significant changes, including one instance in which Factiva almost doubled the number of Associated Press stories returned in search results. The full results of this exploration are in the online appendix.

Figure 2 Variations in the Number of Factiva Search Results over Time

Figure 3 Variations in the Number of Associated Press Stories Retrieved Using the MID Search String

DOES DATABASE INSTABILITY AFFECT INFERENCE?

Instability in online news databases has the potential to affect datasets by preventing real events from appearing in the final dataset. An event that meets the coding criteria might be excluded if all source articles that document it are excluded from a database search. How often does this happen?

We are unable to provide a definitive answer without knowing which articles we have not observed. However, we can offer evidence about which events in published datasets are most at risk of exclusion and failure to replicate. The UCDP Georeferenced Event Dataset (GED) Version 21.1 (Sundberg and Melander Reference Sundberg and Melander2013) records information about political-violence events. For events in the GED collected or updated since 2013, UCDP records the number of source articles used to code that event. Events with multiple sources should be more robust to fluctuations in database availability because the event still will be included, even if any single article is missing. Events with a single source, however, are more susceptible to these fluctuations because the omission of a single article could lead to the exclusion of that event. In the UCDP GED dataset, there are 157,303 events with source information; 114,978 (73.1%) are coded from only a single source.

A second consideration is whether the inclusion or exclusion of events from a dataset affects inferences derived from it. We might conceptualize the exclusion of events due to database instability as stochastic measurement error unlikely to affect substantive conclusions of an analysis. However, there are reasons to believe that such errors are consequential for inference.

First, researchers often study a subset of events within a particular dataset. The omission of a single event from an entire dataset is unlikely to change the substantive conclusions gleaned. If, however, a researcher is performing an analysis of only a single country, then the omission of even a single event can be highly consequential. Such omissions always will have the effect of reducing the power of the analysis and increasing uncertainty around estimates.

Second, some events are more likely than others to be omitted because of database instability, which results in systematic bias. Events with fewer sources are at greater risk of omission from a dataset because there is a greater possibility that all source articles reporting on the event will be omitted from a news database. The number of sources reporting on an event is correlated with politically relevant characteristics such as the country in which the event occurred, the time at which it occurred, and whether it occurred in a rural or urban area. Dietrich and Eck (Reference Dietrich and Eck2020), for example, found that events occurring in Africa are less likely to be reported in media sources.

Third, individual events often are historically and politically significant on their own. Consider, for example, a researcher studying the history of territorial disputes between two countries. The omission of a border-skirmish event might make a year with one border skirmish appear to be a year of peace between the two countries. This is a difference of great narrative significance, despite it being an insufficient sample for statistical analysis.

IS API ACCESS THE SOLUTION?

API access and recent innovations in generating event data could alleviate some of these issues related to the terms of use and functionality of news databases. The MMAD (Weidmann and Rød Reference Weidmann and Rød2019), which uses the LexisNexis Web Service Kit, is a good example of how APIs can be used effectively. MMAD contains data on political-protest events in autocratic countries. Data are generated at the level of an event report and then eventually aggregated into individual events. Aggregating event reports into events allows users to deal more effectively with issues related to inconsistent and missing information about the same event across multiple news reports (Weidmann and Rød Reference Weidmann and Rød2019).

None of the other event datasets described in this article takes such an approach. We could argue that this approach might mitigate the issues discussed previously. For instance, MMAD uses a fairly coarse search string to download many articles and then uses computational methods to reduce the corpus of stories. Manual searches, in contrast, must use more complex search strings to produce a smaller corpus because manually downloading numerous stories is infeasible.

The greater complexity of manual search strings thus could be more vulnerable to the previously mentioned problem of the consequences of syntax choice changing without the user realizing it. Stated another way, the more terms that there are in a search string, the greater the opportunity for error. However, the use of APIs is not a perfect solution to this problem. Indeed, syntactical changes still could affect API users, necessitating periodic validation of search strings by researchers.

Another potential advantage of projects that use APIs (e.g., MMAD) is that sources are assigned numeric IDs, thereby avoiding the problems of changing sources names (e.g., “AFP” no longer working as an abbreviation for Agence France-Presse). Although this feature likely will help researchers with API access, there still are significant issues for those without such resources. Furthermore, API access does not solve the problem of source lists changing over time due to the expiration of contracts with specific news agencies. Thus, even those researchers with API access must remain vigilant about the sources from which their searches are pulling.

API access also can alleviate some of the issues associated with changing terms of use, particularly regarding the ability to download many sources. Again, however, API access often is prohibitively expensive, and the costs only continue to rise. Thus, researchers without such access may have trouble not only producing their event datasets but also replicating existing ones. Overall, whereas API access mitigates some of the issues discussed previously, it still (1) does not alleviate all of the discussed issues; and (2) only further widens the gap between the resources available to scholars at smaller and larger institutions to produce the types of research in which publishers often are interested.

BEST PRACTICES FOR CREATION

Some of these concerns cannot be addressed directly; however, simple best practices can help data gatherers to mitigate the impact of these changes. Although some of our recommended best practices are used by some existing datasets, there still are issues, in that (1) there is no evidence of all major event datasets implementing each of these practices; and (2) some datasets use a few of these suggestions but do not use others. Thus, we advocate for scholars to follow all of these best practices and to document their engagement in them, particularly given the increasing issues with using news databases.

1. 1. Select and Specify Your Sources. Transparency regarding the source of information is critical in evaluating the veracity of data (Hensel and Mitchell Reference Hensel and Mitchell2015; Salehyan Reference Salehyan2015.) Moreover, there are diminishing returns for including additional sources once the corpus reaches an acceptable level of coverage (Palmer et al. Reference Palmer, d’Orazio, Kenwick and Lane2015).

Some event datasets (e.g., the MID, MMAD, and SCAD projects) already provide this information. However, other projects do not, making it difficult to determine how widespread this practice is. Moreover, even when datasets list the sources on which they drew, they do not always justify why these specific sources were selected. The MMAD project, however, provides a detailed justification for its source list.

Given the rising costs of news databases, it is becoming increasingly relevant for researchers to consider the value added from each additional source. As such, all data-gathering projects should provide users with documentation of which sources were used and when they were accessed. In cases in which researchers did not specify sources in the search string, they should document metadata regarding the sources that are used in their final coding.

1. 2. Validate Your Search-String Syntax. It may take multiple iterations to land on the proper specification for a new search string. Researchers should review the results of their search after each adjustment and evaluate whether they are obtaining the expected results. For iterative data-collection projects, verification of the search string should be completed at the beginning of each iteration. Given that interfaces may change, researchers must alter their search procedures to keep up with these changes. One test that researchers could perform is using their former search string to try to replicate a small part of a previous iteration of the dataset. To the best of our knowledge, validating search strings across different iterations of the same datasets is not a common practice—or, at the very least, there is no widespread documentation of such practices.

2. 3. Establish Explicit Procedures and Workflows. Transparent data-collection procedures are crucial for the replicability of scientific research (e.g., Hensel and Mitchell Reference Hensel and Mitchell2015; King Reference King1995). Some databases (e.g., SCAD) clearly explicate the specific steps used to generate the dataFootnote ⁹; other prominent event databases, however, do not. Coupled with the fluctuating nature of news databases, this makes replication difficult if not impossible.Footnote ¹⁰

We suggest a list of relevant information that should be included in the description of data-generation processes. Other studies offer general best practices for replicability (e.g., Hensel and Mitchell Reference Hensel and Mitchell2015; King Reference King1995; Salehyan Reference Salehyan2015). Our advice focuses specifically on best practices regarding news databases. Although some existing datasets apply a few of these practices, we recommend that all event datasets should take each of the following steps and make such information publicly available:

• • Specify which databases are used, if any.

• • List the sources used or that no sources were specified.

• • Provide justification for the sources chosen, including for the decision to search all sources instead of being selective.

• • Include the exact search string(s) in replication materials.

• • Include information on how articles are processed.

• • State the dates of access.

1. 4. Conduct Periodic Checks of Your Results. Researchers should periodically check their search results. For instance, in Nexis Uni, stories from Interfax News Agency are not available past December 31, 2010, even though a content listing provided by Nexis Uni, obtained in September of 2018, listed this source as being available from December 8, 1997, to the present.

Researchers should regularly check that all of their desired sources (1) appear in the data at all; and (2) appear every year from which the researcher(s) wants stories. Additionally, researchers should keep track of the number of results over time; the breadth and quality of coverage can fluctuate significantly (Rekatsinas, Dong, and Srivastava Reference Rekatsinas, Dong and Srivastava2014).

Although fluctuations in the number of stories produced by a single news source are not inherently negative, drastic changes in availability from specific news sources could indicate consequential changes in the coverage of these sources. We are not aware of the practice of periodically checking search results to be widespread across major event datasets.

1. 5. Keep Up to Date on the Terms of Use. Database terms of use regularly change in nontrivial ways. Researchers should keep up to date on whether their coding practices still align with database policy. In cases in which researchers need approval for automatic processing, batch downloading, or other prohibited techniques, they should investigate alternative licenses (e.g., API access). Given that many of the significant changes to the terms of use of news databases are relatively new, there has been little discussion of these issues in existing literature (for an exception, see Palmer et al. Reference Palmer, McManus, D’Orazio, Kenwick, Karstens, Bloch, Dietrich, Kahn, Ritter and Soules2022).