Diversionary cheap talk: economic conditions and US foreign policy rhetoric, 1945-2010

ABSTRACT

This study explains how the economy affects the foreign policy rhetoric used by American presidents. When economic conditions deteriorate, presidents criticize foreign nations to boost their approval ratings. Presidents use this “diversionary cheap talk” in response to the misery index of unemployment plus inflation, which poses a unique threat to their popularity. They target historical rivals, which make intergroup distinctions most salient. Diversionary cheap talk is most influential for and most frequently used by Democratic presidents, whose non-core constituents prefer hawkish foreign policy but already expect it from Republican presidents. I test the observable implications of the theory with the American Diplomacy Dataset, an original record of 50,000 American foreign policy events between 1851 and 2010 drawn from a corpus of 1.3 million New York Times articles.

KEYWORDS:

• Diversionary conflict
• international relations
• domestic politics
• political psychology
• political communication
• social identity theory
• text analysis
• computational social science

Supplementary material

Supplemental data for this article can be accessed on the publisher’s website.

Notes

¹ In the language of the foreign policy substitutability literature (Oakes 2012; Clark, Nordstrom, and Reed 2008; Most and Starr 1984, 1989; Scott and Nordstrom 2000), rhetorical hostility, like the development of new economic policies, may be seen as a substitute for diversionary conflict.

² Group identity is artificially induced with artwork preferences; these are known as minimal groups.

³ See e.g. Baker and Oneal (2001); Mueller (1973, 209).

⁴ There are clear partisan divides. For instance, during the Obama administration, approximately 80% of Democrats (and 50% of Republicans) trusted the federal government to handle international problems. During the Trump administration, approximately 80% of Republicans (and 35% of Democrats) trusted the federal government to do the same.

⁵ This is a large literature but see, for example, Colaresi Rasler, andThompson (2008); Diehl and Goertz (2000); Findley, Piazza, and Young (2012); Goertz and Diehl (1993); Mitchell et al. (2004); Uzonyi and Rider (2017); Wiegand (2011).

⁶ Klein Goertz, and Diehl (2006) include several other countries as rivals to the United States between 1945 and 2001: Canada, Nicaragua, Ecuador, Peru, Czechoslovakia, Yugoslavia, Libya, Egypt, and Syria. I excluded these from my set of rivals because despite some episodic tension, none posed a serious threat to US interests and thus are less relevant for generating a rally, per Mitchell et al. (2004). I then updated Klein Goertz, and Diehl (2006)’s US rivalry data through 2010.

⁷ Some scholars focus on unemployment as a source of diversion (Ostrom and Job 1986). I follow the long tradition of studying the “misery index” which combines unemployment and inflation (DeRouen 1995; James and Hristoulas 1994, James and Oneal 1991; Meernik 1994; Meernik et al. 1996; Ostrom and Job 1986).

⁸ A large literature shows that news coverage appears to shape readers’ attitudes. See for example, Baum (2004); Brewer, Graf, and Willnat (2003); Chiang and Knight (2011); Curran, Iyengar, Lund, and Moring (2007); Gerber, Karlan, and Bergan (2009); Ladd and Lenz (2009); Miller and Albert (2015).

⁹ All documents were downloaded in accordance with terms of service policies. The document texts cannot be released due to copyright reasons, but the event data generated from texts will be made available on my scholar webpage. Each document is a 2–4 sentence summary of the news article. For technical reasons related to event extraction, short summaries are preferable to long articles.

¹⁰ In a recent review, Schrodt calls this a “very difficult problem” (Schrodt 2012b, 554).

¹¹ Technically, “similarity” here reflects string matching, not n-gram similarity. For fuzzing string matching, the Levenshtein ratio is used, the number of changes it would take to make two strings identical, divided by the length of the string. Fuzzy string matching is implemented with the fuzzywuzzy module.

¹² All articles were processed per standard procedure: words were lowercased and stemmed; symbols, numbers and stop words were removed.

¹³ These relationships were captured in a term frequency–inverse document frequency matrix.

¹⁴ Accuracy fell rapidly with the inclusion of additional topic labels; separating political affairs from military affairs, for example, decreased the classification accuracy to 64.7%. The support vector machine classifier performed better than random forest and neural network classifiers, which had accuracy rates of 65.3% and 68.0%, respectively. For more, see Caruana and Niculescu-Mizil (2006).

¹⁵ Event types are taken from the CAMEO ontology, which focuses on interstate behavior (Schrodt 2012a). A list of all 300 event types is available at http://web.ku.edu/$\sim$keds/cameo.dir/CAMEO.SCALE.txt.

¹⁶ DARPA’s Integrated Conflict Early Warning System (ICEWS) covers 1998–2010 and focuses on Asia; the Kansas Event Data System (KEDS) covers 1979–2011 but uses a dated ontology and focuses on the Middle East. There are several event datasets of political violence, but these too date generally from the 1990s, except for the Global Terrorism Database which covers 1970–2010 and Uppsal Conflict Data Program/Peace Research Institute Oslo (UCDP/PRIO) dataset on political instability covering 1946–2011. For a comparison of the American Diplomacy Dataset to other event datasets, see Connellyet et.al. (2019).

¹⁷ On both these problems, see Schrodt (2012b).

¹⁸ This is related to research on elite theory, which finds that partisan media outlets actively shape public opinion (Baum and Philip 2008).

¹⁹ For more discussion, see Schrodt (2012b).

²⁰ The approval rating indicates the fraction of respondents answering “approve” to the question, “Do you approve or disapprove of the way [first & last name] is handling his job as President?” (Gallup 2013).

²¹ Outcome variables are over-dispersed but not zero inflated: only 5% of months have zero material conflicts and only 3% have zero verbal conflicts.

²² I do not use a GMM dynamic panel estimator as Pickering and Kisangani do; this would be inappropriate as the method is designed for small $T$, large $N$ panels.