The Military and Political Information Leaders Need

Richard Neustadt (Reference Neustadt1990, 128–9) wrote, a president’s “first essential need is information.” This certainly applies during international crises when leaders must assess what each side can expect to achieve with force and at what cost. With war as an outside option, optimal decisions—such as how generous to be in diplomacy—hinge upon expected payoffs to fighting. The probability of military victory and costs of fighting are primary determinants of these payoffs and affect which peaceful settlements are preferable to conflict (Fearon Reference Fearon1995). Both the probability of victory and war costs depend on military and political attributes. Although broad, a distinction along these lines—military versus political—elucidates the information leaders must collect to guide their choices during crises. I posit and empirically show that in expectation the political is harder to assess than the military and thus associated with greater uncertainty.

Military attributes include the quantity and quality of an adversary’s military assets and personnel (Friedberg Reference Friedberg1987), defense spending (Lebovic Reference Lebovic1995), doctrine (Biddle Reference Biddle2004), and force posture (Narang Reference Narang2014). Leaders use this information to develop a picture of the crisis landscape. The observable nature of many military characteristics simplifies the assessment process. Greater information allows advisers and leaders to form higher certainty estimates. Although they are imperfectly observable, imperfect signals yield valuable information. Military demonstrations such as North Korean missile tests and public documentation such as Soviet pronouncements of 1950s personnel reductions facilitate assessments. Aerial surveillance reveals opponent positioning, as US detection of Soviet missiles in Cuba exemplifies. Assessing opponent doctrine poses a tougher challenge, but states glean insights from pronouncements and budget allocations when observable. To be sure, information on military traits is not always abundant. Adversaries can strategically opt for secrecy (Meirowitz and Sartori Reference Meirowitz and Sartori2008), and some military traits, such as commander quality, are subject to informational deficits. While acknowledging variation, useful military information is frequently available and obtainable.

An adversary’s political attributes also affect expected war outcomes but are typically more difficult to observe. Limited observability implies a high degree of uncertainty. Although nonexhaustive, three recurring considerations illustrate the broader point regarding political attributes. Domestic unity, which can be a function of institutional arrangements, is one political trait. Great powers regularly attempt to influence which governments hold power abroad (O’Rourke Reference O’Rourke2018). An understanding of the leader’s support helps gauge prospects for regime change. Widespread domestic opposition improves the chances for displacing incumbents. So too do hostile elite factions, particularly within powerful institutions. Military defections against Arbenz’s regime proved vital for the CIA’s 1954 Guatemala operation. Domestic unity is frequently difficult to observe, particularly in autocracies that mask or inflate their domestic support. A second political trait concerns efforts to translate military accomplishments into political objectives. Force is an instrument for achieving desired political end-states. Military success must be translated into policy objectives for which the conflict was fought. Leader beliefs about an adversary’s political landscape inform estimates about translating military gains into political gains. Estimated US payoffs to the Iraq War rested on beliefs about what postwar policies could be installed in Baghdad. This depended on assessments of Iraqi military attributes crucial for displacing Saddam, but also Iraqi political attributes crucial for replacing Saddam. Translation difficulties between military and political outcomes are hard to assess beforehand. Resolve, an adversary’s steadfastness to a policy despite “temptations to back down” (Kertzer Reference Kertzer2016), is a third political attribute that affects crisis payoffs. Resolve encompasses an adversary’s willingness to use force and the costs it will endure during fighting. These depend on political factors: a hostile leader’s preferences and the will of the adversary’s public. Unfortunately for interested parties, resolve is typically unobservable ex ante. A hostile leader’s verbal claims and past actions only yield credible information under circumscribed conditions given incentives to misrepresent (Fearon Reference Fearon1995)Footnote ⁷ and contextual differences versus past interactions (Press Reference Press2005; Weisiger and Yarhi-Milo Reference Weisiger and Yarhi-Milo2015). Assessing resolve is difficult even after conflict begins. Seven years after combat troops arrived, Chairman of the Joint Chiefs Moorer was still surprised that the “North Vietnamese seem to be able to take unlimited losses.”Footnote ⁸

Neither military nor political information is fully observable. I contend there is a difference in the average degree of observability, with political traits being harder to discern. Contextual factors may exacerbate or sometimes invert this expectation. Translation uncertainties are likely elevated in insurgencies compared with conventional conflicts (Simpson Reference Simpson2018) and when a state wishes to install a new government as opposed to hold newly acquired territory (Edelstein Reference Edelstein2008; Sullivan Reference Sullivan2007). Leader resolve or domestic unity are easier to assess in transparent democracies versus closed autocracies. Political information could exceed military information, such as when fighting a conventional war over territory against a democracy with high operational secrecy and previously untested commanders. When true, some theoretical expectations below would not hold. While possible, I empirically demonstrate that typically political information is relatively scarce.

Bureaucratic Position Affects Who Provides Military or Political Information

Leaders use information on the adversary’s military and political attributes to assess what can be achieved through force and at what cost. Time and attention constraints mean leaders cannot do this alone. Instead, leaders (principals) delegate information collection and provision tasks to advisers (agents). Compared with leaders, agents become better, though not necessary completely, informed about adversary attributes. Advisers know more than leaders but are still frequently uncertain about adversary traits because uncertainty permeates international politics. Building on prior literature (Brooks Reference Brooks2008; Feaver Reference Feaver2003; Saunders Reference Saunders2017), I derive new implications for how information asymmetries between a principal and agent affect crisis advisory processes. The theory does not assume that all advisers enjoy equal influence with leaders. Instead, the implications hold across a variety of advising structures, including those that privilege certain advisers, operate via hierarchy, or encourage multiple advocacy (George Reference George1980).

Beyond being better informed, advisers are differentially informed. Senior advisers lead the core bureaucracies of the national security apparatus. A prototypical setup entails a division of labor between bureaucracies to foster expertise and efficiency (March and Simon Reference March and Simon1958). Functional specialization between bureaucracies causes advisers to become differentially informed about adversary attributes (Jost Reference Jost2021). Military advisers study opponents’ doctrines, arms development programs, force locations, and fighting power. Soviet Defense Minister Malinovsky reportedly stressed to Khrushchev that Cuban forces could withstand a US assault for only three or four days during the Cuban Missile Crisis. Foreign policy advisers acquire a distinct information set to perform their tasks. Embassies abroad provide insights into jockeying among domestic factions. US Ambassador to Iran Sullivan’s 1978 cable, “Thinking the Unthinkable,” casting (belated) doubt on the Shah’s viability offers a famous example.Footnote ⁹ Amidst the Dien Bien Phu crisis, British Foreign Secretary Eden warned US officials that “Communism in Asia cannot (repeat not) be checked by military means alone. The problem is as much political as military.”Footnote ¹⁰ Due to their substantive portfolio, foreign policy advisers specialize in adversaries’ political characteristics.

An adviser’s informational expertise and emphasis follows from the substantive mission of her bureaucracy.Footnote ¹¹ An adviser’s assigned task shapes what they emphasize to leaders due to an adviser’s “push” and leader’s “pull.” Limited time with the executive incentivizes advisers to address issues in which they have specialized knowledge. Those most informed on domestic factions within the adversary push this information to leaders. Moreover, leaders pull the information that they expect advisers to collect and reprimand advisers who stray beyond their expertise. President Obama reportedly criticized Secretary of State Kerry for bringing military plans regarding Syria (Goldberg Reference Goldberg2016), which was outside the purview of Kerry’s role. The process of providing information to the leader reinforces information divergences between military and foreign policy advisers.

This expectation may strike readers as intuitive given statutory codes dictating bureaucracies’ functions and prior scholarship (March and Simon Reference March and Simon1958). Portions of Allison’s work touches on related issues. His organizational process model (Model II), which emphasizes organizations’ routines and policy outputs, discusses “factored problems,” or the divvying of information processing across specialized bureaucracies. For Allison, information specialization matters insofar as it guides organizations’ outputs. These outputs, or policy actions, are the actual “happenings of international politics” including the positioning of military units or sending of diplomatic messages (Allison Reference Allison1969, 699). My informational account builds upon, and departs from, Allison’s in three ways. First, it assesses the implication at a broader scale than prior research. Second, information itself is the organizational output of interest. An organization’s information guides leaders’ decisions in my theory as opposed to organizations’ decisions in Allison’s. Third, Model II contends organizations pursue uncertainty avoidance (Allison Reference Allison1969, 700). I instead argue below that bureaucracies convey uncertainty to reflect limits of their knowledge.

In addition to discussing more political content, advisers in bureaucracies specializing in foreign policy convey greater uncertainty when speaking to leaders. Two informational mechanisms underlie this expectation. The first is an informational content composition mechanism: advisers who discuss high-uncertainty topics more often will convey more uncertainty in aggregate. For this mechanism to hold, foreign policy advisers must discuss political content more often than other advisers (as Informational Expectation 1 hypothesizes) and political content must carry greater uncertainty than military content (as posited earlier). If true, then increasing the amount of political content an adviser discusses mechanically increases their uncertainty. The second driver is an informational expertise mechanism: advisers specializing in foreign policy have expertise in recognizing informational gaps on the adversary’s political traits and conveying commensurate uncertainty. Military advisers who stray beyond their competency may convey excess certainty, failing to recognize what they do not know. These two informational explanations stand in contrast to a dispositional/cultural mechanism in which advisers with different dispositions toward uncertainty select into distinct bureaucraciesFootnote ¹² or the bureaucracy’s culture socializes advisers to express more/less uncertainty. The three mechanisms are not mutually exclusive, and later analyses attempt to apportion importance between them. Regardless of its foundations, a hypothesis linking bureaucratic affiliation to adviser uncertainty follows.

Figure 1 summarizes an ideal type arrangement and variation in uncertainty levels for assessing adversary attributes. Consider the US context to fix ideas for the testable claims. State Department officials are a useful proxy for those I term foreign policy advisers. The first expectation claims they are more likely to discuss political dimensions of an adversary, and the second posits that this substantive emphasis, coupled with their expertise in recognizing the inherent difficulty of forming high-certainty assessments on the matter, makes State Department officials express a low degree of certainty. Neither content emphasis nor degree of uncertainty directly affects policy hawkishness. Bureaucracies thus diverge in the information they provide, not the policies they champion.

Figure 1.

Ideal Type Delegation of Information Provision Tasks to Advisers with Differentiated Domains of Expertise

Note: Uncertainty varies across domains. Arrangement does not assume leaders grant equal access to advisers.

Although agnostic on advising structures, the theory partly relies on two leader functions. First, leaders can select and sanction advisers, which limits advisory skew and parochialism. Second, leaders encourage advisers to stick to their competencies, which reinforces content differences across bureaucracies.

Cases

The analysis draws on US crises during the Cold War from the Eisenhower to Carter administrations. Bounding the analysis to US Cold War crises holds potential confounding variables relatively constant. Keeping the state fixed mitigates concerns that variation in institutional structures drives the results. US institutional arrangements remained stable between the 1947 National Security Act and 1986 Goldwater-Nichols Act. Demographically, advisers were remarkably homogeneous. Gender, education, and race, all of which could affect advisory input, were near constants. Finally, bipolarity prevailed throughout this period, albeit with stretches of relative detente. Practical considerations—such as evidence availability—also recommend a US Cold War focus. The analysis requires access to the highest levels of decision making. Classification considerations mandate a lag between events and their study. The US provides a rich documentary record with pertinent materials available through the Carter administration. Moreover, the global ramifications of US behavior make its characteristics of general interest. An emphasis on crises, which inherently involve high-stakes policy choices, helps insure leader engagement in the decision-making process. Leader engagement is necessary for evaluating the theorized expectations that concern the input advisers provide to leaders. Substantively, crises merit attention as pivotal moments prone to the initiation or escalation of conflict. A later section addresses generalizability concerns.

I identify 61 opportunities to use force during which the president and advisers considered and discussed using US force to strike abroad. The sampling criteria is closest to Betts (Reference Betts1991), which similarly studies adviser behavior when both peace and conflict were realistic options. Two sources contribute the majority of cases. First, I include crises from the International Crisis Behavior dataset when the US was an actor (Brecher and Wilkenfeld Reference Brecher and Wilkenfeld1982). Second, I include major military mobilizations from Blechman and Kaplan (Reference Blechman and Kaplan1978).Footnote ¹⁴ Directing many military assets toward hot spots indicates consideration of using force. Importantly, asset mobilization does not imply the actual use of force as some ended peacefully. Together, these two sources contribute 46 cases. Additional observations include some lower-level military mobilizations, some militarized interstate disputes (Ghosn, Palmer, and Bremer Reference Ghosn, Palmer and Bremer2004), and opportunities to strike the Soviet Union and China’s nascent nuclear programs.Footnote ¹⁵

The Supporting Information (SI) §1 lists the opportunities to use force (which I use interchangeably with “crises”) and describes the precipitating context. Some concern potential new uses of force—such as the 1958 turbulence in Lebanon. Others capture opportunities to expand ongoing conflicts—such as invading Laos during the Vietnam War. Sample censorship, in the sense of excluding cases where using force never received leader-level scrutiny poses little inferential threat barring some reason that advisers’ relative areas of specialization switch when under presidential scrutiny.

Corpus

Measuring advisory input requires details about crisis advisory processes. Building on rich qualitative cases that reconstruct advisory processes, (e.g., Hermann and Preston Reference Hermann and Preston1994; Saunders Reference Saunders2017), I collect documentary evidence from presidential-level deliberations to capture the private words of senior advisers at scale. Documents published through the Foreign Relations of the United States (FRUS) series constitute the core sources. The FRUS volumes incorporate materials from presidential libraries, the Departments of State and Defense, National Security Council (NSC), Central Intelligence Agency (CIA), and other agencies and individuals involved in creating US foreign policy. Documents include National Security Council meetings, intraelite memos (such as from Kissinger to Nixon), and minutes from elite conversations in unstructured (ad hoc meetings) and structured (briefings from the Joint Chiefs of Staff [JCS]) settings. Because Carter-era FRUS volumes are incomplete, I gathered documents from the Carter Presidential Library and National Security Archive. These sources provided documents akin to those in FRUS.Footnote ¹⁶ I supplement FRUS with transcripts from the Executive Committee of the National Security Council (ExComm) from the Cuban Missile Crisis. Collectively, these texts provide access to private statements that likely reflect sincere beliefs. Before intervening in Laos in 1964, Johnson told advisers “we should go ahead with the mission but that he had doubts about the action.”Footnote ¹⁷ Presidents rarely voice such reservations publicly.

I select and prune documents to only include portions pertinent to the theory. Retained portions address considerations relevant for assessing potential diplomatic and conflict outcomes. Less cautious approaches that include all FRUS documents risk letting irrelevant documents undermine measurement validity. Given presidential primacy in decision making and the theory’s focus on input provided to leaders, I restrict the corpus to memos sent to the president and transcripts from meetings that the president attended.Footnote ¹⁸

All transcripts are split into speaker-specific texts; thus, a Nixon–Kissinger dialogue produces Nixon texts and Kissinger texts. The corpus includes 5,404 speech acts, defined as uninterrupted communication from a single individual which could be a lengthy memoranda or brief interjection in a meeting. These come from 382 documents, with 176 unique speakers. SI §2 provides corpus details and shows that text volume accurately reflects the prominence of key advisers, such as Dulles, Kissinger, and Brzezinski. The amount of text available varies across crises, typically reflecting crisis salience (e.g., Cuban Missile Crisis has over 20,000 words in the corpus; Shaba II has over 1,000). Tests of the informational expectations include case fixed effects to guard against the possibility that unobserved variables affect crisis salience and adviser content or uncertainty and thereby confound the relationships of interest. Fixed effects guard against apples-to-oranges comparisons that pit State’s Cuban Missile Crisis advice against CIA’s Shaba II advice.

FRUS has several drawbacks as a data source, but these generally pose limited inferential threat. Redacted segments within texts typically concern targeting details or covert sources rather than core assessments. Another concern is that entire documents remain classified or that FRUS historians excluded certain types of documents. Assessing either possibility is difficult at scale. However, SI §2 describes spot-checks that use online finding aids from presidential libraries to gauge how comprehensively FRUS volumes incorporate presidential-level texts. Although far from exhaustive, these checks suggest high coverage of the relevant texts. FRUS volumes privilege the historically salient over the mundane,Footnote ¹⁹ which is advantageous when studying leader-level crisis decision making. That said, slower and more restrictive declassification at the Pentagon or CIA could skew the composition of FRUS volumes. Such discrepancies, however, are sometimes later rectified when especially egregious and only generate bias favoring the hypotheses under precise conditions. We must suspect that State officials emphasize political (vs. military) content and express high uncertainty compared with others in documents included in FRUS and not those excluded from FRUS. Even if FRUS historians harbor a pro-State disposition, it is unclear why this would manifest with retaining high-uncertainty State texts. Presidents often bemoan State’s elliptical responses, which suggests that including documents displaying State’s elevated uncertainty does not flatter the department.

Measuring Advisory Input

I generate measures at the bureaucracy-crisis level by aggregating input from advisers affiliated with five bureaucracies that span the spectrum from specializing in issues of foreign policy to issues of military policy: State, CIA, NSC or White House staff,Footnote ²⁰ the JCS, and non-JCS Defense. The State Department provides the ideal type of foreign policy specialization, as Jervis (Reference Jervis2010, 41) notes, “the Embassy provide[s] the bulk of political reporting” on foreign states. State Department thus serves as the main explanatory variable. The JCS and non-JCS Defense, which I distinguish between for reasons noted below, serve as the ideal type for military policy specialization. The NSC and CIA occupy a middle position between the ideal types. The broad substantive remits of these bureaucracies make the empirical implications for them less stark.

I expect an increase in political content and uncertainty when moving across the continuum from the ideal type military policy bureaucracies to the ideal type foreign policy bureaucracy. Text analysis tools, which provide a replicable means of distilling advisory input from many documents, provide measures of advisory content and certainty. Because these tools are less suited for measuring policy prescriptions, I qualitatively code the policy stances of State and non-JCS Defense officials.

Political versus Military Adviser Content

Theory testing requires a measure of whether advisory input concerned military or political content. Supervised learning techniques, which classify texts into prespecified categories, offer an accurate and scalable method for this measurement task (Grimmer and Stewart Reference Grimmer and Stewart2013). Supervised methods follow a straightforward process. Human coders classify a subset of texts as political or military. An algorithm uses this training set to “learn” which text features (words) are associated with each category. The algorithm uses these features to sort the remaining texts between the categories.

Speaker texts in the full corpus range from a single word spoken during a meeting to memos exceeding 4,000 words. For uniformity in the training set, I use shorter texts and split longer texts into portions containing 30 to 250 words, which proved optimal for hand-classifying texts. The training set includes a random subset of memos and transcripts with 475 texts representing approximately 20% of the corpus’ total word count. Classification rules followed from the theory and were refined through iteratively applying these rules. Military texts address military capabilities, the quality of forces, and force locations. Political texts address the opponent’s domestic political landscape, challenges translating battlefield outcomes to desired political end states, or adversary resolve. Example texts in SI §3.1 typify content from each category. Overall, I coded 61% and 39% of the training set as military and political, respectively. A research assistant produced identical codings for 82% of the 475 texts, with further measures of intercoder reliability provided in the SI. Several standard preprocessing steps—removing numbers and punctuation and stemming to reduce words to common roots (e.g., “operate” and “operation” to “oper”)—preceded the analysis.

I tested a variety of algorithms, including random forest models and support vector machines, and achieved the best performance with a simple naive Bayes classifier, which accurately classified 88% of texts in 10-fold cross-validation checks. SI §3.2 provides additional details on the classifier and performance metrics such as F statistics. Beyond high accuracy, the analysis shows that inaccurate classifications were clustered close to the cut point between the two categories. Figure 2 plots the 25 terms with the highest relative frequency for each category, generated inductively from the training set codings. These terms can be used for both categories (e.g., “believ”), but in practice within the training set they are far more likely to be used with one over the other. Word stems that most distinguish military texts include “forc,” “attack,” “nuclear,” and “general.” Word stems to the right are indicative of political texts. State leaders—“Khrushchev,” “Nasser”—help distinguish these texts, as do references to “govern” and “talk.” Invoking foreign leaders suggests a concern with that state’s domestic politics or the individual’s preferences. Nonetheless, tests in SI §5.2 demonstrate that results hold when excluding leader names from the training set. The most telling term for political texts, “will,” concerns projections of how actors will respond to stimuli and their willpower or resolve. A follow-up coding evaluated how speakers employed “will” in practice. Within a random subset of the training data, “will” preceded political (vs. military) content in 59% of uses, which is high given that only 39% of texts were classified as political.

Figure 2.

Terms with Greatest Frequency Differences between Military and Political Documents

Note: Text size indicates overall frequency in training documents.

I apply the algorithm to the full corpus at the bureaucracy-crisis level of analysis. I aggregate all speech acts from a bureaucracy’s officials during a crisis—for instance, State officials during the Mayaguez seizure—and apply the classifier to the resulting text block. Each bureaucracy-crisis observation receives two scores—a continuous Political Content Score (mean = 0.55) and binary Political content indicator (mean = 0.43). The continuous measure, described further in SI §3.2, standardizes scores to span between zero and one, with greater values indicating more political content.

Bureaucratic Position Affects Advisory Content

The results presentation first shows that State Department officials emphasize political attributes more often than counterparts from other bureaucracies. Differentiated functional responsibilities between bureaucracies encourages specialization. Tasked with guiding foreign policy, State Department officials specialize in adversaries’ political characteristics. Figure 3’s left panel plots the Political Content Score for all observations. For presentational clarity, I pool ideal type military policy bureaucracies (JCS and non-JCS Defense) and mixed remit bureaucracies (NSC and CIA).Footnote ²³ The panel confirms that as a bureaucracy’s mission increasingly concerns foreign (vs. military) policy, advisers increasingly emphasize political content to leaders.

Figure 3.

Text Content by Bureaucracy

Note: Political Content Score observations (left), with diamonds indicating the bureaucracy mean. The defense category includes JCS and non-JCS observations; the Other category includes NSC/White House and CIA observations. Marginal effect of shifting bureaucracy to State (right), based on Models 1–4 in Table 1.

Regression results align with the descriptive patterns. Table 1 reports results from specifications with content as the outcome variable and State Department as the explanatory variable. Models vary the outcome variable coding (Political Content Score in 1–4, Political in 5–8), the inclusion of case controls, and use of case fixed effects, which guards against the possibility that State only participates in cases with elevated political concerns. Models 4 and 8 prune the sample to ideal type bureaucracies—State, non-JCS Defense, and JCS. Divergences should be especially pronounced within this subset.

Table 1.

Bureaucratic Role and Advisory Content

Note: Ordinary least squares with bureaucracy-crisis observations as unit of analysis. Fixed effects and case covariates not shown. Models 4 and 8 limit the sample to State, Defense, and JCS observations. Full results for Models 2 and 6 provided in SI Table A7. *p < 0.10, **p < 0.05, ^*** p < 0.01.

The State Department explanatory variable has a positive coefficient, indicating more political content, across all specifications. The relationship is statistically significant in all models using the continuous measure. Whereas the average bureaucracy input concerns neutral content (Political Content Score = 0.55), State Department officials stress, on average, more political content (coefficients ≥ 0.09). This effect becomes more pronounced when restricting the comparison to Defense and JCS officials, as Model 4 reports. Consider the crisis sparked by North Korea shooting down an EC-121 reconnaissance plane. Defense and JCS had an average Political Content Score of 0.35. Model 4 predicts a State score of 0.51, or substantially more tilted toward political content. State’s actual input was even more politically oriented, with a score of 0.89. These patterns accord with the historical record. Defense officials discussed advantages of airstrikes compared with mining North Korean harbors, whereas State officials considered Soviet reactions and North Korean responses to diplomatic overtures. Figure 3’s right panel plots marginal effects based on Models 1–4. It provides point estimates and 90% and 95% confidence intervals on the increase in Political Content Score when moving to the State Department from bureaucracies that do not specialize in foreign policy.

Using the binary outcome coding produces similar, though less robust results. In Model 5, moving from another bureaucracy to the State Department produces a 60% increase in the relative probability of political content (22% ± 16% at the 95% confidence interval). When including fixed effects, the effect diminishes and is no longer statistically significant. However, when restricting the sample to the ideal types (Model 8), I again find a strong and statistically significant effect. Whereas JCS and non-JCS Defense provide political rather than military content in 21% of cases, shifting to State more than doubles the probability of discussing political content, even when holding the case fixed. Further analysis confirms that although State Department provides the most political content, the gap compared with NSC is limited when using the binary outcome and fixed effects.

Robustness checks in SI §5.2 use randomization inference to address small sample concerns, logistic regression for the binary outcome coding, clustered errors, and the individual speaker as the unit of analysis where individuals are nested within bureaucracies (n = 629) to address aggregation concerns and to account for whether advisers communicated via a memo or meeting. Following the lead of Wasserstein and Lazar (Reference Wasserstein and Lazar2016), another test moves away from p values by comparing Model 2 with one without the State Department variable and produces a Bayes factor of 116. Results hold across all specifications, corroborating that bureaucratic role drives adviser informational specialization.

Bureaucratic Position Affects Adviser Uncertainty

Consistent with Information Expectation 2, foreign policy advisers express greater uncertainty than counterparts during crises. Uncertainty is 12% higher in State Department observations compared with other agencies (5.6% vs. 5.0%). Figure 4’s left panel decomposes the sample by bureaucracy and content to yield additional insights. I again pool JCS and non-JCS Defense together and NSC and CIA together (see SI §5.3 for a plot without pooling). State is the most uncertain of all bureaucracies when averaging across the full sample (black bars). Defense observations offer the starkest contrast with an average score of 4.9%. My informational theory posits that this result stems from the substance that bureaucracies emphasize (State stresses political content) and the requisite expertise to recognize informational deficits. Consistent with the expertise expectation, the lighter gray bars show that State Department officials express greater uncertainty than others when discussing political matters. For military content (darker gray bars), which I posit is prone to greater certainty due to higher observability, Defense officials recognize the relative informational abundance and convey certainty accordingly. Competency widens the political–military uncertainty gap. When the ideal type bureaucracies—State and Defense—discuss their areas of expertise, they provide dramatically different degrees of uncertainty.

Figure 4.

Uncertainty by Bureaucracy

Note: Average uncertainty by bureaucracy (left). The Defense category includes JCS and non-JCS observations; the Other category includes NSC/White House and CIA observations. Color distinguishes between all observations, political content observations, and military content observations. Marginal effect of shifting bureaucracy to State (right), based on Models 1–5 in Table 2.

Models 1 through 3 in Table 2 employ the full sample without and with case controls and case fixed effects. Fixed effects guard against the possibility that State officials only participate in information-poor cases. All specifications produce similar results: State officials express 12% (± 11%) more uncertainty relative to other advisers. Deliberations before the 1965 intervention in the Dominican Republic typify overall patterns. The average bureaucracy Uncertainty score was 5.4%. Model 2 predicts a State Department score of 6.0%. In reality, State officials expressed even more uncertainty (7.8%), representing a 45% relative increase over others. Documents from the crisis corroborate this pattern. Defense officials expressed certainty that two or fewer US divisions could stabilize the island. State Department officials, attuned to political factions on the ground, conveyed a lack of certainty with Undersecretary of State Mann noting that “loyalties of the troops outside the capital are still uncertain.”Footnote ²⁴

Table 2.

Bureaucratic Role and Uncertainty

Note: Ordinary least squares with bureaucracy-crisis observations as unit of analysis. Fixed effects and case covariates not shown. Models 4 and 5 limit the sample to ”ideal” types—State, Defense, and JCS observations. Model 5 further restricts the sample to “expert” observations where “ideal” types discuss their areas of expertise. Full results for Model 2 provided in SI Table A12. *p < 0.10, **p < 0.05, ^*** p < 0.01.

State’s relative uncertainty becomes more pronounced when subsetting the sample to ideal type bureaucracies. JCS and non-JCS Defense serve as the Model 4 baseline. Substantively, State advisers convey 18% (± 14%) more uncertainty than military advisers. Limiting the sample to “expert” observations—that is, State discusses political and Defense/JCS discuss military attributes—produces even larger relative effects (38% ± 27%). Figure 4’s right panel plots marginal effects from Models 1–5, showing the increase in expected uncertainty when moving to the State Department from other bureaucracies that are not specialists in foreign policy.

SI §5.3 shows results hold when employing randomization inference, an alternative uncertainty dictionary measure, clustered errors, and the individual speaker as the unit of analysis which can control for whether advice came from a memo or meeting. Comparing Model 2 with one without the State Department variable produces a Bayes factor of 2.6. Appropriately, the results do not persist in a placebo dictionary test.

Mechanisms. Analyses thus far corroborate Informational Expectation 2 but only hint at underlying mechanisms. This section helps apportion relative importance between three theorized candidate mechanisms. For the informational content composition mechanism to hold, advisers specializing in foreign policy must discuss political content more than other advisers and political content must carry greater uncertainty than military content. The results in Table 1 establish the first point, and the results in Table 3 Models 1 and 2 establish the second. Both models, which include case covariates, find that increasing Political Content Score is associated with increasing Uncertainty score. Model 2 includes bureaucracy fixed effects, indicating that even when holding fixed the bureaucracy (e.g., NSC), increasing uncertainty accompanies political content. The results support the content composition mechanism. If bureaucracies discuss political content more frequently, they convey greater uncertainty.

Table 3.

Sources of Uncertainty Differences across Bureaucracies

Note: Ordinary least squares with bureaucracy-crisis observations as unit of analysis. Case covariates and bureaucracy fixed effects not shown. Full results provided in SI Table A18. *p < 0.10, **p < 0.05, ^*** p < 0.01.

Model 3, which includes an interaction between State Department and political content, helps distinguish between the informational expertise mechanism and a competing dispositional/cultural mechanism. For the expertise mechanism to hold, the uncertainty gap between State and other bureaucracies should increase as the topic becomes increasingly political. State holds relative expertise in political content and thus recognizes blind spots for this content. The positive coefficient on the interaction term offers some support for this mechanism. As content becomes increasingly political, the effect of State Department on uncertainty becomes larger. If the dispositional/cultural mechanism applies, then State officials should convey more uncertainty regardless of the topic under consideration. Model 3’s lower order State Department offers some evidence supporting this mechanism. When the topic is purely military (Political Content Score = 0), State officials still express more uncertainty, though this difference is quite small (0.05). At more realistic content score values indicating strongly military topics, State officials do convey more uncertainty than others but the substantive difference remains relatively limited (0.19). SI §5.3 provides an extended discussion.Footnote ²⁵ Evidence corroborates the two informational mechanisms this study emphasizes while also providing some support for alternative mechanisms linked to dispositional sorting or organizational culture.

Parochialism and Policy?

For Allison and Zelikow (Reference Allison and Zelikow1999, 307), “[k]nowledge of the organizational seat at the table yields significant clues about a likely stand.” Are State Department officials less militarily aggressive than civilian counterparts at Defense? I expect parochialism dissipates in crises because leaders wield tools—adviser selection and sanctioning—that encourage honest communication.

Consistent with my expectation, bureaucratic “seat” does not predict where advisers “stand.” Table 4 shows State and Defense advocating substantively indistinguishable positions in nearly half the cases. When diverging, Pentagon civilians call for more aggressive actions in 30% of cases while State officials do so in 25% of cases. Statistical tests confirm that State Department officials are not especially dovish. A t test that treats the outcome variable as linear fails to reject the null hypothesis of no difference between the bureaucracies.

Table 4.

Counts of Relative Aggressiveness by Bureaucracy

Note: Test does not preclude the null hypothesis of no difference between the bureaucracies. The t test codes outcomes as 0, 1, and 2.

Bureaucratic position did not foreordain an adviser’s policy aggressiveness. What then explains divergent policy stances? Though speculative, advisers’ policy suggestions could reflect their dispositions on using force or the information they collected while executing their reporting functions. Advisers likely hold preexisting views on the nature of international politics that affect their policy prescriptions (Jost et al. Reference Jost, Kertzer, Min and Schub2021). Secretary of State Cyrus Vance, for instance, never adopted the most aggressive posture during a crisis, reflecting his dovish disposition. Other officials though varied depending on the context. Secretary of State Rogers opposed bombing Cambodia in 1969, but Secretary of Defense Laird approved.Footnote ²⁶ They flipped positions three years later. Laird opposed using B-52s against industrial complexes in Hanoi and Haiphong whereas Rogers consented to it.Footnote ²⁷ In such cases, advisers collect different information and the perceived wisdom of aggressive policies depends on what they learn. A State official who discerns waning resolve among the adversary’s elites may endorse aggression while a Defense official who perceives limited chance of stymying the adversary’s flow of supplies may eschew aggression. Though requiring more study, expanding upon Allison’s Model II may offer a promising guide to a bureaucracy’s policy stance. An organization’s specialized information processing could underlie its policy preferences in addition to its policy outputs. This possibility differs from Model III parochialism in which a bureaucracy’s broad substantive remit (e.g., diplomacy) supersedes particulars of the crisis.