«Abstract We examine whether there is systematic evidence that the US intelligence services pandered to their political masters when constructing ...»
Politicization of Intelligence Reporting:
Evidence from the Cold War
(Job Market Paper)
University of Cambridge
We examine whether there is systematic evidence that the US intelligence services pandered to their
political masters when constructing intelligence estimates during the Cold War. We construct a model
which shows how career concerns on the part of intelligence analysts could lead them to distort reports
towards their president’s prior beliefs. We then take the model’s prediction that errors in intelligence reports should be correlated with presidential ideology, to the data by constructing a unique measure of intelligence failures that compares CIA/ORE reports on the Soviet strategic, nuclear arsenal to credible, post-Cold War estimates of the Soviet Union’s actual nuclear capabilities. We ﬁnd that report errors are systematically, positively correlated with both a conventional measure of presidential ideology and a unique, text-based measure of presidential “hawkishness”. This result is robust to controlling for a number of endogeneity issues and alternative mechanisms such as reverse causality, variation in inter- superpower relations, collusion between politicians and the intelligence agencies, and turnover in agency staﬀ. Finally, there is evidence that longer-term forecasts are more sensitive to ideology in a manner that is consistent with our model. (JEL: H56, L82, N42) 1 Introduction To what extent can intelligence agencies be relied upon to provide balanced, accurate reporting to their political masters? In recent years this question has gained greater prominence as the United State’s Central Intelligence Agency (CIA) and the United Kingdom’s Secret Intelligence Service (MI6) have been accused of tailoring information regarding Iraq’s pursuit of Weapons of Mass Destruction to support their respective governments’ preferred policy of regime change in Iraq. (see Barstow et al. (2004) and Norton-Taylor (2011)) ∗ I am grateful to Toke Aidt, Kenneth Benoit, Howard Cobb, Tom Crossley, Bob Evans, Jane Cooley-Fruehwirth, Sanjeev Goyal, Matthew Gentzkow, James Mahon, John McLaren, Sharun Mukand, Andrea Prat, Christoph Vanberg and participants at the Cambridge Empirical Microeconomics Workshop, CAGE Easter School, North American meeting of the Econometric Society and Silvaplana Political Economy Workshop for helpful comments and suggestions. In addition, I am grateful to Christopher Andrew and participants at the University of Nottingham’s Landscapes of Secrecy conference for helping provide the historical context to the paper.
1 If these events are symptomatic of a wider tendency of intelligence agencies to pander to their “consumers” then there are important implications for the reliability of intelligence reporting and hence issues of national security. In this paper we construct a model that builds on the existing economics literature on media bias to argue that career concerns on the part of intelligence analysts could lead them to twist intelligence information to conform to the sitting president’s beliefs. We then test this theory by constructing a unique measure of intelligence failures that compares data from US intelligence reports on the size of the Soviet strategic nuclear arsenal during the Cold War to credible, post-Cold War estimates of the actual number of Soviet strategic weapons. Using this measure as our dependent variable we ﬁnd robust evidence that more “hawkish” presidents received systematically,upwardly-biased reports. We also ﬁnd evidence that longer-term predictions are more sensitive to presidential ideology in a way that is also consistent with our model.
We measure presidential hawkishness using existing measures of presidential ideology from the political science literature and a new measure based on text analysis. This second measure is constructed using the Wordscores algorithm of Laver et al. (2003) and compares the words spoken by each president in defence-related speeches to those spoken by the most liberal and most conservative deciles of congressmen in contemporaneous congressional debates on defence. We exclude the possibility of reverse causality between the material in the intelligence reports and presidential ideology by basing our text-based measure only on speeches made before each president received security clearance.
Institutional evidence also allows us to discount some alternative explanations. Because we use Top Secret reports to which only the President himself, the Joint Chiefs of Staﬀ and the National Security Council had access, it is unlikely that the estimated, positive eﬀect of presidential ideology is due to analysts colluding with the sitting president to produce information to convince political opponents of the President’s preferred policies.
Another potential mechanism we have to account for is the possibility that there could be turnover in Agency staﬀ that is correlated with new presidents taking oﬃce. If, for example, hawkish presidents systematically replace senior analysts with like-minded people, we might estimate a positive eﬀect of ideology on bias, but this would reﬂect changes in the personnel making the intelligence reports rather than changes in their behaviour due to reputational incentives. We provide evidence against this turnover eﬀect by ﬁrstly restricting our sample to years in which historians agree senior appointments to the intelligence services were not politicized and ﬁnd that the positive eﬀect of ideology persists. The result also persists when we control for the ideology of the CIA director, either directly (by assuming that the ideology of the director is equal to that of the president who appointed him) or by using director ﬁxed eﬀects.
One might also be concerned that the estimated eﬀect of presidential ideology could be picking up the eﬀect of changes in the state of relations between the superpowers. The ﬁrst reason this would be a concern is if there is a selection eﬀect: if more hawkish presidents are more likely to be elected in times of greater paranoia about the threat posed by the Soviet Union and this paranoia also causes the intelligence services to be excessively pessimistic when constructing their reports, then we could pick up a positive correlation between presidential ideology and intelligence forecast errors even when no causal eﬀect exists.
The second concern is that hawkish presidents may be more aggressive in their interactions with the Soviets and it is the resulting worsening in relations that causes an increase in errors in intelligence forecasts of Soviet military strength. We provide evidence that our results are robust to these concerns by showing 2 that the positive eﬀect of ideology is robust to controlling for a proxy for inter-superpower relations. While
it is impossible to exclude these eﬀects completely, we argue that the correlations we ﬁnd are still of interest:
whether the positive eﬀect of presidential ideology results from pandering by analysts or an excessive response to changes in the geopolitical situation, the implication is still that the intelligence services made systematic errors when constructing estimates of Soviet nuclear strength.
These results, while interesting in themselves, could well be symptomatic of a wider problem in intelligence reporting. The fact that we ﬁnd evidence of pandering even in reports which were quantitative in nature (and so could be easily contradicted by future events), were about an issue of ongoing importance (so analysts will have expected to be called up on any discrepancies between reports from one year to the next), and were about issues with extremely high stakes (which one might hope would induce analysts to sacriﬁce career concerns for the national good), suggests that this phenomenon may play an even greater role in reports where these constraints are absent.
In terms of the existing economics literature, this paper is most closely related to the literature examining bias in media reporting. In the theoretical literature Mullainathan and Shleifer (2005) show how bias can arise when consumers have behavioural preferences. Of more direct relevance to our analysis, Gentzkow and Shapiro (2006) show how reputational concerns on the part of reporters can lead to bias when consumers are fully rational, but have strong prior beliefs as to the true state of the world. Empirical studies of the determinants of bias include Knight (2011), Groseclose and Milyo (2005), DellaVigna and Kaplan (2007) and Latham (2012). Our model also shares similarities with other models of cheap talk with reputational concerns such as Ottaviani and Sorensen (2006) and Morris (2001), and models of “yes men” such as Prendergast (1993).
This paper also uses techniques from the empirical literature on media bias. In particular, we construct a measure of presidential ideology by using text-analysis methods to compare presidential speeches to speeches by Members of Congress. The concept of using Members of Congress as a reference group is similar to Gentzkow and Shapiro (2010) who constructed a measure of newspaper slant by examining the frequency with which newspapers used phrases which, when spoken in Congress, are highly correlated with Congressmen’s party aﬃliation. However, the actual method used to calculate our text based measure is more closely related to the Wordscores algorithm of Laver et al. (2003).
A second, related, literature studies the motivations for and eﬀects of, foreign interventions by intelligence agencies, like the CIA. However, it should be noted that, while this literature is similar in that it studies the actions of intelligence agencies, it diﬀers in that it is concerned with their secondary role as a paramilitary force whereas we focus on their primary role of intelligence gathering.1 In the theoretical literature on foreign intervention Aidt and Albornoz (2011) demonstrate how the foreign interests of domestic ﬁrms in developed countries might cause them to lobby their governments to suppress democratic movements around the world. In the empirical literature Berger et al. (2012) provide evidence that CIA interventions result in US goods taking up a greater share of trade in the countries in which the interventions took place, while Easterly et al. (2008) provide evidence that CIA interventions have persistent, negative eﬀects on democracy and that these eﬀects are comparable to those caused by 1 Primary in the sense of their original mission statements. In fact some historians (e.g. Weiner (2007)) argue that covert operations became the central concern of the CIA with negative implications for its intelligence gathering capabilities.
3 interventions by the Soviet MGB and KGB. Finally, Dube et al. (2011) provide further evidence that foreign interventions were motivated by economic factors by showing that top secret presidential authorizations for covert actions in foreign countries aﬀect the stock prices of domestic ﬁrms with interests in these countries before the interventions actually take place.
The paper is structured as follows. Section 2 discusses some anecdotal evidence of politicization in intelligence reports, Section 3 describes our model and presents the theoretical results that motivate our empirical analysis. Sections 4 and 5 outline the data and our empirical approach. Sections 6 and 7 present our results and some additional robustness checks. Section 8 concludes.
2 Anecdotal Evidence In this section we brieﬂy outline a number of cases in which the US intelligence agencies’ reports allegedly became politicized. The main aim of this section is to provide historical evidence that career concerns really did cause analysts to distort intelligence reports and to make clear the distinction between two forms of politicization: the ﬁrst, which we argue took place during the Johnson administration, results when analysts with career concerns feel unable to speak truth to power. The second, more sinister, form involves analysts deliberately manufacturing material that can be used by their political masters to advance their policy agenda, an alleged case of which occurred in the run up to the 2003 Iraq War.
2.1 President Johnson and the Vietnamese Order-of-Battle By the mid 1960s American strategy in Vietnam had become a war of attrition. It was believed that, as long as US forces could inﬂict casualties on Communist forces that were suﬃciently in excess of their ability to recruit new members, the war would be won. In order to assess the progress of the war, it was therefore vital to estimate accurately the size of the Communist forces: its “Order-of-Battle”. In the case of the regular North Vietnamese Army this was relatively straightforward. In contrast, estimating the number of guerrilla ﬁghters was much more diﬃcult and responsibility for this fell on the Defense Intelligence Agency (DIA) with oversight from the CIA.
Unfortunately the DIA persistently underestimated the number of guerrilla ﬁghters. “American forces and ﬁrepower, it was believed, must be defeating Asian peasant soldiers and guerillas. Order-of-battle (OB) intelligence must therefore reﬂect that supposed reality,” (Andrew (1995) page 327). This problem was exacerbated by the tendency of the Johnson administration, and the President in particular, to interpret reports which conﬂicted with their existing beliefs as evidence of disloyalty on the part of the authors (Andrew (1995) page 328).
Even once the CIA began to exercise increased oversight over the reports, conﬂicting information was suppressed. When a subordinate CIA analyst, Sam Adams, wrote a memo querying the fact that the DIA’s estimates of guerrilla numbers in one province conﬂicted with captured Vietnamese documents by a factor of ten, this information was suppressed by senior members of the CIA “...After a week Adams went to the seventh ﬂoor and found his memo in a folder marked “indeﬁnite hold”...the oﬃcial in the next-door oﬃce told him that battle-order intelligence was...no concern of the CIA” (Andrew (1995) page 328).