[Jdm-society] an editorial on confidence/probability assessment in national intelligence

[David Budescu]

Editorial from the 2/20/05 Washington Post


What Percent  Is 'Slam Dunk'?
By Michael Schrage

   The controversial decision to reorganize America's sprawling intelligence
   establishment has set in motion the most sweeping bureaucratic change
   for sensors, spies and satellites since the end of World War II.
   Unfortunately, the odds are excellent that this multibillion dollar
   structural shuffle -- capped last week by the appointment of veteran
   diplomat John Negroponte as the new national intelligence director --
   will do little to improve the quality of intelligence analysis
   for this country.

   Why? Because America's intelligence community doesn't like odds. Yet
   the simplest and most cost-effective innovation that community could
   adopt would be to embrace them. It's time to require national  security
   analysts to assign numerical probabilities to their professional
   estimates and assessments as both a matter of rigor and of record.
   Policymakers can't weigh the risks associated with their decisions if
   they can't see how confident analysts are in the evidence and
   conclusions used to justify those decisions. The notion of imposing
   intelligence accountability without intelligent counting -- without
   numbers -- is a fool's errand.

   World-class investment banks, insurance companies and public health
   practitioners are increasingly bringing greater quantitative
   sophistication to their risk analyses. For reasons having chiefly to
   do with custom, culture and practice -- not competence or cost -- the
   CIA, Defense Intelligence Agency, FBI and the federal government's
   other analytic agencies have shied away from simple mathematical
   tools that would let them better weigh conflicting evidence and data.
   That bureaucratic shortsightedness undermines our ability to even see
   the dots, let alone connect them.

   Consider the National Intelligence Estimates, the Presidential Daily
   Briefings or many of the critical classified and unclassified analyses
   flowing through Washington's national security establishment. Key
   estimates and analytic insights rarely come with explicit probabilities
   attached. The nation's most knowledgeable experts on the Middle East,
   counterterrorism, nuclear proliferation, etc., are seldom asked to
   quantify, in writing, precisely how much confidence they have in their
   evidence or their conclusions. Your personal financial planner does a
   better job, on average, of quantitative risk assessment for your
   investments than the typical intelligence analyst does for our
   national security.

   For example, when the State Department's Bureau of Intelligence and
   Research "predicted" terrorism and insurgency in the wake of the
   invasion of Iraq, its forecasts avoided explicit probabilities. But
   precisely how confident were the bureau's experts in their assessments
   of the breadth and intensity of the projected opposition? Did they
   believe that there was a 60 percent or a 40 percent chance that Sunni
   Triangle violence would spread north? Did they foresee a 20 percent or
   a 75 percent chance that car bombings of Shiite mosques would provoke
   widespread retaliation against Sunnis?

   The cultural bias against numbers deprives congressional and White House
   decision-makers of essential metrics to weigh the analytical
   community's own credibility. Naming a national intelligence director
   doesn't change that.

   More than 40 years ago, Sherman Kent -- the godfather of the vital National
   Intelligence Estimates and the man for whom the CIA's analyst school is
   named -- penned a classified memo attempting to describe how vague
   words like "probable" and "serious probability" could be translated
   into meaningful numbers. His "Words of Estimative Probability" proved a
   rhetorically awkward and ultimately futile exercise in encouraging
   more disciplined discussions of probability in the analytic community.

   Passive-aggressive organizational resistance to quantitative rigor
   continues to this day. Former acting CIA director and longtime analyst
   John McLaughlin tried to promote greater internal efforts at
   assigning probabilities to intelligence assessments during the 1990s, but
   they never took. Intelligence analysts "would rather use words than
   numbers to describe how confident we are in our analysis," a senior CIA
   officer who's served for more than 20 years told me. Moreover, "most
   consumers of intelligence aren't particularly sophisticated when it
   comes to probabilistic analysis. They like words and pictures, too.
   My experience is that [they] prefer briefings that don't center on
   numerical calculation. That's not to say we can't do it, but there's
   really not that much demand for it."

   That doesn't mean it shouldn't happen. Fortunately, there's no need
   for a dramatic revolution; subtler measures will do. Here's a
   suggestion: The simplest, easiest, cheapest and most powerful way to
   transform the quality of intelligence would be to insist that
   analysts attach two little numbers to every report they file.

   The first number would state their confidence in the quality of the
   evidence they've used for their analysis: 0.1 would be the lowest level
   of personal/professional confidence; 1.0 would be -- former CIA
   director George Tenet should pardon the expression -- a "slam dunk,"
   an absolute certainty.

   The second number would represent the analyst's own confidence in his
   or her conclusions. Is the analyst 0.5 -- the "courage of a coin toss"
   confident -- or a bolder 0.75 confident in his or her analysis? Or is
   the evidence and environment so befogged with uncertainty that the
   best analysts can offer the National Security Council is a 0.3 level of
   confidence?

   These two little numbers would provoke intelligence analysts and
   intelligence consumers alike to think extra hard about analytical
   quality, creativity and accountability. Policymakers could swiftly
   determine where their analysts had both the greatest -- and the least --
   confidence in their data and conclusions. Decision-makers could quickly
   assess where "high confidence" interpretations were based on
   "low-confidence" evidence and vice versa. That's important information
   for decision-makers to have. Then their ability to push, prod and poke
   the intelligence community would be firmly grounded in their own
   perception of the strength and weakness of the work coming out of it.

  Seeing the areas in which top analysts consistently rate their
  confidence in evidence below a 0.5 might evoke new thinking from the
  covert operations and "sigint" crowds in Langley and Fort Meade as to
  what data they should be procuring. More significantly, these two
  numbers would build a record -- an ongoing audit trail of probabilities
  and odds -- to revisit and review.

  Pushing analysts to weight their intelligence assessments creates a less
  ambiguous standard of accountability, which might explain why the
  consensus in the analytical community is to avoid disclosing their
  odds. But House and Senate intelligence committees seeking greater
  accountability and better quality from the newly reorganized
  intelligence bureaucracies should insist that analyses brought in for
  congressional review -- classified or not, publicly disclosed or not --
  include confidence rankings.

  Yes, analysts and their agencies will attempt to "game" the numbers. Yes,
  policymakers will apply political pressure for analysts and agencies to
  alter their declared odds. All risk-assessment methods are
  corruptible.  But these mechanisms can self-correct. Too many risk analyses
  that hover uselessly below 0.5 or provide too few assessments of 0.7 and
  higher that later turn out to be accurate tend to create their own
  pressures for fundamental change. Better accountability promotes better
  analysis. And better analysis comes from the explicit explanations and
  conversations around probability and risk.

  But even greater analytical accountability isn't good enough. A
  growing number of fields ranging from medical diagnostics to Internet
  spam filtering, for example, increasingly rely upon Bayesian analysis --
  a probability theory that predicts the likelihood of future events
  based on knowledge of prior events -- as a powerful tool to weigh new
  evidence. Bruce Blair, director of the non-partisan Center for Defense
  Information, argues convincingly on the CDI Web site that Bayesian
  analysis goes a long way toward explaining the seemingly flawed risk
  assessments made by the intelligence community during the run-up to
  9/11 and the Iraq war. Nonetheless, although the CIA is familiar with
  Bayesian analysis and its computational cousins, these techniques
  haven't seeped into the national security community's analytical
  mainstream.

  Medical doctors and Wall Street traders today do a better job of challenging
  themselves to explore the growing diversity of analytic options. Even
  Major League Baseball teams, as Michael Lewis documents in his
  best-selling book "Moneyball," are grasping that data-driven analyses
  can lead to better talent acquisition and management decisions. Why
  should professional baseball executives be doing more innovative
  statistical analyses than professional intelligence analysts?

  Mathematics is not a substitute for judgment, nor do equations define
  analysis. But analyzing risk without probabilities is akin to
  discussing art without colors. You can do it, but don't be surprised at
  the sterility of the results. Unfortunately, with evidence I'd weight
  at 0.8 and a conclusion to which I'd assign a confidence rating of 1.0,
  I predict that until the intelligence community overcomes its
  reluctance to go the probability route, it will continue to compromise
  its ability to adequately assess national security risks and threats.

 Author's e-mail: schrage at media.mit.edu

 Michael Schrage is a senior adviser to MIT's Security Studies program. He has participated in non-classified CIA workshops on intelligence analysis.

Would you like to send this article to a friend? Go to
http://www.washingtonpost.com/ac2/wp-dyn/admin/emailfriend?contentId=A37115-2005Feb19&sent=no&referrer=emailarticle