Probability Words ctwardy | blog | SARBayes

NRC Panel: My paper is described in this Oct. blog entry.
SAR
Bob Koester
SAR World
Re-search
Ken C.
ISRID - inactive
D4H

Modeling
Causality
Causal Inference
Gelman
Flowing Data
Nuit Blanche
ACFR
Machine Learning
DecisionScience

Inference
Tim van Gelder
Overcoming Bias
Less Wrong
Zero Intel. Agents
ADVAT
Sources and ...

Φ
Socrates Wake
Computational Φ
Leiter Report
Experimental Φ
PhilPapers

Friends
pfh
njh'
Older njh
jpl-rpl


My Weblogs
SARBayes
Prior Analytics
SCI 410 news
Blogger -unused


"Very good chance" means anything from 50% to 90%. Here is a chart from the 1977 Handbook for Decision Analysis, that I have seen reproduced in other places.

Discussion

The data come from an experiment conducted by NATO intelligence analysts who were concerned about the problem of communication. Several different sentences were constructed in the following manner. "It is highly likely that the Soviets will invade Czechoslovakia," or "It is almost certain that the Soviets will invade Czechoslovakia," or "We believe that the Soviets will invade Czechoslovakia." The basic structure of all the sentences remained constant; only the verbal qualifiers changed.

Twenty-three officers, ranking from squadron leader to lieutenant general, served in the experiment. All participants were familiar with reading intelligence publications. They were asked to indicate the probability (in percent) that they would attribute to each message if they were to read it in that form in an intelligence article." The shaded boxes show Sherman Kent's suggested ranges for probability words. I added the bold in the quotation to emphasize the actual sentences that were used in the experiment.

As a colleague (Tod Levitt) points out, it's actually much worse than this. In some contexts, an informal "highly Likely" could easily be less than 10%. Luke Hope and I ran into this problem when using his software Verbal Reckoner to have people estimate probabilities for a Bayesian network. When there are many options and none is favored, the natural phrase to use was "even chance", or even "50-50". This of course made a global assignment incoherent. The best we could hope for was consistency within a particular Conditional Probability Table.


Background

I have seen this diagram in various places. Here it comes from the Handbook for Decision Analysis. As part of trying to understand the history of Bayesian methods in Intelligence Analysis, I have a copy of the handbook on loan from David Schum. However, I am happy to say I found it online (for $45) at Storming Media and (for $free) at ERIC (link above). Storming Media shows that DDI had many reports between 1979 and early 1986. Relevant to Apollo, several are co-authored by Buede and Sticha.

Source

This is from the DDI Handbook, pp. 67-68, Fig. 3-5. The handbook was written for the Cybernetics Technology Office of DARPA (then, ARPA). It is written as a handbook, and has no references. Does anyone know the original NATO study?



Loading
[æ]