[Jdm-society] Experimenter bias

[Deborah Frisch]

Ken Hanson at Los Alamos National Lab: I believe there is evidence to support the notion that experimenters may be influenced, perhaps subconsciously, by previously published results. Those of us who are or have been involved in physics experiments recognize this possibility.

DF: One of the basic findings of cognitive and decision science is that perception is a combination of top-down and bottom-up processing.  Researchers like Anne Treisman, Barbara Tversky, Mike Posner, Steve Keele and others have demonstrated this for visual perception, Warren & Warren (1970, "phonemic restoration effect") and others have demonstrated it for speech perception and Kahneman, Tversky, Lichtenstein and Slovic have demonstrated it for risk perception.  Like you, social psychologists Lord, Ross and Lepper noted the troubling implications of this fact about human cognition for the process of science. They coined the term "belief polarization" to describe the social pathology that arises from the fact that the magnitude of interpersonal or intergroup disagreement often increases as evidence accumulates, contrary to the predictions and prescriptions of Bayes Theorem, not to mention common sense. 

Scott Plous (1993, p. 260) sums up beautifully the particular irony that exists for decision scientists:   judgment and decision research is subject to a paradox:  If its findings are valid (i.e., biases and errors exist), then its findings are, to some degree, subject to biases and errors.  At the same time, if its findings are not valid (i.e. biases and errors do not exist),then its findings are still in error. Either way, biases and errors are likely to be present.  This paradox is similar in some respects to Austrian logician Kurt Godel's Incompleteness Theorem, which states that any complete, mathematically powerful, and formalized system of logic [like the axioms of EU] must be internally inconsistent.

Bottom line:

People rely on a combination of top-down and bottom-up processing to interpret empirical evidence.

Scientists are people.
Data are empirical evidence.
Therefore, scientists rely on t-d (e.g., previous research) and b-u processing (e.g., statistical analyses) in interpreting evidence. 

KH: Incidentally, I have signed up to talk in a few months about what we can learn about likelihood functions from the history of nuclear-physics measurements. My main purpose is to determine whether the data support long-tailed likelihood functions, which are often invoked to handle outliers.

DF: This sounds (very) vaguely like the distinction between rank-dependent and traditional probability weighting but that might just be my own cognitive bias. 

KH: I am wondering whether you have considered or studied the issue of experimenter bias toward matching previous results. Is this a topic that has been addressed by statisticians or elicitation experts, and if so, what keywords would apply? 

DF: This one's easy.  We call this "confirmation bias" and it was first demonstrated by a British psychologist named Wason in two famous puzzles (2-4-6 and 4-card) although it is not clear what either of them has to do with the decision scientist's definition of confirmation bias (articulated elegantly, albeit not pithily, in the brilliant 1987 Psychological Review paper by Klayman & Ha previously cited. Fischhoff & Beyth-Marom have a comparably top-notch but tough to sum up Psych Review paper on the same general topic).

KH: Some would question themselves, others the previous experimenters.

DF: By question themselves, I trust you mean "trust whether you tested your cherished hypothesis in the right way."  If you truly mean "consider the possibility that your beloved hypothesis is wrong" then either physicists are much better scientists (e.g., take the admonitions of Carl Sagan and Richard Feynman more seriously) or slightly worse psychologists (e.g.,  more self-deceived) than psychologists.  Are you really claiming that physicists question their cherished hypotheses and not just their methodological skill when their results don't pan out?  If so, more power to you!

KH: Anyway, it is an interesting topic.

DF: Roger that.
 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.sjdm.org/mail-archive/jdm-society/attachments/20070511/85c67d25/attachment-0001.htm