This is somewhat tangiential, but I encountered a similar phenomenon
in a study I ran recently, and I'd be interested to hear what people
on this list had to say about it.
Long story short, I gave my subjects a 4-hour window to complete a
task, but found every one either did it within two hours, or failed to
complete it at all within four (despite continuing to work, apparently
understanding the problem, etc). I can't say the distribution is
strictly "bimodal", but they definitely fell into two distinct
The paper is at
and the graph is on page 5 (section 4.4).
The "success breeds success" explanation does sound like an intuitive
fit. I would be very interested to see what others have to say. (Other
comments on the study are, of course, also welcome).
On Fri, Nov 27, 2009 at 6:43 PM, Raymond Lister <raym...@it.uts.edu.au> wrote:
> On Fri, 27 Nov 2009, Richard O'Keefe wrote:
>> Anthony Robins
>> recently noticed
>> that the characteristic bimodal distribution of outcomes in CS1
>> can be explained by a very simple statistical model.
>> (Roughly speaking: a series of topics where failure at any point makes
>> later failure more likely, success makes later success more likely.)
> First, I recommend Anthony Robins' paper as a refreshing alternative
> voice to the repeated claims that passing a zero pre-requisite
> programming course requires an innate talent (the "geek gene). But, as
> with all research, there are a issues with his work that require follow
> First, I'd like to see evidence collected, from multiple institutions,
> examining the grade distribution in CS1. I think the bimodal
> distribution is an urban myth. (Would anyone care to formally define
> what they mean by a bimodal grade distribution? That's a rhetorical
> question, and no correspondence is required.)
> Second, if we are going to make research claims about grade
> distributions, then we need to bring research standards to bear on the
> contruction of our research instrument - the exam. Most exams are
> idiosyncratic constructions, of dubious validity and reliability.
> Third, I regard Anthony's model as a good counter argument to the
> conventional "geek gene" argument, but if I ignore the "geek gene"
> argument and merely examine Anthony's model in isolation, then I
> suspect that the number of free parameters in his model means that
> you can get any distribution you want, by a suitable choice of
> parameter values.
> Dr. Raymond Lister
> Science Teaching and Learning Fellow,
> University of British Columbia,
> Department of Computer Science,
> Vancouver, Canada