Having just written a comment on the proper use of statistics in
determining Risks (for autonomous vehicles) for the mailing list RISKS (
http://www.risks.org) I was inspired to comment on the recent interactions
on this mailing list:
"Improving design methods (was Re: "What is Design Thinking" and
"Improvement In and Through Design Thinking")"
Some of the discussions demonstrated a weak understanding of statistics.
Not surprising: the normal training of designers does not include this.
Worse, when we are taught statistics, it is often the wrong kind. (See the
discussion "*Why designers need a special kind of statistical tests" *at
the end of this note.)
First of all, many fields have developed reliable methods of assessing
reliability of the impact of experimental manipulations. To quote Ali Ilhan:
Education researchers do these types of
analyses all the time with controlled experiments in classrooms, that is,
do a random assignment (or use a sampling strategy), try your "new" method
in one group, do nothing "special" in another group, compare the end
results statistically.
Ali is correct and his description captures the spirit of appropriate
testing. Note that the real test requires more sophistication
than
simple
-
random assignment, but nonetheless,
that is the major
basis.
There are potential other biases, so it is important to control for them.
It is often necessary to do double-blind
studies where neither the recipients nor the people doing the tests know
what condition they are in. It is also important to ensure that the various
test
sites were (statistically) equal prior to the test.
There are several phenomena that can bias results, one of which is called
"The Hawthorne Effect" and another is "Pygmalion." The first refers to the
fact that if people know they are being tested, their performance changes.
The second refers to the fact that if the people doing the test know what
is being tested, they are biased. (In the classic experiment, teachers were
told the names of some students who are "usually gifted." Those students
outperformed the others, even though they were randomly selected and were
not actually special: the teachers' beliefs influenced how the students
were treated and evaluated).
David Sless says:
it’s a bit like clinical practice in medicine where you look for symptoms
of pathology and then apply a treatment. You then look to see if the
symptoms disappear.
Unfortunately, this is a dangerous practice. This kind of test is badly
flawed, even though many physicians follow it. First, it is not blind, so
both physician and patient are biased toward a good result. For the
patient, this is "the placebo" effect. The placebo effect is real -- give
a patient a fake pill, and if they believe it to be a powerful new drug,
they might very well get better (the mechanism for this is still not well
understood). For the physician, it is the Pygmalion effect. And in any
case, a single experiment is statistically unsound: The person might have
gotten better with no treatment (this is the case for many back pain cases).
Most physicians are not scientists (even if the public thinks they are).
Many do not know statistics and do not know how to do proper experiments.
That's not in their training.
David's comments also illustrate what is called N=1 (or" n of 1")
experiments where "n" refers to the number of people being tested: a single
person rather than the hundreds or thousands often used in RCT - Randomized
Clinical Trials, which is today's gold standard. N of 1 trials can be done,
but the best way is to do a sequence of trials.
Consider my situation. For the past several decades, I take a statin pill
daily to treat cholesterol. Statins have as a possible side effect, muscle
weakness or soreness. Now, after years of taking the statin, I have muscle
soreness. So I stop taking the Statin. If the soreness goes away, does it
mean the statin was the cause? No. I have to be careful in assuming the
statin was responsible. So I reintroduce the statin and see if the soreness
comes back. I may have to do this serval times before I can have
confidence. (One of the graduate students in the UCSD Design Lab has
designed a simple method of assisting people in doing n of 1 experiments on
themselves that yield reliable results: doing this that allows people to
run their own trials on themselves.
https://arxiv.org/pdf/1609.05763.pdf
Ali sums it up well:
There are a multitude of factors that may affect the way kids learn reading
and writing (gender, being a minority, problems at home, quality of
teachers, peer effects, age in months etc.), and our design intervention
here, is just one among these many things. Even the fact that they are
using a new digital thing might make kids spend more time working on
reading and writing. But then this is a placebo effect, it is not our
design per se. I cannot envision any scenario that excludes using
statistics in this example, albeit very simple tests, nothing fancy. With
this many different possible sources of variations, five or ten
participants will never help us to understand the role of the app and its
design here.
*Why designers need a special kind of statistical tests *
Designers need a set of simple statistical methods that can inform our
work.
Note that we do NOT need the care and precision normally followed in
science and medicine. Why? Because they are looking for small effects
whereas we are looking for large ones.
To the practicing designer, if the change we are advocating does not make a
large difference (a factor of anywhere between 2 times and 10 times
improvement), it is not worth pursuing.
Scientists look for statistical significance, which does not mean practical
significance. Statistical significance means it is not likely to have
occurred by chance, but it may be a small effect.
We are looking for large effects. Even so, let us not be reckless. Doing
something and seeing a large impact by itself tells us nothing. Try doing
something that has zero relevance and presenting it to the
client/customer/user. It might very well have the same large impact.
Placebo effect.
We need double-blind studies. We need better research methods, ones suited
for looking for large effects (which can, therefore, be simple, quick,
etc.) but which nonetheless controls for factors that could otherwise
confound the results.
We need a good statistician to work with a good designer to develop a set
of methods.
Don
Norman
Prof. and Director, DesignLab, UC San Diego
[log in to unmask] designlab.ucsd.edu/ www.jnd.org <http://www.jnd.org/>
Executive Assistant:
Olga McConnell, [log in to unmask] +1 858 534-0992
-----------------------------------------------------------------
PhD-Design mailing list <[log in to unmask]>
Discussion of PhD studies and related research in Design
Subscribe or Unsubscribe at https://www.jiscmail.ac.uk/phd-design
-----------------------------------------------------------------
|