on the proper use of tools
TRANSCRIPT
Editorial
On the proper use of tools
The Viewpoint section of this issue of EJLST contains a commentary on the use of Response
Surface Methology (RSM) in research and technology. In that I have applied RSM in my
personal research and also have reviewed many manuscripts that used this method, I appreci-
ate the invitation extended by Professor and Editor Uwe Bornscheuer to comment on the
Viewpoint.
In the ‘Odds and Ends’ drawer of my desk is an old multi-bladed pocket knife, made of
high quality steel. Although its two smallest blades are intact and sharp, its larger main blade is
snapped off, a stub compared to its original length. This stub is a testament to the fact that even
very good knives make very poor prybars.
Similarly, in my gun cupboard sits a shotgun, many of whose screws have badly chipped
and rounded slots for the corresponding screwdriver. These are a permanent testimony to the
day that, as a much younger man, I learned that most screwdrivers and their companion screws
have tapered blades and correspondingly tapered slots, but that firearms screws and screw-
drivers have parallel faces. Regular screwdrivers therefore do not properly fit the screws in
firearms, and when used in this improper application can slip out and destroy the screw head.
These two examples illustrate the fact that a good tool used inappropriately is a recipe for
disaster. In a Viewpoint in this issue, Dr. Albert Dijkstra essentially observes that the same goes
for the tools we use in our research. He makes this observation in regard to response surface
methodology (RSM), which, to paraphrase Wikipedia, is a statistics-based experimental
method that explores the relationships between multiple experimental variables and one or
more response variables. Response Surface Methodology involves the use of a sequence of
designed experiments to identify the conditions that will give an optimal response. The
technique has become popular in some types of research because it can produce a maximum
amount of data from a relatively small number of reactions, and can yield powerful predictive
results. In that it allows identification of interactions between multiple variables in an
experimental design, RSM offers a power that is lacking in the more conventional approach
of changing one variable at a time.
An experienced scientist, Dr. Dijkstra has spent a long and productive career conducting
research and managing research teams, projects and entire research centers. He is very good at
what he does, and he is widely regarded as a wise and accomplished man. In his Viewpoint he
presents a multi-faceted commentary on the use, or rather the misuses, of statistically designed
experimental planning and the consequent application of response surface methodology to
the results.
I have a high regard for the author of the Viewpoint, a skilled and wise researcher with a
multitude of accomplishments in science and technology. Also, I agree with several of the
cautions voiced in his Viewpoint. But, as in looking at a glass with water in it and seeing it as
either half empty or half full, I see some of the points raised in the Viewpoint in a somewhat
different light than does the author. After reading the Viewpoint, please consider the following
comments, whose numbers correspond to the similarly numbered items there:
1. Dubious significance: Here the author of the Viewpoint observes that by addition of
extra terms to a data fitting equation one can reduce the apparent standard error of the
fit to the data. It is correctly noted that to do so without increasing the number of
experimental determinations as well is ill advised. Perhaps RSM has become too easy to
apply, and thus too easy to apply incorrectly, such as in without appreciating this fact.
Some contemporary RSM programs will help recognize such a situation as this,
reporting an ‘adjusted’ R2 value to alert the researcher to the presence of an excessive
number of terms in the equation of fit.
The Viewpoint admonishment to conduct multiple repeat determinations in order to
determine the standard error of the measurements is wise advice, best heeded by all who
Michael J. Haas
Eur. J. Lipid Sci. Technol. 2010, 112, 1287–1289 DOI: 10.1002/ejlt.201000530 Editorial 1287
� 2010 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.ejlst.com
implement RSM. Also, the observation is well made that just because your computer will
print results out to seven digits does not mean that the last four or five have any meaning or
should be reported.
2. Reproducibility: This comment alleges that it would be wise to repeat entire RSM
experiments, in order to determine if the original set of data is accurate. It also
speculates that researchers don’t usually do so for fear that their encouraging first
set of results may be thereby proven incorrect. I suspect that this is no greater a reason
for a lack of duplication than with other experimental design methods (a poor
researcher is a poor researcher largely irrespective of his research tools). In fact a well
designed RSM protocol has sufficient internal controls that a single run of the exper-
iment gives a good feeling for the reliability of the data.
3. Repeatability: An argument for the validation of RSM results by the use of multiple labs
to repeat the work. Such validation would add reliability to all research data, not just
that from RSM. I do not see a reason to demand more validation from RSM than we
routinely demand from other types of experimental design.
4. The use of absolute values: An argument for conducting experiments in a varying
order, to minimize the effects of long term fluctuations. Again, this seems to demand
of RSM that which we do not routinely demand of other experimental methods.
In fact, the Viewpoint goes on to acknowledge that the better practitioners of RSM
adopt a random order in conducting their experiments, to minimize the effect of such
fluctuations.
5. The choice of process variables: The author of the Viewpoint observes that that in
applying an RSM approach one may ignore the investigation of some major parameters
affecting the system. I agree, but again cannot see this as a problem unique to RSM.
Failure to correctly perceive and investigate the parameters most likely to impact our
systems is a danger we all face, whether we use RSM or not. Should I choose to explore
whether my tomatoes grow well if I read to them from Shakespeare, Goethe, or Aldo
Leopold, while failing to provide sunlight and water, neither RSM nor any other
experimental design approach will prevent the world from judging me a fool.
6. The value of process variables: And yes, if I should choose to study the effects of water,
but investigate the effects of only two levels, 0 and 25 liters per day per two liter pot, I
should not be surprised if the plants die under both treatment regimes and I fail to
identify water as a requirement in tomato agriculture. The Viewpoint caution here that
it is very important to have a feel for the process before selecting the variables and their
settings is excellent. RSM provides the world no more insulation against shabby
researchers than do other experimental approaches.
7. RSM is superficial: This section faults RSM methods for being applicable in cases
where one has no hypothetical mechanism from which to identify variables, inter-
actions, and relationships. Exercise more of the familiarization with your system that is
suggested in item no. 6 and you may come up with a hypothesis to guide a better RSM
examination. Even so, I am not convinced that the ability to function in the absence of
hypothesis is necessarily fatal. If one designs and conducts an RSM experiment for
which major effecting variables are not examined, the fit of the results to the data will be
poor, and the serious investigator will immediately see the need for more study to flesh
out the real impacting variables in the system.
8. Non-elimination of outliers/ rogue values: Here the author of the Viewpoint relates his
past improvement of the quality of a data set by deleting a data point that his experience
led him to recognize as an outlier. He faults RSM for lacking this capability. In fact,
some RSM programs can conduct outlier searches and alert the experimenter of their
presence. More importantly, RSM does not remove the researcher from the dance, does
not allow him or her to retire their powers of observation from the field of battle. In fact,
I am unaware of any experimental approach in which this would be advisable.
9. No improvement over established methods: The ‘methods’ referred to here are estab-
lished and wise ways of doing research and technology: attention to your system,
attention to feedback from those implementing the reaction at the bench or plant,
1288 Michael J. Haas Eur. J. Lipid Sci. Technol. 2010, 112, 1287–1289
� 2010 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.ejlst.com
and coming to an understanding of the process being studied. RSM does not invalidate
these. I suspect it speeds up the integration of seemingly incongruous observations into
a more complete understanding of the system.
10. Limited usefulness: Here the Viewpoint describes the author’s past use, in industrial
operations, of simple tables that made it easy to optimize the output of a process in the
face of variations in the feedstock. Surely such tables have value; and RSM is not meant
to be all things to all technologies and do away with all else. However, the construction
of those simple guiding tables might have been much easier if guided by the results of an
RSM investigation. The assertion that RSM experiments are done in laboratories and
thus are irrelevant to industrial scale operations is one on which I’d like to have more
data. Taking the assertion as true, one is left wondering why so much of industry would
bother with a method of no practical utility.
11. A method or a methodology? A question of the term, or is it terminology, by which to
refer to RSM. I will leave this decision to those with a stronger sense of English language
usage than I.
In my own research I have found RSM to be a powerful tool. It economizes on the thing I
have the least of: technician time. It also economizes on the use of reagents, which can be in
short supply or of high cost. Most appealing to me, it identifies the existence and magnitude of
interactions between multiple variables in a way that the conventional method of changing one
variable at a time, holding all others at constant values, never can. It also gives one predictive
powers that allow an estimation of how a reaction or system will respond at a chosen set of
conditions not yet examined.
As a journal editor, I have the opportunity to see firsthand the results of the misuse of RSM,
and to make decisions on whether the scientific community should be burdened by the bland
resulting fruits. I am most put off by examples of the poor choice of the ranges of one or more
variables in an experimental design when allegedly seeking to optimize the yield of a reaction.
Poor choice in selecting the ranges of variables leads to predictive surfaces that rise as they
approach their edges, and authors who call these edges ‘maxima’. Local maxima perhaps, but
the rising surface makes it clear that the real optimal conditions lie somewhere off in a space
that the authors were too inattentive or too lazy to search.
Observe any craftsperson, any quality worker in any trade who is able to make a living at
what they do, and you will probably see that success depends on a desire and ability to apply
tools effectively and efficiently, and in knowing when to apply the proper amount of detail and
care. To be a researcher is to be a craftsperson. Success lies in knowing when and how to do a
good job, and in using the tools in ones’ toolbox properly to create a product recognized by
oneself and ones’ peers as of high quality. Do anything less and you will eventually create
faulty, weak data with poor predictive capability. Then you will have the opportunity to ponder
one of the profound questions that all poor craftsmen eventually face: ‘If you don’t have time
to do it right, where will you ever find the time to do it over?’
Michael J. Haas
Eur. J. Lipid Sci. Technol. 2010, 112, 1287–1289 DOI: 10.1002/ejlt.201000530 Editorial 1289
� 2010 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.ejlst.com