Vous êtes sur la page 1sur 4

dia and with government, modern ecological future of news.

Wilson Center Press,


thinking can perhaps become an ingredient in Washington, D.C., USA.
the news on natural disturbances. Farb, N., and A. H. Malcom. 1988. Up from
the ashes. The New York Times Maga-
Literature Cited zine, 6 November 1988, 137:48-53.
Smith, C. In press. Gonzo flame video in S.TA Pickett
Yellowstone: how the media covered a Institute of Ecosystem Studies
natural catastrophe. In P. S. Cook, D. BoxAB
Gomery, and L. W. Lichty, editors. The Millbrook, NY 12545

THE 10 MOST COMMON STATISTICAL ERRORS

Upon finishing his term as a member of the into subsets? Which effects were nested,
Board of Editors of the ESA, Mack (1986) within which effects, and which effects were
provided some advice to authors about writ- crossed, with which effects? What were the
ing with preCision, clarity, and economy. Here denominator mean squares? If the analyses
I follow his example with regard to common are at all complex, consider including a sam-
statistical errors, in the hope that this infor- ple SS table, or at least specify the degrees
mation will also be useful to authors. This ar- of freedom.
ticle is based upon comments made by me
Was the dependent variable a character
and by reviewers on the manuscripts I handled
measured on each individual, or a quadrat
as a member of the Board of Editors. The
mean, or a mean of three subsamples? Was
errors are ordered roughly by relative com-
it the proportion surviving or the number
monness and expressed as suggestions.
surviving? Was a ratio calculated for each
1. Explain clearly what you did. Describe both plant separately, or by dividing means?
the experimental design and the statistical
analyses clearly and precisely, so that the If the editor and reviewers cannot answer
reader can determine exactly what you did. questions like these after reading the manu-
Make sure that questions such as these are script, they do not have the information to
answered: understand and evaluate what you did. Mis-
understood designs and analyses sometimes
How many replicates were there? What were lead reviewers and editors to decide that the
the sample sizes? How were plots, plants, data do not support the conclusions drawn
shoots, seeds located/chosen/arranged? from them. Frequently, unclear descriptions
How many rows and how many columns suggest that the study may have other sta-
were there in each contingency table? What tistical problems (e.g., the degrees of freedom
variable is represented by the rows, what may be apparently inconsistent with the de-
variable by the columns? Was a chi-square sign).
or a G test used to test the table for sig- Standard terminology should always be
nificance? Were any cells pooled? If so, used. Avoid BMDP and SAS acronyms.
which ones and why?
2. Avoid doing lots and lots of separate sta-
Exactly how were the sums of squares (SS) tistical tests. If you do enough statistical tests,
tables and the Ftests constructed for each some are bound to show significance by
analYSis of variance (ANOVA) and each chance alone. In most cases this problem can
analysis of covariance (ANCOVA). Which be avoided by replacing hordes of t tests with
variables were used as effects in the model, the appropriate ANOVA, or by replacing a large
and which were used to partition the data number of one-way ANOVAs with the appro-

161
priate multifactor ANOV A. If large numbers of stances (e.g., Sokal and Rohlf 1981: 284-
separate tests are unavoidable, adjust the 285), but it is a dangerous practice if done
critical Pvalue(s) accordingly (e.g., Rice 1989, without appropriate justification. If it is una-
Sokal and Rohlf 1981 : 779-782). Avoiding the voidable (as it may be in some contingency
pitfalls awaiting those who attempt more com- table analyses), the potential problems should
plex analyses is unfortunately not a sufficient at least be recognized.
justification for dOing lots of simpler analyses.
6. Before you conclude that the null hypoth-
3. When using ANOVAs, ANCOVA S, and esis is true, determine the power of the test.
regressions, be aware of the assumptions and The absence of statistical significance does
test your data for them. ANOVA, in particular, not necessarily mean that the effect did not
is robust to certain departures from assump- occur. Perhaps the sample sizes were too
tions, but not all departures from assumptions small, or the background variation too large,
are permissible in ANOV A or any other anal- to detect any but the largest differences be-
ysis. To avoid some of the commonest prob- tween treatments. Determination of the power
lems in this category, consider the following: of the test (Sokal and Rohlf 1981: 161-168,
If the dependent variable only takes a few Snedecor and Cochran 1989: 68-70, Steel
different values, it may not be legitimate to and Torrie 1980: 113-119) will give you an idea
analyze it as a continuous variable. (A contin- of how small a difference you could have de-
uous dependent variable is an assumption of tected as significant. Don't forget to check
all of these analyses.) Be sure to check the your data for trends. A nonsignificant but ob-
residuals, not the raw data, for normality. If vious trend suggests that the null hypothesis
the dependent variable often takes the value should not be accepted. Of course, in this
0, the distribution of the residuals may not be case the null hypothesis cannot be rejected,
sufficiently close to normal in many cases. either.
Transformations do not correct this problem,
nor do they correct the problem of too few 7. Use multiple comparison tests correctly. A
values, mentioned above. If a transformation posteriori multiple comparison tests should not
is used, for example to increase the homo- follow a nonsignificant main effect. Except in
geneity of variances, check the transformed special cases multiple comparison tests should
variable to see whether it meets each as- not be used as a substitute for a two-way
sumption. ANOV A with interaction term. See the recent
article by Day and Quinn (1989) for updated
4. Interpret interaction effects in multi-way advice on the selection and use of multiple
ANOVAs correctly. In some cases, authors comparison tests.
discussed nonsignificant interaction effects as
if they were significant. A hypothetical ex- 8. Include replication in experimental designs
ample of this error is the reporting of a non- and observational schemes. Despite the at-
significant water· fertilizer effect, followed by tention this issue has recently received (large-
"fertilizer increased yield in the watered plots ly due to the much-cited article by Hurlbert
but not in the unwatered plots." [1984]), some manuscripts still report designs
without replication and do not recognize them
5. Don't pool plots without justification. I use as such. This topic is closely related to the
"plots" in the technical sense; a "plot" can following one:
be a Petri dish containing seeds, a portion of
a greenhouse bench, or a quadrat in the field. 9. Beware of confounded factors. In some
A "plot" can be a block in a blocked design. instances, especially in descriptive studies or
If plot is not included as a term in the analysis, "natural experiments," factors are inherently
it has been pooled with some other term, and confounded. An experimental design can con-
the effects due to differences between differ- found factors, for example by placing all in-
ent plots will be confounded either with the dividuals that receive treatment A in plot 1, all
effects of an experimental treatment, or with individuals that receive treatment B in plot 2,
the residual variation. Both Type I (spurious etc. Mortality or other losses during the ex-
significance) and Type \I (overlooked signifi- periment may result in missing cells, which
cance) errors can arise from such pooling. cause factors to be confounded. A hypothet-
Pooling plots can be justified in some circum- ical example of the latter: An experiment with

162
two levels of watering and two levels of fer- make a good ecological study, but statistical
tilizer was conducted. If all the low water/low errors can fatally damage an otherwise ex-
fertilizer plants died, leaving only three of the cellent study.
four possible treatment combinations, it is not Statistical techniques are some of the most
possible to ascribe differences in final biomass widely used, and most misused, of any tech-
to the separate effects of water, or of fertilizer, niques utilized by ecologists. Over half of the
or to their interaction. Regardless of the source papers that involved a statistical analysis had
of the confounding, problems arise if an au- one or more statistical errors, even excluding
thor fails to recognize the situation. those errors classified in (1) above. That so
many of these errors involved relatively basic
10. Repeated measures of the same individual statistical knowledge is worrying.
are not independent, and cannot be analyzed However, complex questions often require
as if they were. This situation requires special complex designs to answer, and complex de-
analyses (e.g., Gurevitch and Chester 1986). signs usually require more complex statistical
analyses. Perhaps we are asking more com-
Most of these suggestions have been word- plex questions-and our statistical expertise
ed in terms of manipulative, controlled exper- has not kept up. The high frequency of errors
iments. Many of the manuscripts, however, in the more complex analyses may be a sign
described descriptive studies. The frequency of the rate at which the questions we are ask-
and type of errors did not appear to differ ing are becoming more complex and sophis-
between the two types of studies. Each of the ticated.
suggestions applies to descriptive studies as
well as to manipulative experiments. Acknowledgments
There was also a wide variety of less com-
mon errors. Sometimes the author appeared I thank my fellow editors D. Goldberg, L.
to have ignored his/her own statistical results, Miller, P. Morin, M. Rausher, and J. Travis,
and asserted that effects were found when as well as M. SciOli, for their suggestions and
the statistical tests gave nonsignificant re- comments on earlier versions of this manu-
sults, or that no effect was found although the script. I also thank the many reviewers who
statistical test was significant. Sometimes the generously gave their time to evaluate man-
appropriate test of an assertion had not been uscripts and to provide comments for the au-
done. Most of the remaining errors involved thors.
problems with the application, construction,
and interpretation of multi-way ANOVAs,
Literature Cited
nested ANOVAs, or regressions.
No attempt has been made here to provide Day, R. W., and G. P. Quinn. 1989. Com-
complete explanations or references for any parisons of treatments after an analysiS
topic. Most standard introductory texts cover of variance in ecology. Ecological Mono-
most of the topics mentioned here (e.g., Sokal graphs 59:433-463.
and Rohlf 1981 , Snedecor and Cochran 1989, Gurevitch, J., and S. T. Chester. 1986. Anal-
Steel and Torrie 1980). In some instances, ysis of repeated measures experiments.
individual advice from a statistician may be Ecology 67:251-255.
the best source of help. Hurlbert, S. H. 1984. Pseudoreplication and
I encourage all authors to make sure that the design of ecological field experi-
their statistical analyses are correct and that ments. Ecological Monographs 54: 187-
their designs and analyses are clearly de- 211.
scribed, before submitting a manuscript to Mack, R. N. 1986. Writing with precision,
Ecology or Ecological Monographs. Avoiding clarity, and economy. ESA Bulletin 67:
statistical errors, especially the common ones 31-35.
mentioned above, will increase the likelihood Rice, W. R. 1989. Analyzing tables of sta-
that your manuscript will eventually be ac- tistical tests. Evolution 43:223-225.
cepted. By reducing the amount of revision Snedecor, G. W., and W. G. Cochran. 1989.
required, it will also decrease the time be- Statistical methods. Eighth edition. Iowa
tween submission and acceptance. Of course State University Press, Ames, Iowa, USA.
good statistical methodology alone does not Sokal, R. R., and F. J. Rohlf. 1981. Biom-

163
etry. Second edition. Freeman, San Fran- Norma Fowler
cisco, California, USA. Department of Botany
Steel, R. G. D., and J. H. Torrie. 1980. Prin- University of Texas
ciples and procedures of statistics; a bio- Austin, TX 78713
metrical approach. Second edition.
McGraw-Hili, New York, New York, USA.

ON THE RETIREMENT OF JOHN LANGDON BROOKS FROM NSF

After nearly twenty years of service and formed. Its job was to oversee the adminis-
leadership as an administrator of federal sup- tration ofthe U.S. International Biological Pro-
port for research in ecology and systematics, gram. Facilities such as museum collections
John Brooks retired from the National Science were supported in the Systematic Biology
Foundation in June 1989. He was the consis- Program, and John became the Program Di-
tent source of sound judgment during a period rector for the core of ecology, the General
of dramatic changes as the field responded Ecology Program. By 1981 he had risen to
to new tools and new ideas. Molecular meth- the directorship of the entire administrative
ods, developed in microbiology and genetics, unit. In 1975 a new program was created to
were found to be useful in studies of popu- support the population genetics that was for-
lation genetics, systematics, and even behav- merly supported in SystematiC Biology, along
ioral ecology. Systematic data analysis be- with the population ecology and physiological
came dominated by cladistic methods. Highly ecology that was formerly supported in Gen-
complex models of ecosystems were shown eral Ecology, thereby allowing growth in all
to be useful analytical tools. Team studies, three areas. The title of the new program was
labelled Long-Term Ecological Research Population Biology and PhYSiological Ecolo-
(LTER) were incorporated into national net- gy. John became Head of the Ecology and
works of ecological reserves. The differences Population Biology Section and then Deputy
between environmental science, applied ecol- Director (1975) and Director (1981) of the Di-
ogy, and basic research became less clear vision of Environmental Biology. During that
and biologists realized that basic research in time a new program for support of facilities
conservation-related areas should have a was created, Biological Research Resources.
higher priority on their agendas. The job of As Division Director, John was responsible for
NSF has been to respond to these changing the five present programs (Ecology, Popula-
needs, and to offer what support is available tion Biology and Physiological Ecology, Eco-
to the most promising projects. Of course this system Studies, SystematiC Biology, and Bi-
job has been made difficult by the ever-in- ological Research Resources). These
creasing lag in the growth of funds available programs provide most of the federal support
for support compared to the need. The single of basic research in whole organism biology.
person who has had the major responsibility (The exceptions are marine ecology, which is
for making wise decisions in this changing supported in the Biological Oceanography
scene, one that bears importantly on our abil- Program, and behavior, which is supported in
ity to do basic research in ecology and sys- the Biological Basis of Behavior Program.)
tematics and on the health of our discipline, I had the good fortune to work as a program
has been John Brooks. officer in this group from 1973 to 1977. That
One way that NSF responds to changing was long enough to see several "rotators"
needs is by reorganization. When John came (now called "visiting scientists": program di-
to NSF in 1969, the Ecosystem Analysis Pro- rectors on leave from their universities who
gram (now Ecosystem Studies) had just been worked at NSF for one or two years) come

164

Vous aimerez peut-être aussi