Things I Have Learned (So Far)

Jacob Cohen New York University

ABSTRACT: This is an account of what I have learned 20 friend's power was rather worse (.33), but of course
(so far) about the application of statistics to psychology he couldn't know that, and he ended up with nonsignif-
and the other sociobiomedical sciences. It includes the icant resultswith which he proceeded to demolish an
principles "less is more" (fewer variables, more highly important branch of psychoanalytic theory.
targeted issues, sharp rounding off), "simple is better"
(graphic representation, unit weighting for linear com- Less Is More
posites), and "some things you learn aren't so." I have One thing I learned over a long period of time that is so
learned to avoid the many misconceptions that surround is the validity of the general principle that less is more.
Fisherian null hypothesis testing. I have also learned the except of course for sample size (Cohen & Cohen, 1983,
importance of power analysis and the determination of pp. 169-171). I have encountered too many studies with
just how big (rather than how statistically significant) are prodigious numbers of dependent variables, or with what
the effects that we study. Finally, I have learned that there seemed to me far too many independent variables, or
is no royal road to statistical induction, that the informed (heaven help us) both.
judgment of the investigator is the crucial element in the In any given investigation that isn't explicitly ex-
interpretation of data, and that things take time. ploratory, we should be studying few independent vari-
ables and even fewer dependent variables, for a variety
of reasons.
What I have learned (so far) has come from working with If all of the dependent variables are to be related to
students and colleagues, from experience (sometimes bit- all of the independent variables by simple bivariate anal-
ter) with journal editors and review committees, and from yses or multiple regression, the number of hypothesis tests
the writings of, among others, Paul Meehl, David Bakan, that will be performed willy-nilly is at least the product
William Rozeboom, Robyn Dawes, Howard Wainer, of the sizes of the two sets. Using the .05 level for many
Robert Rosenthal, and more recently, Gerd Gigerenzer, tests escalates the experimentwise Type I error rateor
Michael Oakes, and Leland Wilkinson. Although they in plain English, greatly increases the chances of discov-
are not always explicitly referenced, many of you will be ering things that aren't so. If, for example, you study 6
able to detect their footprints in what follows. dependent and 10 independent variables and should find
that your harvest yields 6 asterisks, you know full well
Some Things You Learn Aren't So that if there were no real associations in any of the 60
One of the things I learned early on was that some things tests, the chance of getting one or more "significant" re-
you learn aren't so. In graduate school, right after World sults is quite high (something like 1 - .9560, which equals,
War II, I learned that for doctoral dissertations and most coincidentally, .95), and that you would expect three spu-
other purposes, when comparing groups, the proper sam- riously significant results on the average. You then must
ple size is 30 cases per group. The number 30 seems to ask yourself some embarrassing questions, such as, Well,
have arisen from the understanding that with fewer than which three are real?, or even. Is six significant signifi-
30 cases, you were dealing with "small" samples that re- cantly more than the chance-expected three? (It so hap-
quired specialized handling with "small-sample statistics" pens that it isn't.)
instead of the critical-ratio approach we had been taught. And of course, as you've probably discovered, you're
Some of us knew about these exotic small-sample statis- not likely to solve your multiple tests problem with the
ticsin fact, one of my fellow doctoral candidates un- Bonferroni maneuver. Dividing .05 by 60 sets a per-test
dertook a dissertation, the distinguishing feature of which significance criterion of .05/60 = 0.00083, and therefore
was a sample of only 20 cases per group, so that he could a critical two-sided t value of about 3.5. The effects you're
demonstrate his prowess with small-sample statistics. It dealing with may not be large enough to produce any
wasn't until some years later that I discovered (mind you, interesting /s that high, unless you're lucky.
not invented) power analysis, one of whose fruits was the Nor can you find salvation by doing six stepwise
revelation that for a two-independent-group-mean com- multiple regressions on the 10 independent variables. The
parison with n = 30 per group at the sanctified two-tailed amount of capitalization on chance that this entails is
.05 level, the probability that a medium-sized effect would more than I know how to compute, but certainly more
be labeled as significant by the most modern methods (a than would a simple harvest of asterisks for 60 regression
t test) was only .47. Thus, it was approximately a coin coefficients (Wilkinson, 1990, p. 481).
flip whether one would get a significant result, even In short, the results of this humongous study are a
though, in reality, the effect size was meaningful. My n = muddle. There is no solution to your problem. You

1304 December 1990 American Psychologist

Copyright 1990 bv the American Psychological Association. Inc. 0O03-066X/W$00.75
Vol. 45. No. 12. 1304-1312
 wouldn't, of course, write up the study for publication as sand words, in describing a distribution, a frequency
if the unproductive three quarters of your variables never polygon or, better still, a Tukey (1977, pp. 1-26) stem
existed. . . . and leaf diagram is usually worth more than the first four
The irony is that people who do studies like this moments, that is, the mean, standard deviation, skewness,
often start off with some useful central idea that, if pur- and kurtosis. I do not question that the moments effi-
sued modestly by means of a few highly targeted variables ciently summarize the distribution or that they are useful
and hypotheses, would likely produce significant results. in some analytic contexts. Statistics packages eagerly give
These could, if propriety or the consequences of early them to us and we dutifully publish them, but they do
toilet training deemed it necessary, successfully withstand not usually make it possible for most of us or most of the
the challenge of a Bonferroni or other experimentwise- consumers of our products to see the distribution. They
adjusted alpha procedure. don't tell us, for example, that there are no cases between
A special case of the too-many-variables problem scores of 72 and 90, or that this score of 24 is somewhere
arises in multiple regression-correlation analysis with in left field, or that there is a pile-up of scores of 9. These
large numbers of independent variables. As the number are the kinds of features of our data that we surely need
of independent variables increases, the chances are that to know about, and they become immediately evident
their redundancy in regard to criterion relevance also with simple graphic representation.
increases. Because redundancy increases the standard er- Graphic display is even more important in the case
rors of partial regression and correlation coefficients and of bivariate data. Underlying each product-moment cor-
thus reduces their statistical significance, the results are relation coefficient in an acre of such coefficients there
likely to be zilch. lies a simple scatter diagram that the r presumes to sum-
I have so heavily emphasized the desirability of marize, and well it might. That is, it does so if the joint
working with few variables and large sample sizes that distribution is more-or-less bivariate normalwhich
some of my students have spread the rumor that my idea means, among other things, that the relationship must be
of the perfect study is one with 10,000 cases and no vari- linear and that there are no wild outlying points. We know
ables. They go too far. that least squares measures, like means and standard de-
A less profound application of the less-is-more prin- viations, are sensitive to outliers. Well, Pearson correla-
ciple is to our habits of reporting numerical results. There tions are even more so. About 15 years ago, Wainer and
are computer programs that report by default four, five, Thissen (1976) published a data set made up of the heights
or even more decimal places for all numerical results. in inches and weights in pounds of 25 subjects, for which
Their authors might well be excused because, for all the the r was a perfectly reasonable .83. But if an error in
programmer knows, they may be used by atomic scien- transcription were made so that the height and weight
tists. But we social scientists should know better than to values for one of the 25 subjects were switched, the r
report our results to so many places. What, pray, does an would become -.26, a rather large and costly error!
r .12345 mean? or, for an IQ distribution, a mean of There is hardly any excuse for gaps, outliers, cur-
105.6345? For N = 100, the standard error of the r is vilinearity, or other pathology to exist in our data un-
about . 1 and the standard error of the IQ mean about beknownst to us. The same computer statistics package
1.5. Thus, the 345 part of r = .12345 is only 3% of its with which we can do very complicated analyses like
standard error, and the 345 part of the IQ mean of quasi-Newtonian nonlinear estimation and multidimen-
105.6345 is only 2% of its standard error. These super- sional scaling with Guttman's coefficient of alienation
fluous decimal places are no better than random numbers. also can give us simple scatter plots and stem and leaf
\They are actually worse than useless because the clutter diagrams with which we can see our data. A proper mul-
they create, particularly in tables, serves to distract the tiple regression/correlation analysis does not begin with
eye and mind from the necessary comparisons among the a matrix of correlation coefficients, means, and standard
meaningful leading digits. Less is indeed more here. deviations, but rather with a set of stem and leaf diagrams
and scatter plots. We sometimes learn more from what
Simple Is Better we see than from what we compute; sometimes what we
I've also learned that simple is better, which is a kind of learn from what we see is that we shouldn't compute, at
loose generalization of less is more. The simple-is-better least not on those data as they stand.
idea is widely applicable to the representation, analysis, Computers are a blessing, but another of the things
and reporting of data. I have learned is that they are not an unmixed blessing.
If, as the old cliche has it, a picture is worth a thou- Forty years ago, before computers (B.C., that is), for my
doctoral dissertation, I did three factor analyses on the
This invited address was presented to the Division of Evaluation, Mea- I1 subtests of the Wechsler-Bellevue, with samples of 100
surement, and Statistics (Division 5) at the 98th Annual Convention of cases each of psychoneurotic, schizophrenic, and brain-
the American Psychological Association in Boston, August 13, 1990. damaged patients. Working with a pad and pencil, 10-to-
I am grateful for the comments on a draft provided by Patricia the-inch graph paper, a table of products of two-digit
Cohen, Judith Rabkin, Raymond Katzell, and Donald F. Klein.
Correspondence concerning this article should be addressed to Jacob numbers, and a Friden electromechanical desk calculator
Cohen, Department of Psychology, New York University, 6 Washington that did square roots "automatically," the whole process
P!., 5th Floor, New York, NY 10003. took the better part of a year. Nowadays, on a desktop

December 1990 American Psychologist 1305

 computer, the job is done virtually in microseconds (or which unfortunately shrinks to a value smaller than the
at least lickety-split). But another important difference shrunken multiple correlation. For N = 100 cases, using
between then and now is that the sheer laboriousness of Rozeboom's (1978) formula, that comes to .67. Not bad.
the task assured that throughout the entire process I was But using unit weights, we do better: .69. With 300 or
in intimate contact with the data and their analysis. There 400 cases, the increased sampling stability pushes up the
was no chance that there were funny things about my cross-validated correlation, but it remains slightly smaller
data or intermediate results that I didn't know about, than the .69 value for unit weights. Increasing sample
things that could vitiate my conclusions. size to 500 or 600 will increase the cross-validated cor-
I know that I sound my age, but don't get me relation in this example to the point at which it is larger
wrongI love computers and revel in the ease with which than the unit-weighted .69, but only trivially, by a couple
data analysis is accomplished with a good interactive sta- of points in the third decimal*. When sample size is only
tistics package like SYSTAT and SYGRAPH (Wilkinson, 50, the cross-validated multiple correlation is only .63,
1990). I am, however, appalled by the fact that some pub- whereas the unit weighted correlation remains at .69. The
lishers of statistics packages successfully hawk their wares sample size doesn't affect the unit weighted correlation
with the pitch that it isn't necessary to understand sta- because we don't estimate unstable regression coefficients.
tistics to use them. But the same package that makes it It is, of course, subject to sampling error, but so is the
possible for an ignoramus to do a factor analysis with a cross-validated multiple correlation.
pull-down menu and the click of a mouse also can greatly Now, unit weights will not always be as good or better
facilitate with awesome speed and efficiency the perfor- than beta weights. For some relatively rare patterns of
mance of simple and informative analyses. correlation (suppression is one), or when the betas vary
A prime example of the simple-is-better principle is greatly relative to their mean, or when the ratio of sample
found in the compositing of values. We are taught and size to the number of predictors is as much as 30 to 1
teach our students that for purposes of predicting a cri- and the multiple correlation is as large as .75, the beta
terion from a set of predictor variables, assuming for sim- weights may be better, but even in these rare circum-
plicity (and as the mathematicians say, "with no loss of stances, probably not much better.
generality"), that all variables are standardized, we achieve Furthermore, the unit weights work well outside the
maximum linear prediction by doing a multiple regression context of multiple regression where we have criterion
analysis and forming a composite by weighting the pre- datathat is, in a situation in which we wish to measure
dictor z scores by their betas. It can be shown as a math- some concept by combining indicators, or some abstract
ematical necessity that with these betas as weights, the factor generated in a factor analysis. Unit weights on
resulting composite generates a higher correlation with standardized scores are likely to be better for our purposes
the criterion in the sample at hand than does a linear than the factor scores generated by the computer program,
composite formed using any other weights. which are, after all, the fruits of a regression analysis for
Yet as a practical matter, rnpst of the time, we are that sample of the variables on the factor as criterion.
better off using unit weights: +1 for positively related Consider that when we go to predict freshman grade
predictors, - 1 for negatively related predictors, and 0, point average from a 30-item test, we don't do a regression
that is, throw away poorly related predictors (Dawes, analysis to get the "optimal" weights with which to com-
1979; Wainer, 1976). The catch is that the betas come bine the item scoreswe just add them up, like Galton
with guarantees to be better than the unit weights only did. Simple is better.
for the sample on which they were determined. (It's al- We are, however, not applying the simple-is-better
most like a TV set being guaranteed to work only in the principle when we "simplify" a multivalued graduated
store.) But the investigator is not interested in making variable (like IQ, or number of children, or symptom
predictions for that samplehe or she knows the criterion severity) by cutting it somewhere along its span and mak-
values for those cases. The idea is to combine the pre- ing it into a dichotomy. This is sometimes done with a
dictors for maximal prediction for future samples. The profession of modesty about the quality or accuracy of
reason the betas are not likely to be optimal for future the variable, or to "simplify" the analysis. This is not an
samples is that they are likely to have large standard errors. application, but rather a perversion of simple is better,
For the typical 100 or 200 cases and 5 or 10 correlated because this practice is one of willful discarding of infor-
predictors, the unit weights will work as well or better. mation. It has been shown that when you so mutilate a
Let me offer a concrete illustration to help make the variable, you typically reduce its squared correlation with
point clear. A running example in our regression text other variables by about 36% (Cohen, 1983). Don't do
(Cohen & Cohen, 1983) has for a sample of college faculty it. This kind of simplification is of a piece with the practice
their salary estimated from four independent variables: of "simplifying" a factorial design ANOVA by reducing
years since PhD, sex (coded in the modern manner1 all cell sizes to the size of the smallest by dropping cases.
for female and 0 for male), number of publications, and They are both ways of throwing away the most precious
number of citations. The sample multiple correlation commodity we deal with: information.
computes to .70. What we want to estimate is the cor- Rather more generally, I think I have begun to learn
relation we would get if we used the sample beta weights how to use statistics in the social sciences.
in the population, the cross-validated multiple correlation, The atmosphere that characterizes statistics as ap-

1306 December 1990 American Psychologist

 plied in the social and biomedical sciences is that of a (or more often anonymously but as part of his legacy),
secular religion (Salsburg, 1985), apparently of Judeo- but they deserve cataloging (Gigerenzer & Murray. 1987;
Christian derivation, as it employs as its most powerful Oakes, 1986). Over the years, 1 have learned not to make
icon a six-pointed cross, often presented multiply for en- errors of the following kinds:
hanced authority. I confess that I am an agnostic. When a Fisherian null hypothesis is rejected with
an associated probability of, for example. .026, it is not
The Fisherian Legacy the case that the probability that the null hypothesis is
When I began studying statistical inference, I was met true is .026 (or less than .05, or any other value we can
with a surprise shared by many neophytes. I found that specify). Given our framework of probability as long-run
if, for example, I wanted to see whether poor kids esti- relative frequencyas much as we might wish it to be
mated the size of coins to be bigger than did rich kids, otherwisethis result does not tell us about the truth of
after I gathered the data, I couldn't test this research hy- the null hypothesis, given the data. (For this we have to
pothesis, but rather the null hypothesis that poor kids go to Bayesian or likelihood statistics, in which probability
perceived coins to be the same size as did rich kids. This is not relative frequency but degree of belief.) What it
seemed kind of strange and backward to me, but I was tells us is the probability of the data, given the truth of
rather quickly acculturated (or, if you like, converted, or the null hypothesiswhich is not the same thing, as much
perhaps brainwashed) to the Fisherian faith that science as it may sound like it.
proceeds only through inductive inference and that in- If the p value with which we reject the Fisherian null
ductive inference is achieved chiefly by rejecting null hy- hypothesis does not tell us the probability that the null
potheses, usually at the .05 level. (It wasn't until much hypothesis is true, it certainly cannot tell us anything
later that I learned that the philosopher of science, Karl about the probability that the research or alternate hy-
Popper, 1959, advocated the formulation of falsifiable re- pothesis is true. In fact, there is no alternate hypothesis
search hypotheses and designing research that could fal- in Fisher's scheme: Indeed, he violently opposed its in-
sify them.) clusion by Neyman and Pearson.
The fact that Fisher's ideas quickly became the basis Despite widespread misconceptions to the contrary,
for statistical inference in the behavioral sciences is not the rejection of a given null hypothesis gives us no basis
surprisingthey were very attractive. They offered a de- for estimating the probability that a replication of the
terministic scheme, mechanical and objective, indepen- research will again result in rejecting that null hypothesis.
dent of content, and led to clear-cut yes-no decisions. Of course, everyone knows that failure to reject the
For years, nurtured on the psychological statistics text- Fisherian null hypothesis does not warrant the conclusion
books of the 1940s and 1950s, I never dreamed that they that it is true. Fisher certainly knew and emphasized it,
were the source of bitter controversies (Gigerenzer & and our textbooks duly so instruct us. Yet how often do
Murray, 1987). we read in the discussion and conclusions of articles now
Take, for example, the yes-.no decision feature. It appearing in our most prestigious journals that "there is
was quite appropriate to agronomy, which was where no difference" or "no relationship"? (This is 40 years
Fisher came from. The outcome of an experiment can after my TV = 20 friend used a nonsignificant result to
quite properly be the decision to use this rather than that demolish psychoanalytic theory.)
amount of manure or to plant this or that variety of wheat. The other side of this coin is the interpretation that
But we do not deal in manure, at least not knowingly. accompanies results that surmount the .05 barrier and
Similarly, in other technologiesfor example, engineer- achieve the state of grace of "statistical significance."
ing quality control or educationresearch is frequently "Everyone" knows that all this means is that the effect is
designed to produce decisions. However, things are not not nil, and nothing more. Yet how often do we see such
quite so clearly decision-oriented in the development of a result to be taken to mean, at least implicitly, that the
scientific theories. effect is significant, that is, important, large. If a result is
Next, consider the sanctified (and sanctifying) magic highly significant, say p < .001, the temptation to make
.05 level. This basis for decision has played a remarkable this misinterpretation becomes all but irresistible.
role in the social sciences and in the lives of social sci- Let's take a close look at this null hypothesisthe
entists. In governing decisions about the status of null fulcrum of the Fisherian schemethat we so earnestly
hypotheses, it came to determine decisions about the ac- seek to negate. A null hypothesis is any precise statement
ceptance of doctoral dissertations and the granting of re- about a state of affairs in a population, usually the value
search funding, and about publication, promotion, and of a parameter, frequently zero. It is called a "null" hy-
whether to have a baby just now. Its arbitrary unreason- pothesis because the strategy is to nullify it or because it
able tyranny has led to data fudging of varying degrees means "nothing doing." Thus, "The difference in the
of subtlety from grossly altering data to dropping cases mean scores of U.S. men and women on an Attitude To-
where there "must have been" errors. ward the U.N. scale is zero" is a null hypothesis. "The
product-moment r between height and IQ in high school
The Null Hypothesis Tests Us students is zero" is another. "The proportion of men in
We cannot charge R. A. Fisher with all of the sins of the a population of adult dyslexics is .50" is yet another. Each
last half century that have been committed in his name is a precise statementfor example, if the population r

December 1990 American Psychologist 1307

 between height and IQ is in fact .03, the null hypothesis of Biometrika, which Karl Pearson edited. After old-man
that it is zero is false. It is also false if the r is .01, .001, Pearson retired, efforts by Egon Pearson and Neyman to
or .000001! avoid battling with Fisher were to no avail. Fisher wrote
A little thought reveals a fact widely understood that they were like Russians who thought that "pure sci-
among statisticians: The null hypothesis, taken literally ence" should be "geared to technological performance"
(and that's the only way you can take it in formal hy- as "in a five-year plan." He once led off the discussion
pothesis testing), is always false in the real world. It can on a paper by Neyman at the Royal Statistical Society
only be true in the bowels of a computer processor run- by saying that Neyman should have chosen a topic "on
ning a Monte Carlo study (and even then a stray electron which he could speak with authority" (Gigerenzer &
may make it false). If it is false, even to a tiny degree, it Murray, 1987, p. 17). Fisher fiercely condemned the Ney-
must be the case that a large enough sample will produce man-Pearson heresy.
a significant result and lead to its rejection. So if the null I was of course aware of none of this. The statistics
hypothesis is always false, what's the big deal about re- texts on which I was raised and their later editions to
jecting it? which I repeatedly turned in the 1950s and 1960s pre-
Another problem that bothered me was the asym- sented null hypothesis testing a la Fisher as a done deal,
metry of the Fisherian scheme: If your test exceeded a as the way to do statistical inference. The ideas of Neyman
critical value, you could conclude, subject to the alpha and Pearson were barely or not at all mentioned, or dis-
risk, that your null was false, but if you fell short of that missed as too complicated.
critical value, you couldn't conclude that the null was When I finally stumbled onto power analysis, and
true. In fact, all you could conclude is that you couldn 't managed to overcome the handicap of a background with
conclude that the null was false. In other words, you could no working math beyond high school algebra (to say
hardly conclude anything. nothing of mathematical statistics), it was as if I had died
And yet another problem I had was that if the null and gone to heaven. After I learned what noncentral dis-
were false, it had to be false to some degree. It had to tributions were and figured out that it was important to
make a difference whether the population mean difference decompose noncentrality parameters into their constit-
was 5 or 50, or whether the population correlation was uents of effect size and sample size, I realized that I had
.10 or .30, and this was not taken into account in the a framework for hypothesis testing that had four param-
prevailing method. I had stumbled onto something that eters: the alpha significance criterion, the sample size, the
I learned after awhile was one of the bases of the Ney man- population effect size, and the power of the test. For any
Pearson critique of Fisher's system of statistical induction. statistical test, any one of these was a function of the
In 1928 (when I was in kindergarten), Jerzy Ney man other three. This meant, for example, that for a signifi-
and Karl Pearson's boy Egon began publishing papers cance test of a product-moment correlation, using a two-
that offered a rather different perspective on statistical sided .05 alpha criterion and a sample size of 50 cases,
inference (Neyman & Pearson, 1928a, 1928b). Among if the population correlation is .30, my long-run proba-
other things, they argued that rather than having a single bility of rejecting the null hypothesis and finding the sam-
hypothesis that one either rejected or not, things could ple correlation to be significant was .57, a coin flip. As
be so organized that one could choose between two hy- another example, for the same a = .05 and population r =
potheses, one of which could be the null hypothesis and .30, if I want to have .80 power. I could determine that I
the other an alternate hypothesis. One could attach to needed a sample size of 85.
the precisely denned null an alpha risk, and to the equally Playing with this new toy (and with a small grant
precisely defined alternate hypothesis a beta risk. The from the National Institute of Mental Health) I did what
rejection of the null hypotheses when it was true was an came to be called a meta-analysis of the articles in the
error of the first kind, controlled by the alpha criterion, 1960 volume of the Journal ofAbnormal and Social Psy-
but the failure to reject it when the alternate hypothesis chology (Cohen, 1962). I found, among other things, that
was true was also an error, an error of the second kind, using the nondirectional .05 criterion, the median power
which could be controlled to occur at a rate beta. Thus, to detect a medium effect was .46a rather abysmal re-
given the magnitude of the difference between the null sult. Of course, investigators could not have known how
and the alternate (that is, given the hypothetical popu- underpowered their research was, as their training had
lation effect size), and setting values for alpha and beta, not prepared them to know anything about power, let
one could determine the sample size necessary to meet alone how to use it in research planning. One might think
these conditions. Or, with the effect size, alpha, and the that after 1969, when I published my power handbook
sample size set, one could determine the beta, or its com- that made power analysis as easy as falling off a log, the
plement, the probability of rejecting the null hypothesis, concepts and methods of power analysis would be taken
the power of the test. to the hearts of null hypothesis testers. So one might think.
Now, R. A. Fisher was undoubtedly the greatest sta- (Stay tuned.)
tistician of this century, rightly called "the father of mod- Among the less obvious benefits of power analysis
ern statistics," but he had a blind spot. Moreover, he was was that it made it possible to "prove" null hypotheses.
a stubborn and frequently vicious intellectual opponent. Of course, as I've already noted, everyone knows that one
A feud with Karl Pearson had kept Fisher's papers out can't actually prove null hypotheses. But when an inves-

1308 December 1990 American Psychologist

tigator means to prove a null hypothesis, the point is not UPI dispatch under the headline Children's Height
to demonstrate that the population effect size is, say, zero Linked to Test Scores." The article described a study that
to a million or more decimal places, but rather to show involved nearly 14,000 children 6 to 17 years of age that
that it is of no more than negligible or trivial size (Cohen, reported a definite link between height (age- and sex-ad-
1988, pp. 16-17). Then, from a power analysis at, say, justed) and scores on tests of both intelligence and
a = .05, with power set at, say, .95, so that 0 = .05, also, achievement. The relationship was described as signifi-
the sample size necessary to detect this negligible effect cant, and persisting, even after controlling for other fac-
with .95 probability can be determined. Now if the re- tors, including socioeconomic status, birth order, family
search is carried out using that sample size, and the result size, and physical maturity. The authors noted that the
is not significant, as there had been a .95 chance of de- effect was small, but significant, and that it didn't warrant
tecting this negligible effect, and the effect was not de- giving children growth hormone to make them taller and
tected, the conclusion is justified that no nontrivial effect thus brighter. They speculated that the effect might be
exists, at the P = .05 level. This does, in fact, probabil- due to treating shorter children as less mature, but that
istically prove the intended null hypothesis of no more there were alternative biological explanations.
than a trivially small effect. The reasoning is impeccable,
but when you go to apply it, you discover that it takes Now this was a newspaper story, the fruit of the ever-
enormous sample sizes to do so. For example, if we adopt inquiring mind of a science reporter, not a journal article,
the above parameters for a significance test of a correlation so perhaps it is understandable that there was no effort
coefficient and r = . 10 is taken as a negligible effect size, to deal with the actual size of this small effect. But it got
it requires a sample of almost 1,300 cases. More modest me to wondering about how small this significant rela-
but still reasonable demands for power of course require tionship might be. Well, if we take significant to mean p
smaller sample sizes, but not sufficiently smaller to matter < .001 (in the interest of scientific tough-mindedness), it
for most investigatorseven .80 power to detect a pop- turns out that a correlation of .0278 is significant for
ulation correlation of . 10 requires almost 800 cases. So 14,000 cases. But I've found that when dealing with vari-
it generally takes an impractically large sample size to ables expressed in units whose magnitude we understand,
prove the null hypothesis as I've redefined it; however, the effect size in linear relationships is better compre-
the procedure makes clear what it takes to say or imply hended with regression than with correlation coefficients.
from the failure to reject the null hypothesis that there is So, accepting the authors' implicit causal model, it works
no nontrivial effect. out that raising a child's IQ from 100 to 130 would require
giving the child enough growth hormone to increase his
A salutary effect of power analysis is that it draws or her height by 14 ft (more or less). If the causality goes
one forcibly to consider the magnitude of effects. In psy- the other way, and one wanted to create basketball players,
chology, and especially in soft psychology, under the sway a 4-in. increase in height would require raising the IQ
of the Fisherian scheme, there has been little conscious- about 900 points. Well, they said it was a small effect.
ness of how big things are. The very popular ANOVA (When I later checked the journal article that described
designs yield F ratios, and it is these whose size is of this research, it turned out that the correlation was much
concern. First off is the question of whether they made larger than .0278. It was actually about . 11, so that for a
the sanctifying .05 cut-off and are thus significant, and 30-point increase in IQ it would take only enough growth
then how far they fell below this cut-off: Were they perhaps hormone to produce a 3.5-ft increase in height, or with
highly significant (p less than .01) or very highly significant the causality reversed, a 4-in. increase in height would
(less than .001)? Because science is inevitably about mag- require an increase of only 233 IQ points.)
nitudes, it is not surprising how frequently p values are I am happy to say that the long neglect of attention
treated as surrogates for effect sizes. to effect size seems to be coming to a close. The clumsy
One of the things that drew me early to correlation and fundamentally invalid box-score method of literature
analysis was that it yielded an r, a measure of effect size, review based on p values is being replaced by effect-size-
which was then translated into a t or F and assessed for based meta-analysis as formulated by Gene Glass (1977).
significance, whereas the analysis of variance or covari- The effect size measure most often used is the standard-
ance yielded only an F and told me nothing about effect ized mean difference d of power analysis. Several book-
size. As many of the variables with which we worked length treatments of meta-analysis have been published,
were expressed in arbitrary units (points on a scale, trials and applications to various fields of psychology are ap-
to learn a maze), and the Fisherian scheme seemed quite pearing in substantial numbers in the Psychological Bul-
complete by itself and made no demands on us to think letin and other prestigious publications. In the typical
about effect sizes, we simply had no language with which meta-analysis, the research literature on some issue is
to address them. surveyed and the effect sizes that were found in the rel-
In retrospect, it seems to me simultaneously quite evant studies are gathered. Note that the observational
understandable yet also ridiculous to try to develop the- unit is the study. These data do not only provide an es-
ories about human behavior with p values from Fisherian timate of the level and variability of the effect size in a
hypothesis testing and no more than a primitive sense of domain based on multiple studies and therefore on many
effect size. And I wish I were talking about the long, long observations, but by relating effect size to various sub-
ago. In 1986, there appeared in the New York Times a stantive and methodological characteristics over the stud-

December 1990 American Psychologist 1309

 ies, much can be learned about the issue under investi- lead you to rethink and possibly revise your original
gation and how best to investigate it. One hopes that this specifications. This process ends when you have a credible
ferment may persuade researchers to explicitly report ef- and viable set of specifications, or when you discover that
fect sizes and thus reduce the burden on meta-analysts no practicable set is possible and the research as originally
and others of having to make assumptions to dig them conceived must be abandoned. Although you would
out of their inadequately reported research results. In a hardly expect it from reading the current literature, failure
field as scattered (not to say anarchic) as ours, meta-anal- to subject your research plans to power analysis is simply
ysis constitutes a welcome force toward the cumulation irrational.
of knowledge. Meta-analysis makes me very happy. Next, I have learned and taught that the primary
Despite my career-long identification with statistical product of a research inquiry is one or more measures
inference, I believe, together with such luminaries as of effect size, not p values (Cohen. 1965). Effect-size mea-
Meehl (1978) Tukey (1977), and Gigerenzer (Gigerenzer sures include mean differences (raw or standardized),
& Murray, 1987), that hypothesis testing has been greatly correlations and squared correlation of all kinds, odds
overemphasized in psychology and in the other disciplines ratios, kappaswhatever conveys the magnitude of the
that use it. It has diverted our attention from crucial is- phenomenon of interest appropriate to the research con-
sues. Mesmerized by a single all-purpose, mechanized, text. If, for example, you are comparing groups on a vari-
"objective" ritual in which we convert numbers into other able measured in units that are well understood by your
numbers and get a yes-no answer, we have come to neglect readers (IQ points, or dollars, or number of children, or
close scrutiny of where the numbers came from. Recall months of survival), mean differences are excellent mea-
that in his delightful parable about averaging the numbers sures of effect size. When this isn't the case, and it isn't
on football jerseys, Lord (1953) pointed out that "the the case more often than it is. the results can be translated
numbers don't know where they came from." But surely into standardized mean differences (d values) or some
we must know where they came from and should be far measure of correlation or association (Cohen, 1988). (Not
more concerned with why and what and how well we are that we understand as well as we should the meaning of
measuring, manipulating conditions, and selecting our a given level of correlation [Oakes, 1986, pp. 88-92]. It
samples. . has been shown that psychologists typically overestimate
We have also lost sight of the fact that the error vari- how much relationship a given correlation represents,
ance in our observations should challenge us to efforts to thinking of a correlation of .50 not as its square of .25
reduce it and not simply to thoughtlessly tuck it into the that its proportion of variance represents, but more like
denominator of an F o r t test. its cube root of about .80, which represents only wishful
thinking! But that's another story.)
How To Use Statistics Then, having found the sample effect size, you can
So, how would I use statistics in psychological research? attach a p value to it, but it is far more informative to
First of all, descriptively. John Tuke"y's (1977) Exploratory provide a confidence interval. As you know, a confidence
Data Analysis is an inspiring account of how to effect interval gives the range of values of the effect-size index
graphic and numerical analyses of the data at hand so as that includes the population value with a given proba-
to understand them. The techniques, although subtle in bility. It tells you incidentally whether the effect is sig-
conception, are simple in application, requiring no more nificant, but much moreit provides an estimate of the
than pencil and paper (Tukey says if you have a hand- range of values it might have, surely a useful piece of
held calculator, fine). Although he recognizes the impor- knowledge in a science that presumes to be quantitative.
tance of what he calls confirmation (statistical inference), (By the way, I don't think that we should routinely use
he manages to fill 700 pages with techniques of "mere" 95% intervals: Our interests are often better served by
description, pointing out in the preface that the emphasis more tolerant 80% intervals.)
on inference in modern statistics has resulted in a loss of Remember that throughout the process in which you
flexibility in data analysis. conceive, plan, execute, and write up a research, it is on
Then, in planning research, I think it wise to plan your informed judgment as a scientist that you must rely,
the research. This means making tentative informed and this holds as much for the statistical aspects of the
judgments about, among many other things, the size of work as it does for all the others. This means that your
the population effect or effects you're chasing, the level informed judgment governs the setting of the parameters
of alpha risk you want to take (conveniently, but not nec- involved in the planning (alpha, beta, population effect
essarily .05), and the power you want (usually some rel- size, sample size, confidence interval), and that informed
atively large value like .80). These specified, it is a simple judgment also governs the conclusions you will draw.
matter to determine the sample size you need. It is then In his brilliant analysis of what he called the "infer-
a good idea to rethink your specifications. If, as is often ence revolution" in psychology, Gerd Gigerenzer showed
the case, this sample size is beyond your resources, con- how and why no single royal road of drawing conclusions
sider the possibility of reducing your power demand or, from data is possible, and particularly not one that does
perhaps the effect size, or even (heaven help us) increasing not strongly depend on the substantive issues concerned
your alpha level. Or, the required sample may be smaller that is, on everything that went into the research besides
than you can comfortably manage, which also should the number crunching. An essential ingredient in the re-

1310 December 1990 American Psychologist

 search process is the judgment of the scientist. He or she The problem is that, as practiced, current research
must decide by how much a theoretical proposition has hardly reflects much attention to power. How often have
been advanced by the data, just as he or she decided what you seen any mention of power in the journals \ou read,
to study, what data to get, and how to get it. I believe that let alone an actual power analysis in the methods sections
statistical inference applied with informed judgment is a of the articles? Last year in Psychological Bulletin, Sedl-
useful tool in this process, but it isn't the most important meier and Gigerenzer (1989) published an article entitled
tool: It is not as important as everything that came before "Do Studies of Statistical Power Have an Effect on the
it. Some scientists, physicists for example, manage without Power of Studies?". The answer was no. Using the same
the statistics, although to be sure not without the informed methods I had used on the articles in the 1960 Journal
judgment. Indeed, some pretty good psychologists have of Abnormal and Social Psychology (Cohen. 1962), they
managed without statistical inference: There come to performed a power analysis on the 1984 Journal of Ab-
mind Wundt, Kohler, Piaget, Lewin, Bartlett, Stevens, normal Psychology and found that the median power un-
and if you'll permit me, Freud, among others. Indeed, der the same conditions was .44. a little worse than the
Skinner (1957) thought of dedicating his book Verbal Be- .46 I had found 24 years earlier. It was worse still (.37)
havior (and I quote) "to the statisticians and scientific when they took into account the occasional use of an
methodologists with whose help this book would never experimentwise alpha criterion. Even worse than that, in
have been completed" (p. 111). I submit that the proper some 11% of the studies, research hypotheses were framed
application of statistics by sensible statistical methodol- as null hypotheses and their nonsignificance interpreted
ogists (Tukey, for example) would not have hurt Skinner's as confirmation. The median power of these studies to
work. It might even have done it some good. detect a medium effect at the two-tailed .05 level was .25!
These are not isolated results: Rossi. Rossi, and Cottrill
The implications of the things I have learned (so far) (in press), using the same methods, did a power survey
are not consonant with much of what I see about me as of the 142 articles in the 1982 volumes of the Journal of
standard statistical practice. The prevailing yes-no de- Personality and Social Psychology and the Journal of Ab-
cision at the magic .05 level from a single research is a normal Psychology and found essentially the same results.
far cry from the use of informed judgment. Science simply
doesn't work that way. A successful piece of research A less egregious example of the inertia of method-
doesn't conclusively settle an issue, it just makes some ological advance is set correlation, which is a highly flex-
theoretical proposition to some degree more likely. Only ible realization of the multivariate general linear model.
successful future replication in the same and different I published it in an article in 1982, and we included it in
settings (as might be found through meta-analysis) pro- an appendix in the 1983 edition of our regression text
vides an approach to settling the issue. How much more (Cohen, 1982; Cohen & Cohen, 1983). Set correlation
likely this single research makes the proposition depends can be viewed as a generalization of multiple correlation
on many things, but not on whether p is equal to or greater to the multivariate case, and with it you can study the
than .05: .05 is not a cliff but a convenient reference point relationship between anything and anything else, con-
along the possibility-probability continuum. There is no trolling for whatever you want in either the anything or
ontological basis for dichotomous decision making in the anything else, or both. I think it's a great method: at
psychological inquiry. The point was neatly made by least, my usually critical colleagues haven't complained.
Rosnow and Rosenthal (1989) last year in the American Yet, as far as I'm aware, it has hardly been used outside
Psychologist. They wrote "surely, God loves the .06 nearly the family. (The publication of a program as a SYSTAT
as much as the .05" (p. 1277). To which I say amen! supplementary module [Cohen. 1989] may make a dif-
Finally, I have learned, but not easily, that things ference.)
take time. As I've already mentioned, almost three de- But I do not despair. I remember that W. S. Gosset,
cades ago, I published a power survey of the articles in the fellow who worked in a brewery and appeared in print
the 1960 volume of the Journal of Abnormal and Social modestly as "Student," published the t test a decade before
Psychology (Cohen, 1962) in which I found that the me- we entered World War I, and the test didn't get into the
dian power to detect a medium effect size under repre- psychological statistics textbooks until after World
sentative conditions was only .46. The first edition of my War II.
power handbook came out in 1969. Since then, more These things take time. So, if you publish something
than two dozen power and effect-size surveys have been that you think is really good, and a year or a decade or
published in psychology and related fields (Cohen, 1988, two go by and hardly anyone seems to have taken notice,
pp. xi-xii). There have also been a slew of articles on remember the / test, and take heart.
power-analytic methodology. Statistics textbooks, even
some undergraduate ones, give some space to power anal- REFERENCES
ysis, and several computer programs for power analysis Borenstein, M., & Cohen, J. (1988). Statistical power analysis: A com-
are available (e.g., Borenstein & Cohen, 1988). They tell puter program. Hillsdale. NJ: Erlbaum
me that some major funding entities require that their Children's height linked to test scores. (October 7. 1986). Ai'ir York
grant applications contain power analyses, and that in Times, p. C4.
Cohen. J. (1962). The statistical power of abnormal-social psychological
one of those agencies my power book can be found in research: A review. Journal of Abnormal and Social Psychology. 65.
every office. 145-153.

December 1990 American Psychologist 1311

Cohen. J. (1965). Some statistical issues in psychological research. In certain test criteria for purposes of statistical inference: Part II. Bin-
B. B. Wolman (Ed.). Handbook of clinical psychology (pp. 95-121). metrika, 20A. 263-294.
New York: McGraw-Hill. Oakes. M. (1986). Statistical inference: A commentary for the social and
Cohen, J. (1982). Set correlation as a general multivariate data-analytic behavioral sciences. New York: Wiley.
method. Multivariate Behavioral Research, 17, 301-341. Popper. K. (1959). The logic of scientific discovery. New York: Basic
Cohen, J. (1983). The cost of dichotomization. Applied Psychological Books.
Measurement, 7, 249-253. Rosnow. R. L.. & Rosenthal. R. (1989). Statistical procedures and the
Cohen. J. (1988). Statistical power analysis for the behavioral sciences justification of knowledge in psychological science. American Psy-
(2nd ed.). Hillsdale, NJ: Erlbaum. chologist. 44. 1276-1284.
Cohen, J. (1989). SETCOR: Set correlation analysis, a supplementary Rossi, J. S., Rossi, S. R., & Cottrill. S. D. (in press). Statistical power in
module for SYSTAT and SYGR.4PH. Evanston, IL: SYSTAT. research in social and abnormal psychology. Journal of Consulting
Cohen, J.. & Cohen, P. (1983). Applied multiple regression/correlation and Clinical Psychology.
analysis for the behavioral sciences (2nd ed.). Hillsdale, NJ: Erlbaum. Rozeboom, W. W. (1978). Estimation of cross-validated multiple cor-
Dawes, R. M. (1979). The robust beauty of improper linear models in relation: A clarification. Psychological Bulletin. 85, 1348-1351.
decision making. American Psychologist, 34. 571-582. Salsburg, D. S. (1985). The religion of statistics as practiced in medical
Gigerenzer, G.. & Murray, D. J. (1987). Cognition as intuitive statistics. journals. The American Statistician, 39, 220-223.
Hillsdale. NJ: Erlbaum. Sedlmeien P., & Gigerenzer. G. (1989). Do studies of statistical power
Glass, G. V. (1977). Integrating findings: The meta-analysis of research. have an effect on the power of studies? Psnhological Bulletin, 105,
In L. Shulman (Ed.), Review of research in education (Vol. 5, pp. 309-316.
351-379). Itasca. IL: Peacock. Skinner, B. F. (1957). Verbal behavior. New York: Appleton-Century-
Lord, F. M. (1953). On the statistical treatment of football numbers. Crofts.
American Psychologist, 8, 750-751. Tukey, J. W. (1977). Exploratory data analysts. Reading. MA: Addison-
Meehl, P. E. (1978). Theoretical risks and tabular asterisks: Sir Karl, Sir Wesley.
Ronald, and the slow progress of soft psychology. Journal ofConsulting Wainer. H. (1976). Estimating coefficients in linear models: It don't make
and Clinical Psychology, 46, 806-834. no nevermind. Psychological Bulletin, 83. 213-217.
Neyman, J., & Pearson, E. (1928a). On the use and interpretation of Wainer, H., & Thissen, D. (1976). When jackknifing fails (or does it?).
certain test criteria for purposes of statistical inference: Part I. Bio- Psychometrika, 41. 9-34.
metrika, 20A. 175-240. Wilkinson, L. (1990). SYSTAT: The system for statistics. Evanston. IL:
Neyman, J., & Pearson, E. (1928b). On the use and interpretation of SYSTAT.

December 1990 American Psychologist