CIA-RDP96-00792R000100130006-7

Bol
Approved For Release 2000108/10 RCJA-RRP96-00792R000100130006-7
Harvard University
My talk today is designed in part both to comfort the afflicted and to afflict the
comfortable. The afflicted are those of us who work in the softer, wilder areas of our
field -- the areas where the results seem ephemeral and unreplicable, and where the
,2's seem always to be approaching zero as a limit. These softer, wilder areas include
those of social, personality, clinical, developmental, educational, organizational, and
health psychology. They also include parts of psychobiology and cognitive psy-
chology. These softer, wilder areas, however, may not include too much of
psychophysics.
My message to those of us toiling in these muddy vineyards will be that we are
doing better that we might have thought. My message to those of us in any areas in
which we feel we have pretty well nailed things down will be that we haven't, and
that we could be doing a whole lot better.
How Large Must an Effect Be, To Be Important?
There is a bit of good news-bad news abroad in the land. The good news is that
more sophisticated editors, referees, and researchers are becoming aware that
reporting the results of a significance test is not a sufficiently enlighteningprocedure
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

150000
eport or me magirrruac of
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
the'effect accompanying the p level. The bad news is that we are still not quite sure
what to do with such a report of the magnitude of the effect, for example, a
correlation coefficient.
There is one bit of training that all psychologists have undergone. From under-
graduate days onward we have all been taught that there is only one proper, decent
thing to do whenever we see a correlation coefficient -- we must square it. For most of
the softer, wilder areas of psychology, squaring the correlation coefficient tends to
make it go away -- vanish into nothingness as it were. That is one of the sources of
malaise in the social and behavioral sciences. It is sad and quite unnecessary, as we
shall soon see.
The Physician's Aspirin Study
At a special meeting held on December 18, 1987, it was decided to end
prematurely, a randomized double blind experiment on the effects of aspirin on
reducing heart attacks (Steering Committee of the Physicians' Health Study
Research Group, 1988). The reason for this unusual termination of such an experi-
ment was that it had become so clear that aspirin prevented heart attacks (and
deaths from heart attacks) that it would be unethical to continue to give half the
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

of thApproved For Release 2000108/10 - CIA-RDP96-00792R000100130006- This
research? Was ,2 .90, or .80, or .70, or .60, so that the corresponding r's would 'Lave
been .95, .89, .84, or .77? No. Well, was r2 50, .40, .30, or even .20, so that the
corresponding r's would have been .71, .63, .55, or .45? No. Actually, what r2 was,
was.0011, with a corresponding r of .034.
Insert Table 1 about here
Table 1 shows the results of the aspirin study in terms of raw counts, per-
centages, and as a Binomial Effect Size Display (BESD). This display is a way of
showing the practical importance of any effect indexed by a correlation coefficient.
The correlation is shown to be the simple difference in outcome rates between the
experimental and the control groups in this standard table which always adds up to
column totals of 100 and row totals of 100 (Rosenthal & Rubin, 1982b).
This type of result seen in the physicians' aspirin study is not at all unusual in
biomedical research. Some years earlier, on October 29, 1981, the National Heart,
Lung, and Blood Institute discontinued its placebo-controlled study of propranolol
because results were so favorable to the treatment that it would be unethical to
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

continue withholding the life-saving drug from the control patients. And what Was
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7.
the magnitude of this effect? Once again the effect size r was .04, and the leading
digits of the /2 were .00! As behavioral researchers we are not used to thinking of r's
of .04 as reflecting effect sizes of practical importance. But when we think of an r of
.04 as reflecting a 4% decrease in heart attacks, the interpretation given r in a
Binomial Effect Size Display, the r does not appear to be quite so small; especially if
we can count ourselves among the 4 per 100 who manage to survive.
Insert Table 2 about here
Additional Results
Table 2 gives three further examples of Binomial Effect Size Displays. In a
recent study of 4,462 Army veterans of the Vietnam War era (1965-1971), the
correlation between having served in Vietnam (rather than elsewhere) and having
suffered from alcohol abuse or dependence was .07 (Centers for Disease Control,
1988). The top display of Table 2 shows that the difference between the problem
rates of 53.5 and 46.5 per 100 is equal to the correlation coefficient of .07.
The center display of Table 2 shows the results of a study of the effects of AZT
on the survival of 282 patients suffering from AIDS or AIDs-related complex (ARC)
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

AZApprovedl.For)Release 20001084100 GLARDIR96:00792R0001001300067the
clinical trial on the ethical grounds that it would be improper to continue to give
placebo to the control group patients.
As a footnote to this display let me add the result of a small informal poll I took
a few weeks ago of some physicians spending the year at the Center for Advanced
Study in the Behavioral Sciences. I asked them to tell me of some medical break-
through that was of very great practical importance. Their consensus was that the
breakthrough was the effect of cyclosporine in increasing the probability that the
body would not reject an organ transplant and that the recipient patient would not
die. A multi-center randomized experiment was published in 1983 (Canadian
Multicentre Transplant Study Group, 1983). The results of this breakthrough
experiment were less dramatic than the results of the AZT study. For the dependent
variable of organ rejection the effect size r was .19 (+2 = . 036); for the dependent
variable of patient survival the effect size r was .15 (r2 =. 022).
The bottom display of Table 2 shows the results of a famous meta-analysis of
psychotherapy outcome studies reported by Smith and Glass (1977). An eminent
critic (Rimland, 1979) believed that the results of their analysis sounded the "death
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

knell" for psychotherapy because of the modest size of the effect. This modest effect
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
size was an r of .32 accounting for "only 10% of the variance."
Examination of the bottom display of Table 2 shows that it is not very realistic
to label as "modest indeed" an effect size equivalent to increasing a success rate from
34% to 66% (for example, reducing a death rate or a failure rate from 66% to 34%).
Indeed, as we have seen, the dramatic effects of AZT were substantially smaller (r =
.23), and the "breakthrough" effects of cyclosporine were smaller still (r = . 19).
Telling How Well We're Doing
The Binomial Effect Size Display is a useful way to display the practical magni-
tude of an effect size regardless of whether the dependent variable is dichotomous or
continuous (Rosenthal & Rubin, 1982b). An especially useful feature of the display
is how easily we can go from the display to an r (just take the difference between the
success rates of the experimental versus the control group) and how easily we can go
from an effect size r to the display (just compute the treatment success rate as .50
plus one-half of r and the control success rate as .50 minus one-half of r).
One effect of the standard use of a display procedure such as the Binomial
Effect Size Display to index the practical value of our research results would be to
give us more useful and more realistic assessments of how well we are really doing as
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
.

researchers in the social and behavioral sciences. Employment of the Binomial
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
Effect Size Display has, in fact, shown that we are doing considerably better in our
"softer, wilder" sciences than we may have thought we were doing.
So far, our conversation has been intended to comfort the afflicted. In what
follows the intent is a bit more to afflict the comfortable. We begin with the topic of
replication.
The Meaning of Successful Replication
There is a long tradition in psychology of our urging one another to replicate
each other's research. Indeed, there seems to be something nearly scriptural about
it -- I quote: "If a scholar's work be deemed unreplicable then shall ye gladly cast that
scholar out." (That's from either Referees I or Editors II, I believe.)
Now, while we have been very good at calling for replications we have not been
too good at deciding when a replication has been successful. The issue we now
address is: When shall a study be deemed successfully replicated?
Successful replication is ordinarily taken to mean that a null hypothesis that
has been rejected at time 1 is rejected again, and with the same direction of outcome,
on the basis of a new study at time 2. The basic model of this usage can be seen in
Table 3. The results of the first study are described dichotomously as p < . 05 or p >
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

Insert Table 3 about here
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
.05 (or some other critical level, e.g., .01). Each of these two possible outcomes is
further dichotomized as to the results of the second study as p <. 05 or p>.05. Thus,
cells A and D of Table 3 are examples of failure to replicate because one study was
significant and the other was not. Let us examine more closely a specific example of
such a "failure to replicate."
Pseudo-Failures to Replicate
The saga of Smith and Jones. Smith has published the results of an experiment
in which a certain treatment procedure was predicted to improve performance. She
reported results significant at p <. 05 in the predicted direction. Jones publishes a
rebuttal to Smith claiming a failure to replicate.
Insert Table 4 about here
Table 4 shows the results of these two experiments in greater detail. Smith's
results were more significant than Jones's, to be sure, but the studies were in perfect
agreement as to their estimated sizes of effect as defined either by Cohen's d [(Mean,
- Mean 2) / o] or by r, the correlation between group membership and performance
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

score (Cohen, 1977; 1988; Rosenthal, 1984). Not only did the effect sizes of the two
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
studies agree, but even the significance levels of .03 and .30 did not differ very
significantly: (Z 03 -Z 30)/ V2 = (2.17-1.03) / V2 = Z = . 81, p = . 42; for details on
the comparison of significance levels and effect sizes see Rosenthal and Rubin (1979;
1982a) or a summary in Rosenthal (1984). Table 4 shows very clearly that Jones was
very much in error when he claimed that his study failed to replicate that of Smith.
Such errors are made very frequently in most areas of psychology and the other
behavioral and social sciences. The final column of Table 4 shows that the combined
result of both experiments is associated with a more significant t and with a smaller
confidence interval (for the difference between the means and for the effect size r)
than is either of the individual studies.
On the odds against replicating significant results. A related error often found
in the behavioral and social sciences is the implicit assumption that if an effect is
"real," we should therefore expect it to be found significant again upon replication.
Nothing could be further from the truth.
Suppose there is in nature a real effect with a magnitude out there in the world
of d = . 50 (i.e., [Mean, - Mean2] / o = . 50 o units), or, equivalently, r = . 24 (a
difference in success rate of 62% versus 38%). Then suppose an investigator studies
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

MAIS CHICCO WERE CITY OF OF SUBJECTS Of SO, giving the researcher a fever of Statistical
power'S Sveg tor. Release 2000/08/10 : CJA-RDP96-00792R0001001300963730
common Tevel of power for behavioral researchers of
years (Cohen, 1962; Seulmeier & Gigerenzer, in press). Even though a d of .50 or an
r of .24 is a very important effect (as we saw earlier in this paper), there is only one
chance in four that both the original investigator and a replicator will get results
significant at the .05 level. If there were two replications of the original study there
would be only one chance in eight that all three studies would be significant, even
though we know the effect in nature is very real and very important.
If five studies investigated this phenomenon, there is only a 50:50 chance that
three or more of them would find significant results. In short, given the levels of
statistical power at which we normally operate, we have no right to expect the
proportion of significant results that we typically do expect, even if in nature there is
a very real and very important effect.
Pseudo-Successful Replications
Returning now to Table 3, we focus attention on cell B, the cell of "successful
replication." Suppose that two investigators both rejected the null hypothesis at
p <. 05 with both results in the same direction. Suppose further, however, that in one
study the effect size r was .90 while in the other study the effect size r was only .10,
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

significantly smaller than the r of .90 (Rosenthal & Rubin, 1982a). In this case our
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
interpretation is more complex. We have indeed had a successful replication of the
rejection of the null hypothesis but we have not come even close to a successful
replication of the effect size.
"Successful Replication" of Type 2 Error
Cell C of Table 3 represents the situation in which both studies failed to reject
the null hypothesis. Under those conditions investigators might conclude that there
was no relationship between the variables investigated. Such a conclusion could be
very much in error, the more so as the power of the two studies was low (Cohen, 1977;
1988; Rosenthal, 1986). If power levels of the two studies (assuming medium effect
sizes in the population) were very high, say .90 or .95, then two failures to obtain a
significant relationship would provide evidence that the effect investigated was not
likely to be a very large effect. If power calculations had been made assuming a very
small effect size, two failures to reject the null while not providing strong evidence
for the null would at least suggest that the size of the effect in the population was
probably quite modest.
If sample sizes of the two studies failing to reject the null were small so that
power to detect all but the largest effects was low, very little could be concluded from
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

two failures to reject except that the effect sizes were unlikely to be enormous. For
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
example, two investigators with N's of 20 and 40, respectively, find results not
significant at p <. 05. The effect sizes phi (i.e., « for dichotomous variables) were .29
and .20, respectively, and both p's were approximately .20. The combined p of these
two results, however, is .035 [(Z, + Z.) / V2 = Z], and the mean effect size in the
mid- . 20's is not trivial, as we saw earlier in this paper.
Contrasting Views of Replication
The traditional, not very useful view of replication modeled in Table 3 has two
primary characteristics:
(1) It focuses on significance level as the relevant summary statistic of a study,
and
(2) It makes its evaluation of whether replication has been successful in a
dichotomous fashion. For example, replications are successful if both or neither
p <. 05 (or .01, etc.), and they are unsuccessful if one p <. 05 (or .01, etc.) and the other
p>.05 (or .01, etc.). Psychologists' reliance on a dichotomous decision procedure
accompanied by an untenable discontinuity of credibility in results varying in p
levels has been well documented (Nelson, Rosenthal, & Rosnow, 1986; Rosenthal &
Gaito, 1963, 1964).
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

The newer, more useful views of replication success have two primary
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
characteristics:
1. A focus on effect size as the more important summary statistic of a study
with only a relatively minor interest in the statistical significance level, and
2.
An evaluation of whether replication has been successful made in a con-
tinuous fashion. For example, two studies are not said to be successful or unsuccess-
ful replicates of each other, but rather the degree of failure to replicate is specified.
Insert Table 5 about here
Table 5 shows three sets of replications. Replication set A shows two results
both rejecting the null but with a difference in effect sizes of .30 in units of r or .35 in
units of Fisher's Z transformation of r (Cohen, 1977, 1988; Rosenthal & Rosnow,
1984; Snedecor & Cochran, 1980). That difference, in units of r or Fisher's Z is the
degree of failure to replicate. The fact that both studies were able to reject the null
and at exactly the same p level is simply a function of sample size. Replication set B
shows two studies with different p values, one significant at <. 05, the other not
significant. However, the two effect size estimates are in excellent agreement. We
would say, accordingly, that replication set B shows more successful replication than
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

does replication set A. Replication set C shows two studies differing markedly in
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
both level of significance and magnitude (and direction) of effect size. Replication set
C, then, is a not very subtle example of a clear failure to replicate.
It should be noted that the values of Table 5 were chosen so that the combined
probability of the two studies of sets A, B, and C would all be identical to one
another; (Z1 + Z2) / V2 = Z of 2.77. p =. 0028, one-tailed.
Some Metrics of the Success of Replication
Once we adopt a view of the success of replication as a function of similarity of
effect sizes obtained, we can become more precise in our assessments of the success of
replication.
Insert Figure 1 about here
The replication diagonal. Figure 1 shows the "replication plane" generated by
crossing the results of the first study conducted (expressed in units of the effect size r)
by the results of the second study conducted. All perfect replications, those in which
the effect sizes are identical in the two studies, fall on a diagonal rising from the
lower left corner (-1.00, -1.00) to the upper right corner (+1.00, + 1.00). The results
of replication set B from Table 5 are shown to fall exactly on the diagonal of perfect
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

replication (+.26, +.26). The results of replication set A are shown to fallsomewhat
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
above the line representing perfect replication. Figure 1 shows that although set B
reflects more successful replication than set A, the 'atter is also located fairly close to
the line and is, therefore, a fairly successful replication set as well. The results of
replication set C, however, are shown to fall rather far from the diagonal of perfect
replication.
Cohen's q. An alternative to the indexing of the success of replication by the
difference between obtained effect size r's is to transform the r's to Fisher's Z's before
taking the difference. Fisher's Z metric is distributed nearly normally and can thus
be used in setting confidence intervals and testing hypotheses about r's, whereas r's
distribution is skewed and the more so as the population value of r moves further
from zero. Cohen's q is especially useful for testing the significance of difference
between two obtained effect size r's. This is accomplished by means of the fact that
N -3
1
+
1
N2-3
is distributed as Z, the standard normal deviate (Rosenthal, 1984; Rosenthal &
Rubin, 1982a, Snedecor & Cochran, 1980). When there are more than two effect size
r's to be evaluated for their variability (i.e., heterogeneity) we can simply compute
the standard deviation (S) among the r's or their Fisher Z equivalents. If a test of
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

significance of heterogeneity of these Fisher Z's is desired, the three references above
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
all provide the appropriate formula for computing the X2 test of heterogeneity as
does Hedges (1982).
Meta-analytic metrics. As the number of replications for a given research
question grows, a full assessment of the success of the replicational effort requires
the application of meta-analytic procedures. An informative but slightly unwieldy
summary of the meta-analysis might be the stem-and-leaf display of the effect sizes
found in the meta-analysis (Tukey, 1977). A more compact summary of the effect
sizes might be Tukey's (1977) box plot, which gives the highest and lowest obtained
effect sizes along with those found at the 25th, 50th, and 75th percentiles. For single
index values of the consistency of the effect sizes, one could employ (a) the range of
effect sizes found between the 75th (Q3) and 25th (Q1) percentile, (b) some standard
fraction of that range (e.g., half or three-quarters), (c) S, the standard deviation of
the effect sizes, or (d) SE, the standard error of the effect sizes.
As a slightly more complex index of the stability, replicability, or clarity of the
average effect size found in the set of replicates, one could employ the mean effect
size divided either by its standard error (S/Vk where k is the total number of
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

replicates), or simply by S. The latter index of mean effect size divided by its
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
standard deviation (S) is the reciprocal of the coefficient of variation or a kind of
coefficient of robustness.
The coefficient of robustness of replication. Although the standard error of the
mean effect size along with confidence intervals placed around the mean effect size
are of great value (Rosenthal & Rubin, 1978), it will sometimes be useful to employ a
robustness coefficient that does not increase simply as a function of the increasing
number of replications. Thus, if we want to compare two research areas for their
robustness, adjusting for the difference in number of replications in each research
area, we may prefer the robustness coefficient defined as the reciprocal of the
coefficient of variation.
The utility of this coefficient is based on two ideas -- first, that replication
success, clarity, or robustness depends on the homogeneity of the obtained effect size;
and second, that it depends also on the unambiguity or clarity of the directionality of
the result. Thus, a set of replications grows in robustness as the variance of the
effect sizes decreases and as the distance of the mean effect size from zero increases.
Incidentally, the mean may be weighted, unweighted, or trimmed (Tukey, 1977).
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

Indeed, it need not be the mean at all but any measure of location or central
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
tendency (e.g., the median).
Insert Tables 6 & 7 about here
Table 6 has been prepared to give some feel for the practical meaning of several
degrees of variability (S) for seven sets of five replicates each, assuming a mean
effect size of zero. For our effect size indicator we have employed the Fisher Z trans-
formation of the correlation coefficient r. When the range of the five Zr's is only from
-. 02 to +.02, S =. 016; when the range is from -1.00 to +1.00, S =. 791. Table 7
shows the replication robustness coefficients for each of the seven degrees of
variability (S) for each of four levels of mean effect size (Zr): . 10, .30, .50, and .70.
There are no intrinsic meanings to any particular robustness coefficients.
Instead, they are intended to be used to compare different research domains for their
replicational robustness in a merely heuristic way.
What Should Be Reported?
If we are to take seriously our newer view of the meaning of the success of
replications, what should be reported by authors of papers seen to be replications of
earlier studies? Clearly, reporting the results of tests of significance will not be
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
1

sufficient. The effect size of the replication and of the original study must be
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
reported. It is not crucial which particular effect size is employed, but the same
effect size should be reported for the replication and the original study. Complete
discussions of various effect sizes and when they are useful are available from Cohen
(1977, 1988) and elsewhere (e.g., Rosenthal, 1984). If the original study and its
replication are reported in different effect size units these can usually be translated
to one another (Cohen, 1977, 1988; Rosenthal, 1984; Rosenthal & Rosnow, 1984;
Rosenthal & Rubin, in press).
Especially if the results of either the original study or its replication were not
significant, the statistical power at which the test of significance was made (assum-
ing, for example, a population effect size equivalent to the effect size actually
obtained) should be reported (Cohen, 1988). In addition to reporting the statistical
power for each study separately, it would be valuable to report the overall
probability that both studies would have yielded significant results given, for
example, the effect size estimated from the results of the original and the replication
study combined.
As an illustration of this procedure, consider the data of Table 4. Employing
Cohen's power tables tells us that given an effect size of d =. 50, Smith's power to
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

reject at p =. 05, two-tailed was .60 while Jones's power was .18. Table 8 shows that
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
given these two levels of power there were only 11 chances in a hundred that both
studies would reject the null hypothesis given the effect size d= . 50. Indeed, the odds
were three times greater (p =. 33) that neither study would reject the null hypothesis
than that both would reject!
Insert Table 8 about here
Such results are not at all unusual. It has often been documented that behav-
ioral researchers are far fonder of making type II errors than of making type I errors
(Cohen, 1962, 1988; Rosenthal & Rosnow, in press; Rosenthal & Rubin, 1985;
Sedlmeier & Gigerenzer, in press). It has been suggested that it is part of our Judeo-
Christian-Shinto tradition that we be deeply troubled that somewhere out there
someone might be having a good time, could be getting a free ride, a significant
result they don't deserve, an .05 asterisk that was actually intended for someone
else.
.
A marvelous suggestion has been made by Donald Rubin that would go a long
way toward helping us get over our problem with the relative risks of type II versus
type I errors. Don has suggested that whenever we conclude that there is "no effect"
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

we report the effect size along with that confidence interval around the effect size
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
that ranges from the effect size of zero to the equally likely effect size greater than
the one we obtained.
To return to Table 4, the "failure to replicate" by Jones provides a good
example. Jones did not reject the null but obtained an effect size of d = . 50. If Jones
had been required to report that his d of.50 was just as close to a d of 1.00 as it was to
a d of zero, Jones would have been less likely to draw his wrong conclusion.
Meta-Analytic Procedures: An Evaluation
Of course it was bound to happen. No discussion of replication and of the
evaluation of the success of a particular replication could long avoid a more formal
consideration of meta-analytic procedures.
In the years 1980, 1981, and 1982 alone, well over 300 papers were published
on the topic of meta-analysis (Lamb and Whitla, 1983). Does this represent a giant
stride forward in the development of the behavioral and social sciences or does it
signal a lemming-like flight to disaster? Judging from reactions to past meta-
analytic enterprises, there are at least some who take the more pessimistic view.
Some three dozen scholars were invited to respond to a meta-analysis of studies of
interpersonal expectancy effects conducted by Don Rubin and myself (Rosenthal &
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

Rubin, 1978). Although much of the commentary dealt with the substantive topic of
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
interpersonal expectancy effects, a good deal of it dealt with methodological aspects
of meta-analytic procedures and products. Some of the criticisms offered were
accurately anticipated by Glass (1978) who had earlier received commentary on his
meta-analytic work (Glass, 1976) and that of his colleagues (Smith & Glass, 1977;
Glass, McGaw, & Smith, 1981). In the present discussion, the criticisms of our
commentators are grouped into several conceptual categories, described, and
discussed.
Sampling Bias and the File Drawer Problem
This criticism holds that there is a retrievability bias such that studies
retrieved do not reflect the population of studies conducted. One version of this
criticism is that the probability of publication is increased by the statistical signifi-
cance of the results so that published studies may not be representative of the studies
conducted. This is a well-taken criticism, though it applies equally to more
traditional narrative reviews of the literature. Procedures that can be employed to
address this problem have been described elsewhere (Rosenthal, 1979a; 1984,
Chapter 5; Rosenthal & Rubin, 1988).
·
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

Loss of Information
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
Overemphasis on Single Values
The first of two criticisms relevant to information loss notes the danger of
trying to summarize a research domain by a single value such as a mean effect size.
This criticism holds that defining a relationship in nature by a single value leads to
overlooking moderator variables. When meta-analysis is seen as including not only
con.bining effect sizes (and significance levels) but also comparing effect sizes in both
diffuse and, especially, in focused fashion, the force of this criticism is removed
(Rosenthal, 1984, Chapter 4).
Overlooking negative instances. A special case of the criticism under discussion
is that, by emphasizing average values, negative cases are overlooked. There are
several ways in which negative cases can be defined; e.g., p>.05, r=0, r negative, r
significantly negative, and so on. However we may define negative cases, when we
divide the sample of studies into negative and positive cases we have merely
dichotomized an underlying continuum of effect sizes or significance levels, and
accounting for negative cases is simply a special case of finding moderator variables.
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

Glossing over Details
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
Although it is a curate to say that meta-analyses gloss over details, it is
equally as accurate to say that traditional narrative reviews do so, and that data
analysts do so in every study in which any statistics are computed. The act of sum-
marizing requires us to gloss over details. If we describe a nearly normal distribu-
tion of scores by the mean and o we have nearly described the distribution perfectly.
If the distribution is quadrimodal, the mean and o will not do a good job of summar-
izing the data. It is the data analyst's job in the individual study, and the meta-
analyst's job in meta-analysis, to "gloss well.' Providing the reader with all the raw
data of all the studies summarized avoids this criticism but serves no useful review
function. Providing the reader with a stem-and-leaf display of the effect sizes
obtained, along with the results of the diffuse and focused comparisons of effect sizes,
does some glossing, but it does a lot of informing besides.
There is, of course, nothing to prevent the meta-analyst from reading each
study as carefully and assessing it as creatively as might be done by a more
traditional reviewer of a literature. Indeed, we have something of an operational
check on reading articles carefully in the case of meta-analysis. If we do not read the
results carefully, we cannot obtain effect sizes and significance levels. In traditional
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

reviews, results may have been read carefully or not read at all, with the abstract or
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
the discussion section providing "the results" to the more traditional reviewer.
Problems of Heterogeneity
Heterogeneity of Method
The first of two criticisms relevant to problems of heterogeneity notes that
meta-analyses average over studies in which the independent variables, the depen-
dent variables, and the sampling units are not uniform. How can we speak of inter-
personal expectancy effects, meta-analytically, when some of the independent
variables are operationalized by (a) telling experimenters that tasks are easy versus
hard; or by (b) telling experimenters that subjects are good versus poor task per-
formers? How can we speak, meta-analytically, of these expectancy effects when
sometimes the dependent variables are reaction times, sometimes IQ test scores, and
sometimes responses to inkblots? How can we speak of these effects when sometimes
the sampling units are rats, sometimes college sophomores, sometimes patients,
sometimes pupils? Are these not all vastly different phenomena? How can they be
pooled together in a single meta-analysis?
Glass (1978) has eloquently addressed this issue -- the apples and oranges issue.
They are good things to mix, he wrote, when we are trying to generalize to fruit.
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

Indeed, if we are willing to generalize over subjects within studies, why should we
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
not be willing to generalize over studies? If subjects behave very differently within
studies we block on subject characteristics to help us understand why. If studies
yield very different results from each other, we block on study characteristics to help
us understand why. It is very useful to be able to make general statements about
fruit. If, in addition, it is also useful to make general statements about apples, about
oranges, and about the differences between them, there is nothing in meta-analytic
procedures to prevent us from doing so.
Heterogeneity of Quality
One of the most frequent criticisms of meta-analyses is that "bad" studies are
thrown in with good. This criticism must be broken down into two questions: (a)
What is a "bad" study?, and (b) What shall we do about "bad" studies?
-
Defining "bad" studies. Too often, deciding what is a "bad" study is a procedure
richly susceptible to bias or to claims of bias (Fiske, 1978). "Bad" studies are too often
those whose results we don't like, or, as Glass, McGaw, and Smith (1981) have put it,
the studies of our "enemies." Therefore, when reviewers of research tell us they have
omitted the "bad" studies, we should satisfy ourselves that this has been done by
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

1
criteria we find acceptable. A discussion of these criteria (and the computation of
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
their reliability) can be found elsewhere (Rosenthal, 1984, Chapter 3).
Dealing with "bad" studies. The distribution of studies on a dimension of
quality is, of course, not really dichotomous (good versus bad), but continuous with
all possible degrees of quality. The fundamental method of coping with "bad" studies
or, more accurately, variations in the quality of research, is by differential weighting
of studies. Dropping studies is merely the special case of zero weighting.
The most important question to ask relevant to study quality is that asked by
Glass (1976): Is there a relationship between quality of research and effect size
obtained? If there is not, the inclusion of poorer quality studies will have no effect on
the estimate of the average effect size though it will help to decrease the size of the
confidence interval around that mean. If there is a relationship between the quality
of research and effect size obtained, we can employ whatever weighting system we
find reasonable (and that we can persuade our colleagues and critics also to find
reasonable).
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

Problems of Independence
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
Responses Within Studies
The first of two criticisms relevant to problems of independence notes that
several effect size estimates and several tests of significance may be generated by the
same subjects within each study. This can be a very well-taken criticism under some
conditions and the problem has been dealt with elsewhere in some detail (Rosenthal,
1984, Chapter 2; Rosenthal & Rubin, 1986).
Studies Within Sets of Studies
Even when all studies yield only a single effect size estimate and level of
significance, and even when all studies employ sampling units that do not also
appear in other studies, there is a sense in which results may be nonindependent.
That is, studies conducted in the same laboratory, or by the same research group,
may be more similar to each other (in the sense of an intraclass correlation) than
they are to studies conducted in other laboratories or by other research groups (Jung,
1978; Rosenthal, 1966, 1969, 1979b). The conceptual and statistical implications of
this problem are not yet well worked out.
,
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

Exaggeration of Significance Levels
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
Truncating Significance Levels
It has been suggested that all p levels less than .01 (Z values greater than 2.33)
be reported as .01 (Z=2.33) because p's less than .01 are likely to be in error
(Elashoff, 1978). This truncating of Z's cannot be recommended and will, in the long
run, lead to serious errors of inference (Rosenthal & Rubin, 1978). If there is reason
to suspect that a given p level <. 01 is in error it should, of course, be corrected before
employing it in the meta-analysis. It should not, however, be changed to p =. 01
simply because it is less than .01.
Too Many Studies
It has been noted as a criticism of meta-analyses, that, as the number of studies
increases, there is a greater and greater probability of rejecting the null hypothesis
(Mayo, 1978). When the null hypothesis is false and, therefore, ought to be rejected,
it is indeed true that adding observations (either sampling units within studies or
new studies) increases statistical power. However, it is hard to accept as a legitimate
criticism of a procedure, a characteristic that increases its accuracy and decreases its
. error rate -- in this case, type II errors.
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

A related feature of meta-analysis appears to be that it may, in general, lead to
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
a decrease in type II errors even when the number of studies is modest. Empirical
support for this is provided in a study conducted by Cooper and Rosenthal (1980).
Procedures requiring the research reviewer to be more systematic and to use more of
the information in the data seem to be associated with increases in power, i.e.,
decreases in type II errors.
Some Benefits of Meta-Analysis
From what has been said of the various criticisms of meta-analysis it will
surprise no one to learn that I strongly support the increasing use of meta-analytic
procedures. My reasons for that support go beyond the fact that the various
criticisms of meta-analysis can be readily addressed. In the time that remains I
want to note a number of special benefits of meta-analysis. Some of these benefits
are well known, but some are not -- indeed, some are almost secret benefits.
Most Obvious Benefits
Completeness. Meta-analytic consideration of a research domain is more
complete and exhaustive though this does not mean that all studies found are
. weighted equally. Indeed, every study should be weighted from zero to any desired
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

number. These weights, of course, must be defensible. (It will not do to weight all
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
my results + 1.00 and all my enemies' results 0.00).
Explicitness. The quantitative nature of the process of obtaining effect sizes,
standard normal deviates, and weights, forces explicitness on the analyst. Vague
terms like "no relationship," "some relationship," a "strong relationship." "very
significant," are replaced by numerical values.
Power. Empirical work has shown that meta-analytic procedures increase
power and decrease type 2 errors (Cooper & Rosenthal, 1980).
Less Obvious Benefits
Moderator variables. These are more easily spotted and evaluated in a context
of a quantitative research summary. This aids theory development and increases
empirical richness.
Cumulation problems. Meta-analytic procedures address, in part, the chronic
complaint that social sciences cumulate so poorly compared to the physical sciences.
It should be noted that recent historical and sociological investigations have sug-
gested that the physical sciences may not be all that much better off than we are
. when it comes to successful replication (Collins, 1985; Hedges, 1987; Pool 1988). For
example, Collins (1985) has described the failures to replicate the construction of
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

TEA-lasers despite the availability of detailed instructions for replication.
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
Apparently TEA-lasers could be replicated dependably only when the replication
instructions were accompanied by a scientist who had actually built a laser.
Least Obvious Benefits
Decrease in overemphasis on single studies. One not so obvious benefit that will
accrue to us is the gradual decrease in the overemphasis on the results of a single
study. There are good sociological grounds for our monomaniacal preoccupation
with the results of a single study. Those grounds have to do with the reward system
of science where recognition, promotion, reputation, and the like depend on the
results of the single study, also known as the smallest unit of academic currency.
The study is "good," "valuable," and above all, "publishable" when p ≤.05. Our disci-
plines would be further ahead if we adopted a more cumulative view of science in
which the impact of a study were evaluated less on the basis of p levels, and more on
the basis of its own effect size and on the revised effect size and combined probability
that resulted from the addition of the new study to any earlier studies investigating
the same or a similar relationship. This, of course, amounts to a call for a more meta-
·analytic view of "doing science."
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

B. F .Skinner has been eloquent in his comments on the overvaluation of the
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
single study: "In my own thinking, I try to avoid the kind of fraudulent significance
which comes with grandiose terms or profound 'principles.' But some psychologists
seem to need to feel that every experiment they do demands a sweeping reorgan-
ization of psychology as a whole. It's not worth publishing unless it has some such
significance. But research has its own values, and you don't need to cook up spurious
reasons why it's important." (Skinner, 1983, p. 39).
"The new intimacy." This new intimacy is between the reviewer and the data.
We cannot do a meta-analysis by reading abstracts and discussion sections. We are
forced to look at the numbers and, very often, compute the correct ones ourselves.
Meta-analysis requires us to cumulate data, not conclusions. "Reading" a paper is
quite a different matter when we need to compute an effect size and a fairly precise
significance level -- often from a results section that never heard of effect sizes, precise
significance levels (or the APA publication manual)!
The demise of the dichotomous significance testing decision. Far more than is
good for us, social and behavioral scientists operate under a dichotomous null
hypothesis decision procedure in which the evidence is interpreted as anti-null if p ≤
.05 and pro-null if p > .05. If our dissertation p is < . 05 it means joy, a Ph.D., and a
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

tenure-track position at a major university. If our p is > .05 it means ruin, despair,
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
ard our advisor's suddenly thinking of a new control condition that should be run.
That attitude really must go. God loves the .06 nearly as much as the .05. Indeed, I
have it on good authority that she views the strength of evidence for or against the
null as a fairly continuous function of the magnitude of p. As a matter of fact, two.06
results are much stronger evidence against the null than one .05 result; and 10 p's of
.10 are stronger evidence against the null than 5 p's of .05.
The overthrow of the omnibus test. It is common to find specific questions
addressed by F tests with df > 1 in the numerator or by X2 tests with df > 1. For
example, suppose the specific question is whether increased incentive level improves
the productivity of work groups. We employ four levels of incentive so that our
omnibus F test would have 3 df in the numerator or our omnibus X2 would be on at
least 3 df. Common as these tests are, they reflect poorly on our teaching of data
analytic procedures. The diffuse hypothesis tested by these omnibus tests usually
tells us nothing of importance about our research question. The rule of thumb is
unambiguous: Whenever we have tested a fixed effect with df > 1 for X2 or for the
numerator of F, we have tested a question in which we are almost surely not
interested.
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

·
The situation is even worse when there are several dependent variables as well
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
as multiple d'f for the independent variable. The paradigm case here is canonical
correlation and special cases are MANOVA, MANCOVA, Multiple discriminant
function, multiple path analysis, and complex multiple partial correlation. While all
of these procedures have useful exploratory data analytic applications they are
commonly used to test null hypotheses which are scientifically almost always of
doubtful value. The effect size estimates they yield (e.g., the canonical correlation)
are also almost always of doubtful value.
This is not the place to go into detail, but one approach to the problem of
analyzing canonical data structures is to reduce the set of dependent variables to
some smaller number of composite variables using the principal-components-
follow: ed-by-unit-weighting approach. Each composite can then be analyzed serially.
Meta-analytic questions are basically contrast questions. F tests with df> 1 in
the numerator or X2's with df > 1 are useless in meta-analytic work. That leads to
an additional scientific benefit:
The increased recognition of contrast analysis. Meta-analytic questions require
precise formulation of questions and contrasts are procedures for obtaining answers
to such questions, often in an analysis of variance or table analysis context.
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

Although most textbooks of statistics describe the logic and the machinery of
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
contrast analyses, one still sees contrasts employed all too rarely. That is a real pity
given the precision of thought and theory they encourage and (especially relevant to
these times of publication pressure) given the boost in power conferred with the
resulting increase in .05 asterisks (Rosenthal & Rosnow, 1985).
A probable increase in the accurate understanding of interaction effects.
Probably the universally most misinterpreted empirical results in psychology are
the results of interaction effects. A recent survey of 191 research articles involving
interactions found only two articles that showed the authors interpreting inter-
actions in an unequivocally correct manner (i.e., by examining the residuals that
define the interaction) (Rosnow & Rosenthal, 1989). The rest of the articles simply
compared means of conditions with other means, a procedure that does not
investigate interaction effects but rather the sum of main effects and interaction
effects.
Most standard textbooks of statistics for psychologists provide accurate
mathematical definitions of interaction effects but then interpret not the residuals
that define those interactions but the means of cells that are the sums of all main
effects and all interactions.
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

In addition, users of SPSS, SAS, BMDP, and virtually all other data-analytic
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
software are poorly served in the matter of interactions since virtually no programs
provide convenient tabular output giving the residuals defining interaction. The
only exception to that of which I am aware is a little-known package called Data-
Text developed by Arthur Couch and David Armor for which William Cochran and
Donald Rubin provided the statistical consultation.
Since many meta-analytic questions are by nature questions of interaction (for
example, that opposite sex dyads will conduct standard transactions more slowly
than will same sex dyads), we can be hopeful that increased use of meta-analytic
procedures will bring with it increased sophistication about the meaning of
interaction.
Meta-analytic procedures are applicable beyond meta-analyses. Many of the
techniques of contrast analyses among effect sizes, for example, can be used within a
single study (Rosenthal & Rosnow, 1985). Computing a single effect size from
correlated dependent variables, or comparing treatment effects on two or more
dependent variables serve as illustrations (Rosenthal & Rubin, 1986).
. The decrease in the splendid detachment of the full professor. Meta-analytic
work requires careful reading of research and moderate data analytic skills. We
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

cannot send an undergraduate research assistant to the library with a stack of 5× 8
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
cards to bring us back "the results." With narrative reviews that seems often to have
been done. With meta-analysis the reviewer must get involved with the actual data
and that is all to the good.
Conclusion
I hope that this paper has provided some comfort to the afflicted in showing
that many of the findings of our discipline are neither as small nor as unimportant
from a practical point of view as we may have feared. Perhaps I hope, too, that there
may have been some affliction of the comfortable in showing that in our views of
replication and of the cumulation of the wisdom of our field there is much yet
remaining to be done.
.
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

References
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
Barnes, D. M. (1986). Promising results halt trial of anti-AIDS drug. Science, 234,
15-16.
Canadian Multicentre Transplant Study Group. (1983). A randomized clinical trial
of cyclosporine in cadaveric renal transplantation. New England Journal of
Medicine, 309, 809-815.
Centers for Disease Control Vietnam Experience Study. (1988). Health status of
Vietnam veterans: 1. Psychosocial characteristics. Journal of the American
Medical Association, 259, 2701-2707.
Cohen, J. (1962). the statistical power of abnormal-social psychological research: A
review. Journal of Abnormal and Social Psychology, 65, 145-153.
Cohen, J. (1988). Statistical Power Analysis for the Behavioral Sciences, 2nd ed.,
Hillsdale, NJ: Lawrence Erlbaum Associates.
Collins, H. M. (1985). Changing Order: Replication and Induction in Scientific
Practice. Beverly Hills, CA: Sage.
Cooper, H. M., & Rosenthal, R. (1980). Statistical versus traditional procedures for
summarizing research findings. Psychological Bulletin, 87, 442-449.
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

Elashoff, J. D. (1978). Box scores are for baseball. The Behavioral and Brain
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
Sciences, 3, 392.
Fiske, D. W. (1978). The several kinds of generalization. The Behavioral and Brain
Sciences, 3, 393-394.
Glass, G. V. (1976). Primary, secondary, and meta-analysis of research.
Educational Researcher, 5, 3-8.
Glass, G. V. (1978). In defense of generalization. The Behavioral and Brain
Sciences, 3, 394-395.
Glass, G. V., McGaw, B., & Smith, M. L. (1981). Meta-Analysis in Social Research.
Beverly Hills, CA: Sage.
Hedges, L. V. (1982). Estimation of effect size from a series of independent
experiments. Psychological Bulletin, 92, 490-499.
Hedges, L. V. (1987). How hard is hard science, how soft is soft science? American
Psychologist, 42. 443-455.
Jung, J. (1978). Self-negating functions of self-fulfilling prophecies. The Behavioral
and Brain Sciences, 3, 397-398.
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

Lamb, W. K., & Whitla, D. K. (1983). Meta-Analysis and the Integration of Research
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
Findings: A Trend Analysis and Bibliography Prior to 1983. Unpublished
manuscript, Harvard University, Cambridge.
Mayo, R. J. (1978). Statistical considerations in analyzing the results of a collection
of experiments. The Behavioral and Brain Sciences, 3, 400-401.
Nelson, N., Rosenthal, R., & Rosnow, R. L. (1986). Interpretation of significance
levels and effect sizes by psychological researchers. American Psychologist, 41,
1299-1301.
Pool, R. (1988). Similar experiments, dissimilar results. Science, 242, 192-193.
Rimland, B. (1979). Death knell for psychotherapy? American Psychologist, 34,
192.
Rosenthal, R. (1966). Experimenter Effects in. Behavioral Research. New York:
Appleton-Century-Crofts.
Rosenthal, R. (1969). Interpersonal expectations. In R. Rosenthal and R. L. Rosnow
(Eds.), Artifact in Behavioral Research (pp. 181-277). New York: Academic
Press.
Rosenthal, R. (1979a). The "file drawer problem" and tolerance for null results.
.
Psychological Bulletin. 86, 638-641.
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

Rosenthal, R. (1979b). Replications and their relative utilities. Replications in
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
Social Psychology, 1(1), 15-23.
Rosenthal, R. (1984). Meta-Analytic Procedures for Social Research. Beverly Hills,
CA: Sage.
Rosenthal, R. (1986). Nonsignificant relationships as scientific evidence.
Behavioral and Brain Sciences, 9, 479-481.
Rosenthal, R., & Gaito, J. (1963). The interpretation of levels of significance by
psychological researchers. Journal of Psychology, 55, 33-38.
Rosenthal, R., & Gaito, J. (1964). Further evidence for the cliff effect in the
interpretation of levels of significance. Psychological Reports, 15, 570.
Rosenthal, R., & Rosnow, R. L. (1984). Essentials of Behavioral Research: Methods
and Data Analysis. New York: McGraw-Hill.
Rosenthal, R., & Rosnow, R. L. (1985). Contrast Analysis: Focused Comparisons in
the Analysis of Variance. New York: Cambridge University Press.
Rosenthal, R., & Rosnow, R. L. (in press). Essentials of Behavioral Research:
Methods and Data Analysis, 2nd ed., New York: McGraw-Hill.
Rosenthal, R., & Rubin, D. B. (1978). Interpersonal expectancy effects: The first 345
studies. The Behavioral and Brain Sciences, 3, 377-386.
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

Rosenthal, R., & Rubin, D. B. (1979). Comparing significance levels of independent
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
studies. Psychological Bulletin, 86. 1165-1168.
Rosenthal, R., & Rubin, D. B. (1982a). Comparing effect sizes of independent
studies. Psychological Bulletin, 92, 500-504.
Rosenthal, R., & Rubin, D. B. (1982b). A simple, general purpose display of
magnitude of experimental effect. Journal of Educational Psychology, 74, 166-
169.
Rosenthal, R., & Rubin, D. B. (1985). Statistical analysis: Summarizing evidence
versus establishing facts. Psychological Builetin, 97, 527-529.
Rosenthal, R., & Rubin, D. B. (1986). Meta-analytic procedures for combining
studies with multiple effect sizes. Psychological Bulletin, 99, 400-406.
Rosenthal, R., & Rubin, D. B. (1988). Comment: Assumptions and procedures in the
file drawer problem. Statistical Science, 3, 120-125.
Rosenthal, R., & Rubin, D. B. (in press). Effect size estimation for one-sample
multiple-choice-type data: Design, analysis, and meta-analysis. Psychological
Bulletin.
Rosnow, R. L., & Rosenthal, R. (1989). Definition and interpretation of interaction
effects. Psychological Bulletin, 105, 143-146.
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

Sedlmeier, P., & Gigerenzer, G. (in press). Do studies of statistical power have an
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
effect on the power of studies? Psychological Bulletin.
Skinner, B. F. (1983, August). On the value of research. APA Monitor, p. 39.
Smith, M. L., & Glass, G. V (1977). Meta-analysis of psychotherapy outcome
studies. American Psychologist, 32, 752-760.
Snedecor, G. W., & Cochran, W. G. (1980). Statistical Methods (7th ed.). Ames: Iowa
State University Press.
Steering Committee of the Physicians Health Study Research Group. (1988).
Preliminary report: Findings from the aspirin component of the ongoing
physicians' health study. The New England Journal of Medicine, 318, 262-264.
.
Tukey, J. W. (1977). Exploratory Data Analysis. Reading, MA: Addison-Wesley.
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

Author Notes
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
This paper was presented as an EPA Distinguished Lecture at f
Meeting of the Eastern Psychological Association, Boston, April 2, 19/
tion of this paper was supported in part by the National Science Found
the author was a Fellow at the Center for Advanced Study in the Behavioral
Sciences. I am grateful for financial support provided by the John D. and Catherine
T. MacArthur Foundation, and for improvements suggested by Lynn Gale, Deanna
Knickerbocker, Harold Luft, and Lincoln Moses.
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
Effects of Aspirin on Heart Attacks Among 22,000 Physicians
Heart Attack
No Heart
Attack
Total
I. Raw Counts
Aspirin
104
10,933
11,037
Placebo
189
10,845
11,034
Total
293
21,778
22,071
II. Percentages
Aspirin
0.94
99.06
100
Placebo
1.71
98.29
100
Total
1.33
98.67
100
III. Binomial Effect Size Display
Aspirin
48.3
51.7
100
Placebo
51.7
48.3
100
Total
100
100
200
-
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
Other Examples of Binomial Effect Size Displays
I. Vietnam Service and Alcohol Problems (r = . 07)
Problem
No Problem
Total
Vietnam Veteran
53.5
46.5
100
Non-Vietnam Veteran
46.5
53.5
100
Total
100
100
200
II. AZT in the Treatment of AIDS (r = . 23)
Death
Survival
Total
AZT
38.5
61.5
100
Placebo
61.5
38.5
100
Total
100
100
200
III. Benefits of Psychotherapy (r = . 32)
Less Benefit
Greater
Benefit
Total
Psychotherapy
34
66
100
Control
66
34
100
Total
100
100
200
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

TabApproved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
Common Model of Successful Replication: Judgment is Dichotomous and Based on
Significance Testing
First Study
p >.05*
p < . 05
p <. 05ª
A
Failure to
Replicate
B
Successful
Replication
Second
Study
C
p > .05
Failure to
Establish Effect
D
Failure to
Replicate
*By convention .05 but could be any other given level, e.g, .01.
ªIn the same tail as the results of the first study.
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

Table 4
IlApproved Fes Release52000/08/10an6 [As RD290-100792R000100130006-7
Investigator
I. Smith
II. Jones
Combined
Treatment Mean
.38
.36
.376
Control Mean
.26
.24
.256
Difference
.12
.12
.120
t
2.21
1.06
2.45
df
78
18
96
two-tail p
.03
.30
.02
effect size da
.50
.50
.50
effect size rb
.24
.24
.24
standard normal Z
2.17c
1.03€
2.40
95% Confidence intervals
From:
.01
-. 12
.02
Mean differences
To:
.23
.36
.22
From:
.02
-. 23
.04
Effect size r's
To:
.44
.62
.42
ªObtained from 2tV df .
Obtained from Vt2 /(12 + df).
" These significance levels differ at Z = . 81, p =. 42 from
(Z,-Z2)/V2.
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

Table 5
CoApprovedof on Reisase 2009/08/19 % CIA-RDP96-00792R000100130006-7
Replication Sets
B
C
A
Study 1 Study 2
Study 1 Study 2
Study 1
Study 2
N
96
15
98
27
12
32
p (two-tail)
.05
.05
.01
.18
.000001
.33
Z(p)
1.96
1.96
2.58
1.34
4.89
-0.97
r
.20
.50
.26
.26
.72
-. 18
Z(r)
.20
.55
.27
.27
.90
-. 18
Cohen's q (Zr Zr2)
.35
.00
1.08
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

Table 6
SeveApprovedFor Retense(3900:00/10Si&IA-RPR96-09792RQOB100130006-0.00
Degree of Variability
Replicate
Set 1
Set 2
Set 3
Set 4
Set 5
Set 6
Set 7
.02
.10
.20
.40
.60
.80
1.00
.01
.05
.10
.20
.30
.40
.50
.00
.00
.00
.00
.00
.00
.00
-. 01
-. 05
-. 10
-. 20
-. 30
-. 40
-. 50
-. 02
-. 10
-. 20
-. 40
-. 60
-. 80
-1.00
S
.016
.079
.158
.316
.474
.632
.791
Range
.04
.20
.40
.80
1.20
1.60
2.00
Equal Steps of
.01
.05
.10
.20
.30
.40
.50
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

Table 7
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
Replication Robustness Coefficients for Four Levels of Mean Effect
Size (Zr) and Six Degrees of Varicoility of Effect Size (S)
Mean Effect Size (Zr)
S
.10
.30
.50
.70
.016
6.25
18.75
31.25
43.75
.079
1.27
3.80
6.33
8.86
.158
0.63
1.90
3.16
4.43
.316
0.32
0.95
1.58
2.22
.474
0.21
0.63
1.05
1.48
.632
0.16
0.47
0.79
1.11
.791
0.13
0.38
0.63
0.88
.
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7

Table 8
ProbApproved For Release. 2009/08/10ReGIA-RPR96-09782R800199139006-7.
Studies of Table 4
-
Study I: Smith
Probability of Not
Rejecting False Null
(Type II Error Rate =
.40)
Probability of
Rejecting False Null
(Power = . 60)
Σ
Study II: Jones
Probability of
Rejecting False Null
(Power = . 18)
.07
.11
.18
Probability of Not
Rejecting False Null
(Type II Error Rate
= . 82)
.33
.49
.82
Σ
.40
.60
1.00
.
1
.
-
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
-

Figure I
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
The Replication Plane
1.00
.80
.60
SET A
*
.40
-
.20
SET B
RESULTS
.00
OF SECOND
.. 20
SET C
STUDY (r)
-. 40
-. 60
-. 80
-10
1
1
-. 10
-. 80
-. 60
-. 40
.. 20
.00
.20
.40
.60
.80
1.00
Approved For Release 2000/08/10 : CIA-RDP96-00792R000100130006-7
DIAGONAL OF SUCCESSFUL REPLICATION