Document Type: 
Document Number (FOIA) /ESDN (CREST): 
Release Decision: 
Original Classification: 
Document Page Count: 
Document Creation Date: 
November 4, 2016
Document Release Date: 
May 18, 2000
Sequence Number: 
Case Number: 
Content Type: 
PDF icon CIA-RDP96-00789R002200520001-0.pdf842.64 KB
Approved For Release 2000/08/08 :CIA-RDP96-007898002200520001-0 Vol. 19, No. IZ, December 1989 Printed in Belgium Evidence for Consciousness-Related Anomalies in Random Physical Systems Dean I. Rodin' and Roger D. Nelson2 Received May 6, 1988; revised June 12, 1989 Speculations about the role of consciousness in physical systems are frequently observed in the literature concerned with the interpretation of quantum mechanics. While only three experimental investigations can be found on this topic in physics jou_ rRals, more than $OO relevant experiments have been reported in the literature of parapsychology. Awell-defined body of empirical evidence from this domain was reviewed using meta-analytic techniques to assess methodological quality and overall effect size. Results showed effects conforming to chance expectation in control conditions and unequivocal non-chance effects in experimental conditions. This quantitative literature review agrees with the findings of two earlier reviews, suggesting the existence of some form of consciousness-related anomaly in random physical systems. The nature of the relationship between human consciousness and the physical world has intrigued philosophers for millenia. In this century, speculations about mind-body interactions persist, often contributed by physicists in discussions of the measurement problem in quantum mechanics. Virtually all of the founders of quantum theory-Planck, de Broglie, Heisenberg, Schrodinger, Einstein-considered this subject in depth, ~' ~ and contemporary physicists continue this tradition.~2-'~ ~ Department of Psychology, Princeton University, Princeton, New Jersey 08544. Present address: Contel Technology Center, 15000 Conference Center Drive, P.O. Box 10814, Chantilly, Virginia 22021-3808. 2 Department of Mechanical and Aerospace Engineering, Princeton University, Princeton, New Jersey 0$544. 1499 OO15-9018/89/1200-1499f06.00/0 R?^, 1989 Plenum Publishing Corporation Approved For Release 2000/08/08 :CIA-RDP96-007898002200520001-0 Approved For Release 2000/08/08 :CIA-RDP96-00789R0022005200Q1-0 IS00 Radin and Nelson The following expression of the problem can be found in ~ recent interpretation of quantum theory: If conscious choice can decide what particular observation I measure. and kh~re- fore into what states my consciousness splits. might not conscious choice also be able to influence the outcome of the measurement? One possible place where mind may influence matter is in quantum effects. Experiments on whether i~t is possible to affect the decay rates of nuclei by thinking suitable thoughts would presumably be easy to perform, and might be worth doing.18~ Given the distinguished history of speculations about the:':. role of consciousness in quantum mechanics, one might expect that the; physics literature would contain a sizable body of empirical data on this :topic. A search, however, reveals only three studies. The first is in an article by Hall, Kim, McElroy, and Shimoity, who reported an experiment "based upon taking seriously the proposalthat the reduction of the wave packet is due to a mind-body interaction; ~n which both of the interacting systems are changed."'9' This experiment; e~Camined whether one person could detect if another person had previously bbserved a quantum mechanical event (gamma emission from sodium-22 atoms). The idea was based on the supposition that if person A's observation actually changes the physical state of a system, then when person obser- ves the same system later, B's experience may be different acco ding to whether A has or has not looked at the system. Hall er al.'s results, based on a total of 554 trials, d:id not support the hypothesis; the bbserved number of "hits" obtained in their experiment was precisely the.; number expected by chance (277), while the variance of their measurements was significantly smaller than expected (p < 0.05 ). c9 ~ The second study is referred to by Hall et al., who end their article by pointing out that a similar, unpublished experiment using cobalt-~7 as the source was successful (40 hits out of 67 trials ).'10' The third study is a more systematic investigation- repdrted by Jahn and Dunne,"' who summarize results of over 25 million binary trials collected during seven years of experimentation with random-event generators. These experiments, involving long-term data colleeti~on with 33 unselected individuals, provide persuasive, replicable evidence of an anomalous correlation between conscious intention and the o>~rtput of random number generators. Thus, of three pertinent experiments referenced in mainstream} physics journals, one describes results statistically too close to chance expectation and two describe positive effects.'9-"' Given the theoretical implications of such an effect, it is remarkable that no further experiments of this ;type can be found in the physics literature; but this is not to say that .no such experiments have been performed. In fact, dozens of researchers have Approved For Release 2000/08/08 :CIA-RDP96-007898002200520001-0 Approved For Release 2000/08/08 :CIA-RDP96-007898002200520001-0 reported conceptually identical experiments in the puzzling and uncertain domain of parapsychology. Perhaps because of the insular nature of scientific disciplines, the vast majority of these experiments are unknown to most scientists. A few critics who have considered this literature have dismissed the experiments as being flawed, nonreplicable, or open to fraud,~12-16~ but their assertions are countered by at least two detailed reviews which provide strong statistical support for the existence of anomalous consciousness-related effects with random number , _. _ generators.~l'~'$~ In this paper, we describe the results of a comprehensive, quantitative meta-analysis which focused on the questions of methodoiogi- cal quality and replicability in these experiments. The experiments involved some form of microelectronic random number generator (RNG), a human observer, and a set of instructions for the observer to attempt to "influence" the RNG to generate particular numbers, or changes in a distribution, solely by intention. RNGs are usually based upon a source of truly random events such as electronic noise, radioactive decay, or randomly seeded pseudorandom sequences. Feedback about the distribution of random events is often provided in the form of a digital display, but audio feedback, computer graphics, and a variety of other mechanisms have also been used. Some of the RNGs described in the literature are technically sophisticated, the best devices employing electromagnetic shielding, environmental failsafe mechanisms triggered by deviant voltages, currents, or temperature, automatic computer-based data recording on magnetic media, redundant hard copy output, periodic randomness calibrations, and so on.~is.zo~ RNGs are typically designed to produce a sequence of random bits at the press of a button. After generating a sequence of say, 100 random bits (0's or 1's ), the number of 1's in the sequence may be provided as feedback. In an experimental protocol using a binary RNG, a run might cansist of an observer being asked to cause the RNG to produce, in three successive button presses, a high number (sum of 1's greater than chance expectation of 50), a low number (less than 50), and a control condition with no direc- tional intention. An experiment might consist of a group of individuals each contributing a hundred such runs, or one individual contributing several thousand runs. Results are usually analyzed by comparing high aim and low aim means against a control mean or theoretical chance expectation. Approved For Release 2000/08/08 :CIA-RDP96-007898002200520001-0 Approved For Release 2000/08/08 :CIA-RDP96-007898002200520001-0 tgp~ Rsdin land Nelson The quantitative literature review, also called meta-analysis, has become a valuable tool in the behavioral and social sciences. ~=11 Meta-analysis is analogous to well-established procedures usead in the physical sciences to determine parameters and constants. The technique assesses replication of an effect within a body of studies by examining the distribution of effect sizes.~z7-2?' In the present context, the null- hiypothesis (no mental influence on the RNG output) specifies an expected mean effect size of zero. A homogeneous distribution of effect sizes with-nonzero mean indicates replication of an effect, and the size of the deviation of the mean from its expected value estimates the magnitude of the effect. Meta-analyses assume that effects being compared are sirnil'ar across different experiments, that is, that all studies seek to estimate the same pop- ulation parameters. Thus the scope of a quantitative review must be strictly delimited to ensure appropriate commonality across the different studies that are combined.~21?ZS' This can present a nontrivial problem; in meta- analytic reviews because replication studies typically investigate a number of variables in addition to those studied in the original experiments. In the present case, because different subjects, experimental protocols, aid RNGs were employed within the reviewed literature, some heterogeneity attributable to these factors was expected in the obtained distribution of effect sizes. However, the circumscription for the review required tihat every study in the database have the same primary goal or hypothesis; end hence estimate the same underlying effect. Experiments selected for review examined the following hypothesis: The statistical output of an electronic RNG is correlated wixh' observer intention in accordance with prespecified instructions, as indicated by the directional shift of distribution parameters (usually the mean} from expected values. Because this "directional shift" is most often reported as a~ standard normal deviate (i.e., Z score) in the reviewed experiments, we ;determined effect size as a Z score normalized by the square root of the sample size (N), e = Z/~, where N was the total number of individual random events (with probability of a hit at p = 0.5, p = 0.25, etc. ). This effect size measure is equivalent to a Pearson product. moment correlation. ~Z" 3.1. Unit of Analysis To avoid redundant inclusion of data in ameta-analysis, ~ "units of analysis" are often specified. We employed the following method: If an author distinguished among several experiments reported in a single ;Approved For Release 2000/08/08 :CIA-RDP96-007898002200520001-0 Approved For Release 2000/08/08 :CIA-RDP96-007898002200520001-0 Consciousness in Physical Systems 1503 article with titles such as "pilot test" or "confirmatory test," or provided independent statistical summaries, each of these studies was coded and quality-assessed separately. If an experiment consisted of two or more conditions comparing different intentions or types of RNG devices, the data were split into separate units of analysis to allow the results to be coded unambiguously. In general, within a given reviewed report, the largest possible aggregation of nonoverlapping data collected under a single intentional aim was defined as the unit of analysis (hereafter called an experiment or study). For each experiment, a Z score was assigned corresponding to whether the observed result matched the direction of intention. Thus, a negative Z obtained under intention to "aim low" was recorded as a positive score. When sufficient data were provided in a report, Z was calculated from those data and compared with the reported results; the new calculation was used if there was a discrepancy. If only probability levels were reported, these were transformed into the corresponding Z score. For experiments reported only as "nonsignificant," aconservative value of Z = 0 was assigned; if the outcome was reported only as "statisti- cally significant," Z =1.645 was assigned; and if sample size was not repor- ted or could not be calculated from the information provided, a special code of N =1 was assigned. 3.2. Assessing Quality Because the hypothesized anomalous effect is not easily accom- modated within the prevailing scientific world-view, it is particularly important to assess the trustworthiness of each reviewed experiment. Unfortunately, estimating experimental quality tends to be a subjective task confounded by prior expectations and beliefs.~2s~Z') Estimates of inter- judge reliability in assessing the quality of research reports, for example, rarely exceed correlations of 0.5.28) We addressed this problem by assigning to each experiment a single quality weight derived from a set of sixteen binary (present/absent) criteria. The first author coded and double-checked the coding for all studies; the second author independently coded the first 100 studies. Inter-judge reliability for quality criteria was r = 0.802 with 98 degrees of freedom. These criteria were developed from published criticisms about random-number generator experiments~'a.~s.z9-33) and from expert opinion on important methodological considerations when performing. studies involving human behavior.~2o.3a.3s> Collectively, these criteria form a measure of credibility by which to judge the reported data. The criteria assess the integrity of the experiment in four categories-procedures, Approved For Release 2000/08/08 :CIA-RDP96-007898002200520001-0 Approved For Release 2000/08/08 :CIA-RDP96-007898002200520001-0 Radin and Nelson statistics, the data, and the RNG device-and they cover virtually all methodological criticisms raised to date. They are (1) control teSits noted, (2) local controls conducted, (3) global controls conducted, (4~ controls established through the experimental protocol, (5) randomness calibrations conducted, (6) failsafe equipment employed, (7) data automatically recor- ded, (8) redundant data recording employed, (9) data double checked, (10) data permanently archived, (11) targets alternated on successive trials, (12) data selection prevented by protocol or equipment, (13') Fixed run lengths specified, (14) formal experiment declared, (15)tamper=resistant RNG employed, and (16) use of unselected subjects. Each criterion was coded as being present or absent in the ,;report of an experiment, specifically excluding consideration of previously published descriptions of RNG devices or control tests. This strategy was ermployed to reflect lower confidence in such experiments since, for exarnple;~ random- ness tests conducted once an an RNG do not guarantee acceptable perfor- mance in the same RNG in all future experiments. As a result; assessed quality was conservative, that is, lower than the "true" quality -'for some experiments, especially those reported only as abstracts or cpnference proceedings. Using unit weights (which have been shown to be !Robust in such applications~361) on each of the sixteen descriptors, the quality rating for an individual experiment was simply the sum of the descripte}rs. Thus, while a quality score near zero indicated a low quality or poorly reported experiment, a score near sixteen reflected a highly credible experimment. 33. Assessing Effect Size Assume that each of K experiments produces effect size estimates e of a parameter E, based on N samples, and that each a has a known ~ standard error s. The weighted mean effect size is calculated as e. _ ~ cetre;/~ w;, where w; = 1 /s = N;, and i ranges from 1 to K. The standard error of e. is se = (~ w;) - `~'-. A test for homogeneity for the K estimates of e; is given by HK = ~ w; (e; - e. )', where Hx has achi-square distribution with K -1 degrees of freedom.~37 The same procedure can be followed tc~ test for homogeneity of effect size across M independent investigators. In ;this case, e. ~ and seJ are calculated per investigator, and the test for homogeneity is performed as H,,,=~ w;(e.;-e.M)=, where e.; and w; are meanweighted effect size and 1/sP per investigator, respectively, e. M = ~ w;e.;/~ w;, and j ranges from 1 to M. HM has M - 1 degrees of freedom. For a quality-weighted analysis, we may determin(: e. Q = ~ (Q; cv; e; )/~ (Q; w; ), where Q; is the quality assessed for experiment i. The standard error associated with eQ is seQ = (~ (Q?w;)/(~ QIw;)?)''~2; the test for homogeneity is similar to that described above. Finally, :Following Approved For Release 2000/08/08 :CIA-RDP96-007898002200520001-0 Approved For Release 2000/08/08 :CIA-RDP96-007898002200520001-0 the practice of reviewers in the physical sciences,~23~24~ we deleted potential "outlier" studies to obtain a homogeneous distribution of effect sizes and to reduce the possibility that the calculated mean effect size may have been spuriously enlarged by extreme values. The procedure used was as follows: If the homogeneity statistic for all studies was significant (at the p < 0.05 level ), the study that would produce the largest reduction in this statistic was deleted; this was repeated until the homogeneity statistic had become nonsignificant. On-line bibliographic databases for psychology and physics journals were searched, as was a specialized database covering parapsychological articles, technical reports, conference proceedings and manuscripts. Altogether 152 references were found from 1959 to 1987. These reports described 832 studies conducted by 68 different investigators (597 experimental studies and 235 control studies). Fifty-four experimental and 33 control studies reported only as nonsignificant were assigned Z = 0. Six experiments and two control studies coded as (N= 1, Z> 0) were eliminated from further meta-analysis because effect size could not be accurately estimated (this required the elimination of one investigator who reported a single study). Figures 1 and 2 show the distributions of Z scores reported for control and experimental studies, respectively. m a 3 w THEORY FIT 00000 SHIFT ----- -2 -1 0 1 2 3 4 5 2-SCORES Fig. 1. Distribution of Z scores reported in 235 control studies. Thirty-three of these studies were reported only as "nonsignificant" and were assigned Z scores of zero. To replace the spurious spike at Z=O, those 33 studies were recast as normally distributed Z scores, bounded by ? 1.64, averaging Z = 0. Approved For Release 2000/08/08 :CIA-RDP96-007898002200520001-0 Approved For Release 2000/08/08 :CIA-RDP96-007898002200520001-0 106 Rsdin snd Nelson Fig. 2. Distribution of Z scores reported in 597 experimental studies. Fifty-four of these studies were reported as "nonsigni4'icant" and were assigned Z scores of zero.'Ag in Fig. 1, those 54 studies were recast as normally distributed Z scores, bounded by ? 1.64; averaging d O 2 x m N .y U ~ d -3 -2 -1 '--'0 1 2-SCORES N . 58 Fig. 3. Mean effect size point estimates ? 1 standard error for (a)contro] studies and (b)individual experiments; (c) mean effect size per investigator, (d) homogeneous mean effect size for experiments, (el homogeneous mean effect size per investigator, (f) mean effect size for quality-weighted experiments, and (g) mean effect size for homogeneous quality-weighted experiments. THEORY FIT o a o jo 0 ~4pproved For Release 2000/08/08 :CIA-RDP96-007898002200520001-0 Approved For Release 2000/08/08 :CIA-RDP96-007898002200520001-0 These results, expressed as overall mean effect sizes, show that control studies conform well to chance expectation {Fig. 3a), and that experimental effects, whether calculated for studies or investigators, deviate significantly from chance expectation (Fig. 3b, 3c). To obtain a homogeneous distribu- tion of effect sizes, it was necessary to delete 17 % of individual outlier studies (Fig. 3d) and 13 % of mean effect sizes across investigators (Fig. 3e). This may be compared with exemplary physical and social science reviews, where it is sometimes necessary to discard as many as 45 % of the studies to achieve a homogeneous effect size distribution.~19' Of individual studies deleted, 77 % deviated from the overall mean in the positive direction, and of investigator means deleted, all were positive (i.e., supportive of the experimental hypothesis). 4.1. Effect of Quality Some critics have postulated that as experimental quality increases in these studies, effect size would decrease, ultimately regressing to the "true" value of zero, i.e., chance results. suggesting that the present database is not compromised by poor experimental methodology. Another assessment of the effect of quality was obtained by comparing unweighted and quality-weighted effect sizes per experiment (Fig. 3b vs. 3f). These are nearly identical, and the same is true after deleting outliers to obtain a homogeneous quality-weighted distribution (Fig. 3d vs. 3g), confirming that differences in methodological quality are not significant predictors of effect size. It might be argued that the quality assessment procedure employed here was nonoptimal because some quality criteria are more important than others, so that if appropriate weights were assigned, the quality-weighted effect size might turn out to be quite different. This was tested by Monte Carlo simulation, using sets of 16 weights, one per criterion, randomly selected over the range 0 to 6. Aquality-weighted effect size was calculated for the 597 experiments as before, now using the random weights instead of unit weights, and this process was repeated one thousand times, yielding a distribution of possible quality ratings. The average effect size from the simulation was 3.18 x 10 -4 ? 0.15 x 10-4, indicating that in this particular database coded by these sixteen criteria, Approved For Release 2000/08/08 :CIA-RDP96-007898002200520001-0 Approved For Release 2000/08/08 :CIA-RDP96-007898002200520001-0 1508 Rodin end Nelson the probable range of the quality-weighted mean effect size clearlyexcludes chance expectation of zero. Although accounting for differences in assessed quality does npt nullify the effect, it is well known in the behavioral and social sciences that non- significant studies are published less often than significant studied (this is called the "filedrawer" problem~21?41~3j). If the number of nonsignificant studies in the filedrawer is large, this reporting bias may serious~y inflate the effect size estimated in ameta-analysis. We explored several procedures for estimating the magnitude of this problem and to assess the possibility that the filedrawer problem can sufficiently explain the observed t'esults. The filedrawer hypothesis implicitly maintains that all or nearly all significant positive results are reported. If positive studies are not balanced by reports of studies having chance and negative outcomes, the empirical Z score distribution should show more than the expected prap~rtion of scores in the positive tail beyond Z = 1.645. While no argument can be made that all negative effects are reported, it is interesting to notel that the database contains 37 Z scores in the negative tail, where only 30 would be expected by chance. On the other hand, there are 152 scores in thq positive tail, about five times as many as expected. The question is whe'.ther this excess represents a genuine deviation from the null hypothesis or' a defect in reporting or editorial practices. This question may be addressed by modeling based on the assumption that all significant positive results are reported. Afour-parameter.~fit mini- mizing the chi-square goodness-of--fit statistic was applied to all :observed data with Z ,1.645, using the exponential to simulate the effect of skew or kurtosis in producing the dspropor- tionately long positive tail. This exponential is a probability distribution with- the same mean and variance as the normal distribution; .but with kurtosis = 3.0. To begin, the null hypothesis of a (0, 1) normal distribution: with no kurtosis was considered. To account for the excess in the pgsi';tive tail, "~~ N= 585,000 filedrawer studies were required, and the chi-squared! statistic ~~ remained far too large to indicate a reasonable fit (see Table I). This large 1V, in comparison with the 597 studies actually reported together with the .poor goodness-of--fit statistic, suggests that the assumption of a (0, 1) normal distribution is inappropriate. Approved For Release 2000/08/08 :CIA-RDP96-007898002200520001-0 Approved For Release 2000/08/08 :CIA-RDP96-007898002200520001-0 'Consciousness in Physical Systems 1509 Table I. Four-Parameter Fit (E:N, N, Mean, sd) Minimi2ing Chi-Square (lOdf) Goodness-of--Fit Statistic to the Positive Tail of the Observed Z Score Distribution, for Several Exponential:Nonnal Ratios? Normal distribution 0 585,000 0 1 57,867.84 0 (null hypothesis) 1 5,300 0 1 220.97 0 2 4,800 0 1 167.84 0 3 4,600 0 1 148.45 0 10 4,400 0 1 119.69 0 Empirical distribution 0 700 0.145 2.10 23.94 0.008 1 747 0.345 1.90 16.32 0.091 2 757 0.445 1.80 14.21 0.164 3 777 0.445 1.80 11.08 0.226 10 807 0.445 1.80 11.08 0.351 ? The null hypothesis is tested by clamping the mean at 0 and the standard deviation at 1, allowing N and E:N to vary. The empirical database is addressed by allowing all four parameters to vary. Adding simulated kurtosis to a (0, 1) normal distribution by mixing exponential [Eq. (1) ] and normal distributions in a 1:1 ratio reduced N by two orders of magnitude, and ratios of 2:1, 3:1, and 10:1 exponential to normal (E:N) yielded further small improvements. However, the chi- squared statistic still indicated a poor fit to the empirical data. Applying the same mixture of exponential and normal distributions, but starting from the observed values of N = 597, mean Zscore = 0.645, and standard deviation =1.601, with the constraint that the mean could only decrease from 0.645, resulted in much better fits to the data. Table I shows the results. This procedure shows that the null hypothesis is unviable, even after allowing a huge filedrawer. The chi-square fit vastly improves with the addition of kurtosis, but only becomes a reasonably good fit when mean and standard deviation are allowed to approximate the empirical values. The filedrawer estimate from this model depends on a number of assump- tions (e.g., the true distribution is generally normal, but has a dispropor- tionately large positive tail). It suggests a total number of experimental studies on the order of 800, of which three-fourths have been formally reported. A somewhat simpler modeling procedure was applied to the data assuming that all studies with significant Z scores in either the positive or negative tail are reported. The model is based on the normal distribution with a standard deviation =1, and estimates the mean and N required to Approved For Release 2000/08/08 :CIA-RDP96-007898002200520001-0 Approved For Release 2000/08/08 :CIA-RDP96-00789R0022005200a1-0 1510 Raclin and Nelson account for the 152 Z scores in the positive tail and 37 Z scores in the negative tail. This mean-shift model, which ignores the shape of the observed distribution, results in an N =1,580 and amean Zscore = 0.34. These modeling efforts suggest that the number of unreported or unretrieved RNG studies falls in the range of 200 to 1,000. A remaining question is, how many filedrawer studies with an average null resu~t would be required to reduce the effect to nonsignificance (i.e., p "Evaluating psychological research reports: Dimensions, reliability, and correlates of quality judgments," Am. Psychol. 33, 920-934 (1978). 29. C. Akers, "Methodological criticisms of parapsychology." In Advances in Parapsychologi- cal Research, Vol. 4, S. Krippner, ed. (McFarland, Jefferson, North Carolina, 1984); "Can meta-analysis resolve the ESP controversy?" In A Skeptic's Handbook of Parapsychology, P. Kurtz, ed. (Prometheus Books, Buffalo, New York, 1985). 30. J. E. Alcock, "Parapsychology: Science of the anomalous or search for the soul," Behav. Brain Sci. Y0, 553-565 (1987). 31. P. Diaconis, "Statistical problems in ESP research," Science 201, 131-136 (1978). 32. C. E. M. Hansel, ESP and Parapsychology: ACritical Reevaluation (Prometheus gooks, Buffalo, New York, 1980). 33. R. Hyman, "The ganzfeld psi experiment: A critical appraisal," J. Parapsychol. 49, 3-50 (1985 ). 34. T. X. Barber, Pitfalls in Human Research: Ten Pivotal Points (Pergamon Press, Elmsford, New York, 1976). Approved For Release 2000/08/08 :CIA-RDP96-007898002200520001-0 Approved For Release 2000/08/08 :CIA-RDP96-007898002200520001-0 1514 Radio ~ad Nelson , 35. J. B. Rhine, "Comments: 'A new case of experimenter unreliability,"' J. Para~sychol. 38, 215-255 (1974). 36. R. M. Dawes, "The robust beauty of improper linear models in .decision making," Am. Psychol. 34, 571-582 (1979 }. 37. L. V. Hedges, "How hard is hard science, how soft is soft science?" Am, P.,sychol. 42,- 443-455 (1987). 38. C. E. M. Hansel, ESP: A Scientific Evaluation (Charles Scribner's Sons, New York, 1966), p. 234. 39. R. Rosenthal and D. B. Rubin, "Interpersonal expectancy effects: The first 3;45 studies," Behm;. Brain Sci. 3, 37715 (1978). L~ 40. G. V. Glass, B. McCaw, and M. L. Smith, Meta-analysis in Social Research. f~Sage Publi- cations, $cverly Hills, California, 1981). 41. Q. McNemar, "At random: Sense and nonsense," Am. Psychol. 15, 295-300 (,960). 42. S. Iyengar and J. B. Greenhouse, "Selection models and the filc-drawer problem," Technical Report 394, Departrnent of Statistics, Carnegie-Mellon University (~uly, 1987). 43. L. V. Hedges, "Estimation of effect size under nonrandom sampling: The effects of censoring studies yielding statistically insignificam mean differences," J. Edi+c. Stat. 9, 61-86 (1984 ). 44. H. H. Collins, Changing Order: Replication and Induction in Scientific Prdctice (Sage Publications, Beverly Hills, California, 1985). 45. S. Epstein, "The stability of behavior, II: Implications for psychological reserarch," Am. Psychol. 35, 790-806 (1980). 46. D. Druckman and J. A. Swets, eds. Enhancing Human Performance: Issues, Tfteories, and Techniques (National Academy Press. Washington, D.C., 1988),-p. 207. 47. A. Neher, The Psychology of Transcendence (Prentice-Hall, Englewood Cliffs, ~iew Jersey, 1980), p. 147. j Printed by Catherine Press, Ltd., Tempelhof 41, B-8000 Brugge, Belgium Approved For Release 2000/08/08 :CIA-RDP96-007898002200520001-0