PEAR, Gender Differences In Human-Machine Anomalies.pdf

Journal of Scienti ® c Exploration, Vol. 12, No. 1, pp. 3± 55, 199 8

0892-3310 /98

Gender Differences in Human/Machine Anomalies

B RENDA J. D UNNE

Princeton Engineering Anomalies Research, School of Engineering and Applied Sciences,

Princeton University, Princeton, NJ 08544-526 3

Abstract Ð Assessment of 270 individual databases produced by 135 human

operators in five local and four remote human/machine an omalies experi-

ments conducted in t he PEAR laboratory between 1979 an d 1993 reveals

several significant gender-related differences in performance. Alt hough t he

140 databases produced by 62 females are much larger on average than t he

130 produced by 73 males, t he male average results display sign ificantly

stronger correlations wit h the operators’ pre-recorded intentions to shift t he

output distribution means of a variety of random devices to higher or lower

values. Bot h groups demonstrate greater success in t he high-intention ef f orts

than in t he low, but whereas a majority of t he males succeed in bot h direc-

tions of eff ort, producing intentional results that are relatively symmetrical in

com parison wit h t heir empirical baselines, m ost of th e females’ low-inten-

tion results are opposite to intention. The baseline data generated by t he

males largely concur wit h calibration and theoretical expectations, while t he

females tend to higher than chance values. The female data also frequently

display larger score distribution variances. These disparities are more pro-

nounced in five local experiments than in four remote databases. No gender

differences appear in two experiments that yield null overall results, suggest-

ing t hat t he gender-related patterns observed in th e successful experiments

may be indicative characteristics of the primary human/machine an omalies.

Keyword s : gender Ð human/machine interactions Ð engineering an omlies

research

Introduction

The Princeton Engineering Anomalies Research (PEAR) program was estab-

lished in 1979 to assess the potential vulnerabili ty of sensitive engineering

sy stems and information processors to an omalous influences associated with

the consciousn ess of their human o perators. This engineering orientation has

f ocused mainly on the physical parameters of t hese human/machine interac -

tions, rather than on possible psychologica l or physiologi cal correlates, ot her

than t he primary variable of o perator intentio n. All of these human/machine

experiments involve carefully calibrate d devices based on well-under stood

physical processes, each capable of rapidly generating, displaying, and record-

ing extensive sequenc es of random events. Volunteer human o perators at-

tem pt, solely t hrough consciou s ef f ort, to shift t he output distribution means

of these devices to higher or lower counts, or to generate an undisturbed base-

line, in accordance with pre-recorded intentions, and t he data are t hen

B. J. Dunne

exami ned for statistical correlations between those intentions and device per-

f ormance. Although t hese databases are extraordinaril y large, con sistent with

t he need for reliable statistical estimates of minuscule ef fects, they have been

pr oduced by a relatively small number of o perators. S pe cifically, nearly 20

million exp erimental data points, generated between 1979 and 1993 by some

135 o perators on a variety of such physical systems, have provided persuasive

s tatistical eviden ce for small but repeatable shifts of the output distributio n

means t hat correlate with the o perator intentions.

Pr evious examination of the individual operator contributions to these data -

base s established t hat t heir ef fects distributed normally around the shifted

means, im plying that a majority of the o perators contributed incrementally to

t he overall results, in contrast to any dominating perf ormances by a few excep-

tional o perators [1]. While some qualitativ e indications of characteristic d if-

fe r ences in individual perf ormance were noted, particularly among the more

pr olific o perators, t hese proved diff icult to assess quantitativel y because of

t he small signal-to- noise ratios involved. Nonetheless, since t he o perator

pool i s fairly evenly com posed of 72 males and 62 females, there is an ade-

quate basi s for exploring possible collective diff erences in perf ormance as a

function of gender .

The study reported here was also motivated by a body of so-called ª co-o per-

atorº experimen ts, wherein pairs of operators addres sed t he tasks with shared

intentions [2]. Beyond providing further confirmation of an omalous correla-

tions between o perator intentions and mean shifts, t hese studies showed no

evidence of any sim ple additive ef fects o f individual o perator perf ormance ,

but did provide strong indications that o perator gender may be an im portant

contributing factor. For exam ple, o perator pairs of the same sex tended to pro-

duce null results, trending insigni ficantly in the directions o pposite to inten -

tion. Opposite-sex pairs, on the ot her hand, produced significant overall re-

sults in t he desired directions, with ef fects considerably larger than those

generated by these same individuals working alone, and this enhancem ent of

ef fe ct size was strongest when t he two operators shared a deep emotional bond

wit h each ot her. Anot her curiosity of t hese o pposite-sex data was a relative

symmetry between t he high- and low-going achievements, unlike the asym-

metrical yields frequently observed in the single-o perator experiments where

one intention was typicall y found to produce considerably stronger results

than t he ot her. Prom pted by these f indings, a com prehensive evaluatio n of all

of PEAR’s existing databases has been undertaken to assess the relative per-

f ormance of its male and female o perators over nine diff erent experiments

whose design, protocols, and overall results have been detailed previously [3-

8].

Methodology

The most direct assessment of male/female differe nces in perf ormance

would appear to be a sim ple com parison of t he com posite results of t he two

Gender Dif ferences

gr oups for each experimen t via a sim ple z -scor e calculatio n for t he differ -

ence s. However, these com posite values are strongly weighted by substantial

di sparities in the sizes of t he individual o perator databases, which can easily

di stort their interpretation. More informative indications of the relative contri -

butions of t he male and female o perators can be obtained by examining their

r esults on an individual basis and t hen com paring the average yields and t he

pr oportions of o perators in each gender group who produce results correlating

w it h intention. This proportional approach also permits comparisons across

div erse databases where calculat ions of ef fects are necessarily based on differ -

ent scale s.

In t he sections to follow, t he results of each of nine distinct experiments are

pr esented by gender, both in terms of their composite and average results, and

as summarie s of the proportional yields of t he individual o perators. (Full de-

tails of t he individual results are available in a Technical Report [10]). It

should be n oted at the outset that most of t he experimental databases are rela-

tiv ely small in terms of the numbers of contributi ng o perators, and thus t he

s tatistical results based on these proportions frequently entail large error bars .

It should also be n oted t hat many of t he operators participated in more t han

one exp eriment, but since all of t he experiments are independent of each ot her,

each operator-experime nt database is treated as a separate entity. This ap-

proach results in a total of 270 individual contributions over nine separate ex-

periments, 130 from male o perators and 140 from female, com prising a more

robust base for overall statistical assessment of gender contributi ons.

1. Random Event Generator Experiments

The most extensive PEAR databases have utilized a microelectronic random

event generator (REG) as the target device [3-6]. The ª benchmark º experi -

ment comprises more than 2.5 million trials, each consisting of 200 random bi-

nary sam ples. These data were generated over a 12-year period by 91 o pera-

tors in 522 independent experimental series ranging in size from 1000 to 5000

trials per intention, depending on the protocol involved. (In all PEAR experi -

ments, a ª seriesº is the pre-established evaluative unit, each constituting an in-

dependent replication of the basic experiment.) The benchmark REG databas e

was accumulated over three distinct experimental phases which differe d in

terms o f series size, run lengt h (t he number of trials produced automatically as

a result of a single initiating button push), and the number of secondary o p-

tions available to t he o perator, e.g., run lengt h, automatic or manual o peration ,

volitional or instructed assignment of intention, or t he available modes o f vi-

su al feedback on the machine face and its accom panying com puter screen.

How ever, all the experiments followed the same basic tri-polar protocol in

which the o perator was seated in the same room as the device and generated

data under three distinct intentions: attem pts to shift t he mean of t he output

distributions in the positive directio n (HI), in t he negative direction (LO), or to

B. J. Dunne

produce a baseline (BL) under no directional intention, with all ot her condi -

tions held constant for the duration of a given series .

Of the 91 o perators who contributed to this database, 50 males produced a

total of 228 series, or approxim ately 327,000 trials per intention , and 41 fe-

males generated 294 series, or approxim ately 506,000 trials per intention .

(These numbers are approximate because in some of t he earlier series a ran-

domly assign ed instruction for the direction of each run resulted in unequal

numbers of trials per intention; a later modification of the program guarantee d

equal numbers of trials per intention in this ª Instructedº mode.) The com pos-

ite results of t his database , as well as the relative contributi ons by male and fe-

male o perators, are summarized in Table 1. The ª normalized deviation, º d c ,

utilized here is sim ply the deviation of the com posite experimental mean from

the theoretical expectation of 100, multiplie d by 10 0 for convenienc e of tabu -

lation. It provides an indication of t he magnitude of the deviation achieved,

but is vulnerabl e to statistical uncertainty for small data sets. 1 The ª z -scoreº ,

or z c , def ined as the deviation of t he com posite experimen tal mean from t he

theoretical expectation n ormalized by t he standard error, , where s 0 is the t he-

oretical trial standard deviation and N is the number of trials in t he given data

set, provides a more reliable indication of t he statistical signifi cance of t he

achieved deviation over databases of varying sizes but, as noted above, can ob-

scure t he absolute magnitude of achievement in the smaller data sets. These

two indicator s, d c and z c , thus com plement one anot her for interpretation of

the results.

Ta ble N ot es. In t his and subsequent tables, achievements in the direction of

ef f ort in the high intentions (HI) and in the high-low diff erences (HI-LO) are

indicated by positive deviations and z -scores, and in the low intentions (LO)

by negativ e numbers. Positive numbers in t he baselines (B L) indicate results

higher than the theoretical mean. Results o pposite to intention, or lower t han

the t heoretical expectation in the baselines, are indicated by parentheses .

Those z-scores exceeding t he one-tailed p <.05 criterion (> ± 1.6 5 ) in the direc-

tion of intention, and their associated probabilities, are noted by asterisk s ( *) ;

z -scores > ± 1.65 o pposite to intention are indicated by daggers (² ). Probabili -

ties for intentional ef f orts are calculated on a one-tailed basis; those for base-

lines, where there is no directional expectation, are two-tailed, wit h a p <.05

criterion of z > ± 1.9 6.

These com posite results suggest t hat while both groups produce com para -

ble results in the LO and BL, t he female operators are collectively more suc-

This n ormalized deviation is similar to the standardized ª effect size º def ined by Rosent hal [11], ex-

cept that his version is normalized by the t heoretical trial standard deviation, s 0 , while ours is

normalized by an arbitrary constant for convenien ce in tabulation. Since s 0 is itself a constant of t he ex-

periment, the normalized deviations, d c , and standard ef fect sizes , e , are related by the constant ratio of

d c / e = 100 s 0 = 707.1.

Gender Dif ferences

TABLE 1

Composite Results of A ll Local REG Experiment s

HI - LO

All Operators (522 Series )

Number of Trials ( N )

839,800

820,75 0

836,650

~837,8 25

Distribution Mean (µ)

100.026

100.013

99.984

.042

Normalized Deviation ( d c )

2.6

1.3

± 1.6

4.2

Std. Dev. Trial Scores ( s )

7.070

7.074

7.069

9.998

z-Score ( z c )

3.369

1.71 3

± 2.016 *

3.809*

Probability ( p )

4 ´ 10 - 4 *

0.086

0.022*

7 ´ 10 - 5 *

50 Male Operators (228 S eri es )

Number of Trials ( N )

331,650

316,75 0

331,300

~331,4 75

Distribution Mean (µ)

100.015

100.011

99.983

0.032

Normalized Deviation ( d c )

1.5

1.1

- 1.7

3.2

Std. Dev./Trial Scores ( s )

7.060

7.064

9.987

z-Score ( z c )

1.228

0.865

- 1.424

1.875*

Probability ( p )

0.110

0.386

0.077

0.030*

41 Female Operators (294 S eries )

Number of Trials ( N )

508,150

504,00 0

505,350

~506,75 0

Distribution Mean (µ)

100.033

100.015

99.986

0.047

Normalized Deviation ( d c )

3.3

1.5

- 1.4

4.7

Std. Dev./Trial Scores ( s )

7.077

7.080

7.073

10.00 6

z-Score ( z c )

3.339*

1.500

- 1.441

3.382*

Probability ( p )

4 ´ 10 - 4 *

0.134

0.075

4 ´ 10 - 4 *

* Ð see Table Notes o n p . 6.

cessfu l t han the males in t he HI ef f orts, resulting in a corresponding advantag e

in the HI - LO. How ever, as noted above, this im pression is misleading because

of t he considerabl e variability among individual operator perf ormances and in

th e sizes of their respective databases. The female average database is nearly

twice as large as the male average and includes three exceptionally large indi -

vidual databases wit h strong positive results. Even excluding t he most prolific

female database consisting of some 120,000 trials per intention, the average

female database still remains nearly a third larger than the average male’s .

While t his d iff erence clearly cannot be regarded as an experimental result, it

bears n oting because of its im pact on the statistical representation of the com-

posite results; it may also reflect diff erent o perational strategies em ployed by

the two groups.

The individual o perator performances are summarized by gender in Table 2,

wherein are displayed the averages of the individual n ormalized deviation s

and t he average z -scores for the HI, BL, and LO ef f orts, along with t hose of t he

HI - LO diff erences for each group. The number and proportion of o perators of

each gender who produce results consistent wit h their intentions (or above 100

in t he baselines ) , relative to t he 50% who might be expected to do so by

chance, and t he number and proportion of individuals who achieve results be-

yond t he one-tailed .05 chance expectatio n (two-taile d for baselines) are also

provided, with the proportions in the o pposite tail in parentheses, along with

Plik z chomika:

Inne pliki z tego folderu:

Inne foldery tego chomika: