AgapÈ : checking the stacking-effect
Bernard M. Auriol1, Frédérick Garcia2, Laetitia Puech3, Sylvie Lagrange3, Corinne Morer3, Olivier Rabat3, Sophie Valentin3, Eve Leconte3 and Olivier Perrin3
1Institut Métapsychique International, IMI-Paris,
France
3Université Toulouse I and Université Toulouse III,
France
Abstract
From December 1993 to January 2001, we carried out 240 ESP group sessions with majority vote. Both the number of individual successes and the number of collective successes (result of majority vote) proved to be compatible with the null hypothesis. The variance presented interesting anomalies but their interpretation is awkward.
The redundancy offered by the majority vote, didn’t significantly increase the
psi-hitting as we expected under the hypothesis of an improvement of the ratio
Signal / Noise. On the contrary, it seems that its impact made the results
random: the number of individual hits as well as the number of collective hits
(result of majority vote) proved to be compatible with the hypothesis of
results occurring by chance alone.
In case the seemingly random data belong to a second kind error, we initiated an analysis of the covariance, carried out for exploratory purpose only, in order to discern the protocol parameters which could have had an impact on the results. We work at the trial level. A trial is made of each participant’s vote for a word or a picture and for a given emission of a message from the transmitters. We confined ourselves to studying the trials with results either significantly higher or lower than chance expectation. In order to be able to test the influence of each modality of the variables, we resorted to a transformation of the variable "percentage of successes” that enables us to compare percentages of successes obtained with 2 pictures, 3 words or 5 words. We build for each trial the percentage of right answers given by the receivers; this will be the variable to be explained. The significant variables have been selected with a step by step procedure, and kept above a threshold of 5%.
All this leads to a rather puzzling result since, it appears that it’s nearly the same protocol parameters that hold the results away from chance expectation in both cases (hits and misses). These parameters are, in particular: a higher number of transmitters in relation to the number of recipients, the active involvement of the transmitters in the transmission, and their social links (assessed with a sociometric survey) with the receivers. But too strong social links between the transmitters result in bringing the outcome closer to chance.
But because these interesting results could be linked to a "stacking effect" (Caroline Watt, 2003), we used a resampling method to control that "stacking effect". Indeed, the replacements of the genuine targets by fake targets does not fundamentally change the results of the covariance analysis: The creation of faked targets allowed us to select new significant collective trials. The regression conducted on these gathered results very similar to those excerpted from the genuine data. We observe the same thing about the sociometric data.
Thus, we are most probably facing a stacking effect.
|
Introduction
The Agapè project,
ESP group sessions with majority vote (Auriol, 2003) gave puzzling
results: Both the number of individual successes and the number of collective
successes (result of majority vote) proved to be compatible with the null
hypothesis but the variance presented interesting (though awkward) anomalies.
The seemingly randomness of our data could
belong to the second kind error. This is why, on a purely exploratory basis,
and under the assumption that ESP would be (within the framework of our
protocol) a scarce phenomenon with a weak impact and/or unspecified sign
(sometimes towards Psi-Hitting, sometimes towards Psi-Missing) we undertook a
covariance analysis on the collective trials which deviated significantly from
expectation (p < .05). We are well aware that we can expect 5% of the trials to be different
from chance under the threshold p = 0.05, but we don’t foresee (Ho)
that a regression on these trials leads to identify parameters which would
increase or decrease the difference from chance (H1) .
If any interesting results were released they could be
due to a stacking effect. An available parade would be to use the Greville
correction (1944). However, this tool, in the context of our study, is very
laborious (Cf. Thouless & Brier, 1970), nearly impossible to deal with. We
thus decided to use a resampling technic.
Methods
Protocol
structure
The main steps were :
1. A Personal Computer selects a target (word or picture) randomly and displays it in the room of the transmitters.
2. This target is also displayed with decoys in the room of the receivers. Each receiver has a keypad and selects a number corresponding to one of the possible targets.
3. Then, results are recorded and the PC gives (or not) a collective and/or individual feedback.
4. Then, there is a new trial. After a certain number of trials, there is a break and the next series starts. And so forth until the participants are tired or the time is over.
Several sub-protocols
Three main sub-protocols have been tested:
- Two pictures: A picture which is randomly picked by a computer among a directory of 1 500 pictures is proposed with a decoy to the transmitting group. The members of the receiving group must choose between the two pictures which are displayed randomly on two screens (one is the emitted picture, the other one is a decoy).
- Three words: A very common word is randomly picked among a large list and is proposed to the transmitting group. It is proposed to the receiving group with two decoys.
- Five words: The participants brainstorm a list
of five words repeatedly used during the whole session. The transmitting group
sends one of the five words, randomly selected by the PC, and the receivers
make their choice among the words of the list.
In order to determinate if there are conditions
favoring ESP results we varied some other details in the protocol (Auriol, 2001) :
The «white screen latency» : that is a period between the beginning of the transmission by the transmitters and the display of the list of possible targets on the screens of the receiving group. This period of time can last during 0, 3, 5, 9, 10, 18, 20 or 60 seconds. The instruction given to the receiving group was to let all sensations, words or pictures come to their consciousness. The members of the receiving group were encouraged to use this material afterwards for guessing the target.
The «maximum deadline left» to the receiving group to answer can vary between 18, 20, 25, 30, 60 or 120 seconds.
This period can be cut short: When 50% of the
receiving group has answered, the software leaves to the rest of the group only
the elapsed time (x2=x1) or twice the elapsed time (x2+x3=2x1) to give their
answer. If the total time exceeds the maximum deadline, then, the answers done
after the deadline are not taken into account (Fig. 1).
|
Fig. 1
After the judging is closed a feedback is given.
- Individual feedback: each receiver is informed of his/her personal hit or miss, by displaying on the screen either his/her name or the target.
- Feedback for the group: when the trial is over, if the vote gives a hit, the lights are switched on (traffic lights’ style: see on the right of the displays, fig 1, 2 and 3) in the transmitters’ room and the receivers’. For some sessions, one light is switched on even if the majority gives a miss provided that the number of right individual answers be above the mean chance expectation. The number of lights on depends on the strength of the majority.
- Instructions may be given to the transmitters, or they may be free to send the message their own way. When there were instructions, they consisted of: for some sessions, to focus on the target to transmit, for others, to focus especially on one (or several) receiver(s) chosen, as much as possible by each transmitter.
Before many sessions, the receivers had to answer on their keypad to the following question about each of the other participants (either transmitter or receiver): “Do you know this person?” [mark from «not at all» (1) to «very well» (5)].
The participating groups were "open" groups: each individual joins in as he/she wishes. Each participant chooses to be transmitter or receiver at the beginning of session and he/she keeps doing until at the end of this session. The number of the transmitters varied from 0 to 15 people and the number of receivers from 1 to 16 people.
Results
Collective trials significantly departing
from chance
In order to test the effect of each variables
modality, we used a transformation of the « percentage of hits » to
be able to compare the results for the protocols with two pictures, three or
five words.
where:
- is the percentage of right answers in the trial
- p is the expected percentage (0.50 for two pictures, 0.33 for three words and 0.20 for five words).
If we look into the quality of the trials, we can test if the percentage of hits got for each trial is significantly lower than chance, significantly higher than chance, or equal to chance thanks to a test of Khi2. The statistic of this test, calculated for each trial, is
~Khi2 (one df)
|
Number of
trials |
Percentage
of trials |
Higher than
chance |
1079 |
3.87% |
Lower than
chance |
413 |
1.48% |
Equal to
chance |
26 353 |
94.64% |
Table 1 : Collective trials selected with a
threshold of 5%
We focus on the trials where the target collects either significantly more votes or significantly fewer votes than expected. The effect of different variables of “protocol” on the answers will be outlined with a covariance analysis. The significant variables will be selected thanks to a stepwise procedure and kept under a threshold of 5%. We took into account the trials significantly lower than chance as well as the trials higher than chance because the former could be a manifestation of “Psi-Missing” and we could hope to single out the factors involved in favouring success from other factors facilitating failure.
To check the stability of the selected variables
we carried out, with comparable results (Auriol et al., 2004b), the same
covariance analysis regarding the collective trials tendencially significant (p<0.10 and p<0.20).
We consider the case where the variable « answer » (transformed percentage h) is significantly higher or lower than the expected value under the null hypothesis. We get significant variables with a p-value close to 0.0001 (Table 2).
|
Variable |
Estimate |
|
|||||
|
Collective
trials |
Higher
than chance |
Lower
than chance |
|
||||
|
Constant
|
+1.06 |
-1.12 |
|
||||
Qualitative
Variables |
The transmitters were given
an instruction |
+0.10 |
… |
|
||||
The participants chose the
targets’ list |
-0.34 |
+0.26 |
|
|||||
There was some kind of group
reward |
-0.13 |
… |
|
|||||
The time left once half the
receivers answered = the time already spent |
+0.27 |
-0.02 |
||||||
The time left once half the
receivers answered = twice the time already spent |
+0.30 |
-0.12 |
||||||
Quantitative Variables |
Time to answer |
-0.01 |
+0.01 |
|
||||
Ratio {nb of transmitters /
nb of participants} |
-1.46 |
… |
||||||
Ratio {nb of transmitters /
nb of receivers} |
+0.96 |
… |
||||||
Table 2: significant variables- (calculated on
the genuine data)
Fig. 2 |
Parameters linked to getting the results closer to chance (Fig.2):
The participants could choose the potential targets’ list (-0.34)(+0.26)
The session included some kind of readable group feedback (-0.13) (…)
A longer time granted to answer (-0.01) (+0.01).
Parameters linked to the results departing from chance (Fig.2):
When 50% of the receiving group has answered, the software leaves to the rest of the group only the elapsed time (x2=x1) or twice the elapsed time (x2+x3=2x1) to give their answer. (+ 0.27 & +0.30) (-0.02 & -0.12)
Protocol in which the transmitters received an instruction (to focus either on the receivers, either on the target) (+ 0.10) (…).
·
Number of transmitters and receivers: The two ratios
{ number of transmitters / number of participants }
[(k*NbT)/(NbT+NbR)] (-1.46) (…) {closer to chance} and { number of
transmitters / number of receivers } [(g*NbT)/NbR] (+0.96)(…) {away
from chance} must be interpreted together because the same variables are
involved in both..
Effect of the relations between participants
We made two ratio based on the data from the sociometric questionnaire:
|
Receivers’ Mark |
|
|
Transmitters’ Mark |
RecGr = |
---------------------------- |
|
TransGr = |
-------------------------- |
|
Group’s Mark |
|
|
Group’s Mark |
where “Receivers’ mark” is the mean of the marks given by the receivers to the
other receivers, and “Transmitters’ Mark” is the mean of the marks given by the
receivers to the transmitters. The « Group’s Mark » is the mean
of all the marks.
For the
significant trials (p<0.05) with higher results than expected (Psi-Hitting),
the table had 1,079 observations with 438 missing data; the study of simple
regression was therefore carried out on 641 observations. The adjusted R²
equals 10.8% (« small » effect according to Cohen’s convention).
Variable |
Estimate |
Success |
Distance from chance |
P-value |
RecGr |
-1,54 |
- |
- |
<0.0001 |
TransGr |
+0,58 |
+ |
+ |
0.0003 |
Table 3
When the receivers know one another better than they know the transmitters, the results come closer to chance (less success). When the receivers know the transmitters better than they know one another, we move away from chance (more success) (Table 3).
We find the same tendency, although not as clearly (effect very small), if we take into account significant trials towards Psi-Missing (in that case, when the receivers know one another well, the failure is lessened, while when they know the transmitters better, the failure is strengthened) .
The variable
linked to better results are also linked to worse results. In that way we could
conclude «the significance is increasing only because the receivers give
more similar answers, independently of the hypothetical transmission process
originated by senders».
Checking
the stacking effect
Covariance analysis is valid only if one admits independence of the answers of the subjects; this condition was partially met because each subject had an individual keyboard. However occurrence of a "stacking effect" it cannot be cleared out insofar as somebody could have been "copying". A telepathic influence of a leader on the choice of other receivers cannot be drawn aside; when this leader misses the target; he/she could transmit to some of the other receivers the decoy he believes the right answer (Auriol, 2004a). Mostly, the receivers can have collective preferences / dislikes for some possible targets.
In addition, the subjects could have adopted more or less common strategies (Krugel, 1978), based on the feedback: to play back a guess which succeeded at the preceding trial, or on the contrary to avoid it. In fact, this phenomenon had significant effects on the protocols with three and five words (Garcia, 2003).
A stacking effect could meet the same results as our covariance analysis. In such a case, the selection of the variables would depend only on the answers of the percipients. This selection should be about the same one if we replace the genuine targets by faked target ones. We thus have remake calculations by carrying out a permutation of the targets:
Permutation 1 consists to replacing the genuine target by a target shifted of one row.
Permutation 2 consists to replacing the genuine target by a target shifted of two rows. And so on for permutations 3 and 4.
Permutation 5 consists to replacing the genuine target (target number i) by the fictitious target (target number i+1). If i is maximum, we replaced i by 1.
|
Genuine data |
First permutation |
Second permutation |
Third permutation |
Fourth permutation |
Fifth permutation |
|
Number of trials |
Number of trials |
Number of trials |
Number of trials |
Number of trials |
Number of trials |
Higher than chance |
1079 |
1047 |
1028 |
1052 |
1011 |
1073 |
Lower than chance |
410 |
426 |
423 |
435 |
454 |
460 |
Table 4 : Collective trials selected with a threshold of 5%
We focus on the trials where the target collects either significantly more votes or significantly fewer votes than expected. The effect of different variables of “protocol” on the answers will be outlined with a covariance analysis. The significant variables will be selected thanks to a stepwise procedure and kept under a threshold of 5%.
1.1 Analysis of the
collective trials significantly higher than expected :
|
Names |
Abbrev. |
Qual. Variables |
The transmitting group has had an instruction |
Instruction |
The participants (agents and percipients
together) have chosen the list of potential targets |
Choice |
|
There was a group feedback |
G feedback |
|
There
was an individual feedback |
I
feedback |
|
Maximum
time for the vote = 20 seconds |
20Max |
|
Maximum
time for the vote = 60 seconds |
60Max |
|
Maximum
time for the vote = 120 seconds |
120Max |
|
The time left to vote was
shorted when the half of the percipients
had voted |
TimeLeft |
|
Quant.Var. |
Ratio
: number of agents / number of participants |
A/(A+P) |
Ratio
: number of agents / number of percipients |
A/P |
Table 5 : names of variables and abbreviations
|
|
Genuine
data |
First
permutation |
Second
permutation |
Third
permutation |
Fourth
permutation |
Fifth
permutation |
Variable
|
Estimate |
Estimate |
Estimate |
Estimate |
Estimate |
Estimate |
|
Intercept |
+ 0.92 |
+
0.91 |
+
0.82 |
+
0.78 |
+ 0.78 |
+ 0.93 |
|
Qualitative Variables |
Instruction |
+
0.10 |
+
0.06 |
|
|
+
0.10 |
|
Choice |
-
0.29 |
-
0.32 |
-
0.30 |
-
0.32 |
- 0.30 |
- 0.32 |
|
G feedback |
-
0.11 |
-
0.08 |
-
0.09 |
-
0.11 |
- 0.10 |
- 0.06 |
|
I
feedback |
+
0.04 |
|
+
0.06 |
|
|
|
|
20Max |
+
0.13 |
+
0.20 |
+
0.19 |
+
0.21 |
+ 0.19 |
+ 0.15 |
|
60Max |
+
0.19 |
+
0.21 |
+
0.24 |
|
|
+0.17 |
|
120Max |
+
1.08 |
+
1.04 |
+
1.05 |
|
|
+ 0.98 |
|
TimeLeft |
+
0.07 |
+
0.13 |
+
0.07 |
+
0.12 |
+ 0.09 |
+ 0.13 |
|
Quantitative |
A/(A+P) |
-
1.06 |
-
1.31 |
-
0.81 |
-
1.16 |
- 1.50 |
- 1.58 |
A/P |
+
0.86 |
+
1.02 |
+
0.82 |
+
0.93 |
+ 1.13 |
+ 1.13 |
Table 6 : Analysis of the collective trials significantly higher than expected
1.2 Analysis of the collective trials significantly lower than expected
|
|
Genuine
data |
First
permutation |
Second
permutation |
Third
permutation |
Fourth
permutation |
Fifth
permutation |
Variable
|
Estimate |
Estimate |
Estimate |
Estimate |
Estimate |
Estimate |
|
Intercept |
- 1.21 |
- 1.43 |
- 1.43 |
- 1.42 |
- 1.24 |
- 1.45 |
|
Qualitative Variables |
Instruction |
- 0.10 |
- 0.12 |
- 0.15 |
- 0.15 |
- 0.09 |
- 0.15 |
Choice |
+ 0.22 |
+ 0.24 |
+ 0.31 |
+ 0.31 |
+ 0.26 |
+ 0.34 |
|
G feedback |
|
+ 0.12 |
+ 0.06 |
+ 0.07 |
|
+ 0.07 |
|
I
feedback |
+ 0.10 |
+ 0.10 |
|
|
|
|
|
20Max |
|
|
|
|
|
|
|
60Max |
|
- 0.14 |
|
|
|
|
|
120Max |
|
|
|
|
|
|
|
TimeLeft |
- 0.08 |
|
|
|
- 0.0433 |
|
|
Quantitative |
A/(A+P) |
|
|
+ 1.37 |
+ 1.06 |
+ 1.02 |
+ 1.16 |
A/P |
- 0.13 |
- 0.15 |
- 0.65 |
- 0.49 |
- 0.49 |
- 0.49 |
Table 7: Analysis of the collective trials significantly lower than expected
We asked to
the percipients, „do you know this people ?“ (ranked from „not at all“ to „very
well“) and we calculated two ratios :
Resgr = |
|
|
Emetsgr = |
|
|
|
(“Receiver score“ = mean of the marks that the receivers have allocated to the other receivers; “Transmitter score” = mean of the marks that the receivers have allocated to the transmitters; “Group score” = mean of all the marks)
2.1 Analysis of the collective trials significantly higher than expected (p <.05):
|
Genuine data |
First permutation |
Second permutation |
Third permutation |
Fourth permutation |
Fifth permutation |
Total number of
observations |
1079 |
1047 |
1028 |
1052 |
1011 |
1073 |
Usables observations |
662 |
624 |
620 |
644 |
616 |
654 |
Table 8
|
Genuine
data
|
First permutation |
Second permutation |
Third permutation |
Fourth permutation |
Fifth permutation |
Variable |
Estimate |
Estimate |
Estimate |
Estimate |
Estimate |
Estimate |
Resgr |
- 1.48 |
- 1.49 |
- 1.32 |
- 1.86 |
- 1.34 |
- 1.39 |
Emesgr |
+ 0.59 |
+ 0.38 |
+ 0.36 |
+ 0.35 |
+ 0.41 |
+ 0.40 |
R² |
0.11 |
0.11 |
0.07 |
0.13 |
0.09 |
0.09 |
Table 9
2.2 Analysis of the collective trials significantly lower than expected (p<.05):
|
Genuine data |
First permutation |
Second permutation |
Third permutation |
Fourth permutation |
Fifth permutation |
Total number of
observations |
410 |
426 |
423 |
435 |
454 |
460 |
Usables observations |
352 |
363 |
369 |
373 |
383 |
387 |
Table 10
Results
|
Genuine data |
First permutation |
Second permutation |
Third permutation |
Fourth permutation |
Fifth permutation |
Variable |
Estimate |
Estimate |
Estimate |
Estimate |
Estimate |
Estimate |
Resgr |
+ 1.77 |
+ 2.18 |
+ 1.93 |
+ 1.91 |
+ 2.04 |
+ 1.62 |
R² |
0.12 |
0.13 |
0.15 |
0.13 |
0.13 |
0.11 |
Table 11
Indeed, the replacements of the genuine targets by fake targets does not fundamentally change the results of the covariance analysis: The creation of faked targets allowed us to select new significant collective trials. The regression conducted on these gathered results very similar to those excerpted from the genuine data. We observe the same thing about the sociometric data.
Thus, we are most probably facing a stacking effect.
Discussion
Every
association of variable levels was not present. Consequently, confusional
effects could have been at work. For example, participants can have chosen the
list of possible targets only during the first sessions (5 words protocol), the
maximum deadline “120 seconds” was tested only for the 3 words protocol …
Is Psi-Missing an anti-Psi defence?
Yet, if we assume that there is Psi in our experiment, we are forced to acknowledge that we have not been able to find any variable that would facilitate Psi-Hitting against Psi-Missing. Couldn’t it be the need for each individual to avoid his own dissolution in order to exist as an individual with a psychic frontier? Some subjects, who got significant results within other experiments involving pairs, just had results close to chance in the collective frame defined by Agapè’s protocol. This phenomenon could be enhanced if the individual belongs to a group, situation which favors the fusion among the members!
References
Auriol, B.M. (2001), Variations de Protocole pour le Projet Agape
[On-line] Available: http://auriol.free.fr/parapsychologie/Agape/protocoles.htm
Auriol,
B.M. (2003), Agape Experiment: Further Statistical Studies (In Progress), EuroPA meeting, 14th-17th November 2003,
Port-Royal, France. [On-line]
Available : http://cabinet.auriol.free.fr/parapsychologie/agape/further-stat-study.htm
Auriol, B.M. (2004a), Connivence et Emetteurs Pirates, [On-line] Available
http://auriol.free.fr/parapsychologie/Agape/connivence.htm
Auriol etc (2004b) Agape : Extension de l'analyse de covariance à des essais non significatifs, [On-line] Available: http://auriol.free.fr/parapsychologie/Agape/regression-ns.htm.
Auriol, B.M. (2003),
Lagrange S., Morer C., Rabat O., Valentin S., Leconte E. & Perrin O.,
Agape : Group Telepathy, A long-term experimental series [On-line]
Available: http://cabinet.auriol.free.fr/parapsychologie/agape/AgapePA-Engl.htm
Bion, W. R. (1961). Experiences in
Groups, London: Tavistock.
Garcia, F. (2003), Agape : analyse des séries succès des essais
collectifs, [On-line] Available: http://cabinet.auriol.free.fr/parapsychologie/agape/agape-garcia.pdf
Greville, T.N.E.
(1944). On Multiple Matching with one Variable Deck, Annals
of Mathematical Statistics, vol. 15 pp.432-434. [On-line] Available :
http://auriol.free.fr/parapsychologie/redondance/Greville2.htm
Krugel, W. (1978), A
measure for call-strategies, Statistical Issues, Research in Parapsychology, Twenty-first
Annual Convention of the Parapsychological Association, 121-125.
Le Bon, G. (1895) Psychologie des foules, Paris, Alcan. [On-line] Available:
http://www.uqac.uquebec.ca/zone30/Classiques_des_sciences_sociales/classiques/le_bon_gustave/
Lewin,
K. (1935) A Dynamic Theory of Personality, New-York: Mc-Graw-Hill.
Pagès, M. (1968) La
vie affective des groupes, Paris: Dunod ; [german
translation (1974). Das affektive Leben der
Gruppen: Stuttgart: Klett ]
Thouless,
R.H. & Brier, R.M. (1970). Stacking Effect and Methods of Correcting for
it, Journal of Parapsychologie,
34, pp. 126-128 [On-line] Available:
http://cabinet.auriol.free.fr/Documents/stacking-thouless.htm
Watt, C. (2003), unpublished discussion at EuroPA meeting, 14th-17th November 2003, Port-Royal, France
Softwares : Excel, S-Plus (MathSoft) and
SAS.
Acknowledgments
Grant : Fondation Odier de
Psycho-Physique
Enregistré comme =>
http://cabinet.auriol.free.fr/parapsychologie/agape/stack_covar.htm