Back from the Future: Parapsychology and the Bem Affair

James Alcock

0 Shares

Psychologist Daryl Bem has reported data suggesting that individuals’ future
experiences can influence their responses in the present. Careful scrutiny of
his report reveals serious flaws in procedure
and analysis, rendering this interpretation untenable.

This paper will also appear in the March/April 2011 issue of Skeptical Inquirer.

A flurry of media attention is being
directed toward the prepublication distribution of Daryl Bem’s forthcoming
research paper “Feeling the Future: Experimental Evidence for Anomalous
Retroactive Influences on Cognition and Affect.”1 Bem claims to have
found evidence of marvelous psychic abilities that transcend time and
allow the future to reach backward to change the past. Both the academic
stature of its author, a respected emeritus professor of psychology
at Cornell University, and the fact that it was to be published in the
American Psychological Association’s (APA) Journal
of Personality and Social Psychology
have made this report particularly newsworthy.

Daryl J. Bem

Parapsychology
has long struggled, unsuccessfully, for acceptance in the halls of science.
Could this article be its breakthrough? After all, the article apparently
provides evidence compelling enough to persuade the editors of the world’s
preeminent social-psychology journal of its worthiness. However, this
is hardly the first time that there has been media excitement about
new “scientific” evidence of the paranormal. Over the past eighty-odd
years, this drama has played out a number of times, and each time parapsychologists
ultimately failed to persuade the scientific world that parapsychological
phenomena (psi) actually exist. Recalling George Santayana’s now-clichéd
dictum, “Those who cannot remember the past are condemned to repeat
it,” we should approach Bem’s work using a historical framework
to guide us. Consider the following:

1. In 1934,
Joseph Banks Rhine published Extra-Sensory
Perception (Rhine &
McDougall, 1934/2003), summarizing his careful efforts to bring parapsychology
into the laboratory through application of modern psychological methodology
and statistical analysis. Based on a long series of card-guessing experiments,
Rhine wrote: “It is independently established on
the basis of this work alone that Extra-Sensory Perception is an actual
and demonstrable occurrence” (p. 210). Elsewhere, he wrote: “We
have, then, for physical science, a challenging need for the discovery
of the energy mode involved. Some type of energy is inferable and none
is known to be acceptable . . .” (166).

Despite
Rhine’s confidence that he had established the reality of extrasensory
perception, he had not done so. Methodological problems with his experiments
eventually came to light, and as a result parapsychologists no longer
run card-guessing studies and rarely even refer to Rhine’s work.

2. Physicist
Helmut Schmidt conducted numerous studies throughout the 1970s and ’80s
that putatively demonstrated that humans (and animals) could paranormally
influence and/or predict the output of random event generators. Some
of his claims were truly extraordinary: for example, that a cat in a
garden shed, which was heated only by a lamp controlled by a random
event generator, was able—through psychokinetic manipulation—to
turn the lamp on more often than would be expected by chance. Schmidt’s
claim to have put psi on a solid scientific footing garnered considerable
attention, and his published research reported very impressive p values.2
In my own extensive review of his work, I concluded that Schmidt had
indeed accumulated impressive evidence that something other than chance
was involved (Alcock 1988). However, I found serious methodological
errors throughout his work that rendered his conclusions untenable,
and the “something other than chance” was attributable to methodological
flaws.

As
with Rhine, excitement about Schmidt’s research gradually dwindled
to the point that his work became virtually irrelevant, even within
the field of parapsychology itself.

3. The 1970s
gave rise to “remote viewing,” a procedure through which an
individual seated in a laboratory can supposedly receive psychic impressions
of a remote location that is being visited by someone else. Physicists
Russell Targ and Harold Puthoff claimed that their series of remote-viewing
studies demonstrated the reality of psi. This attracted huge media attention,
and their dramatic findings (Targ and Puthoff 1974) were published in Nature, one
of the world’s top scientific journals. At first, their methodology
seemed unassailable; years later, when more detailed information became
available, it became obvious that there were fundamental flaws in procedure
that could readily account for their sensational findings. When other
researchers repeated Targ and Puthoff’s procedure with the flaws
intact, significant results were obtained; with the flaws removed, outcomes
were not significant (Marks and Kamman 1978, 1980).

Add
Targ and Puthoff to the list of “breakthrough” psi researchers whose
work is now all but forgotten.

4. In 1979,
Robert Jahn, then dean of the School of Engineering and Applied Science
at Princeton University, established the Princeton Engineering Anomalies
Research (PEAR) unit to study putative paranormal phenomena such as
psychokinesis. Like Schmidt, Jahn was particularly interested in the
possibility that people can predict and/or influence purely random subatomic
processes. Given his superb academic and scientific credentials, his
claims of success drew particular attention within the scientific community.
When his laboratory closed in 2007, Jahn concluded that “over
the laboratory’s 28-year history, thousands of such experiments, involving
many millions of trials, were performed by several hundred operators.
The observed effects were usually quite small, of the order of a few
parts in ten thousand on average, but they compounded to highly significant
statistical deviations from chance expectations” (PEAR, n.d.).

However,
parapsychologists themselves were among the most severe critics of his
work, and their criticisms were in line with my own (Alcock 1988). More
importantly, several replication attempts have been unsuccessful (Jeffers
2003), including a large-scale international effort led by Jahn himself
(Jahn et al. 2000).

5. In the 1970s,
the ganzfeld, a concept borrowed from contemporaneous psychological
research into the effects of sensory deprivation, was brought into parapsychological
research. Parapsychologists reasoned that psi influences may be
so subtle that they are normally drowned out by information carried
through normal sensory channels. Perhaps if a participant were in a
situation relatively free of normal stimulation, then extrasensory information
would have a better opportunity to be recognized. The late Charles Honorton
carried out a large number of ganzfeld studies and claimed that his
meta-analysis3 of this work substantiated the reality of psi. Hyman
(1985) carried out a parallel meta-analysis that contradicted that conclusion.
Hyman and Honorton (1986) subsequently published a “joint communiqué”
in which they agreed that the ganzfeld results were not likely to be
due to chance, but they thought that replication involving more rigorous
standards was essential before final conclusions could be drawn.

Daryl
Bem subsequently published an overview of ganzfeld research in the prestigious Psychological Bulletin (Bem and Honorton 1994), claiming
that the accumulated data were clear evidence of the reality of paranormal
phenomena. That effort failed to be convincing, in part because a number
of meta-analyses have been carried out since then with contradictory
results (e.g., Bem et al. 2001; Milton and Wiseman 1999). Recently,
the issue was raised again in the pages of Psychological
Bulletin, with papers
from Storm et al. (2010a, 2010b) and Hyman (2010). While Storm and coworkers
argued that their meta-analyses demonstrate paranormal influences, Hyman
pointed to serious shortcomings in their analysis and reminded us that
the ganzfeld procedure has failed to yield data that are capable of
being replicated by neutral scientists.

Because
of the lack of clear and replicable evidence, the ganzfeld procedure
has not lived up to the promise of providing the long-sought breakthrough
that would lead to the acceptance of psi by mainstream science.

Add
Honorton (and Bem the first time around) to the list.

The
lesson in this history is that new claims of impressive evidence for
psi should give one pause. Early excitement is often misleading, and
as Ray Hyman has pointed out, it often takes up to ten years before
the shortcomings of a new approach in parapsychological research become
evident.

One
must also keep in mind that even the best statistical evidence cannot
speak to the causes of observed statistical departures.
Statistical deviations do not favor arbitrary pet hypotheses, and statistical
evidence cited in support of psi could as easily support other hypotheses
as well. For example, if one conducted a parapsychological experiment
while praying for above-chance scoring, statistically significant outcomes
could be taken as evidence for the power of prayer just as readily as
for the existence of psi.

Another
key consideration is that parapsychology’s putative phenomena are
all negatively defined: to claim that psi has been detected, all possible
normal influences must be ruled out. However, one can never be certain
that all normal influences have been eliminated; the reader of a research
report has only the experimenter’s word for it.

This
point brings us to a related concern. Research reports involve an implicit
social contract between experimenter and audience. The reader can evaluate
only what has been put into print and must presume that the researcher
has followed the best practices of good research. We assume that the
participants did actually participate and that they were not allowed
to use their cellular telephones during the experiment or to chat with
other participants. We assume that they were effectively shielded from
cues that might have inappropriately influenced their responses. We
assume that the data were as reported—that none were thrown out because
they did not suit the experimenter—and that they were analyzed
appropriately and in the manner indicated. We assume that equipment
functioned as described and that precautions reported in the experimental
procedure were carefully followed. We take for granted that the researcher
set out to test particular hypotheses and did not choose the hypotheses
after looking at the data. We must take all this on faith, for otherwise
any research publication might simply be approached as a blend of fact,
fantasy, skill, and error, possibly reflecting little more than the
predilections of the researcher. Obvious methodological or analytical
sloppiness indicates that the implicit social contract has been violated
and that we can no longer have confidence that the researcher followed
best practices and minimized personal bias. As Gardner (1977) wrote,
when one finds that the chemist began with dirty test tubes, one can
have no confidence in the chemist’s findings and must wonder about
other, as yet undetected, contamination. So, when considering Bem’s
present research, not only do we need to look at the data, but—following
the metaphor—we need to assess whether Bem used clean test tubes.

Bem’s Research

Bem describes
a series of nine experiments that “test for retroactive influence
by ‘time-reversing’ well-established psychological effects
so that the individual’s responses are obtained before the putatively
causal stimulus events occur.” His stated goal is “to provide well-controlled demonstrations of psi
that can be replicated by independent investigators.” He defines psi
as denoting “anomalous processes of information or energy transfer
that are currently unexplained in terms of known physical or biological
mechanisms.”

EXPERIMENT 1: Precognitive
Detection of Erotic Stimuli

Each trial
in this experiment involved the presentation of an erotic, negative,
or neutral picture. The participant sat in front of a computer screen
and was tasked to predict which of two curtains had a picture behind
it. Only after the participant had chosen a curtain by depressing a
key did the computer select a picture at random and present it behind
either the left or the right curtain.

Each
participant was presented with thirty-six of these trials and was given
feedback on each one. The erotic pictures were considered to be “explicit
reinforcement for correct ‘precognitive’ guesses,” although
no effort was made to determine whether they were indeed reinforcing
anything. The main hypothesis was that participants would be able to
identify the position of the hidden erotic picture significantly more
often than by chance.

So
far, clear enough. But then things become quite messy: we learn that
“most” of the pictures used in the experiment were selected
from a databank, the International Affective Picture System. Bem then
states that each session (a “session” refers to all the trials
of an individual participant) involved thirty-six trials of randomly
intermixed erotic and non-erotic pictures (eighteen of each). However,
we soon learn that not all sessions were conducted in this way: the
first forty of the one hundred sessions (that is, those of the first
forty participants) involved twelve trials of erotic pictures, twelve
of negative pictures, and twelve of neutral pictures! (The distinction
between the “non-erotic” pictures seen by the majority of the participants
and the “neutral” pictures seen by only the first forty is unclear.)
To muddle things even more, Bem then states that the remaining sixty
sessions involved “18 trials of erotic pictures and 18 trials of non-erotic
positive pictures with
both high and low arousal ratings. These included pictures featuring
couples in romantic but non-erotic situations.
. .” (emphasis added). How many were of high or low arousal weighting,
or what those terms even mean, he does not say.

What
is going on here? Setting aside the confusion about the stimulus, no
competent researcher dramatically modifies an experiment two-fifths
of the way into it! To do so is to seriously compromise any subsequent
analysis and interpretation.

But
that is not all. Bem next indicates that in all the experiments using
highly arousing erotic or negative stimuli, “a relatively large number
of non-arousing trials must be included to permit the participant’s
arousal level to ‘settle down’ between critical trials. This
requires including many trials that do not contribute directly to the
effect being tested.” This leaves us not knowing how many trials were
actually run and wondering by what method the researcher determined
the number of non-arousing trials that were needed to ensure that even
the most randy of participants would “settle down.”

So
by this point, it is not clear how many trials were actually presented
to each participant or even whether they all received an equal number
of trials. It is unclear just what the stimulus materials were, and
we are faced with a procedure that was changed partway through the experiment.

Just
when one thinks that this study cannot be made any more confusing, Bem
informs us that he discovered in Experiment 5 (which turns out to have
been conducted prior to Experiment 1!) that “women showed psi effects
to highly arousing stimuli but men did not.” In light of this odd
complication, Bem states that “we introduced different erotic and
negative pictures for men and women in subsequent studies, including this one using stronger and more explicit
images from Internet sites for the men. We also provided two additional
sets of erotic pictures so that men could choose the option of seeing
male-male erotic images and women could choose the option of seeing
female-female erotic images” (emphasis added).

By
now, a careful reader is totally confused as to what went on in this
experiment. Now, we find that participants were allowed to choose their
target set! This is one of the most baffling descriptions of research
materials and procedures that I have ever encountered.

In
reflecting on the extremely unusual change in procedure during the experiment—when
the appropriate course would be to run two different experiments—one
cannot help but wonder if two experiments were indeed run, and when
each failed to produce significant results the data from them were combined
with the focus shifted to only the erotic pictures common to all participants.
Surely that was not done, for such an action would make a mockery of
experimental rigor.

Data
Analysis: Bem states that
“the main psi hypothesis was that participants would be able to identify
the position of the hidden erotic picture significantly more often than
chance (50%).” At first, this claim is puzzling. Although sixty of
the participants completed eighteen trials with erotic pictures and
eighteen trials with “non-erotic positive pictures”—therefore
making the chance outcome 50 percent of the thirty-six trials—the
other forty participants received twelve trials with erotic pictures,
twelve with negative pictures, and twelve with neutral pictures. For
them, the chance outcome would be 33.3 percent. However, it turns out
that Bem combined the data for success or failure, but on the erotic
pictures only, from all one hundred sessions (i.e., from all one hundred
participants) and then applied t-tests4 to assess whether identification
of the future position of erotic pictures occurred significantly more
frequently than the 50 percent rate expected by chance. We are also
informed that the hit rate for non-erotic pictures—whether they were
neutral, negative, positive, or romantic and non-erotic—did not differ
significantly from chance. (This is the first mention of “romantic
but non-erotic,” which adds to the confusion.)

Now
we have learned that the focus of the experiment is on the erotic pictures
presented to the participants, but no information is provided regarding
how participants with three choices scored on erotic pictures as compared
with those who had only two choices; one wonders why this is so.

The
data analysis was conducted through multiple t-tests without any correction
for that multiplicity. We are informed that there were at least seven
such t-tests, and the only significant outcome was that the one hundred
participants “identified the future position of erotic pictures significantly
more frequently than the 50% hit rate expected by chance: 53.1%.”
This was stated to be statistically significant at p = .01. However,
that significance level is simply incorrect. This kind of error (Type
I)5 increases with the number of t-tests conducted, and given that there
were at least seven such t-tests with a criterion of p ≤ .01, the
actual probability associated with each of these t-tests is 1 –(.99)7=.06
one-tailed.6 Thus, none of these t-tests is actually statistically significant,
not even at a more generous .05 p value. It is simply unacceptable that
Bem did not correct for multiple testing, despite indications later
in his report that he is familiar with one such correction technique,
the Bonferroni t-test.7

Another
reason for concern is Bem’s deliberate use of one-tailed t-tests,
which provide a simpler criterion to meet than the two-tailed tests
generally employed by parapsychologists. (Parapsychologists typically
interpret both above-chance and below-chance scoring as indicative of
psi, and thus they do not make specific predictions about the direction
of the extra-chance scoring.) When we say that something is significant
at the .01 level two-tailed, this means that we would expect these results
to occur by chance alone only 1 percent of the time. But, given that
either above-chance or below-chance results are considered to be meaningful,
this 1 percent must be distributed in both directions. Thus, above-chance
results would be significant at the .01 level two-tailed only if they
are so extreme that they would be expected to occur by chance only half
of 1 percent, or 0.5 percent, of the time. The same applies for below-chance
results.

Bem
also reports that he carried out a nonparametric binomial test8 on the
overall proportion of hits on erotic targets across all trials and sessions,
but he offers no adequate rationale for using more than one type of
significance test for the same data. The test is redundant and offers
nothing beyond the t-test.

Then,
after having examined the data, he introduces the possibility that introversion/extroversion
may play a role in presumed precognitive ability. He suggests that it
may be an extrovert’s “susceptibility to boredom and the tendency
to seek out stimulation” that underlies observed correlations between
extroversion and psi performance reported in the literature. However,
rather than using existing, well-documented measures of stimulus-seeking,
he constructed his own such scale comprising two statements, reversed
in scoring: “I am easily bored” and “I often enjoy seeing movies
I’ve seen before.” The content and construction of this scale is
bewildering. Proper scale construction involves precise and often difficult
work, including operationalization of the construct, finding items that
can be demonstrated to relate to the construct, endeavoring to ensure
that the increments on the response measure are of approximately equal
size, and establishing satisfactory reliability and validity of the
final scale. Bem has ignored these considerations. As a result, the
arbitrary assignment of numbers to participants’ responses on this
“scale” is unjustified and misleading.

Nonetheless,
Bem correlates responses on the scale with the participants’ “psi
scores” and reports a significant correlation, but only for those
participants whose scores on his scale fall above the midpoint. Participants
who score below the midpoint on the scale did not score significantly
above chance on either erotic or non-erotic trials.

Overall
Evaluation: Just about
everything that could be done wrong in an experiment occurred here.
And even if one chooses to overlook that methodological mess, Bem’s
data still do not support the claimed above-chance effect because of
the multiple-testing problem.

It
is difficult to have confidence that the other eight experiments, some
of which were carried out earlier than the one just described, were
conducted with appropriate attention to experimental rigor: We have
toured the laboratory; we have found the dirty test tubes and the mislabeled
vials; we have observed inappropriate methodology and analysis. We have
lost confidence in the chemist, and there seems little need to poke
about further.

Nonetheless,
go on we must.

EXPERIMENT 2: Precognitive
Avoidance of Negative Stimuli

This study
involved 107 female and forty-three male undergraduate students. Using
a computer, each participant first responded to Bem’s two-item stimulus-seeking
scale and then completed a sequence of thirty-six trials in which a
“low arousal affectively neutral” picture was presented side by
side with its mirror image. The participant depressed a key to indicate
which picture he or she liked better. Only after the preference was
registered did the computer randomly choose which of the two pictures
would be considered the “target.” If the participant had chosen
this target, the computer thrice flashed a reportedly subliminal “positively
valenced picture.” If the participant chose the non-target, then a
“highly arousing negatively valenced picture” was flashed three
times.

A
hit was defined as choosing the “target-to-be.” However, as
in Experiment 1, the description of the situation is difficult to
unravel: For the first one hundred sessions (the first one hundred participants),
“the flashed positive and negative pictures were independently selected
and sequenced randomly.” Then there was a change in procedure. For
the next fifty participants, “the negative pictures were put into
a fixed sequence, ranging from those that had been successfully avoided
most frequently during the first 100 sessions to those that had been
avoided least frequently.” When the participant selected what was
to later be designated as the target picture, the positive picture was
flashed, subliminally as before, and the negative picture was retained
for the next trial. However, when the participant selected the non-target,
“the negative picture was flashed and the next positive and negative
pictures in the queue were used for the next trial.”

This
presents the same problem as before—the procedure has been changed
partway through the experiment. Bem states that this was done to evaluate
the possibility that “the psi effect may be stronger if the most successfully
avoided negative stimuli were used repeatedly until they were eventually
invoked.” It is difficult to get one’s head around this justification,
and in any case, this should have been examined in a separate study.
Again, given the inherent unreasonableness of changing the procedure
in an ongoing experiment, one cannot help but wonder if two separate
experiments were run and then combined after neither produced significant
results on its own.

As
in the first experiment, simple t-tests were used to compare participants’ hit rates against the chance
hit rate of 50 percent, and a nonparametric binomial test was used to
assess the proportion of hits across all sessions. (A third statistic
was also calculated; it is said to correct for unequal frequencies of
left/right target positions within each session.) In this instance,
we are not told how many other t-tests were carried out; if there were
other tests, as is likely, this again would have required a correction
for multiple comparisons. Of course, because all the data were pooled,
we have no information about how many participants actually scored at
a level significantly above chance. It seems odd that this information
was not of interest.

Bem
reports a significant correlation between the score on his two-item
stimulus-seeking test and psi performance, but once again the effect
was non-significant for “low stimulus seekers.” (Could it be
that Bem has serendipitously invented a two-item scale that predicts
psi ability?)

EXPERIMENT 3: Retroactive
Priming I

This experiment
involved a “priming” paradigm borrowed from contemporary psychological
research: participants indicate as quickly as possible whether a picture
is pleasant or unpleasant, and their response time is measured. Just
prior to the presentation of the picture, a positive or negative word
(a “prime”) is presented briefly (“subliminally”) on the screen.
This prime has been shown to have an effect in that participants usually
respond more quickly when a positive picture is preceded by a positive
word, or a negative picture is preceded by a negative word, than when
picture and word are incongruent. Bem refers to this as a contrast effect.

Bem
has taken this procedure and changed it so that the prime is presented
after the participant has responded. He reports a significant contrast
effect. His data analyses are very complex, involving two transformations
as well as outlier9 cutoff criteria; without access to the actual data,
it is difficult to evaluate the adequacy of the analysis. However, it
is obvious once again that multiple comparisons were carried out without
any control for multiple testing.

EXPERIMENT 4: Retroactive
Priming II

Experiment
4 is described as a replication of Experiment 3 “with one major change
and two timing changes.” Similar positive results were reported.
Again, one would need access to all the data, including the discarded
outliers, before one could properly evaluate the stated conclusions.

EXPERIMENT 5: Retroactive
Habituation I

After the
presentation of the previous four experiments, we are now informed that
Experiment 5 was a pilot for the basic procedures used in the other
experiments in this article. Why it is presented as the fifth experiment
is not explained.

The
experiment employed a mere-exposure protocol10 borrowed from experimental
psychology, but Bem “runs it backwards.” The participant is presented
with two pictures side by side. One is the “habituation target”
and the other is “closely matched” to the habituation target. The
participant is then instructed to indicate which picture he or she likes
better. Only then is the participant repeatedly exposed, subliminally,
to a picture of the “habituation target.” If it turns out that the
habituation target is the one that was earlier chosen, this is considered
a “hit;” it is assumed that the effect of the repeated subliminal
exposure to the target after the participant had made a choice
operated backward in time to influence that original choice.

The
habituation target was chosen 53.1 percent of the time, which is reported to be significant at the .014 level one-tailed. However,
once again multiple t-tests (six) are reported, which means that the
actual p values need to be adjusted. (Suppose that Bem had begun with
.014 as the criterion value; then the actual Type I error would be 12(12.014)6=.08,
which is not significant).

Incidentally,
Bem reports that the hit rate was significantly above chance for women
but not for men. Nonetheless, he also states that there was not
a significant sex difference! Though this seeming contradiction can
arise statistically, it is up to the researcher to make sense of it—which
Bem does not.

EXPERIMENT 6: Retroactive
Habituation II

Experiment
6 is described as a replication and extension of Experiment 5. Trials
with erotic picture pairs were added, and it was hypothesized that the
outcome for erotic pictures would be the opposite of that for the negative
pictures and that the participants would prefer the habituation picture
in fewer than 50 percent of the trials. Bem does not explain his reasoning.
There was also another change: on the basis that men may have simply
been less aroused than women by the erotic pictures in Experiment 5,
thus leading them to not produce a significant effect, it was decided
to use stronger and more explicit negative and erotic images obtained
from Internet sites for male subjects. Men were also given the choice
of male-male erotic images and women the choice of female-female erotic
images. (The reader will recall that this was also done in Experiment
1, which was run after Experiment 5.) Such matters should be investigated
in further pilot studies rather than incorporated into what is billed
as a replication experiment.

Bem
also tells us that he had not yet introduced by this point his two-item
stimulus-seeking scale into his series of experiments (remember, Experiments
5 and 6 were at the beginning of this series of nine). Instead, he constructed
another ad hoc scale by converting two items from
Zuckerman’s (1974) well-known Sensation-Seeking Scale into true-false
statements: “I enjoy watching many erotic scenes in movies” and
“I prefer to date people who are physically exciting rather than people
who share my values.” He gives no reason for choosing only these statements,
but he does not hesitate to treat them as a reliable and valid measure.
While showing no concern for the psychometric properties of these two
statements, he then arbitrarily defines only those who endorse both
statements as “erotic stimulus seekers.” Thus, an individual who
enjoys “many erotic scenes in movies” but prefers to date people
who share his/her values was not considered to be an erotic stimulus
seeker. This is purely an ad
hoc and unacceptable procedure,
again suggesting a cavalier attitude about the rigors of proper experimentation.

As
for the data analysis, once again there were numerous t-tests without
any control for multiple testing, thereby rendering erroneous the claimed
significance levels.

EXPERIMENT 7: Retroactive Induction of Boredom

The hit rate
was not reported to be significant in this experiment. The reader is
therefore spared my deliberations.

EXPERIMENT 8: Retroactive Facilitation of Recall I

This experiment
was an attempt to test the hypothesis that the future rehearsal of a
set of words can make them easier to recall in the present. (Students
would be delighted if this effect could be verified and harnessed, for
they could then do further study following a difficult exam and presumably
improve their performance on the examination already taken). The design
was simple. Participants were shown a set of words and then were tested
for their recall of the words. Subsequently, they were given practice
exercises with a randomly selected subset of those words, and the hypothesis
was that as a result of this subsequent practice, their performance
on the test (in the past, remember) would be enhanced and they would
have (in the past) recalled more of the words that were practiced in
the present.

The
participants were one hundred undergraduates. Again, they first responded
to the two stimulus-seeking statements. Next, forty-eight common nouns
were presented serially for three seconds each. The participant was
then asked to type out all the words he or she could recall. The computer
then selected twenty-four words at random, and the participant was now
instructed to type each of the selected words. It was hypothesized that
these practiced words would turn out to be the ones that had been better
recalled (before the practice).

Each
recalled word was deemed to be a trial and was scored as either a practice
or a control word. The actual difference—recall of practice words
minus recall of control words—was not analyzed; only a weighted score
was given, which was that difference multiplied by the participant’s
overall score (on both practice and control words). We are told that
this was done to give more weight to the scores of those participants
who recalled more words. No appropriate justification is given for this
awkward analysis; an analogy is drawn with the practice of weighting
studies by their sample size in a meta-analysis, but this is a spurious
analogy. The apparently arbitrary weighting of scores, when the more
direct-difference scores would offer less ambiguity, renders these findings
extremely difficult to interpret. Making choices about data analysis
after the data are collected introduces unacceptable opportunity for
bias and allows selection of a method that suits one’s hypothesis.

Making
matters more complicated, Bem then informs us that another twenty-five
“control” sessions were run, similar to the sessions outlined above
but without any practice sessions. These control sessions were interspersed
among the experimental sessions. The overall recall of words in his
control sessions was no different than that in the experimental sessions,
and so he concluded that “the enhanced recall of practice words came
at the expense of diminished recall of control words.”

Again,
it was found that participants who scored low in terms of his stimulus-seeking
scale scored at the chance level in the recall test, while those high
in stimulus seeking scored above chance.

EXPERIMENT 9: Retroactive Facilitation of Recall II

This is described
as a replication of Experiment 8, with one procedural change: a new
practice exercise was introduced “immediately following the recall
test in an attempt to further enhance the recall of the practice words.”
Again, weighted scores were calculated, and on this basis a significant
result was obtained. However, on this “replication,” the stimulus-seeking
questions did not correlate with psi success. My concerns
about the data analysis in Experiment 8 similarly apply in this case.

Overall,
then, this is a very unsatisfactory set of experiments that does not
provide us with reason to believe that Bem has demonstrated the operation
of psi. All that he has produced are claims of some significant departures
from chance, and these claims are flimsy given the many methodological
and analytical problems that I have touched on in this review. Moreover,
Ray Hyman has noted (in my personal communication with him) that the
correlation of effect size (as well as significance level) with sample
size is highly significant across this set of Bem’s experiments, but
it is in the wrong direction! “Effect size,” simply put, refers
in this case to the magnitude of the difference between the observed
scoring rate and the chance rate. Larger samples provide a better opportunity
to detect such a difference if it is truly there, and thus effect size
should increase with increased sample size. However, in Bem’s experiments,
the effect size correlates negatively (−.91) with sample size, indicating
that the claimed effect is smaller when the sample size is larger.

Statistical
power is a related concept that refers to the ability to detect an effect when
it is actually there. Hyman notes that while power (he uses the log
of significance probability as a proxy for power) should be positively
correlated with sample size (technically with the square root of sample
size), in this series of studies the correlation is approximately .80—in the wrong direction once again. This raises a bright red
flag and further erodes confidence with regard to the conduct of this
research.

* * *

Having presented
his nine experiments, Bem then discusses a number of general issues
in parapsychology research and then turns to quantum mechanics! Even
if one were to take his interpretation of his results at face value,
the claimed results are small and hardly justify an incursion into quantum
mechanical theory in the pursuit of accommodation of psi phenomena within
modern scientific theory.

While
it may seem puzzling that this distinguished psychologist has produced
such flawed research, anyone who has read his “Writing the Empirical
Journal Article” (published on his website at http://dbem.ws/WritingArticle.pdf)
would not be surprised. There he provides advice to students regarding
the conduct of research. A few revealing selections (all emphasis
added):

Once upon
a time, psychologists
observed behaviour directly, often for sustained periods of time. No
longer. Now, the higher the investigator goes up the tenure ladder,
the more remote he or she typically becomes from the grounding observations
of our science. If you are already a successful research psychologist,
then you probably haven’t seen a participant for some time. Your graduate
assistant assigns the running of a study to a bright young undergraduate
who writes the computer program that collects the data automatically.
And like the modern dentist, the
modern psychologist rarely sees the data until they have been cleaned
by human or computer hygienists.

To
compensate for this remoteness from our participants, let us at least
become intimately familiar with the record of their behaviour: the data.
Examine them from every angle. Analyze the sexes separately. Make up
new composite indexes. If a datum suggests a new hypothesis, try to
find additional evidence for it elsewhere in the data. If you see dim
traces of interesting patterns, try to reorganize the data to bring
them into bolder relief. If there are participants you don’t like,
or trials, observers, or interviewers who gave you anomalous results,
drop them (temporarily). Go on a fishing expedition for something—anything—interesting…

When you are
through exploring, you may conclude that the data are not strong enough
to justify your insights formally, but at least you are now ready to
design the ‘right’ study. . . . Alternatively, the data may be strong enough
to justify re-centering your article around the new findings and subordinating
or even ignoring your original hypotheses…

Your overriding
purpose is to tell the world what you have learned from your study.
If your research results suggest a compelling framework for their presentation,
adopt it and make the most instructive findings your centerpiece. Think
of your data set as a jewel. Your
task is to cut and polish it, to select the facets to highlight, and
to craft the best setting for it. Many experienced authors write the
results section first.

But
before writing anything, Analyze Your Data!

Reflections
of this advice appear to be writ large throughout Bem’s research article.

* * *

The publication
of this set of experiments will serve no one well. Parapsychology
is not honored by having this paper accepted by a mainstream psychology
journal. Neither does Bem’s paper serve the public well, for it only
adds to confusion about the scientific case for the existence of psi.
And it does no service to the reputation of the Journal
of Personality and Social Psychology.
Although Bem has failed to demonstrate the existence of mysterious
intellectual powers that defy the normal constraints of time and space,
there seem nonetheless to have been mysterious intellectual powers at
play here. I refer to the decision by the editors of an esteemed psychology
journal to publish this badly flawed research article.

“Think
of your data set as a jewel,” Bem instructs. However, with these nine
experiments, Bem did not end up with a polished jewel. Rather, to extend
his metaphor, the jewel cracked under the intense pressure used to try
to shape it to fit expectation. One is left with nothing but useless
fragments that reflect not the light of knowledge but the biases of
the researcher.

Rhine,
Schmidt, Targ, Puthoff . . . the list grows on. Plus ça
change, plus c’est la même chose.

Acknowledgements

Thanks to
Ray Hyman, Scott O. Lilienfeld, Timothy Moore, and Benjamin Wolozin
for their very sage comments on an earlier draft of this article.

Notes

1.
My discussion is based on the pre-publication version of Professor Bem’s
article that appears on his website at http://www.dbem.ws/FeelingFuture.pdf.

2.
The p value is the likelihood of having concluded
that there is a significant effect when in fact there is not. The lower
the p value, the less likely it is that the null hypothesis (that there
is no effect) is true.

3. Meta-analysis
is a statistical process for testing the combined results of a number
of studies that were based on similar research hypotheses.

4.
The t-test is a statistical test typically used
either to compare two means or to compare a mean with a theoretical
expectation—for example, to assess the difference between an observed
average success rate and a hypothetical chance rate of 50 percent.

5.
A Type I error occurs when the null hypothesis (that
there is no effect) is rejected when it is in fact true.

6.
In a two-tailed
test, one assesses the
data to see whether they significantly differ from what would be expected
by chance in either direction, that is, whether they are greater than
or less than what would be expected by chance alone. When we say that
something is significant at the .01 level two-tailed, this means that
we would expect these results to occur by chance alone only 1 percent
of the time. But, given that either above-chance or below-chance results
are considered to be meaningful, this 1 percent must be distributed
in both directions. Thus, above-chance results would be significant
at the .01 level (two-tailed) only if they are so extreme that they
would be expected to occur by chance half of 1 percent, or 0.5 percent,
of the time or less. The same would apply for below-chance results.

With
a one-tailed test, one also assesses the data to see
whether they significantly differ from what would be expected by chance
but in only one direction, that is, whether they are either greater
than expected by chance or less than expected by chance, but not both.
A one-tailed test is properly used if one has good reason to predict
the direction of the data in advance. Again, using the example of the
.01 level, for a one-tailed test the data only need be extreme enough
that they would be expected by chance alone 1 percent of the time or
less (compared to 0.5 percent with a two-tailed test). This makes it
much easier to claim statistical significance.

Parapsychologists
normally employ two-tailed tests because results that are either significantly
above chance or significantly below chance are taken to reflect psi.
Although Bem indicates that he predicted that the erotic pictures would
lead to above-chance scoring, which could justify using a one-tailed
test, what would he have done had the participants scored at a below-chance
rate that would have been significant had he predicted that the results
would indeed be below chance? Apparently committed to a one-tailed test
and having made only the above-chance prediction, he properly would
have had to ignore those data—something that parapsychologists do
not want to do. By using two-tailed tests, parapsychologists avoid the
problem and also avoid any suspicion of having changed the direction
of their prediction after having examined the data.

7.
The Bonferroni
t-test is a modified t-test
that adjusts for the number of tests being carried out so that the overall
likelihood that one of them produces significance by chance alone is
kept at a specified level, such as 5 percent.

8.
A nonparametric
binomial test deals with
data divided into two categories and examines the statistical significance
of deviations from a theoretically expected distribution. It is referred
to as “nonparametric” because it does not rely on the parameters
of a distribution, such as the mean.

9.
An outlier is a datum that is numerically distant
from all the other data in the sample, either as a result of measurement
error or because the data are not distributed in the manner that was
assumed.

10. Mere-exposure protocol is a research approach in which participants’
responses are assessed with the assumption that having simply been exposed
(perhaps subliminally) to a stimulus object will cause an effect.

References

Alcock, J.E.
1988. A comprehensive review of major empirical studies in parapsychology
involving random event generators and remote viewing. In Commission
on Behavioral and Social Sciences and Education, Enhancing
Human Performance: Issues, Theories and Techniques, Background Papers. Washington, D.C.: National Academy
Press, 601–719. Available online at http://books.nap.edu/openbook.php?record_id=778&page=601.

Bem, D.J.,
and C. Honorton. 1994. Does psi exist? Replicable evidence for an anomalous
process of information transfer. Psychological
Bulletin 115: 4–18.

Bem, D.J.,
J. Palmer, and R. Broughton. 2001. Updating the ganzfeld database: A
victim of its own success.
Journal of Parapsychology 65:
1–6.

Gardner, M.
1977. ESP at random. New
York Review of Books,
July 14.

Hyman, R.
1985. The ganzfeld psi experiments: A critical appraisal. Journal of Parapsychology 49: 3–49.

———. 2010. Meta-analysis that conceals
more than it reveals: Comment on Storm et al. [2010a]. Psychological
Bulletin 136(4): 486–90.

Hyman, R.,
and C. Honorton. 1986. A joint communiqué: The psi ganzfeld controversy. Journal of Parapsychology 50: 351–64.

Jahn, R.,
B. Dunne, G. Bradish, Y. Dobyns, A. Lettieri, R. Nelson, J. Mischo,
E. Boller, H. Bosch, D. Vaitl, J. Houtkooper, and B. Walter. 2000. Mind/machine
interaction consortium: PortREG replication experiments. Journal of Scientific Exploration 14: 499–555.

Jeffers, S.
2003. Physics and claims for anomalous effects related to consciousness.
In J.E. Alcock, J. Burns, and A. Freeman (Eds.), Psi
Wars: Getting to Grips with the Paranormal.
Exeter, UK: Imprint Academic, 135–52.

Marks, D.,
and R. Kamman. 1978. Information transmission in remote viewing experiments. Nature 274:
680–81.

———.
1980. The Psychology
of the Psychic. Buffalo,
NY: Prometheus Books.

Milton, J.,
and R. Wiseman. 1999. Does psi exist? Lack of replication of an anomalous
process of information transfer. Psychological
Bulletin 125: 387–91.

PEAR (Princeton
Engineering Anomalies Research). n.d. Experimental research. Available
online at http://www.princeton.edu/~pear/experiments.html.

Rhine, J.B.,
and W. McDougal. 1934/2003. Extra-Sensory
Perception. Whitefish,
MT: Kessinger Publishing.

Storm, L.,
P.E. Tressoldi, and L. Di Risio. 2010a. Meta-analysis of free-response
studies, 1992–2008: Assessing the noise reduction model in parapsychology. Psychological Bulletin 136: 471–85.

———.
2010b. A meta-analysis with nothing to hide: Reply to Hyman. Psychological Bulletin 136: 491–94.

Targ, R.,
and H. Puthoff. 1974. Information transmission under conditions of sensory
shielding. Nature 251: 602–4.

Zuckerman,
M. 1974. The sensation seeking motive. In B.A. Maher (Ed.), Progress in Experimental Personality
Research (Vol. 7). New
York, NY: Academic Press, 79–148.

James Alcock

James E. Alcock is professor of psychology, Glendon College, York University, Toronto.