The feeling of being stared at: An analysis and replication

Abstract

Experiments conducted where the possibility of such cues were reduced showed similar results, arguing against sensory artifacts. In a computer-based replication attempt, nearly significant results (p = 0.07) were obtained that were consistent with the estimates of previous effect sizes. A strong decline effect was also observed across runs of 20 trials.

Full text

Feeling of being stared at Dean Radin
Page 1/9
The feeling of being stared at: An analysis and replication
Dean Radin
Institute of Noetic Sciences
101 San Antonio Road
Petaluma, CA 94952
DeanRadin@Noetic.org
When data generated in “the feeling of being stared at” experiments are adjusted
to account for response biases, hit rates associated with responding “yes” when
being stared at, and “no” when not being stared at, are virtually identical.
Experiments conducted where the possibility of such cues were reduced showed
similar results, arguing against sensory artifacts. In a computer-based replication
attempt, nearly significant results (p = 0.07) were obtained that were consistent
with the estimates of previous effect sizes. A strong decline effect was also
observed across runs of 20 trials.
Introduction
Consider a simple form of “the feeling of being stared at” experiment. For the sake of exposition
let us call the starer(s) in such experiments “Jack” and the staree(s) “Jill.” Jack and Jill sit within
a few meters of each other, Jill with her back to Jack. Jack follows a random schedule which
determines on each successive trial whether he should stare or not stare at the back of Jill’s head.
Cued with a clicking tone, Jill responds “yes,” she believes she Jack is staring at her, or “no,” he
is not. The result of each such trial may be compiled into one of four categories: hit, miss, false
alarm, and correct rejection, as shown in Table 1.

Feeling of being stared at Dean Radin
Page 2/9
response
yes no
staring hit miss
condition not
staring false alarm correct rejection
Table 1. Types of trial outcomes in a staring experiment.
Sheldrake (1998, 1999, 2000a,b, 2003) reported a series of experiments based on this design,
some involving trial-by-trial feedback under casual conditions, such as tests conducted by pairs
of children in classrooms, and others involving blindfolds, no feedback, and more secure
conditions such having Jack stare at Jill through a window. From these published reports, plus
those of Coover (1913) and Portmann (1957), I was able to retrieve a total of 33,357 trials for
analysis.
The overall success rate in this database, i.e., the proportion of hits plus correct rejections, is
54.5%. On average, this would require only one additional hit over the chance-expected 10 hits
in a typical run of 20 trials, but given the large statistical power afforded by 33,357 trials, there is
little doubt that this success rate excludes chance as an explanation (z = 16.3, effect size per trial
= z/√(N) = 0.09, p << 10-10).
As shown in Figure 1, Sheldrake (2003) cited a consistent pattern of responses across these tests:
Jill was able to tell when Jack was staring (57.8% hit rate), but not when he was not staring
(51.1%). Sheldrake suggested that the existence of this pattern argues against a subliminal cuing
explanation, because with a combination of subliminal cues and trial-by-trial feedback one would
expect Jill to be able to learn to discriminate between staring and not staring conditions, and thus
the hit rates on both staring and not staring trials should be about the same. Given that they are
not, Sheldrake argued that the data suggest an anomaly consistent with a genuine “staring effect.”

Feeling of being stared at Dean Radin
Page 3/9
40%
42%
44%
46%
48%
50%
52%
54%
56%
58%
60%
hit false alarm miss correct reject
success rate
Figure 1. Observed success rates for the four types of outcomes in a staring experiment.
However, to properly interpret the statistical significance of these hit rates, they must be adjusted
for Jill’s response biases.1 While the stare vs. not stare conditions were distributed randomly in
these studies to preclude Jill from easily inferring the next target (50.1% vs. 49.9%, respectively),
Jill cannot be expected to respond at random. In fact, in these experiments, she said “yes” 53.4%
of the time and “no” 46.5%. This response bias inflates the statistical significance associated
with the observed success rate for hits, and deflates the success rate for correct rejections.
To see why, consider the case where Jill is feeling paranoid and responds “yes” on every trial.
Jill’s response bias will not inflate her overall success rate, but her hit rate and false alarm rates
will both be 100%, and her miss and correct rejection rates will be 0%. A less extreme bias
would be reminiscient of the pattern noted by Sheldrake, namely a higher percentage of hits and a
lower percentage of correct rejections.
A simple way to re-interpret the hit rates shown in Figure 1 is to use the formula for a z score of a
difference in proportions:
()
Nqpppz // 0001 −= , where p1 is the observed success rate, p0 is
the expected rate, q0 = 1-p0, and N is the number of trials under consideration. If p0 is taken to be
chance expectation of 0.5, and N = number of staring trials, then the overall “raw” z score is
z(hit) = 19.9. But if p0 is the observed response bias of 53.4% then the adjusted z score is z(hit)
= 11.5. This is still far from chance, but it is also significantly smaller than the unadjusted or raw

Feeling of being stared at Dean Radin
Page 4/9
z score (see Figure 2). From this perspective it is clear that the adjusted success rates for hits and
correct rejections are virtually identical, as are the rates for false alarms and misses. This is
more in alignment with what one might expect if the results were due to subliminal perception.
-25
-20
-15
-10
-5
0
5
10
15
20
25
hit false alarm miss correct reject
z score
raw score
adjusted
Figure 2. Raw and adjusted z scores for the four types of outcomes.
To explore the subliminal hypothesis in more detail, I partitioned Sheldrake’s data into settings
with progressively stronger controls for sensory cuing. Of the 33,357 trials, a total of 21,168
were collected with the pairs in close proximity – within two or three meters – and with trial-by-
trial feedback; 5,580 trials were collected in close proximity with no feedback; 4,800 were
collected with the pair separated by a window.

Feeling of being stared at Dean Radin
Page 5/9
-0.15
-0.10
-0.05
0.00
0.05
0.10
0.15
hit false alarm miss correct reject
effect size
feedback
no feedback
through windows
replication
Figure 3. Adjusted effect sizes ( Nze /=), and one standard error bars, for three testing
conditions and a new replication conducted as part of this study.
Figure 3 shows that study outcomes declined as testing conditions increasingly shielded against
subliminal cues. However, the effects did not decline to zero, nor were the results significantly
different from one another, as indicated by the error bars. This provides no support for the idea
that the staring effect is solely due to sensory artifacts.
A remaining limitation of most of the previous studies is that the outcome of each trial was
recorded manually, and this is known to be vulnerable to both mistakes and motivated biases. To
explore what effect recording errors may have had on the overall outcome, I conducted a new
experiment using an automated, computer-based design.

Feeling of being stared at Dean Radin
Page 6/9
Method
A person being stared at (Jill) and a person assigned to stare (Jack) sat two meters apart, with
Jill’s back to Jack, and Jack sitting in front of a laptop computer. Jill wore a blindfold to block
visual cues, and in her hands she held a gamepad (Microsoft’s “Sidewinder USB,” a computer
peripheral used to control video games). This gamepad provides a left and right-hand trigger
button used by the two index fingers, and several buttons on top of the gamepad designed used
by the right thumb.
Jill initiated each trial by pressing a button on the gamepad, whereupon a computer-synthesized
voice announced “Prepare for trial #”, where # was the current trial number. As the same time,
the words “Stare” or “No Stare” silently appeared on Jack’s laptop screen. Jack followed this
instruction either by gazing intently at the back of Jill’s neck, or by closing his eyes and thinking
about something else. Five seconds later, the computer sounded a click tone. This signaled Jill
to respond at will by pressing her right index finger button to indicate “I’m being stared at” or
pressing her left index finger button to indicate “I’m not being stared at.” After Jill responded,
the computer provided feedback by speaking one of four phrases: “Stare, correct” if Jack was
staring and Jill’s response was staring; or similarly “Stare, incorrect,” “No stare, correct,” or “No
stare, incorrect,” depending on the outcome. This sequence constituted one trial, and one run
consisted of 20 such trials.
Jill was required to sit directly in front of Jack, facing away from Jack at all times. Both Jack
and Jill listened to the computer’s prompts and sounds over headphones. The sequence of stare
and not-stare trials was determined randomly, with p(stare)= p(not-stare) = 0.5. The Microsoft
Visual Basic 6 pseudorandom function was used to make these random selections; the algorithm
was seeded by the computer system clock time when the program began.
Results
A total of 31 sessions of 20 trials, and one session of 5 trials, were contributed by 12 pairs of
participants (not including the author), or 625 trials. A total of 331 successful responses were

Feeling of being stared at Dean Radin
Page 7/9
recorded, or 53%, z = 1.48, p = 0.07. Figure 3 shows that the adjusted effect sizes across the four
possible outcomes were closely in alignment with previously reported results.
Figure 4 shows the z score associated with the success rate per trial and cumulatively. Each of
the first six trials were above chance expectation, accumulating to about 3 standard errors over
chance at trial 6 and then declining to a terminal z score of 1.48. This pattern is reminscient of
similar decline effects often observed in forced-choice tests.
-3
-2
-1
0
1
2
3
4
1234567891011121314151617181920
trial
z score
score per trial
cumulative score
Figure 4. Z score associated with success rate per trial and cumulatively.
Examination of autocorrelations in the target sequence indicated that none of the correlations
through lag 20 were statistically significant. A similar examination of the sequence of responses
showed a significant negative correlation for lag +1, r = -0.14, z = -3.89, p = 0.0001. This was
expected and is due to a common response bias in which subjects avoid giving the same answer
or response twice in a row. Lag +7 was also positive, r = 0.10, z = 2.80, p= 0.003, but given the
20 multiple tests this result is on the cusp of significance and may not be meaningful.

Feeling of being stared at Dean Radin
Page 8/9
Discussion
The adjusted statistical assessment of Sheldrake’s data and the present replication data show a
clear symmetry between hits and successful rejections, and misses and false alarms. These
outcomes are consistent with both a subliminal cuing effect and a genuine staring effect.
Arguing against the cuing explanation is the fact that staring without feedback and through
windows resulted in effect sizes that are statistically indistinguishable from staring close-up with
feedback.
In addition, two features of the present experiment suggest that Sheldrake’s reported results
reflect a genuine effect: (a) The computer-based recording method obviated motivated or
inadvertent errors, and (b) success rates in the first few trials of the average 20-trial run were
much better than in later trials. The latter finding is contrary to expectations about cuing
artifacts, which would presumably have resulted in progressively increasing success rates over
the course of runs of 20 trials.
In conclusion, the feeling of being stared at in these experiments appears to be a genuine, small
magnitude effect. Assuming a modest per-trial overall success rate of 53% where 50% is
expected by chance, a power analysis indicates that a p = 0.01 result can be achieved with power
= 0.90 with N = 3,632 trials. This may be achieved with 181 runs of 20 trials each, although
given the observation of a potential decline effect after six trials, it might be better to conduct
362 runs of 10 trials each.
If the reader is interested in using the Windows-based program I have developed to run this
experiment, please contact me.

Feeling of being stared at Dean Radin
Page 9/9
Bibliography
Coover, J. E. (1913). The feeling of being stared at – experimental. American Journal of
Psychology, 24: 570-575.
Poortman, J. J. (1959). The feeling of being stared at. Journal of the Society for Psychical
Research, 40, 4-12.
Sheldrake, R. (1998) The sense of being stared at: Experiments in schools. Journal of the
Society of Psychical Research 62: 311-323.
Sheldrake, R. (1999). The 'sense of being stared at' confirmed by simple experiments Biology
Forum 92, 53-76.
Sheldrake, R. (2000a). The 'sense of being stared at' does not depend on known sensory clues.
Biology Forum 93, 209-224.
Sheldrake, R. (2000b). The sense of being stared at: Effects of blindfolding subjects and giving
them feedback. Journal of the Society of Psychical Research.
Sheldrake, R. (2003). The sense of being stared at. New York: Crown Publishers.
1 Stefan Schmidt has independently noted that such an adjustment is required.