manuscript_sections_current_pipeline.txt

================================================================================
DRAFT — Aligned with current analysis scripts (tube_analysis.py,
within_subject_exp2.py). Exp 2 threat analysis: ID015 (low) vs ID017 (high);
ID030 excluded from main figures. Angle validity: 3 deg < tilt < 40 deg;
cascade balancing of toward/away pairs; behavioral bot screening optional
(BOT_CLEAN = True in tube_analysis.py).
================================================================================

ABSTRACT
--------

People appear to treat others' gaze as if it exerts a small physical influence
on nearby objects. Prior work used an online tube-tipping task to show a
directional bias when a face was visible versus blindfolded. We replicated that
core pattern in a large online sample (Experiment 1). In Experiment 2 we asked
whether perceived facial threat modulates the same directional measure (D =
mean tilt toward the face minus mean tilt away), using two face identities
rated low versus high on perceived threat (ID015 vs ID017). Data were cleaned
with explicit worker quarantine (repeat submissions), trial angle filters, and
pairwise cascade invalidation so D is computed from symmetric toward/away
trials. Between-subject one-sample tests of D against zero showed no reliable
bias for either face at conventional levels; a within-subject paired analysis
(N = 249 participants with complete data for both faces) likewise found no
significant difference between high- and low-threat faces on D. These results
do not support a simple hypothesis that higher perceived threat strengthens the
gaze-related tilt bias in this paradigm, though they leave open more nuanced
accounts (e.g., individual differences, cutoff sensitivity, or factors other
than threat).


BACKGROUND
----------

Empirical vision science treats the eyes as receptors, yet people often feel
"watched" even when the glance could not geometrically reach them. Guterstam et
al. (2019) reported that, in a virtual tube-tipping task, visible faces biased
participants' judgments of the tipping point relative to the same face shown
blindfolded---as if gaze imparted a small force. That work framed the effect as
a socially grounded misapplication of physical intuition rather than a belief
in literal "eye rays."

The present project extends that question by asking whether perceived social
threat---operationalized by varying face identity with independent threat
ratings---amplifies the same directional bias. The core prediction would be a
larger toward-minus-away difference (D) for higher-threat than lower-threat
faces, holding tube geometry and task structure constant. Experiment 1 was
included to verify replication of the sighted-versus-blindfold dissociation at
scale before interpreting threat manipulations in Experiment 2.


METHODS
-------

Participants. Adults were recruited via Amazon Mechanical Turk. Experiment 1
used portrait stimuli of a single identity under sighted versus blindfold
conditions. Experiment 2 used separate face identities selected for low versus
high perceived threat (ID015 vs ID017). The main analysis pipeline excludes a
third high-threat face (ID030) used in broader data collection, so reported
threat contrasts are restricted to ID015 and ID017.

Task. On each trial, participants judged the tipping point of a paper tube
shown in short video-style presentation. Trials varied tube type, tip
direction, and lateral placement of the face relative to the table. In
Experiment 2, face identity carried the threat manipulation.

Data handling (current scripts). Raw submissions were organized with a file-level
quarantine that removes repeat workers (more than one session per workerId).
Primary loading merges Experiment 2 JSON archives (post-quarantine user-data)
with a legacy CSV (run 1) keyed consistently for participant identity.
Sessions flagged by behavioral bot heuristics were removed before analysis.
Trials were marked valid if recorded tilt magnitude satisfied 3 deg < angle <
40 deg (exclusive bounds). A single-pass cascade then invalidated the paired
toward/away partner whenever one member of the pair failed validity, so per-participant D is never computed from asymmetric pair loss. The primary outcome
D was computed per participant as mean angle on toward trials minus mean angle
on away trials (sign convention as implemented in code).

Experiment 1 analysis. One-sample t-tests of each participant's D against zero
for Eyes Open versus Eyes Covered.

Experiment 2 analysis. (i) Between-subject reference: separate one-sample t-tests
of D against zero for each face identity, including all participants with usable
data for that face after filtering (ID015: N = 256; ID017: N = 254 in the
current run). (ii) Within-subject / paired analysis: restrict to participants
who have valid toward and away data for both ID015 and ID017 (N = 249); repeat
one-sample tests of D vs zero at each face on this matched sample, and a paired
t-test of the difference D_ID017 - D_ID015. All tests are two-tailed; error bars
on figures denote standard error of the mean unless noted otherwise.


FINDINGS
--------

Experiment 1 (replication). With the current cleaning pipeline, Experiment 1
yields a pattern consistent with the original double dissociation: Eyes Open
participants show a small but statistically reliable positive D (N = 577,
D = +0.237 deg, SE = 0.114, p = 0.039), whereas Eyes Covered does not (N = 579,
D = +0.011 deg, SE = 0.110, p = 0.921). The absolute magnitude of D in Eyes
Open is modest and smaller than some published benchmarks, but the qualitative
contrast with the blindfold condition is clear under these analysis rules.

Experiment 2 (threat faces, ID015 vs ID017). Between-subject one-sample tests
on the full per-face samples are not significant for either identity: ID015
N = 256, D = -0.024 deg, SE = 0.199, p = 0.904; ID017 N = 254, D = +0.157 deg,
SE = 0.205, p = 0.447. The numerical ordering (somewhat higher D for the
higher-rated threat face) does not reach conventional significance and the
low-threat mean is very close to zero.

Within-subject sample: same participants for both faces; N = 249.

POSTER — Experiment 2: within-subject analyses (copy/paste friendly)
--------------------------------------------------------------------
• Sample: N = 249 (valid toward/away data for both ID015 and ID017).

• (1) EACH FACE VS. ZERO — one-sample t-tests on D (matched sample)
      ID015: D = -0.036 deg, SE = 0.200, p = 0.857
      ID017: D = +0.114 deg, SE = 0.201, p = 0.574
      Takeaway: no reliable bias vs. zero for either face.

• (2) DIRECT CONTRAST — ID017 vs ID015 within person
      Difference score: D_ID017 - D_ID015 per participant.
      Test: paired t(248); same inference as one-sample t on those differences
            (H0: mean paired difference = 0).
      Mean diff = +0.150 deg, SE = 0.270, t = 0.554, p = 0.580
      Takeaway: small, non-significant high- minus low-threat contrast.

(Long-form sentence version for manuscript: On the intersection sample,
one-sample tests of D vs. zero were non-significant for ID015 and ID017;
the paired contrast D_ID017 - D_ID015 was likewise non-significant.)

The reduction from the per-face unpaired Ns (256 and 254) to the paired N of 249
reflects a small number of participants who lack usable toward-and-away data for
one face or the other after the same filtering rules; those individuals are
excluded from paired inference but retained in the single-face between-subject
counts where applicable.


CONCLUSION
----------

In the current preregistered-style pipeline, Experiment 1 replicates the
sighted versus blindfold dissociation at large N: visible face information is
associated with a small reliable tilt bias, while the blindfold control is
flat. Experiment 2 does not provide evidence that swapping between a low- and
a high-threat-rated male identity (ID015 vs ID017) produces a statistically
reliable change in D, either in unpaired per-face tests or in a strictly paired
within-subject comparison (N = 249). Together, these findings temper hypotheses
that perceived threat simply scales the strength of the gaze-as-force illusion
in this exact task. Alternative explanations---including limited power for
subtle identity effects, heterogeneity across the two runs merged in the data,
and sensitivity to angle-cutoff choices (explored in separate sensitivity
figures)---should be acknowledged in the general discussion. The results are
most conservatively read as: the core gaze-induced bias is robust when the face
is visible versus obscured (Exp 1), but a coarse threat manipulation by face
identity (Exp 2, ID015 vs ID017) does not produce a clear, statistically
supported escalation of that bias under the present preprocessing and tests.

================================================================================
END OF DRAFT
================================================================================