The Registration Stack

The mean estimate made by the targets in this study (48%) was virtually identical to that made by the targets in the previous study who were wearing an embarrassing T-shirt (46%). The actual accuracy of the observers, in contrast, was dramatically different (8% vs. 23% in the present and previous studies, respectively).

Study 1, the Barry Manilow T-shirt. The running text of Study 1 puts both figures in Figure 1 alone, so the quote comes from footnote 3 of Study 2, where 46% is the Study 1 estimate and 23% the Study 1 observer accuracy. The 48% and 8% in the same footnote belong to Study 2. Fifteen of the 109 participants were targets, 64 were observers and 30 were controls. Anyone re-checking Study 1 meets 23% a second time in its own results section, where the figure names the mean discrepancy between each target's estimate and the accuracy of the observers in that session, a quantity that coincides with observer accuracy here only because the average estimate came to exactly twice it.

Across the eight sessions, liars estimated that an average of 48.8% of the participants would correctly peg them as the liar when, in fact, only 25.6% did so—an accuracy rate indistinguishable from chance.

Study 1a, a round-robin lie detection game. Analyses were run at the level of the eight sessions rather than the individual. Studies 1b and 1c repeated the procedure to rule out alternative readings.

Tasters estimated that an average of 4.91 observers would correctly identify the foul-tasting drinks, which was significantly higher than the average of 3.56 observers who actually did make correct identifications on these trials, t(14) = 3.63, p < .005.

Study 2a. Fifteen tasters sampled 15 cups, 5 of them containing vinegar brine, while trying to keep a neutral face for a videotape later shown to 10 observers. The figures count observers out of 10, against a chance rate of 3.33. Study 2b repeated the procedure with a yoked partner.

with participants reporting liking their conversation partner (actual: M = 5.82, 95% CI = [5.49, 6.14]) significantly more than they perceived their conversation partner to like them (perceived: M = 5.17, 95% CI = [4.85, 5.49])

Study 1a, strangers paired for a getting-acquainted conversation. Because liking given and liking received are the same quantity across a set of dyads, the 5.82 is the measured half. Both figures are predicted marginal means from a linear mixed model taking rating type as its fixed effect, with an intercept for each participant and each dyad, a note the paper attaches to every mean it reports. Thirty-six people took part and one dyad was excluded as close friends. Studies 3 and 5 found the gap again in conversations running from 2 to 45 minutes and among suite mates followed across an academic year. Study 4 is a weaker parallel, since it operationalised liking as how interesting partners found each other on a 5-point scale at public workshops on talking to strangers.

targets expected to be judged more harshly (M = 55.0) than they actually were (M = 62.5)

Study 4. Each target read a computer-generated introduction of themselves that mentioned occasional difficulties with bedwetting, then predicted the impression an observer would form on a 101-point scale where 50 is the average student. The 120 participants worked in target and observer pairs. What the scale measured is a global favourability impression formed after an embarrassing disclosure, and the row is filed under appearance only because the taxonomy has no domain for self-presentation.

participants overestimated the number of people they thought they would have to approach by approximately twice the number of people they actually needed to approach to complete the task (predicted: M = 20.5, SD = 12.3; actual: M = 10.5, SD = 3.9)

Study 1. Participants asked strangers on campus to fill in a questionnaire until five agreed, and the 23 who first predicted how many they would have to ask expected twice the number they needed. Quote taken from the author manuscript in Cornell eCommons, whose pagination differs from the published article.

at the other university, the figures for estimated and actual (reported) prevalence were 54% and 68% for feeling depressed, 42% and 62% for feeling lonely, 41% and 59% for crying, and 82% and 92% for feeling overwhelmed.

Pilot study, the East Coast sample of 225 introductory psychology students. The stored pair is the first item, feeling really depressed for a day. The same procedure run on 122 students at a West Coast university gave 52% estimated against 78% reported, and Studies 1B and 3 found the same shortfall on a different set of experiences.

As shown in Fig. 1, average SPV was approximately 10 on a 100-point scale among both Democrats (M = 9.3, SD = 17.2) and Republicans (M = 10.3, SD = 21.7). Importantly, 4.5% of Democrats and 9.8% of Republicans responses to the SPV composite fell above the midpoint of the scale, indicating a small—yet concerning—proportion of participants supporting violence at a relatively high level. We next tested whether partisans’ metaperceptions of out-partisans’ support for violence were accurate. Democrats’ estimates of Republicans’ SPV (M = 35.5, SD = 32.2) significantly exceeded actual levels of support for violence among Republicans (t[620] = 12.2, P < 0.001), constituting a 245% overestimate.

Study 1, a preregistered nationally representative survey run in October 2020. SPV is support for partisan violence, measured on four slider items from 0 to 100. The quote runs contiguously from the sentence introducing Figure 1 to the end of the accuracy test, with nothing elided, so the stored pair spans the four sentences the paper needs to reach it. Republicans made the mirror error, putting Democrats at 37.1 against an actual 9.3.

However, nonpatients’ ratings for the lung disease scenario (M = 54.4, SD = 23.9) were significantly lower than the self-reported QoL for the lung disease group as a whole, t(1,087) = 10.20, p < .001, and for the 115 lung disease patients who indicated that their own health condition was similar to the lung scenario (M = 72.3, SD = 21.0), t(569) = 7.32, p < .001.

456 nonpatients rated a scenario describing a lung condition on the same 0 to 100 scale every participant used for their own quality of life, anchored at 0 for a quality of life no better than death and 100 for an ideal one. The 72.3 is what the 115 lung disease patients who judged their own condition similar to that scenario reported about themselves, and the 569 degrees of freedom are those two groups together, which is where n comes from. What the paper sets out to compare is patients' and nonpatients' ratings and rankings of the patients' own disease, which is the comparison the stored pair makes for the lung group. Anyone re-checking 54.4 meets it first as the mean age of the patients in the same paper, 54.4 years against the nonpatients' 49.3, a collision of the same kind the spotlight row's note describes. The scenario is written in the second person and describes becoming short of breath after walking briskly for two blocks or up one flight of stairs, so the 54.4 stored here is a rating of life under a described condition. Riis and colleagues found the same divergence in 2005 by sampling dialysis patients' mood as they lived it.

Similarly, nonpatients’ ratings for the diabetes scenario (M = 63.6, SD = 24.0) was significantly lower than the self-reported QoL for the diabetes group as a whole, t(1,086) = 9.39, p < .001, and for the 80 lung disease patients who indicated that their own health condition was similar to the diabetes scenario (M = 75.9, SD = 20.5), t(534) = 4.33, p < .001.

The second of the two conditions the same survey covered. 456 nonpatients rated a scenario describing diabetes on the same 0 to 100 scale every participant used for their own quality of life, and the 75.9 is what the 80 patients who judged their own condition similar to that scenario reported about themselves. The 534 degrees of freedom are those two groups together, which is where n comes from. The sentence calls those 80 patients lung disease patients, which the parallel lung sentence and the arithmetic both show to be a slip for diabetes patients, since 456 and the 80 matched patients are the two groups the test compares. The scenario is written in the second person and describes checking blood sugar by pricking a finger several times a day with insulin injections three times a day and no complications of the disease, so the 63.6 stored here is a rating of life under a described condition.

For control subjects, the average estimate of mood while imagining life under the hemodialysis scenario was −0.38 (SE = 0.11). This was a very large and significant underestimation of the EMA average mood of the hemodialysis patients (M = 0.70, SE = 0.07)

Forty-nine haemodialysis patients carried a Palm Pilot for a week and rated the mood they had been in just before each prompt, and forty-nine healthy controls matched to them on sex, race, age and years of education did the same, which is where n comes from. The stored pair sets the controls' estimate of mood under a scenario describing a haemodialysis patient's life against what the patients recorded through that week, on the five-level scale Table 1 describes as running from 2 for very pleasant to −2 for very unpleasant. Controls imagined having been on dialysis either for a year or for as long as their matched patient, whichever was longer, and the patients rating that same scenario for themselves came to 0.63. Anyone re-checking meets 0.70 twice in Table 1, once as the patients' measured average and once as their recall of the week they had just spent carrying the device, where the match is a coincidence of two independently reported means.

Across all issues, liberals and conservatives overestimated opponents’ approval of unambiguously immoral behavior. For example, while 8.75% of Democrats, and 11.32% of Republicans in our sample labeled cheating on a spouse as acceptable, 27.36% of Republicans (and 37.91% of Democrats) reported that the average opponent viewed this behavior as acceptable.

Study 2, a Mechanical Turk survey of 240 Democrats and 106 Republicans, which is where n comes from. The example the paper works through is cheating on a spouse, one of seven acts it treats as unambiguously wrong, and the two halves of the stored pair come from the two partisan subsamples of that single survey: 37.91% is the share of Democrats who judged that the average Republican finds the act acceptable, and 11.32% is the share of Republicans who judged it acceptable themselves. The two are not quite the same quantity, since the estimate is a binary judgement about a modal opponent and the measured figure is an endorsement rate, and this sentence is where the paper sets them against each other. Four percentages appear in it, and 27.36% against 8.75% is the mirror pair with Republicans reading Democrats. Study 3 ran the design again on a sample representative of age, gender and ethnicity, asked for exact percentages and paid half the participants for accuracy, which did not move their estimates.

Participants who agreed to perform the escort task predicted they would need to ask an average of 7.2 people to walk them to the gym before 1 person agreed, but they actually had to ask only 2.3 people on average, t(11) = 3.01, p = .01

Study 2, which repeated Study 1 on two more natural favours. Twelve participants predicted how many strangers on the Columbia campus they would have to stop before one agreed to walk them to the university gym, then went and asked, with the experimenter following at a distance to keep the tally. Every one of them completed the task, which meant walking at least two blocks with an escort. The other favour was borrowing a cell phone, where 17 participants predicted 10.1 approaches against an actual 6.2. Quote taken from the author manuscript in Cornell eCommons, whose pagination differs from the published article.

Participants predicted they would need to approach significantly more people (M = 210.3, SD = 271.2) than they actually needed to approach (M = 122.2, SD = 118.6) in order to reach their fundraising goal, paired t(87) = 3.13, p = .002

Study 3, the field version, run on volunteers raising money for Team in Training, a division of the Leukemia and Lymphoma Society whose runners must bring in between $2,100 and $5,000 before their race or make up the difference themselves. 124 volunteers predicted how many people they would have to approach and 91 answered the second questionnaire once their deadline had passed, 88 giving both figures, which is where n comes from. Both are counts the volunteers reported, and the authors compared the sponsor counts of 82 of them against Team in Training's accounting records and found no significant difference. Quote taken from the author manuscript in Cornell eCommons, whose pagination differs from the published article.

In Study 1, participants recruited on Prolific—with quotas for age, race, and gender—vastly underestimated their peers’ agreement with 15 pro-diversity statements. Averaged across the 15 statements, actual support was 82.26%, whereas perceived support as 55.09%, a 27.17% difference

Study 1, 1,001 US adults recruited on Prolific with quotas for age, race and gender. Each of them said whether they agreed with 15 statements favouring diversity and inclusion, among them that racial diversity benefits the country and that state taxes going to college scholarships for students from minority groups is acceptable, then estimated the percentage of US adults who would agree with each. The 82.26% is that same sample's own agreement rate, so the figure the estimate is set against is the agreement rate of these 1,001 respondents, when what they were asked to estimate was the US adult population. The paper records 50.50% of them identifying as Democrats. Study 2 found a 27.61-point gap on a second Prolific sample and Study 3 found 12.73 points on a Qualtrics sample representative of the US population. The sentence reads "perceived support as 55.09%" and the missing verb is the paper's own.

Participants in the cell phone group predicted they would have to approach an average of 10.1 people to complete their task of getting 3 people to agree to their request. In fact, they needed to ask only 6.2 people on average, t(16) = 2.1, p = .05.

Study 2's other favour. Seventeen participants predicted how many strangers they would have to stop on the Columbia campus before three agreed to lend a cell phone for a call to the laboratory, then carried a tally sheet and went and asked. Each borrowed phone was confirmed by the experimenter answering the call. The 3 in the sentence is the number of successful calls each participant needed, the 16 the degrees of freedom and the 2.1 the t value. Four of those who began the task failed to complete it. Quote taken from the author manuscript in Cornell eCommons, whose pagination differs from the published article.

As before, liars overestimated the detectability of their lies: On average, they predicted that 44.3% of their fellow players would detect their deception; in fact, only 32.4% did so, paired t(7) = 2.91, p < .05.

Study 1c, which yoked each of the five players in a session to a personal observer who received the same card at the same moment and so knew whether it called for a truth or a lie. Eighty Cornell students took part across eight sessions, 40 as players and 40 as yoked observers, and the stored pair belongs to the 40 players. Those observers put detection at 25.3%, well below the liars' own 44.3%, which is what rules out an abstract belief that lies are easy to spot and rules out the curse of knowledge. Analyses were run at session level, so the 7 in the sentence is the degrees of freedom across the eight sessions.

A paired samples t-test revealed a significant effect of rating type on liking, mean difference = 0.75, 95% CI = [0.39, 1.11], t (35) = 4.20, p< 0.001, d=0.70, with participants estimating their conversation partners liked them (M=4.71, SD=1.00) significantly less than their conversation partners actually liked them (M=5.46, SD=0.93), which indicated that a negative deviation effect existed

Study 1a, an independent replication of the liking gap run on 72 Chinese students in 36 pairs of strangers, using items taken from Boothby's 7-point scale. After a five-minute conversation one member of each pair rated how much they liked their partner while the other estimated how much they were liked, so the 5.46 is one person's report of liking set against the other person's estimate of it. The 4.71 and the 5.46 are the two liking means and every other figure in the sentence belongs to the paired t test. Later studies in the same paper repeated the design on fresh samples, and the gap widened to 3.96 against 5.40 among pairs who communicated without speaking.