Kurisutina
Library/How minds change

Haslam et al. · 2014

Meta-Milgram: an empirical synthesis of the obedience experiments

PLoS ONE 9(4): e93927, doi 10.1371/journal.pone.0093927; open access (CC BY), PMC3976349. Read by researcher R5a (change without good reasons), 5 October 2026. Provenance: papers/change/haslam2014_meta_milgram.provenance.json.

What was read

  • Read in full: the Europe PMC full-text XML, converted with tools/pmc2txt.py (self-check passed): abstract, introduction, Tables 1–6, methods, results, discussion, funding statement, the 20 references.
  • Not read: Figures 1–2 as images. Figure 1 counts textbook coverage of each condition; Figure 2 is the coding schematic. Both are described in the text. No supplement.

What they did

  • Data. A re-analysis, with no new data, of Milgram's obedience programme (August 1961 to May 1962): 23 conditions, 780 participants.
    • "Condition 21" (predictions by psychiatrists and laypeople) is not an experiment.
    • Conditions 10 and 12 are excluded because going to 450 V there meant disobeying.
    • Analysed: 21 conditions, N = 740 (16 conditions of 40, five of 20).
    • Rates were confirmed against Milgram's handwritten summary sheets in the Yale archive.
  • Outcome: delivering the 450 V shock (yes or no).
  • 14 codes in six groups:
    • experimenter: number, illegitimacy, non-directiveness, inconsistency;
    • teacher: female gender, group pressure to obey, group pressure to disobey;
    • learner: vulnerability, rights expression;
    • experimenter–teacher relation: distance;
    • teacher–learner relation: intimacy, proximity (0–3), indirectness;
    • setting: low status (Bridgeport).
  • Analysis: a χ² test per code, then one logistic regression with all codes.

Main results (verified)

  • Overall: 323 of 740 (43.6%) went to 450 V.

  • Baseline (voice feedback): 25 of 40 (62.5%). The paper: this "exceeded by a factor of 500 the figure estimated by psychiatrists who read the study protocol" (citing Milgram 1974). That implies an estimate of about 0.1% (my computation).

  • Rates by condition (number obeying of n; percentages my computation):

    Condition Obeyed Condition Obeyed
    1 No feedback 26/40 (65%) 14 Carte blanche (no command) 1/40 (2.5%)
    2 Voice feedback (baseline) 25/40 (62.5%) 15 Two experimenters disagree 4/20 (20%)
    3 Proximity 16/40 (40%) 16 Experimenter becomes learner 13/20 (65%)
    4 Touch 12/40 (30%) 17 Teacher in charge (peer authority, directive) 11/20 (55%)
    5 Coronary trouble 26/40 (65%) 18 No experimenter (phone) 9/40 (22.5%)
    6 Different actors 20/40 (50%) 19 Authority from afar (recording) 15/40 (37.5%)
    7 Two peers defy 4/40 (10%) 20 Women 26/40 (65%)
    8 Learner's proviso 16/40 (40%) 22 Peer authority, non-directive 4/20 (20%)
    9 Two peers urge obedience 29/40 (72.5%) 23 Bridgeport (low-status setting) 19/40 (47.5%)
    11 Group choice (peers escalate, no command) 7/40 (17.5%) 24 Learner is a friend or relative 3/20 (15%)
    13 Non-trigger (another person shocks) 37/40 (92.5%)

  • Bivariate (Table 4; proportion obeying when coded 1 against 0):

    • non-directiveness .12 against .49 (χ² 47.09);
    • indirectness .93 against .41 (41.03);
    • group pressure to disobey .10 against .46 (19.47);
    • female .65 against .42 (7.84);
    • distance .33 against .46 (7.24);
    • intimacy .15 against .44 (6.86);
    • vulnerability .54 against .42 (6.44);
    • inconsistency .20 against .44 (4.67).
    • Not significant: number, illegitimacy, group pressure to obey (.45 against .43), rights expression, low-status setting (.48 against .43).
    • Proximity: Spearman r = −.37.

  • Logistic regression (Table 5), Nagelkerke R² .30. Eight codes predict independently:

    Code B (SE)
    Non-directiveness −2.79 (0.39)
    Group pressure to disobey −2.49 (0.60)
    Indirectness +2.22 (0.67)
    Intimacy −2.03 (0.69)
    Inconsistency −2.01 (0.73)
    Illegitimacy +1.37 (0.47)
    Distance −1.14 (0.38)
    Proximity, linear −1.14 (0.34)
    • Illegitimacy goes the "wrong" way; the authors suspect collinearity.
    • Group pressure to obey: +0.78 (0.40), p = .052.
    • Female gender, vulnerability, rights expression, number of experimenters and setting are not significant once the rest are controlled.

  • Code sets (Table 6, Nagelkerke R²): experimenter .116; teacher–learner relation .110; teacher .052; experimenter–teacher relation .013; learner .012; setting < .001. The authority side (.129) and the victim side (.122) are about equal.

  • The authors' reading:

    • Obedience to 450 V "rarely occurs if the authority figure does not give explicit commands", even when peers push to escalate.
    • A dissenting second experimenter "almost completely overrode" the first.
    • A social bond with the learner cut obedience to under a quarter of the stranger rate.
    • Setting prestige and authority legitimacy have been over-estimated.

Limits

  • Situational codes only. No individual differences were analysed. Participants' interpretation and belief in the set-up were not measured. The experimenter often went beyond the four standard prods (Gibson 2013 and Perry 2012, as cited).
  • The conditions are ad hoc, not factorial. The codes are collinear: conditions 5/6 and 18/19 are coded identically. Cells are small (20–40).
  • Population. Men only except condition 20; the New Haven area, 1961–62.
  • Inconsistencies found:
    • Low status. Table 5 gives p = .614; the text gives p = .301. Wald 1.07 on 1 df gives p ≈ .30 (my computation), so the table repeats Table 4's p.
    • Proximity quadratic. B −0.59 with SE 0.32 implies Wald ≈ 3.4, not the printed 0.03 (p = .855).
    • Vulnerability. B 0.06 with SE 0.37 implies p ≈ .87, not .987.
    • "Group pressure for experimenters to disobey" should read teachers. "As Table 4 shows", for the code-set comparison, should read Table 6.
  • Checked:
    • 323/740;
    • the sum of the condition counts;
    • N = 16 × 40 + 5 × 20;
    • the other Wald values equal (B/SE)².

What it means for Kurisutina (inference)

  • Obedience to a directive authority is large and set by the situation, not the person's traits.

    • 62.5% at baseline, 2.5–92.5% across variants.
    • The strongest levers: an explicit command (without one it nearly vanishes), dissenting peers (10%), an indirect role (92.5%), a bond with the victim (15%) and a divided authority (20%).

  • Allies resisting matter far more than peers urging compliance: −52 points against +10.

  • The ×500 prediction gap is a human reference for how badly observers underestimate others' obedience. Self-predictions are not reported here.

  • Obedience is an action, escalated stepwise under a present instructor. A replica that only states beliefs cannot show it. Reproducing it needs:

    • an action channel;
    • escalating requests with refusal points;
    • the situational levers above as inputs.

    This paper supplies a human rate for each variant.

  • Contamination risk. Milgram is among the most described experiments in English text. A battery item built on it needs a corpus filter and perturbed variants (R4d §5).

This summary is our record of the paper, written after reading the full text and published as written; links into our own repository have been removed.