Citron (1952), Chapters Four and Five: training experiments

These files transcribe the training experiments in Citron, A. F. 1952. Experiments on Answering Anti-Minority Remarks. PhD dissertation, Columbia University. Chapter Four ("The Experiment on Training Volunteers to Answer", typed pp. 97-136) and Chapter Five ("The Experiment on Training Trainers", typed pp. 137-163) supply the material, together with the appendix outline of the training course (typed pp. 194-200). Every value was read from the page images at 110 to 300 dpi. The OCR text served only to find pages. Page numbers in every file are the typed page numbers; the PDF page is the typed page plus six.

Citron published group-level aggregates only. No individual trainee's scores survive in the dissertation.

Check before use: python3 scripts/verify_training.py recomputes 110 printed totals, means, ranks and test statistics from these files. All 110 pass. Three printed figures in Citron's text cannot be reproduced from his own tables; the script lists them as known discrepancies and the data keep the printed values.

The two experiments

Chapter Four (New York volunteers). The Commission on Community Interrelations of the American Jewish Congress trained six groups of women, all volunteers from New York City chapters of the Congress's Women's Division, in a five-session course (about 1 hour 50 minutes per session, usually a week apart). Chapter Four gives no year. The six groups formed three pairs by neighbourhood. Within each pair one group, chosen at random, received "cultural analysis" (C.A.), a discussion of the cultural sources of people's inhibitions about answering in public, which took about one-fifth of course time. The other group (N.C.A.) spent that time on more role practice. One experienced group worker trained each pair: Trainer 1 had Groups Ac and A, Trainer 2 had Bc and B, Trainer 3 had Cc and C (a lower-case c marks the C.A. group). Session attendance ran from 9 to 31.

Citron measured four outcomes. A ten-item incident-situations questionnaire asked trainees to rate printed answers on a four-point scale (start of session 2 and end of session 5). A "reality test" staged an incident: six women drawn by lot answered a bigot played by the trainer on an imagined bus (session 2), and later a professional actress, a stranger, playing a bigot on an imagined subway (session 5). An end-of-course questionnaire collected self-rated confidence (1 to 10, with "before" and "end" marks made at one sitting) and an evaluation of the course. Observers also rated each trainer's security and each group's morale several times per session.

Chapter Five (Hartford trainers). In spring 1948 ten Hartford organizations each sent a lay trainer (eleven trainers, eight women and three men; the American Jewish Congress chapter sent two) to a central "Institute". The Institute met on Tuesday and Thursday evenings, with sessions on April 22, April 29 and May 6, 1948. Each trainer then ran a four-session workshop with his or her own group; the first two workshop sessions fell between the last Institute sessions. Outcomes were quasi-reality tests of the trainees (scored out of 50), retention, and a field test 60 days after training. In the field test trainees answered an actress's anti-Negro remark in a radio-station waiting room, and 15 untrained members of similar organizations formed a comparison group.

Files

Columns and keys

The eight table files share one long layout: table, page, row_label, column_label, trainer (1, 2, 3), group (Ac, A, Bc, B, Cc, C), cultural_analysis (yes for Ac, Bc, Cc; no for A, B, C), value (exactly as printed, so 8.00 keeps its zeros), unit, n (only where the text gives it), is_printed_total (yes for every total or mean Citron printed, so the verifier can recompute it), and notes. A blank value means the page prints a dash. No cell was illegible.

train_text_numbers.csv uses id as its key. Its experiment column is ny_volunteers (Chapter Four) or hartford_trainers (Chapter Five). value holds the number as printed, statistic names it (t, F, chi-square, mean, correlation), df holds the printed degrees of freedom, and significance_printed holds Citron's significance wording. quote is a short extract of the sentence.

train_course_outline.csv has session (1 to 5), track (all for sessions 1 and 2, cultural_analysis or no_cultural_analysis for the two columns of sessions 3 to 5, both for a session heading), code (Citron's outline numbering, for example III.B.2), level (0 = session, 1 = lettered heading, 2 = numbered item, 3 = lettered sub-item), text and page. Lettered and numbered sub-items below level 2 sit inside their parent's text.

Units and scales

• Table 34: mean rating of a printed answer on a 1-4 scale, 4 highest. All six groups are pooled.
• Table 35: reality-test score, 0-40. The observer rated content, poise, timing, sincerity and non-aggressiveness 0-5 each; the score is four times content plus the other four ratings, so content and manner carry equal weight (p. 122).
• Table 37: self-rating, 1 ("completely unable to answer") to 10 ("completely confident").
• Tables 38 and 39: counts of observer ratings. Both tables total 116 ratings with identical group counts.
• Table 40: ranks, 1 = best; ties share the average rank (2.5, 1.5).
• Chapter Five quasi-reality and field tests: 50 points, content 1-5 multiplied by five plus five manner qualities rated 1-5 (poise, non-aggressiveness, clarity, timing, sincerity).

What the verifier recomputes

The script checks every printed total in Tables 34, 36, 38 and 39 and the text's attendance range, losses and percentages. It recomputes the mean reality and confidence gains and their t statistics (as a range over printing precision), the four attendance and morale chi-squares, and three columns of Table 40 ranks. It also checks Table 33 against the text's descriptions of Groups Cc and A, the Hartford organization list, and the field test's percentage and Fisher p. Details worth knowing:

• Citron's morale chi-squares (32.1 and 1.57, p. 132) reproduce when Table 39 is collapsed into fair-or-below against good-or-above (32.5 and 1.58).
• His attendance chi-squares use session 1 as base and session 2 for Group C, as p. 126 says. They reproduce at 11.9, 3.94 and 5.93 against printed 11.8, 3.87 and 5.91.
• Table 40's attendance ranks follow the share of the base still present at session 5, and its morale ranks follow the mean of the five categories scored 1 to 5.
• Table 40's adjusted reality ranks come from an analysis of covariance whose inputs Citron did not print. Ranks 3 to 6 equal the ranks of the unadjusted final means; Groups Bc and Ac swap first and second after adjustment, which fits Bc's very low initial mean (2.5).
• The "critical variance" of 9.00 on p. 125 equals 18 x 18 / 36 x (27.5 - 26.5) squared, the mean square for a difference between two sets of 18 final-test cases.
• Table 33 can be checked only against text claims. A wrong age or Lentz cell would pass the script undetected.

Known discrepancies in Citron's text

1. p. 117, t = 4.51 (9 D.F.) for the overall drop in ratings. The ten situation means in Table 34 reproduce all three printed general means, and their paired t is 2.88 (p near .02). The printed 4.51 cannot be reproduced from the table.
2. p. 119, correlation of .75 between the initial ratings and the author's. The Table 34 columns give a Pearson r of .82. The final-rating correlation (.84) does reproduce. The OCR text reads the final figure as .81; the image reads .84.
3. p. 123, "thirty-one per cent of the women were completely blocked". The same page gives 33 initial cases, and no whole number of women out of 33 gives 31 per cent (10 gives 30.3, 11 gives 33.3).

Cautions

• The New York design compares C.A. with N.C.A. within three trainer pairs, so only three independent contrasts exist. Citron's t tests for gains use the six group means as units (5 D.F.).
• The reality tests sampled six women per group (three for Group Cc initially), and the two draws were separate, so the gains compare samples that need not contain the same women.
• Initial confidence was rated retrospectively at the end of the course (p. 129).
• Attendance fell unevenly: Trainer 3 lost half his starting participants. Women who dropped out answered no end-of-course questions. The F test on final confidence (5 and 71 D.F.) implies 77 respondents, while Table 36 shows 94 at the fifth session; the text does not explain the gap.
• Group C held no first session (Table 36 prints a dash) and started a week late with a longer second session covering two sessions of material.
• The text says all three background tests were given at the first meeting (p. 110). The appendix outline places the Otis test in session 1, the Allport test in session 3 and the Lentz test in session 4.
• Chapter Five calls the Institute "a total of six sessions" over three weeks (p. 142) and then describes a seventh and an eighth (final) session (pp. 142-153).
• The Hartford 84 per cent retention and its chi-square against New York (p. 158) cannot be checked; Citron prints no Hartford counts. The New York 68 per cent equals 94 of 139 in Table 36.
• The New York (0-40) and Hartford (50-point) test scales differ in range and in which qualities they rate. Citron's claim that the Hartford gain (20 to 43) is "almost exactly the same" as the New York gain (p. 156) rests on his judgment, and the files give no common metric.
• The Hartford quasi-reality t (11.06, 9 D.F.) implies ten units, presumably the ten workshop groups; the group means are not printed.
• The field test drew 16 trainees from 7 of the 10 groups. Two groups that Citron called among the best had dispersed for the summer and are absent (p. 161).
• The right margin of appendix p. 197 is cut off in the scan. Bracketed letters in train_course_outline.csv complete clipped words and were not read from the image.
• One Table 34 cell (Situation 4, author's rating, 3) is faint; a 300 dpi render reads 3, and the column's printed mean of 1.70 confirms it. One Table 38 cell (Group Bc, insecure, 5) reads 3 in the OCR text; the image and the printed total of 17 both give 5.