Citron (1952), Chapter Two: audience preference data

Levi & Goldberg, transcribed 2026-10-13 from the page images of

• Citron, A. F. 1952. Experiments on Answering Anti-Minority Remarks. PhD dissertation, Columbia University. Chapter Two, "The Experiments on Audience Preference", typed pp. 32-57, and the appendix, typed pp. 176-187 and 191. PDF: (held by the authors) (typed page = PDF page minus 6).

All page numbers in these files are the typed page numbers. Every number was read from the page image; the OCR text was used only to find pages. No cell was illegible, so no value is blank for that reason. Citron published no respondent-level data, and every file here holds a published aggregate or a sentence of his text.

Check before use: python3 scripts/verify_preference.py recomputes 389 printed totals, percentages, chi-squares and cross-table identities from these files. All pass. The 28 inconsistencies inside Citron's own typescript are listed in the script's KNOWN dict and in section 4 below; each was confirmed on the page image, and each entry is pinned to the value the transcribed cells give, so a later change to a cell cannot hide behind it.

1. The experiments

Citron ran six "rounds" of short plays. In each, "Jones" loses his turn in an employment-office waiting room to "Goldstein" and says that Jews are all alike, and a second man ("Stevenson") answers him. Rounds I to III were pilots, and Chapter Two reports no data from them. Round IV is the preference experiment proper: 537 people in 33 audiences (ten church and similar groups, eleven union groups, twelve groups of Times Square passers-by) each saw the scene twice with two different answers and then said which answer was better. The answers crossed two contents (American tradition, individual differences) with two manners (low: calm, poised; high: emotional, aggressive, militant), and each audience saw one of four pairings in which the first and second answers differed on both content and manner. About six people per audience (144 in all) were interviewed privately; the rest marked a discussion ballot. Rounds V and VI were run mainly for attitude change (Chapter Three). In those rounds a four-minute game separated the two answers, and the answerer was Stevenson (Round V, 293 Times Square passers-by) or "Levy", who says he is Jewish (Round VI, 121 Times Square and 91 Newark passers-by). Chapter Two gives no dates for any round.

Rounds V and VI appear to be the Stevenson and Levy days of the 1950 article's Experiment III (October to December 1946) in the main dataset; Chapter Two does not say so, and the match rests on the shared game, answerers and Newark sessions.

2. Files

Columns

pref_tables.csv: table (Citron's name, e.g. "Table Two"), page, row_label and column_label (Citron's labels, lightly normalized), value (as printed: 43 means 43 per cent), unit (count or percent), is_printed_total (yes for any printed row or column total, including Table Three's "combined" rows, which sum two reply rows), notes.

pref_t02_preference.csv: presentation (A to D as lettered in Table Two), position (first, second), content (american_tradition, individual_differences), manner (low, high), answer (low_american, high_american, low_individual, high_individual, the keys of the main dataset), group (union, church_y, street), n_preferring (people in that group and presentation who preferred that answer), page.

pref_sample.csv: table (e.g. "Item Eight Table 4"), title, page, category, value, unit, is_printed_total, describes, notes. A blank percent is a printed "-".

pref_text_numbers.csv: id (used by the verifier), page, statistic, value (a number, or a phrase such as "one-fourth" when Citron gives a fraction in words), unit (count, percent, chi_square, p_value, fraction, ratio, statement), df for chi-squares, refers_to, quote (a short extract of the sentence), notes.

pref_design.csv: round, item, value, page, notes.

pref_instruments.csv: instrument, item, wording, response_options, page, describes (which chapter's experiments the instrument belongs to), notes.

3. Things a user should know

1. Presentation letters. The text (pp. 34-35) letters the four pairings A to D one way, and Table Two (p. 40) letters them another way: Table Two's A is the text's C, and B is the text's D. Table Two's row labels agree with the p. 39 account of the one interview-ballot discrepancy, so the files keep Table Two's letters and labels. Use content, manner and position, not the letters.
2. Interviews and ballots are pooled. Citron compared interview and ballot preferences in each of twelve presentation-by-group cells and found one significant difference (union audiences, individual differences low first: 4 of 11 interviewees preferred the first answer, against 0 of 36 ballots). He then pooled the two sources. Table Two holds the 532 of 537 subjects who chose; the other five gave no opinion.
3. Each Table Two subject counts once, for one of the two answers he or she saw. A preference for the second answer is also a preference against the first, so order, manner and content cannot be read off the row totals separately. Citron separates content from manner and order by comparing presentations A and C, where low manner came first (p. 41).
4. Table Three does not reconcile with Table Two. Table Three counts 512 choices, not 532, and its cells run a few below the matching Table Two cells (low American tradition seen last: 117 of 131 against 121 of 137). Citron gives no reason. Table Eleven's Round IV column is built from Table Three, not Table Two: 350 low-manner choices = 236 (low seen last) + 114 (low seen first).
5. Tables Four and Five use the same interviewees. Table Four's first column counts 142 interviewees who described the two answers, and its second column counts the 119 of them who gave a reason for their choice. Table Five splits roughly the same 119 by the manner they chose (it sums to 118).
6. Tables Six and Seven total 491, fewer than the 532 with a preference, because they leave out other religions and other occupations. Table Six percentages run down each religion's column; Table Seven percentages run across each occupation's row.
7. Table Eight counts the same 144 interviewees three times, once per question. Citron's chi-square of 72.75 on 6 df treats the three columns as independent samples. Table Nine comes from 142 discussion ballots, which had no "other" category.
8. Percentages forced to 100. Eight printed percentages (four in Tables One to Twelve, four in Item Eight) are one point off the rounded value, and each correction makes its column sum to 100. The data keep the printed value.
9. Chi-squares. Citron mixes methods. The goodness-of-fit tests against a 50:50 split use the Yates correction for order (102.05), manner (71.48) and the religion and occupation subgroups (2.94, 4.19, 4.89, 3.93) and no correction for Tables Four and Five (11.84, 4.57, 27.84). His Fisher p of .004 (p. 39) equals twice the one-sided exact p (.0019). Nine of his chi-squares do not reproduce from his own counts by any standard method tried (section 4). For seven of the nine the printed and recomputed values differ by about one point or less; the other two are Table Ten, printed on 2 df for a 3 × 3 table, and Table Eleven, where a 2 × 2 test gives 19.05 against his 31.90.
10. Peoria. A footnote (p. 41) reports 67 church-group subjects in Peoria, Illinois, of whom 78% chose low manner. They are not in any table.
11. Jewish subjects. Chapter Two says all interviewees were non-Jews but does not say whether Jewish audience members were dropped from the 537. Item Eight's religion table has no Jewish category.
12. Appendix labels. The appendix numbers its sections by experiment, not by chapter. Its "Chapter One" items (pp. 176-187) belong to Chapter Two, and its "Chapter Two" items (pp. 188-193) belong to Chapter Three. The discussion ballot on p. 191 (Item Four) and the sample description on pp. 192-193 (Item Five, Times Square 428 and Newark 92) therefore describe Chapter Three's experiments; the ballot's wording is in pref_instruments.csv with describes set accordingly, and Item Five is in sample_exp3_detail.csv. The Round IV discussion ballot is not reproduced anywhere; its questions are known only from the text (pp. 40, 49, 51).

4. Inconsistencies in Citron's typescript

Each was checked on the page image. The verifier prints all 28.

• Rounding (8): Table One mildly prejudiced total 69.46% printed 70; Table Five content alone, high manner 59.46% printed 60; Table Eight do nothing, most people 44.44% printed 45 (the text on p. 51 repeats 45); Table Ten do nothing, mildly prejudiced 24.49% printed 25; Item Eight Caucasian 85.47% printed 86, sex unknown 0.37% printed 1, central and western U.S. 12.48% printed 13, Protestant 48.42% printed 49.
• Chi-squares (9) and one df: Table One 2.86 (counts give 3.12); union A vs C 26.70 (26.92 with Yates); street A vs C 3.56 (3.48 with Yates); low vs high seen last 76.66 (77.73, or 76.00 with Yates); religion 11.40 (10.87); occupation 15.13 (14.41); Table Eight 72.75 (72.59); Table Ten 10.67 on 2 df (11.26 on 4 df); Table Eleven 31.90 (19.05). The Table Eleven figure lies close to a goodness-of-fit test of Round V against Round IV's 68% (30.70 to 32.37), which may be what Citron computed.
• Text against tables (10): p. 46 says 61 low and "twelve" high of 76 citing manner; Table Five, the 20%, and the chi-square of 27.84 all fit 15, the manner-alone count (12) plus the three who cited both. p. 47 says 57 cited content; Tables Four and Five both give 55. p. 36 gives 49% Catholic and 42% Protestant, the reverse of Item Eight Table 8 and of Table Six. Table One's group sizes (159 church, 196 union, 182 street) differ from Item Eight Table 1 (160, 194, 183), though both total 537. Table Two's 161 church and Y subjects with a preference exceed Table One's 159 church subjects. Table Five's manner-alone count (63) and total (118) fall one short of Table Four's second column (64, 119).

The chi-square of 2.12 for content among interviewees (p. 47) implies, with 57 subjects and no correction, 34 for American tradition and 23 for individual differences. Citron does not print that split, and the files do not record it as data.

5. What the tables report

These summaries describe Citron's printed counts; the inconsistencies above apply.

• Manner. In Round IV, 364 of 532 subjects (68%) preferred the low-manner answer (p. 41). Among answers seen last, the low-manner answer drew 89% of choices (236 of 266) and the high-manner answer 54% (132 of 246) (Table Three, p. 43). In Table Two the low-manner answer drew 66% of union choices (125 of 189), 73% of church and Y choices (117 of 161) and 67% of street choices (122 of 182).
• Order. 383 of 532 (72%) preferred whichever answer came second (p. 41). With a four-minute game between the answers (Round V), 81% of votes went to the second answer and the low-manner share fell to 53% (155 of 293), against 68% in Round IV (Table Eleven, p. 54).
• Answerer identified as Jewish. With the same procedure as Round V, 68% (144 of 212) preferred the low-manner answer when Levy gave it, against 53% when Stevenson did (Table Twelve, p. 56). Citron calls the difference significant and prints no statistic; a 2 × 2 chi-square from his counts is 11.5.
• Content. Overall preference between the two contents was close (p. 57). Where manner and order pulled against each other (presentations A and C), union audiences favoured American tradition (71 of 92 choices) and church and Y audiences favoured individual differences (48 of 73) (Table Two, p. 42). Catholics (57%) and labor (57%) leaned to American tradition; Protestants (56%) and professionals (67%) leaned to individual differences (Tables Six and Seven, p. 48).
• What to do. Of 144 interviewees, 56% said the right thing was to answer the remark, 15% thought most people would, and 35% said they themselves would (Table Eight, p. 50). Among the 26 least prejudiced interviewees, 73% named answering as right; among the 20 most prejudiced, 35% did (Table Ten, p. 52).