Methods in Full

1. Origin of the design

The design predates access to the subject pool. Eleven dated versions run from January 2014 to February 2016:

Entrepreneurs_Jan2014 → _29July14 → _10Dec2014 → Feb.9.2015 → March.10.2015 → March.26.2015 → May29.2015 → Oct11.2015 → Oct17.2015 → Nov.2015 → Entrepreneurs.Feb.2016.docx (fielded)

The January 2014 document already contains the motivation and the measure list. Access to Emlyon was negotiated in August 2014 — that is, the design was written first and the subject pool found for it. Six parallel Tests.*.docx track the instruments. Cooper, Saral, and Villeval are authors on the early versions; Dubard Barbosa joins later, at a point not yet established.

Designed but never fielded: a moonlighting variant of the trust game, a creativity task (Erat and Gneezy rebuses), ambiguity measures, and the DOSPERT scale. The later NSF proposal would have swapped Raven’s matrices for the cognitive reflection test, in a follow-on study that was not funded and never ran.

2. Timeline

3. Sessions

Dates come from the z-Tree output filenames, which are authoritative. Two folder names in the archive are wrong — folder 7 is labelled 26 May and folder 6 is labelled 25 June; the filenames say 24 May and 24 June respectively.

Session 1 stands six weeks before the main block, and is the session whose trust game differs. Recruitment for wave 1 was by Doodle poll, with Dubard Barbosa personally bringing the Emlyon cohort to the laboratory. Survey-to-laboratory matching was performed by hand, on participant names, with at least one case that could not be resolved.

Who is in the analysis sample, and who is not

The z-Tree files hold 131 records. One person took part twice, so those are 130 distinct people. The chapter uses 122.

The eight. All are wave 1, from sessions 2 (1), 3 (1), 4 (1), 5 (4) and 6 (1) — none from session 1. They played all five games and their laboratory data are complete; what is missing is the intake survey, and since the three-way classification is built from the intake responses on entrepreneurial experience and intent, they cannot be placed in a group. Age is missing for all eight and gender for three.

The ninth. Subjects 206 (session 2, 26 April 2016) and 709 (session 7, 24 May 2017) are the same person, recruited first at Emlyon and again thirteen months later through the GATE pool under a second identifier. The second wave ran the same five games in the same order, so her 2017 decisions were made by someone who had already played them. She is also classified differently on the two occasions — future entrepreneur in 2016, non-entrepreneur in 2017 — and answered the 2018 follow-up under both identifiers. We keep the first participation and drop the second.

By wave, the analysis sample is 91 from wave 1 and 31 from wave 2.

This is separate from the trust game exclusion. All twenty session-1 subjects are in the 122 and contribute to every measure except trust, which they are excluded from because that session ran a different version of the game (Section 5.1). That is why trust rests on 102 observations and the other laboratory games on 122. The two exclusions do not overlap: none of the eight unclassified subjects came from session 1.

Earlier presentations of this study used 123, keeping both of the repeat participant’s records. Re-running every specification both ways leaves each coefficient intact to two decimal places and moves no p-value across a conventional threshold — the entrepreneur marginal effect on tournament entry is 0.331 against 0.333, the conditional contribution coefficient in the follow-up 0.940 against 0.944, and the 11–20 request 0.042 against 0.040. Raw means shift only in the non-entrepreneur column, by at most 0.12.

4. The intake survey, by instrument

Wave 1 received two separate instruments; wave 2 received one combined instrument (Survey III, May 2017). Placement below is established from the surviving instruments and confirmed against coverage in the data.

Survey I — demographics; educational and professional background; entrepreneurial history (work_status, year of first business, firm size, sole ownership, reasons for starting, obstacles, failure, parental and peer entrepreneurship); entrepreneurial intentions; entrepreneurial self-efficacy (searching, planning, marshalling subscales); three items on attitudes to risk in choosing and holding a job, and whether fear of failure would deter them from starting a business; and the Cognitive Style Index.

Survey II — six blocks, in order: login; fast risk (Dohmen et al., 2011); fast patience (Vischer et al., 2013); the Big Five (Gosling TIPI, ten items); locus of control (Levenson, 1981, 24 items); grit (Duckworth, 12 items).

Survey III — the wave-2 combined instrument, in French, reproducing both of the above except the risk-perception items, which were not asked.

Evidence for the placement of the Cognitive Style Index: among wave-1 participants, all 80 with a computed index also completed the Survey I self-efficacy block (100%), whereas only 81% also completed Survey II. It therefore belongs to Survey I. The index is administered as 38 true/false/uncertain items (tf1–tf38, each with 117 responses) but scored on 37.

Item 29 was omitted from the French version. tf29 holds no Vrai, Faux or Incertain responses at all — only the 35 English ones, against 96 blanks — while every other item carries both languages. Survey III’s own preamble confirms it, promising “37 déclarations.” The scoring block for item 29 is therefore commented out in cleaningII.do (ll. 1113–1115), and the scored series runs cs1–cs28, cs30–cs38. Scoring it would have restricted the index to the English-language subsample, so the omission is correct and cognitive_style is legitimately a 37-item scale.

One loose end: the variable’s label reads “max possible score of 76,” which is the original CSI’s 0–2 scoring. This implementation scores each item 0/1, so the maximum is 37, and the observed maximum is indeed 37.

Coverage, analysis sample of 122

The wave-1 shortfall on the Survey II block (71 of 91) is the single largest source of missing intake data, and it is an artefact of splitting the battery across two separately-completed instruments.

5. The laboratory session

GATE-Lab, Écully, in the Lyon metropolitan area. z-Tree. Approximately 2.5 hours. Order of play, recorded in ProgramNotes.txt and in no deck or write-up:

login → trust → 11–20 → Raven’s matrices → market entry → tournament entry → public goods → results → questionnaire

11–20 game parameters, from the z-Tree source (strategic_sophistication.ztt, March 2016, and strategic_sophistication2.ztt, March 2017 — the two files are identical): subjects randomly paired, each requests an integer from 11 to 20 points, exchange rate 0.1 so 10 points = €1. Payoff is Profit = request + 20 when the request is exactly one less than the other player’s, Profit = request otherwise. The game was unchanged across all eight sessions.

⚠ The superseded lab_instructions.docx contains two exchange rates in the same document, 10 points to €1 in one paragraph and 4 points to €1 in another. The z-Tree source and the 13 March 2016 instruction file both say 10, so the 4 is a drafting error that never reached the software.

The choice of the 11–20 game over the p-beauty contest is attributed on the slides to p-beauty being too familiar to business students. No contemporaneous source for this has been found; treat it as recollection.

5.1 The session-1 trust game

Verified in the z-Tree source. trust.ztt is dated 9 March 2016, two days before session 1, and its on-screen text reads “Each of you has been given 20 points… Player 1 has the opportunity to give a portion of these 20 points to Player 2.” The session-1 data corroborates this independently: it records player ∈ {1,2} with a receive column, where sessions 2–8 record givegroup instead.

5.2 Instruction delivery changed as well

Early sessions used a verbal script read aloud by the experimenter (lab_instructions.docx, 10 March 2016). On 22 March 2016 all programs were rewritten so that instructions were delivered on screen — ProgramNotes.txt records “instructions are now built into the programs - no verbal script.” Both scripts survive. This is a separate change from the trust redesign three days earlier and affects all five games; it is not the reason session 1 is dropped.

5.3 Why the market entry game failed

Diagnosed 2026-09-22 from lab_instructions_13.3.2016.docx. Subjects kept a €5 stake by staying out. Entering placed them in a market of up to four entrants with capacity two, paying €16.50 to the top-ranked entrant and €8.50 to the second, ranked on twelve Raven’s matrices.

For an entrant equally likely to hold any rank:

Even under full entry the expected return is 25% above the certain €5. We make no claim that entry was a dominant strategy — that depends on risk preferences, since entering is a gamble over €0/€8.50/€16.50 against a certain €5, and on beliefs about relative ability. The defensible statement is simply that the parameters made entry very attractive.

Two features of the procedure reinforce it. The entry decision was taken before the quiz, so subjects had seen one sample matrix and had almost no private information about their own ranking. And the industry payoff was fixed at €25 regardless of entrant count, so the congestion that should discipline entry never bites hard enough. Camerer and Lovallo (1999) set the industry payoff so that entry beyond capacity is unprofitable; that tension is what makes excess entry evidence of overconfidence, and we removed it. Eighty-nine percent entry is close to what these incentives called for.

Boredom is ruled out, and this is a point in the design’s favour. Every subject attempted the twelve-item quiz whether or not they entered — the instructions ask non-entrants to attempt it anyway, and all 122 have a skilltotalcorrect score (13 non-entrants, 110 entrants, 100% coverage in both). Staying out therefore did not mean watching others work, and entering carried no additional effort cost. Both the effort cost of competing and the disutility of enforced idleness are absent, which isolates the payoff miscalibration as the cause rather than muddying it.

An incidental curiosity: those 13 non-entrants completed the task under effectively flat pay — their earnings were the €5 stake whatever they scored — and still solved 4.69 of 12 against 5.28 for the performance-paid entrants (t = -0.76, p = 0.45). Not a clean incentive comparison, since entry was self-selected and n = 13, but they plainly engaged with the task without being paid for performance.

Tournament entry is sound by contrast and should not be tarred with the same brush: piece rate €0.50 per correct answer, tournament €2.00 per correct answer to the winner of a group of four — a standard Niederle & Vesterlund structure.

5.4 Power

Removed 2026-10-05. An ex-post power section stood here, reporting minimum detectable differences against the differences observed. David’s objection was that power calculations belong before a study, not after it, and the corresponding passage was taken out of the chapter at the same time. The calculation is not part of this deposit, since nothing in the chapter reports it.

6. The online experiment

Built in oTree and deployed to Heroku — French on fathomless-wildwood-26204, English on secure-wildwood-80793. The fielded builds are ENT5 (English) and ENT5_fr (French). The five apps were first written by an external developer (GitHub user leliel12) in the repository otree_saral_2; Saral carried that build forward through ENT2–ENT5. A public snapshot sits at github.com/kjsaral/ENT, last pushed 8 June 2016; the fielded builds went to Heroku only. Source code is complete and archived.

Order of play, from the oTree app_sequence: login, time preferences, risk aversion, preference for skew, honesty, eyes test, results.

How many completed — revised 2026-10-05. 95 of the 122 subjects in the analysis sample completed Phase 3, and that is the figure the chapter now uses.

The number depends on which file you read, and the reason is chronology:

Wave 1 did the online experiment about a week after their laboratory session, in 2016. Wave 2 did it in early 2018 — roughly nine months later, using the links in ENT administration/participant tracking and pay/GATE_OnlineExperimentLinks_Feb2018.xlsx, with the cleaned imports dated 11 April 2018. ent_djc_3.dta was built before that collection happened, which is why it holds no wave-2 online data at all.

On the 122-subject analysis sample the counts are 76 (48/12/16) in ent_djc_3.dta against 95 (66/13/16) in ent_final2.dta. All nineteen recovered subjects are non-entrepreneurs or future entrepreneurs; the entrepreneur cell is 16 either way, which is consistent with sessions 7–8 being drawn from the GATE subject pool. The repeat participant never took the online experiment, so she does not enter the arithmetic, and there are no partial completions — all five tasks have data for the same subjects.

This supersedes an earlier note here which reported 78 and 76 and recorded the 98 in ent_final2.dta as unexplained. It is explained: it is the same 78 plus wave 2’s 20. The 2018 NSF pilot writeup’s figure of 78 was correct for the data that existed when it was written.

Do not use playerdecision1–10, playerriskchoice1–4 or playerskewchoice1–3: these columns appear populated on a naive count but contain no numeric data.

7. The follow-up

April 2018, Qualtrics, French and English waves, batch-emailed by Quentin Thevenet of GATE rather than by the authors. Twenty-two questions: current entrepreneurial status and intentions, firm details where applicable, refreshed demographics, repeated fast risk and fast patience, a repeated true/false block, and a new item on preference for working in teams.

The fee was randomised across €5 / €7.50 / €10 — proposed in November 2017 as €5/€10/€15 and changed before fielding, for reasons not on record.

The fee was an instrument for a Heckman selection correction, and it failed. Cooper, mid-fielding, 22 April 2018: “We need something exogenous which is correlated with their entry decision. This is exogenous, obviously, but does not appear to correlate.” There is consequently no selection correction — not for want of trying, but because the instrument was uncorrelated with response.

Response counts. ent_final2.dta holds 68 follow-up records; David’s ent_djc_4.dta holds 67. The difference is subject 507, present only in the former, most likely a late response arriving after the 1 May 2018 dataset was cut. Because one person responded twice (§9), 68 records correspond to 67 people and David’s 67 records to 66 people.

8. Payment

Wave-1 laboratory earnings were paid in cash, in person at GATE, by 15 April 2016, and only on completion of the online phase — the instructions state “your total earnings from today will be paid to you upon completion of the follow-up online portion.” Online and follow-up payments were made by Amazon voucher (.fr, .de, .com) and PayPal in October–December 2016 and April 2018.

Mean payoffs: €27 laboratory, €20 online, €7 follow-up.

9. The repeat participant, in detail

Who she is and why the 2017 record is dropped are in Section 3. Three further details, recorded because they bear on how independent the two participations were.

How it was found. Both identifiers matched a single external profile; nothing in the laboratory or survey records flagged it. Any comparable study recruiting a second wave from a different pool a year later would face the same risk.

Her decisions differ sharply between the two. Tournament entry 1 in 2016 against 0 in 2017; group contribution 54 against 8. Whatever the second participation measures, it is not a stable repetition of the first.

The 2018 follow-up holds two responses from her, one under each identifier. The factual answers agree — status, income, education, marital status — while the attitude sliders differ slightly, as one would expect of the same person answering twice. The chapter counts the response attached to the retained record.

10. Consent and disclosure

Two regimes operated. The laboratory used a signed French paper form (ConsentForm_10March16.doc, dated 11 March 2016 — session 1’s own date), covering voluntary participation, the right to withdraw, anonymity of decisions, absence of physical risk, and the in-person payment deadline. The online components used an English click-through form in US IRB style, including a tax-reporting note and the contact address kjsaral@webster.ch.

The multi-part structure was disclosed at the laboratory: “this is a multi-part experiment and today is the only portion that will take place in a laboratory. The follow-up parts will all be online.” Email addresses were collected on that basis.

11. Funding

• PALSE / UniLyon (MAELYSE), awarded 2015 — €10,000 as awarded, reported as €9,600 the following year. Funded the data collection.
• Webster University St. Louis, 2016 — $7,120, faculty research grants 2015–16 and 2017–18.
• Travel support from the John and Laura Arnold Foundation, $1,000.
• The 2018 NSF proposal (1853498) was not funded.

12. Materials available online

Ours, and publishable without restriction: all z-Tree programs including both versions of the trust game; the full oTree source in English and French; laboratory instructions in both the verbal and self-guided versions; French translations of the instruction screens; screenshots of every laboratory and online screen; both consent forms; and the Qualtrics .qsf files for the follow-up.

⚠ Not ours, and requiring clearance before posting. Shane’s contributor terms require us to “warrant originality and to clear any third-party material.” Several instruments are third-party: Raven’s Progressive Matrices (copyrighted, licensed through Pearson), the Reading the Mind in the Eyes photographs, and the validated scales — TIPI, Levenson’s locus of control, Duckworth’s grit scale, and the Allinson & Hayes Cognitive Style Index. Screenshots of our implementation reproduce some of this material directly, and so does the program source: 37 image files in the oTree eyes_mind app, in both the English and French builds, and 13 in each z-Tree build — the twelve Raven’s matrices plus a sample. Every other part of the code is clean, so publishing the programs means publishing the instruments unless those folders are stripped first, and stripping them leaves the eyes-reading and market-entry tasks unrunnable as published. What can actually be posted needs deciding before we describe the appendix as complete; the fallback is to cite the source papers and post only our own code, instructions and surveys.

Note that the English on-screen text for the redesigned trust game survives only as screenshot images and as literal strings inside trust2.ztt; the French version exists as extractable prose. For quotation, take the English wording from the program source.

13. Every measure tested the same way

The chapter discusses the measures that bear on its argument. This section tests every other measure collected across the three-way classification, on one uniform specification, so that what is null is null on the record rather than by omission. The seven laboratory games are not repeated here — they appear in Tables 2 to 4 of the chapter, each with the estimator appropriate to it.

Specification. Group means, then the coefficients on Future entrepreneur and Entrepreneur from an OLS regression on the classification with the chapter’s demographic controls (age, gender, broad income, broad education), robust standard errors, non-entrepreneur omitted. Sample of 122, as in the chapter.

This is a screen, not the estimator of record. OLS is used throughout so the columns read across measures; it is the wrong estimator for some of them. Where the chapter reports a different number, the chapter governs.

Coverage. The online-experiment measures rest on the 76 subjects with online data, a subsample selected on entrepreneurial classification. The first-wave-only survey items rest on 86 to 89, and the personality, locus-of-control and grit blocks on about 100, because the intake survey was administered in two parts. Entrepreneurial intention items are excluded, since they enter the construction of the classification.

Multiple comparisons. Counting the laboratory games, thirty-eight measures are tested. Four reach the 5% level where about two would be expected by chance. Survey risk tolerance and tournament entry are large enough to survive that arithmetic; openness probably is; the Reading the Mind in the Eyes result is not, and its standardised difference of 0.62 sits below what a sample this size reliably detects.

Risk and related survey items

Online experiment (incentivized)

Big Five personality

Locus of control, grit, cognitive style

Entrepreneurial self-efficacy

Task performance (incentivized)

*** p<0.01, ** p<0.05, * p<0.1. Produced by chapter_analysis.do.

What the sweep turned up

Three measures separate the groups and are discussed in Section 3.4 of the chapter: self-reported risk tolerance, which is the largest cross-sectional difference in the study; openness to experience, the only Big Five dimension that moves; and the Reading the Mind in the Eyes test, on which entrepreneurs score lower. Grit and the searching subscale of self-efficacy lean in the expected direction without reaching conventional significance.

Everything else is null: loss aversion, time preferences, honesty, all three locus-of-control subscales, the remaining four personality dimensions, cognitive style and the job risk items.

A note on the future-entrepreneur column. It reaches the 5% level on one measure of the thirty-eight, which is fewer than chance would give, and on the laboratory games its smallest p-value is 0.16. People who intend to start a business but have not done so are not distinguishable, in this study, from people with no interest in starting one. The exception is risk: the survey items order that group above the non-entrepreneurs on every one of them, though only the first reaches even the 10% level on its own.