HOW GROUPS REACH AGREEMENT IN RISKY CHOICES:AN EXPERIMENT
JINGJING ZHANG and MARCO CASARI∗
This paper studies how groups resolve disagreement in lottery choices. In anexperiment, subjects submit individual proposals, exchange chat messages, and mustreach unanimity. Overall, group choices are more coherent and closer to risk neutralitythan individuals’. The proposal of the minority prevails in about one instance out of five.About one third of the groups do not reach immediate agreement after communication.In these groups, extrovert subjects are more likely to lead the group outcome thanconfused or conscientious subjects. The amount, equality, and timing of chat messageshelp us to predict which choice prevails in the group. (JEL C92, D81)
I. INTRODUCTION
Although economists model decision makersas isolated individuals, within firms and organi-zations decisions are often taken through delib-erations in groups and committees. Many ofthose decisions involve options with differentdegrees of risk. In the last decade economistshave produced a growing number of studies onthis issue.
In an experiment, we study decision-makingprocedures of individuals versus groups in aseries of choices between a safe and a riskyoption. How do groups aggregate individualpreferences when members are initially in dis-agreement? In the laboratory, one can design aclean set up, which is free from external con-founding factors, in order to better answer thisquestion. Eliciting risk attitudes for groups was
*This project was developed while both authors wereat Purdue University, which provided an excellent researchenvironment. We are grateful to Christine Jackson forher support and contribution of ideas. We thank AnyaSavikhin for valuable research assistance, Tim Cason forcomments on an earlier version of the paper as well asseminar participants at the IMEBE meeting in Alicante,Spain, CapuaLabSı meeting, Italy, ESA meeting in Lyon,France. Two anonymous referees made useful suggestions.J.Z. wishes to thank McMaster University for support duringher post-doc when the paper was completed.Zhang: Assistant Professor, Institute for Empirical Research
in Economics, University of Zurich, Blumlisalpstrasse10, Zurich CH-8006, Switzerland. Phone +41-44-634-5562, Fax +41-44-634-4907, E-mail [email protected]
Casari: Associate Professor, Department of Economics,University of Bologna, Piazza Scaravilli 2, Bologna40126, Italy. Phone +39-51-209-8662, Fax +39-51-209-8493, E-mail [email protected]
initiated in management and social psychology(Lamm and Myers 1978; Pruitt 1971; Stoner1961) and recently involved also economists(Baker et al. 2008; Masclet et al. 2009). When agroup decides whether to enter a lottery or not,there is no obvious correct choice and individ-uals may legitimately differ in their proposalsdue to their preferences. For this reason, thepsychological literature on groups and teamswould classify this task as “judgmental.” On thecontrary, “intellective” tasks have a demonstra-bly correct solution. For instance, Cooper andKagel (2005) study a strategic market entry task,which is mostly intellective. The only intellec-tive aspect of our lottery task is that choicesshould be coherent.1 Earlier studies in socialpsychology introduced the concepts of riskyshifts and cautious shifts. “Risky shift” denotessituations where groups make riskier decisionsthan individuals, and “cautious shift” otherwise.Depending on the study, the results reportedin the literature are sometimes in one direc-tion and sometimes in the other. One reason forthis diversity of findings may be the presenceof important, but overlooked, differences in thedesign and methodology among studies. Hence,we first proceed by mapping the approach of
1. Tasks involving other-regarding preferences aremostly judgemental tasks, Cason and Mui (1997) or Luhanet al. (2009). The beauty contest game is a task with bothcomponents (Kocher and Sutter 2005).
ABBREVIATION
CRRA: Constant Relative Risk Aversion
502
Economic Inquiry(ISSN 0095-2583)Vol. 50, No. 2, April 2012, 502–515
doi:10.1111/j.1465-7295.2010.00362.xOnline Early publication February 8, 2011© 2011 Western Economic Association International
ZHANG & CASARI: GROUPS AND RISKY CHOICES 503
some recent experimental studies. In the presentwork, we designed group interaction rules tofacilitate information exchange, to encourageparticipation by all members, and to focus theinteraction on how to aggregate individual pref-erences. The main aim is to understand in detailhow groups of three members deal with dis-agreement. Our design is novel because thereis a written record of the communication amonggroup members to understand internal dynamicsand to correlate with actual differences in out-comes. It is the first, among the studies of grouprisk attitude, where before the discussion, eachparticipant must post her proposal, a feature thatsaves discussion time and prevents shy membersfrom being silenced. This piece of informationallows us to perform an individual-level analysisof preference aggregation. Moreover, in case ofdisagreement, the minority has veto power overthe group decision. Like many other studies, wecall for a unanimous decision but, unlike others,here the sanction for disagreement is severe: nochoice and zero earnings. In the field this ruleis observed in international bodies that do nottake a stand on an issue when they do not reachconsensus, or in organizations that do not partic-ipate in an auction unless the board of directorsagree on a bid. This rule creates a common inter-est within the group to communicate and reacha decision. Other default rules do not generatethis positive group dynamic.
Through the group process, we find that lot-tery choices become more coherent and closerto risk neutrality. In resolving disagreement, theproposal of the majority did not always prevail.It prevailed more often when its proposal wascloser to risk neutrality. There are some interest-ing personality and demographic effects, whichwe report in detail below.
The remainder of the paper is organized asfollows. Section II reviews literature. Section IIIdescribes experimental design and procedures.Section IV reports the results and Section Vconcludes.
II. LITERATURE REVIEW
This section focuses on four recent papersthat examine decisions made by groups fac-ing risky choices, Baker et al. (2008), Harri-son et al. (2010), Masclet et al. (2009), andShupp and Williams (2008). Table 1 presents adesign comparison with the present study. Allstudies, including ours, compare lottery choicesof groups of three members with individuals
TA
BL
E1
Des
ign
Com
pari
son
acro
ssFi
veSt
udie
sof
Gro
upL
otte
ryC
hoic
es
Num
ber
ofG
roup
sin
the
Exp
erim
ent
Incl
udes
Bet
wee
n-Su
bjec
tD
esig
n
Gro
upC
ompo
siti
onac
ross
Cho
ices
Def
ault
Cho
ice
Whe
nN
oG
roup
Una
nim
ity
Max
imum
Att
empt
sto
Rea
chG
roup
Cho
ice
Eve
ryIn
divi
dual
Pos
tsa
Non
bind
ing
Pro
posa
lC
omm
unic
atio
nP
osit
ive
Ince
ntiv
esto
Talk
/Lis
ten
toO
ther
s
Zha
ngan
dC
asar
i(t
his
stud
y)40
No
Fixe
dN
one
(zer
oea
rnin
gs)
3Y
esC
hat
(2m
in)
Yes
/yes
Mas
clet
etal
.20
0936
No
Ran
dom
Ran
dom
5N
oN
one
Yes
/no
Shup
pan
dW
illia
ms
2007
28Y
esFi
xed
Mea
nof
indi
vidu
albi
ds1
No
Face
-to-
face
(20
min
)Y
es/n
oa
Bak
eret
al.
2008
40Y
esFi
xed
Maj
ority
rule
1N
oFa
ce-t
o-fa
ceY
esfo
rm
inor
ity/n
oH
arri
son
etal
.20
1036
No
Fixe
dM
ajor
ityru
le1
No
Non
eY
esfo
rm
inor
ity/n
o
a Itm
aybe
“yes
/yes
,”a
shor
tex
plan
atio
nfo
llow
s.Sh
upp
and
Will
iam
s(2
007)
aske
dto
pric
eea
chlo
ttery
and
then
awar
ded
the
lotte
ryus
ing
anin
cent
ive
com
patib
lem
echa
nism
.T
hede
faul
tbi
dw
ithou
tun
anim
ityis
the
aver
age
ofin
divi
dual
bids
.A
nin
divi
dual
play
erm
ayha
vean
ince
ntiv
eto
man
ipul
ate
the
grou
ppr
ice
byst
rate
gica
llyov
er-
orun
der-
bidd
ing
inor
der
toge
nera
tea
grou
pbi
dcl
oser
tohe
rpr
efer
red
leve
l.
504 ECONOMIC INQUIRY
choosing in isolation. In both treatments, sub-jects face the same set of lottery choices(ranging from 8 to 15) and identical monetaryincentives. At the end of the session, only one ofthe lotteries is randomly selected for payment.Baker et al. (2008), Masclet et al. (2009), andShupp and Williams (2008) all found that groupsare more risk averse than individuals. On thecontrary, Harrison et al. (2010) report no groupeffect.
Existing studies exhibit a significant diver-sity in design along a number of dimensions(Table 1). The most interesting differences per-tain to group interaction. First, Masclet et al.(2009) randomly changed group compositionfor each lottery choice, whereas the otherskept it fixed. This generates different dynamicincentives to “tune-in” with the group. Second,communication ranges from none, to anony-mous chat rooms, to face-to-face interaction. Weknow from experiments on social dilemmas thatcommunication can have profound effects onchoices.2 With lottery choices, the issue is notto overcome free riding but rather to aggregatepreferences. Hence, communication fulfills otheraims.
In all studies in Table 1, the instructions callfor a unanimous group decision except in Harri-son et al. (2010) that employed a majority votingrule. Within the consensus call, there are sub-stantive differences in the default choice whena group does not reach unanimity. Althoughoften downplayed in experiments with groups,this aspect is theoretically extremely important.Among the criteria to resolve disagreement thereare random choice, majority rule, mean choice,and no choice. Each default rule implicitly setsdifferent incentives for group discussion, whichincludes incentives to “talk” and to “listen.”Let us adopt the standard assumption that sub-jects have well-defined preferences toward riskand assume that they are informed about theintended choices of others in their group. Thelast column of Table 1 lists whether a subjectwould benefit from successfully persuading oth-ers to change their intended choice (“talk”).Except Shupp and Williams (2007), all studiesasked the subjects to make a binary decision
2. It has been documented in the experimental lit-erature that pre-play face-to-face communication signif-icantly improves cooperation in public good game (forinstance, Cason and Khan 1999; Isaac and Walker 1988)and common-pool resource experiments under conditions ofheterogeneity in resource endowment and payoffs (Hackettet al. 1994).
between a safe and a risky option. Thus, theinitial opinions must be a majority of two ver-sus a minority of one.3 All default rules exhibitpositive incentives to talk, except majority rule,where if you are already part of the major-ity you do not have any incentive to persuadeothers. Another crucial aspect is the incentiveto “listen.” The default rules implemented inthe previous studies may not generate positiveincentives to listen (Table 1). Of course, theremay be other types of advantages from listeningto others besides those considered in Table 1.Communication may enhance the understand-ing of the task as well as learning about theintended choices of others and so benefit every-one in the group. Table 1 considers incentivesunder the more narrow view of rational subjectsendowed with precise utility functions, whichare common knowledge.
Like some of the experiments, we employeda within-subject design which allows a moredirect comparison of choices in isolation (I) andin group (G) but may exhibit order effects. Tocontrol for order effects, Masclet et al. (2009)run sessions with I–G and G–I sequences anddo not find any. Others employed a between-subject design, which relies instead on anassumption of similar preferences of the twoexperimental samples for I and for G treatments.
Other experimental studies have groups fac-ing more challenging choices under risk. Rock-enbach et al. (2007) compared individuals andgroups with respect to choices among alterna-tive financial investments and found that groupsaccumulate significantly higher expected val-ues at a significantly lower total risk. Charnesset al. (2007) study choice monotonicity over lot-tery and Bayesian updating by individuals andgroups. They find that social interaction reducesviolation rates, and thus groups make substan-tially fewer errors than individuals and the errorrate decreases with group size.
III. EXPERIMENTAL DESIGN AND PROCEDURES
Each session had four parts plus a ques-tionnaire and involved 15 participants. Overall120 students participated in the experiment. Inpart 1, we measured subjects’ risk attitude with15 binary choices between lotteries. In part 2,
3. Shupp and Williams (2007) asked to price each lotteryand then awarded the lottery using an incentive compatiblemechanism. The default bid without unanimity is the averageof individual bids.
ZHANG & CASARI: GROUPS AND RISKY CHOICES 505
subjects were randomly divided into groups ofthree persons and faced the same task as inpart 1. The per-capita expected payoff in part 1was equal to that in part 2. We report resultsof parts 3 and 4, which involved a differenttask, in Casari et al. (2010). The overall incen-tive structure was similar to that in Holt andLaury (2002). Subjects chose between a “safe”Option A and a “risky” Option B. The pay-off for Option A was deterministic (50 tokens)and the payoff for Option B was either 150 or0. On the first decision, the probability of thehigh payoff (150) for Option B was 0. In subse-quent choices, the probability of the high payoffincreased by 1/20 each line, 0, 1/20, . . . , 14/20.A risk-neutral person would choose Option A inlotteries 1 through 7 and then switch to OptionB in lottery 8. Risk seeking agents may switchto Option B earlier than lottery 7 and risk-averseagents may switch later than lottery 7. Any ratio-nal agent should choose Option A over OptionB in the first lottery (50 vs. 0 francs always) andlater on eventually switch to Option B. Multi-ple switches would be a signal of confusion. Wepaid only one of the 15 decisions, chosen ran-domly at the end of the session. Random choiceswere all implemented through drawings from abingo cage.
In part 2, there were five groups in each ses-sion. There was a proposal phase, a chat phase,and a group choice phase. Everyone simultane-ously made an individual proposal about each ofthe 15 lottery choices. Then any line with dis-agreement was highlighted for all three groupmembers to see. At this point, participants couldswitch to a chat window and had 2 minutesto send free-format messages to others in theirgroup. We asked participants to follow two basicrules: (1) to be civil to one another and do notuse profanities and (2) not to identify them-selves in any manner. Messages were recorded.In the chat window, subjects received an id num-ber from 1 to 3 based on the order in whichthey sent messages in that specific period. Afterthe chat stage, everyone had to submit a choicefor the group decision. If the choices of allthree group members were identical for a spe-cific decision line (unanimity), then we had agroup choice. If there was unanimity on all15 choices, then part 2 was over. Otherwise,the line(s) with disagreement was (were) high-lighted and all three group members were askedto submit their new proposals. If there was stilldisagreement, there was another, final round ofproposals. At this point part 2 was over even if
disagreement remained. The design followed adefault rule of “no choice”: if the group reachedno unanimous decision, no decision was placed,so earnings were zero for everyone in the group.Such a default rule generates positive incentivesboth to talk and to listen to others in the chat.In fact, these incentives are the highest amongthe studies listed in Table 1. With disagreementbetween a safe and a risky option, a default rulewith random selection induces a game wherea subject’s dominant strategy is to choose hermost preferred option. Instead, a default rule of“no choice” induces a battle-of-the sexes gamewhere a subject would always switch choice toavoid a disagreement outcome. We paid onlyone of the 15 decisions, chosen randomly atthe end of the session. Random choices wereall implemented through drawings from a bingocage. If for the line selected the group was stillin disagreement, then the group earned 0 forpart 2. Overall, there were 40 groups and 600group decisions taken.
We distributed written instructions and readthem aloud, taking questions as they arose. Theexperiment was performed with a z-tree applica-tion (Fishbacher 2007). No person participatedin more than one experimental session. We guar-anteed a minimum payment of $5. We convertedeach experimental token to an actual dollar atthe rate of $0.03. Including all parts, a sessionlasted on average about 2 hours and averageearnings per person were about $20. We con-ducted eight experimental sessions at PurdueUniversity (USA) between 25 September and 28October, 2007. Participants were recruited fromthe undergraduate campus population by email.
IV. RESULTS
We report five main results.
A. Result 1: The Monotonicity of LotteryChoices Improved from the Individualto the Group Treatment
We employed a table format to elicit riskattitude, where a subject with monotonic riskpreferences would choose Option A in deci-sion 1 and then eventually switch forever toOption B at one later decision. A subject whoswitched from A to B more than once, orwho switched from B back to A, is classifiedas nonmonotonic, which is taken as a proxyof confusion or irrationality.4 Recorded levels
4. Some multiple switches could also be a sign ofpreference indifference over a certain range.
506 ECONOMIC INQUIRY
of monotonicity in the experiment were veryhigh, ranging from 87.5% for individual choices(105/120) to 95.0% for group choices (38/40)(one-sided t test, m = 120, n = 40, p-value =0.034). A small portion of this improvementmay be attributed to task learning, but we findno significant difference in monotonicity lev-els between individual proposals and individ-ual choices (90.0% vs. 87.5%, one-sided t test,m = 120, n = 120, p-value = .27). In part 2,individual proposals do not have significantlydifferent level of monotonicity in lottery choicesthan group final decisions (90.0 vs. 95.0%, one-sided t test, m = 120, n = 40, p-value = .32).
B. Result 2: Group Choices Were Closerto Risk Neutrality Than Individual Choices.In Particular, Group Choices Exhibiteda Risky Shift from Individual Choices
Support for Result 2 comes from Table 2and Figure 1. We discuss separately lotteries1–7 from lotteries 8–15. In lotteries 1–7, onlya risk-seeking agent would choose the riskyOption B. Differences here were rather limitedbecause risk seeking behavior was rare: on aver-age, only 2% of individual choices and 0.4%of group choices were for B. In these lotter-ies, groups were less risk seeking than individu-als. Most of the differences came from lotteries8–15 where a risk-neutral agent would choosethe risky Option B. In these lotteries, groupswere more risky than individuals. On average,57.4% of individual choices and 61.7% of groupchoices were for B. Group choices are morerisky than individual choices (p-value < 0.05).5
Recall that part 1 elicited individual choiceswhile part 2 elicited individual proposals andgroup choices. Although this fixed order mayhave had some impact on results, order effectsare unlikely to explain the risky shift. First, westated in part 1 of the instructions that the tasksin parts 1 and 2 were the same lottery choices.Hence subjects could optimize considering theoverall level of uncertainty. Second, little evi-dence can be traced from a comparison of the
5. The two-sample Kolmogorov–Smirnov test rejectsthe equality of two distributions of the switching pointsfrom A to B with monotonic preference (n = 105, m =39, p-value = 0.0476). About 58.8% of individual pro-posals were for B (1.4 points more than individualchoices). The distribution of the switching points inindividual proposals does not significantly differ fromgroup decisions (Kolmogorov–Smirnov test, n = 108, m =39, p-value = 0.978), neither from individual choices(Kolmogorov–Smirnov test, n = 108, m = 105, p-value =0.998).
individual choices in part 1 with the individualproposals in part 2. We elicited individual pro-posals before any communication could takeplace in the group setting and report only mini-mal differences with part 1 choices, which helpsto rule out large order effects. As mentionedafter Result 1, some of this difference is simplya correction of nonmonotonic behavior. Third,Masclet et al. (2009) explicitly studied ordereffects but did not find any.
Result 2 may be a consequence of the defaultrule adopted in the design. The “no choice” rulemay have generated a different group dynamic.In particular an asymmetry in payoffs betweenrisk-averse and risk-neutral subjects in the nego-tiation over disagreement. More risk-averse sub-jects may have less to lose from switching totheir least preferred choice.6
C. Result 3: When in Disagreement,the Majority Proposal Did Not Always Prevail.It Prevailed More Often When Its ProposalWas Closer to Risk Neutrality. Proposalsfrom Nonmonotonic Subjects Were LessLikely to Prevail
Support for Result 3 comes from Tables 3–5and Figure 2. We focus explicitly on group deci-sions where there was an initial disagreement.We define disagreement as a situation wherenot all three individual proposals were equal.All groups disagreed on at least one decision,77.5% found an agreement on the first round,20% after a second or third round, and only2.5% (1 group) never found complete agree-ment.7 On average, a group disagreed on fourlottery decisions (27% of decisions). The bulkof the disagreement (85%) was in lotteries 8–13,where risk neutrality pointed toward Option B
6. Without pretence of generality, below we illustratethis point. Consider a game with two players with differentrisk attitudes who choose between a safe option S = 50and a risky option R = (150, p; 0, 1 − p). Assume player1 has a CRRA utility function and is risk averse, u(x),u′ > 0, u′′ < 0 and player 2 is risk neutral, v(x) = x.Assume disagreement, that is player 1 prefers S, U [S] >U [R], and player 2 prefers R, V [S] < V [R]: u(50) >pu(150) and 50 < 150 p, which implies that 1/3 < p <u(50)/u(150). One can show that for lotteries 9–15 we havethat V [R]/V [S] > U [S]/U [R], i.e., 3p2 > u(50)/u(150),which holds for p > 1/sqrt(3) and, given the actual values ofp, for lotteries 9–15. There were instances of disagreementalso for lotteries 6–8.
7. The group reached agreement on 12 of the 15 lotterychoices (disagreement over lotteries 8–10). Analyses with159 groups (477 proposals) dropped those three observationswhile analyses with 162 groups (486 proposals) replacedthose three group lottery choices with the individual inputsin the third attempt to reach a group decision.
ZHANG & CASARI: GROUPS AND RISKY CHOICES 507
TA
BL
E2
Lot
tery
Cho
ice
Task
Opt
ion
AO
ptio
nB
Ris
kP
refe
renc
eIn
divi
dual
Cho
ices
Indi
vidu
alP
ropo
sals
Gro
upC
hoic
es
Lot
tery
Num
ber
Pay
offs
Pay
offs
Pro
babi
lity
ofG
etti
ng15
0To
kens
Exp
ecte
dP
ayof
fof
Opt
ion
B
Ran
geof
CR
RA
ifSw
itch
from
Ato
Bat
the
Fol
low
ing
Lot
tery
Fre
quen
cyof
Cho
ices
for
B(%
)
Fre
quen
cyof
Cho
ices
for
B(%
)
Fre
quen
cyof
Cho
ices
for
B(%
)
150
150
or0
00
r<
−1.7
30
0.8
02
5015
0or
00.
057.
5−1
.73
<r
<−1
.10
00
350
150
or0
0.1
15−1
.1<
r<
−0.7
31.
70
04
5015
0or
00.
1522
.5−0
.73
<r
<−0
.47
00
05
5015
0or
00.
230
−0.4
7<
r<
−0.2
71.
70
06
5015
0or
00.
2537
.5−0
.27
<r
<−0
.15
2.5
07
5015
0or
00.
345
−0.1
<r
<0.
045.
86.
72.
58
5015
0or
00.
3552
. 50.
04<
r<
0.16
1520
16.7
950
150
or0
0.4
600.
16<
r<
0.27
24.2
26.7
21.7
1050
150
or0
0.45
67.5
0.27
<r
<0.
3631
.732
.536
.711
5015
0or
00.
575
0.36
<r
<0.
4558
.358
.365
1250
150
or0
0.55
82.5
0.45
<r
<0.
5368
.367
.580
1350
150
or0
0.6
900.
53<
r<
0.6
8080
87.5
1450
150
or0
0.65
97.5
0.6
<r
<0.
6688
.390
.895
1550
150
or0
0.7
105
0.66
<r
93.3
9597
.5
Perc
enta
geof
mon
oton
icde
cisi
onm
aker
s87
.590
95
Not
es:
Eve
ryon
esh
ould
choo
seO
ptio
nA
inde
cisi
on1.
Ris
k-ne
utra
lsu
bjec
tsw
ould
switc
hto
Opt
ion
Bin
deci
sion
8.A
switc
hin
late
rde
cisi
ons
reve
als
risk
aver
sion
and
asw
itch
inea
rlie
rde
cisi
ons
reve
als
risk
-see
king
beha
vior
.C
RR
Ast
ands
for
aut
ility
func
tion
exhi
bitin
gco
nsta
ntre
lativ
eri
skav
ersi
on.
Num
ber
ofin
divi
dual
obse
rvat
ions
for
each
line
is12
0.
508 ECONOMIC INQUIRY
FIGURE 1Individual versus Group Risk Attitude. Fraction of Risky Lottery Choices by Groups and
Individuals
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
lotter
y1
lotter
y2
lotter
y3
lotter
y4
lotter
y5
lotter
y6
lotter
y7
lotter
y8
lotter
y9
lotter
y10
lotter
y11
lotter
y12
lotter
y13
lotter
y14
lotter
y15
Cho
ices
for
B (r
isky
opt
ion)
Individual Choices, part one
Group Choices, part two
Risk Neutral Choices
Notes: N = 120. One group did not agree on three lottery decisions. For those decisions, the graph employed theirindividual third attempt proposals. Lottery numbers are the same as in Table 2.
TABLE 3Risk Neutrality When Disagreement
MajorityPrevailed
MinorityPrevailed
Majority at riskneutrality
79 12 91(57.2%)
Minority at riskneutrality
50 18 68(42.8%)
Totals 129(81.1%)
30(18.9%)
159(100%)
Notes: The unit of observation is a decision a groupmade in a lottery in part 2. This table includes onlygroup decisions with disagreement (159/600 obs.). Thetable compares individual proposals with group choice. Themajority proposal was A when AAB and B when ABB.
while risk-averse subjects may have preferredthe safer Option A (Figure 2).
The analysis of disagreement is particularlyinteresting because one can understand the inter-nal process that led to a decision and shed lighton Result 2. Given that the decision was binary,A or B, and a group comprised three individ-uals, there were only two possible patterns ofdisagreement, a majority for A (AAB) or a
TABLE 4Risky Shift When Disagreement
MajorityPrevailed
MinorityPrevailed
Majority more risky 68 11 79(49.7%)
Minority more risky 61 19 80(50.3%)
Totals 129 30 159
Notes: The unit of observation is a decision a groupmade in a lottery in part 2. This table includes onlygroup decisions with disagreement (159/600 obs.). Thetable compares individual proposals with group choice. Themajority proposal was A when AAB and B when ABB.
majority for B (ABB). Out of a total of 600,there were 159 group decisions with disagree-ment. In order to study how disagreement wasresolved through group interaction, we considertwo possible benchmarks: the outcome with adictator selected at random in the group andthe outcome with majority voting. Followinga random dictator process the proposal of themajority would prevail in 66.7% of the caseswhile following majority voting the proposal of
ZHANG & CASARI: GROUPS AND RISKY CHOICES 509
TABLE 5Probit Regression on How Groups Resolve Disagreement
Sample: Decisions with Disagreement OnlyDependent variable:1 = my proposal equals group choice, 0 = otherwise
All(1)
MajorityPrevails
(2)
MinorityPrevails
(3)
MultipleAttempts
beforeUnanimity
(4)
OneAttemptbefore
Unanimity(5)
Independent variables:My proposal was the risk-neutral choice (1 or 0) 0.20∗ 0.14∗ 0.15∗ 0.09 0.34∗∗
(0.103) (0.060) (0.063) (0.114) (0.121)My proposal was in the majority (1 or 0) 0.65∗∗ — — 0.71∗∗ 0.63∗∗
(0.081) — — (0.160) (0.112)My individual choice was different than my proposal −0.27 −0.40∗ 0.05 −0.38∗∗ −0.33∗
(0.138) (0.191) (0.120) (0.144) (0.158)My proposals were not monotonic (1 or 0) −0.29∗∗ −0.19 −0.07 −0.23 −0.32∗∗
(0.074) (0.103) (0.043) (0.118) (0.116)Number of lottery decisions on which the group disagree −0.00 0.00 −0.00 −0.03 0.01
(0.011) (0.010) (0.019) (0.034) (0.011)Multiple attempts to decide (1 or 0) 0.05 0.04 −0.06 — —
(0.048) (0.036) (0.073) — —
Chat messagesI talked first (1 or 0) 0.08 0.09∗ 0.10 −0.07 0.04
(0.066) (0.036) (0.083) (0.129) (0.074)I talked last (1 or 0) −0.10 −0.06 −0.06 0.14 −0.11
(0.058) (0.030) (0.048) (0.103) (0.074)Number of words I wrote in my group (×100) 0.18 −0.19 0.55∗∗ 0.04 0.15
(0.453) (0.190) (0.210) (0.429) (0.579)Number of words that all other members wrote (×100) −0.38 −0.36∗∗ 0.16 −0.49∗∗ −0.30
(0.253) (0.111) (0.176) (0.181) (0.334)
DemographicsScience and Engineering Major (1 or 0) 0.07 0.06 −0.06 0.29∗∗ 0.03
(0.079) (0.034) (0.071) (0.103) (0.094)Above 75 percentile SAT/ACT (1 or 0) 0.05 0.04 −0.02 −0.05 0.21∗∗
(0.059) (0.035) (0.059) (0.138) (0.057)Below 25 percentile SAT/ACT (1 or 0) 0.09 0.08∗∗ 0.08 −0.00 0.20∗∗
(0.076) (0.028) (0.097) (0.085) (0.068)Male (1 or 0) −0.03 −0.03 0.04 0.17 −0.17∗
(0.080) (0.041) (0.082) (0.187) (0.084)Missing SAT/ACT or demographic data (1 or 0) 0.10 0.06 −0.02 −0.33 0.23∗∗
(0.085) (0.033) (0.056) (0.339) (0.084)Personality traitsAgreeableness 0.00 0.01 0.03 −0.00 −0.05
(0.045) (0.029) (0.037) (0.105) (0.055)Conscientiousness −0.10 −0.08∗ 0.01 0.15 −0.11
(0.059) (0.038) (0.082) (0.129) (0.083)Neuroticism −0.01 −0.03 0.01 0.08 −0.06
(0.066) (0.035) (0.053) (0.205) (0.069)Openness 0.02 −0.00 0.00 0.06 0.04
(0.050) (0.027) (0.052) (0.090) (0.056)Extroversion −0.00 −0.00 0.00 0.20∗∗ −0.03
(0.048) (0.021) (0.056) (0.062) (0.050)Number of observations 477 318 159 150 327Pseudo R2 0.361 0.284 0.205 0.498 0.375Log likelihood −204.7 −110.2 −61.23 −49.97 −137.9
Notes: Marginal effects, robust standard errors in parentheses, clusters on groups. Sample: decisions with disagreementonly. The regression includes lottery decision dummies, which have not been reported in the table. One group did not agreeon three lottery decisions and those decisions are excluded from this table.
Statistical significance ∗∗p < .01, ∗p < .05.
510 ECONOMIC INQUIRY
FIGURE 2Which Lottery Decisions Were Most Controversial. Disagreement within Groups for Each Lottery
(n = 600)
lottery1lottery2
lottery3lottery4
lottery5lottery6
lottery7lottery8
lottery9lottery10
lottery11lottery12
lottery13lottery14
lottery15
0%
20%
40%
60%
80%
100%
The minority prevailed
The majority prevailed
Groups in dis-agreement
Perc
enta
ge o
f gr
oups
Note: Lottery numbers are the same as in Table 2.
the majority would prevail in 100% of the cases.As Table 3 illustrates, when in disagreement, theproposal of the majority prevailed in 81.1% ofthe decisions, while the minoritarian proposalprevailed in the remaining 18.9% (two Pearsonchi-squared tests, p-value < .01, n = 159). Theactual outcome is in-between a random dictatorand a majority rule process and exhibits someinteresting biases in group decision making.
When in disagreement, 52.7% of individ-ual proposals and 61.0% of group choiceswere the same as those of a risk-neutralagent. Table 3 suggests that the proposal of themajority prevailed more often when its pro-posal was the same as that of a risk-neutralagent (79/91 vs. 50/68, Pearson chi-squaredtest, p-value = .011). Hence the group interac-tion generated a shift toward more risk-neutralchoices.
Table 4 presents a breakdown with respectto whether the more risky proposal prevailed.Overall, with disagreement the more risky pro-posal prevailed in 54.7% of the decisions, whichis slightly higher than predicted by a coinflip resolution of disagreement (50.0%) andhigher than what is expected had the proposalof the majority always prevailed (49.7%).8
8. These differences are not statistically significant.
In particular, when the majority prevailed, in52.7% of the decisions it had the more riskyposition; when the minority prevailed, in 63.3%of the decisions it had the more risky position.
Table 5 presents the marginal effect from aprobit regression on individual proposals. Thedependent variable is equal to one when an indi-vidual proposal equals the actual group choice(hence it prevails in case of disagreement).Among the independent variables, we includedsome aspects of lottery choices, demographicand personality traits, and chat activity. Thefocus is on individual proposals in disagree-ment with others in the group. We will postponethe discussion of chat activity to Result 4 anddiscuss the other findings. Demographic regres-sors include skill, gender, and major. Skills areproxied by the ACT/SAT scores obtained fromthe university Registrar’s Office. We have eitherSAT or ACT scores for 92.5% of the subjects(missing data = 0), who are coded using the USnationwide distribution of the SAT-takers (Col-lege Board of Education 2006) and ACT-takers.The threshold for high ability is being in the topquartile of the distribution and for low abilityis being in the lower quartile. The variablesare primarily based on SAT scores and, whenmissing, on ACT scores. The cutoff values are
ZHANG & CASARI: GROUPS AND RISKY CHOICES 511
the average between male and female nationaltables.
Another class of regressors code five per-sonality traits using questionnaire answers. Thepersonality traits are designed based on thebig five inventory by John et al. (1991), agree-ableness, conscientiousness, neuroticism, open-ness, and extroversion. For example one variablemeasures conscientiousness through the aver-age rating on nine statements.9 Subjects cir-cle a number 1 through 5, where 1 stands for“strongly disagree,” 2 for “disagree,” 3 for “neu-tral,” 4 for “agree,” and 5 for “strongly agree.”
Table 5 presents results from the same econo-metric specification run on five data samples.Column 1 includes all decisions with disagree-ment, columns 2 and 3 show a breakdown ofthe sample into cases where the majority orminority prevailed, respectively. We will latercomment on the other columns. The results cor-roborate five points. First, there was a significantshift toward risk neutrality as stated in Result 2(columns 1, 2, and 3). Second, as already dis-cussed, being in a majority substantially raisedthe likelihood to prevail in case of group dis-agreement (column 1). Third, subjects who areconfused with the task are less likely to pre-vail (columns 1 and 2). We proxied a subject’sconfusion using a lottery-specific dummy forher individual choice being different from herproposal and a subject-specific dummy for herproposal not being monotonic. Fourth, personal-ity matters, in the sense that more conscientioussubjects conceded more chances to the proposalof the minority when they were in a majoritythat prevailed (column 2). Fifth, skill sometimesmatter but not in an expected way: low skilledsubjects were more likely to prevail (column 2).
D. Result 4: About One Third of Groups DidNot Find Agreement Immediately afterCommunication. Groups with High Skill andScience and Engineering Members Were MoreLikely to Find an Immediate Agreement as wellas Those with Monotonic and More ExtrovertMembers
Tables 5 and 6 provide support for Result 4.Table 6 presents a probit regression on the dif-ficulty of reaching a group agreement in the
9. I do a thorough job. I do things efficiently. I makeplans and follow through. I am a reliable worker. I persevereuntil the task is finished. I am easily distracted. I canbe somewhat careless. I tend to be lazy. I tend to bedisorganized.
first attempt. Predictably, the higher the numberof lotteries with disagreement the less likelythe groups would resolve disagreement imme-diately. In addition, both skill and personalitymeasures had an impact. Groups with memberswith SAT/ACT score above the 75th percentileand with monotonic proposals were more likelyto find an immediate agreement. Groups withmore extrovert members were also more likelyto find an immediate agreement (also more con-scientious members, albeit at a 10% significancelevel). There was also a strong effect of Scienceand Engineering although no gender effect wasrecorded. We will comment on the impact ofchat activity in Result 5.
When disagreement persists after the commu-nication stage, group processes may change sub-stantially. In column 4 of Table 5 we restrict thesample to those groups who required multipleattempts before converging toward unanimity.Those groups faced an emergency situationsince they would have obtained 0 payoffs ifthey had not reached an agreement after threeattempts. When disagreement is not resolvedimmediately, our previous conclusions needqualification. Putting forward a risk-neutral pro-posal is no longer important in the emergencysituation (risk-neutral proposal prevailed in 95out of 109 in one attempt vs. 30 out of 50 in mul-tiple attempts); different personality traits pre-vailed: extroversion has now a significant impactwhile conscientiousness is no longer important;finally, proposals from Science and Engineeringmajors were more likely to prevail.
E. Result 5: Chat Activity Was Intense, Growingwith the Level of Disagreement and Aimed atFinding Consensus. The Amount and Timing ofChat Messages Help to Predict Group Choices
Figures 3 and 4 and Tables 5 and 6 providesupport for Result 5. All of the 120 subjectsintervened in the 2 minutes of chat time. Onan average, a person intervened 4.3 times andwrote a total of 23.9 words. Hence, the averagelength of an intervention was rather short (5.6words). Interestingly, the higher the number ofdecisions with disagreement the more intensewas the chat activity, suggesting that messageswere aimed at finding a common ground. Withmore disagreements, participants intervened onaverage about the same number of times butwith longer messages (Figure 3).
Figure 4 informs about the content of thecommunication by giving an uncensored list of
512 ECONOMIC INQUIRY
TABLE 6Probit Regression on Group Difficulty of Reaching an Agreement
Sample: Decisions with Disagreement Only
Dependent variable:1 = my group required more than one attempt to decide; 0 = otherwise
Independent variables:My proposal was the risk-neutral choice (0/1) −0.02
(0.068)My individual choice was different than my proposal 0.07
(0.116)My proposals were not monotonic (0/1) 0.40∗∗
(0.139)Number of lotteries with disagreement in the group 0.17∗∗
(0.051)Number of words written overall by group (× 100) −0.57
(0.313)Number of words written in the last intervention (× 100) 0.02
(0.020)Number of words written in the second to last intervention (× 100) 0.04∗∗
(0.012)Difference in words written between the most and the least active individual (× 100) 1.35∗
(0.617)Science and engineering major −0.28∗∗
(0.104)Above 75 percentile SAT/ACT −0.19∗
(0.076)Below 25 percentile SAT/ACT 0.03
(0.111)Male 0.12
(0.088)Missing SAT/ACT or demographic data 0.04
(0.113)Agreeableness −0.01
(0.060)Conscientiousness −0.14
(0.074)Neuroticism 0.06
(0.073)Openness −0.10
(0.090)Extroversion −0.14∗
(0.063)Observations 477Pseudo R2 0.523Log likelihood −141.8
Notes: Marginal effects, robust standard errors in parentheses, clusters on groups. Sample: decisions with disagreementonly.
The regression includes lottery decision dummies, which have not been reported in the table.Statistical significance ∗∗p < 0.01, ∗p < 0.05.
the top 100 words employed. In the figure, thecharacter size is proportional to frequency ofuse. “A” and “B” were the option names andwere among the most frequently used. Noticenumbers 1–15, whose sizes are roughly linkedto how controversial that particular lotterydecision was. “I” and “we” suggest the tension
between individual and group. Overall, thewords employed denote a very practical use ofcommunication to reach consensus or expressopinions for or against a choice. This contentanalysis did not rely on human coders, as Cooperand Kagel (2005) but on quantitative statisticson the text, which delivered interesting results.
ZHANG & CASARI: GROUPS AND RISKY CHOICES 513
FIGURE 3Chat Activity
0
20
40
60
80
100
120
1 2 3 4 5 6 7 8 9
Number of lottery decisions on which groups disagree at the proposal stage
0
1
2
3
4
5
6
7
8
9
10
Average number of words in group (left scale)
Average number of group members sending messages (rightscale)
7.5%
20.0% 17.5%
20.0%
7.5%
12.5%
10.0% 2.5%
2.5%
Notes: n = 40 groups; all groups disagreed on at least one decision; percentages refer to the number of observations,which sums up to 100%.
The probit regressions in Tables 5 and 6 showthat chat activity helps to predict how groupsresolved disagreement. In Table 5 four variablessummarized chat activity: who talked first, whotalked last, number of words written by thesubject, and total number of words written bythe other two people in the group. Even withoutanalyzing the content of the messages, one cansee the effects on whose choice prevailed ingroup decision making. The persuasion effortas measured by the number of words writtenpaid off in the expected direction. Not voicingyour own reasons lowered someone’s chances ofdetermining the group decision. In particular asubject with a majority proposal was more likelyto be the first to express opinions and to renderthe majority proposal to prevail. A subject ina minoritarian position had chances to convincethe other two if she wrote longer messages. Thisevidence is consistent with Eliaz et al. (2007)’stheory which predicts that the majority prevailsthrough greater voice and larger group size and,
whenever the minority prevails, voice more thancompensates for the group size.
In Table 6, four other variables that arebased on the count of the number of wordssummarized the chat activity: overall activityin the group, difference in activity between themost and least active member in the group,length of the last intervention and of the secondto last intervention. We report two major effectsof chat activity on the difficulty of reachinga group agreement in the first attempt. First,groups with more words written in the chat cansort out disagreement more quickly. Second, alarge inequality in chat activity among groupmembers and more words in the second to lastintervention correlate with more difficulties ofreaching a consensual group choice.
V. CONCLUSIONS
We study group decision making with theaim to understand how small groups resolve
514 ECONOMIC INQUIRY
FIGURE 4One Hundred Most Frequent Words in Chat Messages
Notes: “B” is the most frequently used word in the chat (122 times). “DECISION” is the least used (seven times).Character size is proportional to frequency of use.
disagreement when facing a safe versus a riskyoption. We present experimental evidence bothat the aggregate and at the individual level.
In the aggregate, we report that group deci-sions generate a “risky shift” in comparison toindividual decisions. This shift occurs becausegroup choices were 4.3 percentage points morefrequently closer to risk neutrality than indi-vidual choices; groups made choices that wereless risk averse than those of their members.In addition, group choices followed monotonic-ity more often than individual choices. Theseaggregate results contribute to the debate onwhether group decision making generates a riskyor a cautious shift. Baker et al. (2008), Mascletet al. (2009), and Shupp and Williams (2008) allfound that groups are more risk averse than indi-viduals. On the contrary, Harrison et al. (2010)report no group effect. We put forward theexplanation that the attitude of group decisionsover risk depends on the interaction rules andon group size. These conjectures spring from
considering the variety of default rules adoptedin the literature in case of group disagreement.Chat communication alone did not always gen-erate unanimity because individuals may holdgenuinely different stands over what risks totake. In these cases, as Baker et al. (2008) note,the unanimity rule is more likely to induce morepressure toward conformity in groups than themajority rule. We carried out analyses of theincentives set by alternative default rules, whichmakes clear that our design gives the high-est incentives to negotiate and reach consensualdecisions within the group. In addition to for-mal incentives, there may be a behavioral grouppressure to conform, which depends on mem-bers’ personality and group size. We did notexplore differences in group sizes but conjecturethat in a group of three members, a two-against-one situation is qualitatively different than a dis-agreement in a group of two of one-against-one.In our experiment, in situations of two-against-one the minority proposal prevailed on average
ZHANG & CASARI: GROUPS AND RISKY CHOICES 515
in 19% of cases. This fraction is positive butless than one third, as a random selection wouldsuggest, and further reduced to 14% in case thedisagreement persisted over multiple attempts todecide, which signals an even stronger attractiontoward the opinion of the group majority.
Lack of agreement caused an emergencysituation because without unanimity in a lotterychoice, participants’ payoff was zero for thatlottery. Agreement could eventually be reachedwithout further communication in a second orthird attempt. In these emergency situations,the mode of interaction within group memberschanged substantially.
We report evidence that personality and com-munication abilities mattered. In particular, thepresence of extrovert and conscientious mem-bers influences group choices. A conscientioussubject may be more willing to give in to min-imize the chances of no choice in case of dis-agreement. Extrovert subjects were more likelyto push for an immediate agreement or to voicehis or her proposals when in emergency situa-tion. The patterns of communication in terms ofamount, equality, and timing significantly influ-ence the outcome. In the experiment, the moreone writes relative to others, the more likely isone’s opinion to prevail. Moreover, a balancedexchange of messages among members makesimmediate agreement more likely.
To conclude, in a group with clearly outlinedindividual preferences and incentives to solvedisagreements, group decisions exhibited a shifttoward risk neutrality. This “risky shift” was notfound in other studies and likely depends on theincentives to internally negotiate an agreement.We conjecture that the risk attitude of groupdecisions is rule-specific: it depends on theinteraction rules in place within the group.
REFERENCES
Baker, R. J., S. K. Laury, and A. W. Williams. “Compar-ing Small-Group and Individual Behavior in Lottery-Choice Experiments.” Southern Economic Journal,75(2), 2008, 367–82.
Casari, M., J. Zhang, and C. Jackson. “Do Groups Fall Preyto the Winner’s Curse?” Working Paper No. 504, Insti-tute for Empirical Research in Economics, Universityof Zurich, 2010.
Cason, T. N., and F. U. Khan. “A Laboratory Study ofVoluntary Public Goods Provision with ImperfectMonitoring and Communication.” Journal of Develop-ment Economics, 58(2), 1999, 533–52.
Cason, T. N., and V.-L. Mui. “A Laboratory Study of GroupPolarization in the Team Dictator Game.” EconomicJournal, 107(444), 1997, 1465–83.
Charness, G., E. Karni, and D. Levin. “Individual andGroup Decision Making Under Risk: An ExperimentalStudy of Bayesian Updating and Violations of First-Order Stochastic Dominance.” Journal of Risk andUncertainty, 35(2), 2007, 129–48.
College Board of Education. “SAT Percentile Ranksfor Males, Females, and Total Group, College-Bound Seniors—Critical Reading + Mathematics +Writing.” 2006. Accessed December 14, 2010. http://www.collegeboard.com/prod_downloads/highered/ra/sat/SATPercentileRanksCompositeCR_M_W.pdf.
Cooper, D. J., and J. H. Kagel. “Are Two Heads BetterThan One? Team versus Individual Play in SignalingGames.” American Economic Review, 95(33), 2005,477–509.
Eliaz, K., D. Ray, and R. Razin. “Group Decision-Makingin the Shadow of Disagreement.” Journal of EconomicTheory, 132, 2007, 236–73.
Fischbacher, U. “z-Tree-Zurich Toolbox for ReadymadeEconomic Experiments.” Experimental Economics,10(2), 2007, 171–78.
Hackett, S., E. Schlager, and J. Walker. “The Role of Com-munication in Resolving Commons Dilemmas: Exper-imental Evidence with Heterogenous Appropriators.”Journal of Environmental Economics and Manage-ment, 27, 1994, 99–126.
Harrison, G. W., I. L. Morton, E. E. Rutstrom, andM. Tarazona-Gomez. “Preferences over Social Risk.”Working Paper No. 19, J. Mack Robinson College ofBusiness, Georgia State University, 2010.
Holt, C. A., and S. K. Laury. “Risk Aversion and Incen-tive Effects.” American Economic Review, 92, 2002,1644–55.
Isaac, R. M., and J. M. Walker. “Communication and Free-Riding Behavior: The Voluntary Contribution Mecha-nism.” Economic Inquiry, 26(4), 1988, 585–608.
John, O. P., E. M. Donahue, and R. L. Kentle. The Big FiveInventory. Berkeley: University of California, 1991.
Kocher, M. G., and M. Sutter. “The Decision Maker Mat-ters: Individual versus Group Behavior in Experi-mental Beauty-Contest Games.” Economic Journal,115(500), 2005, 200–23.
Lamm, H., and D. G. Myers. “Group-induced Polarizationof Attitudes and Behavior.” Advances in ExperimentalSocial Psychology, 11, 1978, 145–95.
Luhan, W., M. Kocher, and M. Sutter. “Group Polarizationin the Team Dictator Game Reconsidered.” Experimen-tal Economics, 12, 2009, 26–41.
Masclet, D., Y. Loheac, D.-B. Laurent, and N. Colombier.“Group and Individual Risk Preferences: A Lottery-Choice Experiment.” Journal of Economic Behaviorand Organization, 70(3), 2009, 470–84.
Pruitt, D. G.. “Choice Shifts in Group Discussion: An Intro-ductory Review.” Journal of Personality and SocialPsychology, 20, 1971, 339–60.
Rockenbach, B., A. Sadrieh, and B. Mathauschek. “TeamsTake the Better Risks.” Journal of Economic Behaviorand Organization, 63, 2007, 412–22.
Shupp, R., and A. W. Williams. “Risk Preference Differen-tials of Small Groups and Individuals.” The EconomicJournal, 118(525), 2008, 258–83.
Stoner, J. A. F. “A Comparison of Individual and GroupDecisions Under Risk.” Unpublished Master’s Thesis,Massachusetts Institute of Technology, School of Man-agement, 1961.