Uneak White's Personal Brand Exploration Presentation
Complexity in Ambiguous Problem Solution Search: Group Dynamics, Search Tactics and Performance
1. Ambiguous Problem Complexity, Group Synergy
and Performance: An Experiment
Elliot Bendoly
Svenja Sommer
Stylianos (Stelios) Kavadias
2. • An ongoing debate over the benefits of brainstorming (Stroebe and
Diehl 1994; Paulus et al. 1996)
• Unclear group benefits in problem contexts with non-obvious links
between solution details (‘decision’) and solution performance
(apprehension decreased, etc. – Gallupe et al. 1991)
• Non-monotonic role of problem complexity (group can in fact lead to
more or less productive results (Kavadias and Sommer 2009)
“Nominal”
solution or
“best of set”
Group
collaborative
solution
>
(In LOW complexity)
<
(In HIGH complexity)
Motivation
3. Context: an adaptation of Ederer and Manso’s (2012) incentives experiment
Continuous Decisions:
Price, Lemon content, Sugar content
Discrete Decisions:
Color (2), Location (3)
In this problem, subjects are
asked to decide on a set of
parameters associated with a
simple retail context:
a Lemonade Stand
Search Experiment
4. • All participants exposed to 2 solution development settings:
Group vs. Individual (Nominal)
• first 15 min in one, next 15 min in the alternate setting
• 2 levels of control of parameters in first 15 min:
Generalists vs. Specialists
• 2 Initial solution development settings for Generalists:
Group vs. Individual (Nominal)
• 3 Levels of complexity: Low, Medium, High
• Pay for performance scheme used for recruitment and compensation
Experimental Design
7. Study 1: Generalists setting
Subjects population: MBA students
Total number of subjects: 308 ::: 122 Groups
Study 2: Specialists setting (only nominal first)
Subjects population: University students
(diverse, but control variables like age, gender, years in college, background not significant)
Total number of subjects: 168 ::: 56 3-person groups
Low = 16, Medium=20, High =20
The Study
8. Initial Task Exposure (1st 15 min) Re-exposure to Task (2nd 15 min)
Blocking or Freeriding
Study 1:
Performance of Nominal vs. Collaborative Groups
• Complexity matters!
• Nominal” groups settings generally seem to benefit in more complex task settings
• “Nominal group technique” (first individual, then in groups) performs poorly in
very complex task settings
9. Study 2:
Performance of Generalist vs. Specialist Groups
• Collaborative groups of Specialists perform significantly better than collaborative
groups of Generalists. - - “Nominal group technique” benefits from this difference.
Specialist
Generalist
40%
50%
60%
70%
80%
90%
100%
Low
Complexity
Medium
Complexity
High
Complexity
1st 15 Minutes
(Nominal)
TaskPerformance(%ofMax)
Specialist
Generalist
40%
50%
60%
70%
80%
90%
100%
Low Complexity Medium
Complexity
High Complexity
2nd 15 Minutes
(Collaborative)
TaskPerformance(%ofMax)
Low Medium High
Complexity Complexity Complexity
10. What causes these differences?
• Search Process
– Number of “ideas”
– Coverage of space
– Step size
• Engagement
– Affective Award
– Learning
• Groups effects
– Production blocking
– Evaluation Apprehension
– Freeriding
11. First Period Individual Generalists Individual Specialists Collaborative Generalists
Independent Indiv. Performance Indiv. Performance Group Performance
Variables Beta SE Beta SE Beta SE
Constant 0.666 (0.187) *** 0.720 (0.125) *** 1.396 (0.372) ***
High Complexity -0.202 (0.481) *** -0.142 (0.042) *** -0.558 (0.048) ***
Medium Complexity -0.069 (0.494) -0.113 (0.042) *** -0.492 (0.039) ***
NumSolutions -0.005 (0.002) ** -0.001 (0.001) -0.001 (0.002)
Soln. Coverage 0.568 (0.166) *** 0.242 (0.191) 0.018 (0.166)
Avg. Increment -0.411 (0.139) *** -0.175 (0.085) ** 0.123 (0.121)
Affective Reward 0.021 (0.023) 0.003 (0.022) -0.045 (0.034)
Learning 0.032 (0.020) 0.005 (0.021) 0.009 (0.020)
Lack of Blocking -0.051 (0.043)
Evaluation App. -0.045 (0.035)
Freeriding 0.017 (0.032)
R2 0.283 0.093 0.893
Adj R2 0.248 0.053 0.867
N 150 168 62
Impact Factors in First Period Search
Not Number of Solutions, but
space coverage matters!
Need to search intelligently...
12. First Period Individual Generalists Individual Specialists Collaborative Generalists
Independent Indiv. Performance Indiv. Performance Group Performance
Variables Beta SE Beta SE Beta SE
Constant 0.666 (0.187) *** 0.720 (0.125) *** 1.396 (0.372) ***
High Complexity -0.202 (0.481) *** -0.142 (0.042) *** -0.558 (0.048) ***
Medium Complexity -0.069 (0.494) -0.113 (0.042) *** -0.492 (0.039) ***
NumSolutions -0.005 (0.002) ** -0.001 (0.001) -0.001 (0.002)
Soln. Coverage 0.568 (0.166) *** 0.242 (0.191) 0.018 (0.166)
Avg. Increment -0.411 (0.139) *** -0.175 (0.085) ** 0.123 (0.121)
Affective Reward 0.021 (0.023) 0.003 (0.022) -0.045 (0.034)
Learning 0.032 (0.020) 0.005 (0.021) 0.009 (0.020)
Lack of Blocking -0.051 (0.043)
Evaluation App. -0.045 (0.035)
Freeriding 0.017 (0.032)
R2 0.283 0.093 0.893
Adj R2 0.248 0.053 0.867
N 150 168 62
Impact Factors in First Period Search
13. Nominal Group Technique
Impact of individual phase on group performance?
Combined individual
solution coverage
Production
blocking
Group solution
coverage
Nominal Group
Performance
Group
Performance
Generalists Specialists
+ **
+ +
+**
+
+***+
+** +*
+***
***
+***
+
14. Some Evidence from Questionnaires
Low Collaborative Performance
“…We couldn’t agree on whether to raise or lower the lemon
content. One of the group members said they could get better
results if it was lowered. We went back and forth and spent most
time on it, but maybe should have thought more about the other
issues. Still the debating probably helped use avoid bad solutions.”
High Collaborative Performance
“The mouse was mine, so probably did more than the others. They
both wanted to go in two different directions, and not what I
thought was best (who knows). I kind of tuned out early on and
drove. I kept saying I’d “test that after this” but since we had
momentum we usually didn’t go back…”
There seem to be two paths to performance :
• Leadership that ends bickering by blocking and promoting search
• No blocking – seems more likely when combined individual
coverage is contained (shared mental models?)
15. Nominal Group Technique
Impact of individual phase on group performance?
Combined individual
solution coverage
Production
blocking
Group solution
coverage
Nominal Group
Performance
Group
Performance
Generalists Specialists
+ **
+ +
+**
+
+***+
+** +*
+***
***
+***
+
No bickering, since difference in expertise of members
is recognized by everyone.
16. What about
collaboration before individual search?
Impact of group phase on individual performance?
Group solution
coverage
Individual
Solution
Coverage
Individual
Performance
Production
Blocking
Collaborative
Group
Performance
+*
+**
**
+
+
17. Some more evidence from Questionnaire
High Nominal Performance, and Low Blocking
“When we were in the group we all took turns posting
ideas since we were all new to it. Good dynamic
overall. I think we made some progress figuring out
where the best solutions were. I basically picked up
where we left off when the group split up and when I
was working on my own…”
Low Nominal Performance, and High Blocking
“I didn’t have much of a chance to impact things when
working in the group, so I stopped thinking about the
problem after a while. When I was on my own it was
like my first time on the problem I guess. I honestly
don’t know if I missed something the group discovered
earlier, it was hard to connect back.”
18. Some results so far...
• Nominal groups (best of individuals) benefit relatively from
increasing complexity
• Solution space covered (intelligent search) more important than
number of solutions
• “Nominal Group Technique” (first individuals, than in groups)
performs poorly under high complexity – unless expertise
difference of members is recognized by everyone (credentials) .
• Production blocking might be beneficial in case of nominal group
technique: Leadership ends bickering by blocking and promoting
search
• Not blocking also helps, but more likely when combined individual
coverage is contained (shared mental models = less bickering...)
21. Theorized Dynamics in Detail:
Kavadias and Sommer (2009) propose a normative model designed to characterize
distinctions between collaborative and “nominal” group activity under various
conditions of task complexity.
• Brainstorming activity: multi-agent searches on problems where groups cannot
describe the performance function in advance.
• Sufficient initial consideration of these problems, drive meaningful mental
models that link decisions to performance; as a result progress to good solutions
can be made.
• HOWEVER, increases in the complexity (performance interactions) make it
increasingly more difficult for this to happen.
• Group dynamic effects like production blocking and evaluation
apprehension, make group performance to suffer particularly more so than
“nominal” groups.
Editor's Notes
Mention primer here:Participants are also asked to read a two page primer on optimization and the potential for multimodal complexity in objective terrains (for a two decision variable setting). Following their read, they are given comprehension questions (e.g. “based on this graphical depiction of the relationships between decisions and outcomes, what leads to a global optimum?”).
Through the interface, in both the group and “nominal” settings, participants are allowed to make any modifications within the permitted ranges of the decision variables. They can then submit their decision set for consideration to an Market Analyst – a very basic AI that evaluates their solution and provides feedback. In specialist setting...The Analyst provides strictly positive subjective feedback specific to improvements made in solutions. E.g. “I think there’s a good chance what you’re doing with lemon content could push us to higher profitability”When small local improvements are made during the search, but a globally superior solution has previously been examined, the Analyst also makes recommendations to generally look back at prior solutions. A log of solution history is provided for subjects to review if desired. When solutions do not improve performance, no feedback is given. This feedback structure is designed to mimic recommended practice for brainstorming sessions.
In US data some control variables are significant: years in college (-), nonlinear algebra (+), logical reasoning (+)
*Collaborative “on average” never dominates (though comparable to “Nominal” at Low).
% of maximum achievableDifference significant at high complexity setting. Not elsewhere. For individual setting, we manipulated where starting point was so what maximum they could reach, so do not read as much into performance and performance differences on nominal groups...
From individual part mentionNumber of solutions negative - probably means random trial – goes hand in hand with avg. IncrementsSol coverage however very important*** 1%** 5%*10%Individual specialists: same signs... -- here keep in mind very different landscapes, certainly driving main differences (not included hence low R”)Group: ???
From individual part mentionNumber of solutions negative - probably means random trial – goes hand in hand with avg. IncrementsSol coverage however very important*** 1%** 5%*10%Individual specialists: same signs... -- here keep in mind very different landscapes, certainly driving main differences (not included hence low R”)Group: ???
First one low collaborative perofmance low blockingSecond one high collaborative performance, high blocking