Page 1
Experience Goods and Expectational Traps: Bounded
Rationality and Consumer Behavior in Markets for Medical Care
Presented by: Brian Elbel, MPHPhD Candidate
Yale School of Public Health
With: David Stuckler, MPH
Mark Schlesinger, PhD
At: AcademyHealth Health Economics Interest Group Meeting
Page 2
Outline• Background
– Physician Services as Experience Good– Dyadic v. Generalized Expectations
• Consumers’ Evaluation of Experience– Bayes Rule– Representativeness Heuristic– Expectational Traps
• Data—Consumer Experiences Survey• Estimation Strategy—Selection Model• Results• Conclusions
Page 3
Experience Goods
• Nelson first recognized in the 1970’s
• In order to evaluate a good, you must “experience” or try it out
• Switch/Exit if dissatisfied– Provides incentives to improve quality
Page 4
Experience Goods Literature
• Largely focused on how consumers evaluate purchased goods
• No focus on how consumers make inferences about the distribution of goods in the market
• Generally assumed:– Consumers know the distribution– They then act as Bayesians
Page 5
Physician Services as Experience Good/Service
• Not many other means to assess physicians– Few quality measures– Those that do exist aren’t very good
(small n)– Some learning through social networks;
tastes very heterogeneous
Page 6
Model of Evaluating Experience Goods
• When consumers are evaluating their physician/considering switching
– Assessment of Individual Physician— Dyadic Expectations
– Assessment of Physicians as a Class—Generalized Expectations
Page 7
Consumers Compare Expectations
• Consumers compare Dyadic Expectations to Generalized Expectations
• If expectations are sufficiently divergent, they switch
• Problems arise when both expectations closely track each other
Page 8
Expectations in Response to Problem
• Problems relatively common
• Could use information gained from problematic experience in two ways:– They act as Bayesians– They rely on the Representativeness
Heuristic
Page 9
Bayesian Learning
Pr(MDbad | problem) = Pr(problem | MDbad) x Pr(MDbad)
Pr(problem)
• Expectations should diverge as long as consumer believe “bad” physician have more observable problems– Can’t say for certain what that ratio is
• Consumers likely have few “draws” by which to evaluate physicians– Generalized expectations may largely reflect
dyadic
Page 10
Representativeness Heuristic
• Representativeness: assumption of correspondence, generally between an individual and a population
• Taking knowledge of one physician, and assuming it is representative of all physicians
• After experiencing a problem, generalized expectations equal to dyadic
Page 11
Equal Revision of Expectations?
• Representativeness Heuristic would lead to equal revision of expectations
• Bayes rule maybe could
• Leaving little incentive to switch physicians
• Expectational Traps
• Market Doesn’t Punish Poor Physicians
Page 12
We Find:• On average, following a problematic
experience dyadic expectations are revised downward as much as generalized expectations
• This matters: Divergent expectations predicts switching physicians in response to a problem
• Some evidence due to Representativeness Heuristic
Page 13
Data
• Consumer Experiences Survey• N=5,000• We initially restrict sample to:
– Those with ≤ 1 problem (79.8%)– Those who saw MD in last year
(88.6%)– Those that didn’t switch physicians
(only 7.7% switched)
• Final N =3,071
Page 14
Measures of Expectations
• Measured on 5 dimension for both Dyadic and Generalized Expectations– LEARN: take the time to learn about up to date
treatments– TIME: take enough time with their patients– INSURANCE: speak up for their patients in
disputes with their health plan – ERRORS: make too many mistakes in taking
care of the patients– FAIRNESS: treat all patients fairly regardless of
race• Standardized as a Z-score• Sum them then divide by 5
Page 15
Measures of Problematic Experiences
• Asked if they experienced any of 15 problems in the last year
• Asked who was responsible for problem
• Three Groups– No Problem (55.0%)– Problem Blamed on Physician (13.2%)– Problem Not Blamed on Physician
(31.8%)
Page 16
Model Specification I
• Where– PROB_BL = problem blamed on MD– PROB_NO= problem not blamed on
MD– X = vector of controls– λ = selection term– β = terms to be estimated
4321 __ XNOPROBBLPROBExpec
Page 17
Model Specification II• Outcome = Aggregate Expectations
• Outcome = Individual Expacations– 5 Generalized Expectations– 5 Dyadic Expectations
• Specification same for each
• Controlling for:– SES, Health Status, Recency of Problem,
HC Knowledge, Severity of Problem, Social Support, Managed Care
Page 18
Identification
• Potential Endogeneity of Problem Identification
• Control Function/Treatment Effects/Heckman without Truncation
• Need “instruments”– Otherwise identifying off functional form
• Our instruments:– Presence of Mental Illness, COPD, and an
index of “Don’t Know” responses
Page 20
Dyadic Generalized Difference
Blame NoBlame
Blame NoBlame
Blame NoBlame
AggregateExpectation
(n=3028)
-1.49(0.19)
**
-1.09(0.19)
**
-1.40(0.22)
**
-1.27(0.21)
**
0.09(0.24)
-0.19(0.24)
Influence of Problems on Expectations
Page 21
Dyadic Generalized Difference
Blame NoBlame
Blame NoBlame
Blame NoBlame
AggregateExpectation
(n=3028)
-1.49(0.19)
**
-1.09(0.19)
**
-1.40(0.22)
**
-1.27(0.21)
**
0.09(0.24)
-0.19(0.24)
Influence of Problems on Expectations
Page 22
Dyadic Generalized Difference
Blame NoBlame
Blame NoBlame
Blame NoBlame
AggregateExpectation
(n=3028)
-1.49(0.19)
**
-1.09(0.19)
**
-1.40(0.22)
**
-1.27(0.21)
**
0.09(0.24)
-0.19(0.24)
Influence of Problems on Expectations
Page 23
Influence of Problems on Dyadic and General Expectations of Physicians
Dyadic GeneralizedDifference
Expec. Blame No Blame Blame No Blame Blame No Blame
Learn (n=3047)
-1.69 (.32)**
-1.28 (.32)**
-1.56 (.36)**
-1.51 (.35)**
0.12 (.41)
-0.23 (.41)
Time (n=3099)
-0.86(.34)*
-0.61(.32)
-1.16(.38)**
-1.09(.38)**
-0.31(.42)
-0.48(.41)
Insurance (n=2824)
-1.51(.32)**
-0.98(.34)**
-1.90(.39)**
-1.73(.37)
-0.39(.41)
-0.75(.41)
Errors (n=3034)
-1.24(.25)**
-0.80(.35)**
-0.99(.35)**
-1.01(.34)**
0.26(.50)
-0.21(.49)
Fairness (n=3071)
-1.90(.28)**
-1.37(.30)**
-1.28(.34)**
-1.11(.32)**
0.62(.46)
0.26(.47)
* - denotes p-value significant at 0.05 level; ** - denotes p-value significant at 0.01 level.
Page 24
Both Expectations Revise Equally
• On the whole, expectations tend to revise equally
• But, does this really matter?
Page 25
Switching• DV = Switch MD in response to problems• Selection Model
– Only those with a problem– This time, with truncation
• Made new variable to capture divergence of expectations=Generalized – Dyadic Expectations
• Three categories– >0 and ≤ 2 (34.0%)– >2 and ≤ 3 (4.2%)– > 3 (2.1%)– Excluded Category—Dyadic higher than
Generalized (59.7%)
Page 26
Probability of Switch on Z-Score Difference across Generalized and Dyadic Expectations
=Generalized - Dyadic
>0 ≤ 2 > 2 and ≤ 3 > 3
Aggregate Expectation
0.02 (0.01)** 0.05 (0.01)** 0.09 (0.02)**
* - P-value significant at 0.05 level; ** - at 0.01 level
Page 27
Probability of Switch on Z-Score Difference across Generalized and Dyadic Expectations
Difference between Expectations
Expectation >0 ≤ 2 > 2 and ≤ 3 > 3
Learn 0.05 (0.02)** 0.10 (0.03)** 0.63 (0.11)**
Time -0.01 (0.02) 0.03 (0.04) -0.01 (0.04)
Insurance 0.02 (0.02) 0.06 (0.06) 0.04 (0.13)
Errors 0.03 (0.02) 0.10 (0.04)** 0.14 (0.05)**
Fairness 0.05 (0.02)** -0.21 (0.29) 0.18 (0.05)**
* - P-value significant at 0.05 level; ** - at 0.01 level
Page 28
Switching Responds to Expectations
• More divergent expectations leads to more switching for 3 of 5 expectations
• Divergent Expectations Matter
Page 29
Bayesian Learning or Representativeness Heuristic?• Some Differences in Categories of
Expectations
• Responses of those with a greater sense of the base rate—Long-Term Medical Condition
• Blame v. No_Blame Results
Page 30
Limitations
• Cross-Sectional Data
• Omitted Variables
• Noisy Measures
Page 31
Conclusions
• Consumers revise Generalized Expectations essentially as much as Dyadic Expectations
• Expectational Traps: This likely does explain some of the low-switching rates
• Physician’s lack market incentive to improve quality
• Some consumers likely not acting as Bayesians