research frontiers nicolai v. kuminoffnkuminof/lec5.pdf · 2017-06-13 · challenge for future...
TRANSCRIPT
RESEARCH FRONTIERS
DEPARTMENT OF ECONOMICS
Nicolai V. Kuminoff
(ASU Economics & NBER)
Copy of slides: www.public.asu.edu/~nkuminof/lec5.pdf
1. Spatial dispersion of amenities
2. Strategic regulatory behavior
3. Heterogeneity in information and beliefs
4. Revealed preferences analysis when some choices don’t reveal preferences
5. Can features of spatial equilibria inform the VSL?
TOPICS
Spatial dispersion of air
pollution is non-uniform
This matters for hedonic
estimation
Strategic Regulatory Behavior
Strategic Regulatory Behavior
Strategic Regulatory Behavior
Strategic Regulatory Behavior
My Question: Is the projected improvement in
air quality simply caused by diverting auto
traffic away from air quality monitors?
AZDOT Answer: “Siting, operation, and
recording information from monitoring sites are
the responsibility of the Maricopa Air Quality
Department…According to EPA guidance, new
monitors are not necessary to analyze air
quality impacts.”
Broader Evidence of Strategic Behavior:
Aufhammer, Bento and Lowe (JEEM 2009)
Kahn and Mansur (JPUBE 2013)
Grainger, Schreiber and Chang (2016)
If people are fully informed
about amenities, then
information disclosures will
have no effect on prices.
Disclosure increases price
discount by 37 percentage points
1. Strategic behavior on the part of firms and regulators may weaken the
observable assignment of pollution to people
2. There is heterogeneity in information and beliefs
3. Pollution may create heterogeneity in cognition and decision making
Question: what does this mean for revealed preferences and policy analysis?
Summary
Frictions and Nudges
Consumers leave money on the table
• Automobiles: Busse et al. (QJE 2015), Lacetera et al. (AER 2012)
• Home loans: Agarwal & Mazumder (AEJ 2013), Woodward & Hall (AER 2012)
• Health insurance: Ketcham et al. (AER 2012,2014), Handel & Kolstad (AER 2015)
• Retirement savings: Bernheim et al. (AER 2015), Chetty et al. (QJE 2015)
• Energy: Allcott & Taubinsky (AER 2015), Gillingham & Palmer (REEP 2014)
Explanations: costly information, cognitive ability, psychological biases,
latent product quality
Federal policies increasingly manipulate choice architecture
• Decision support: EPA air quality index; CMS Medicare plan finder
• Default options: HHS (2014) proposal for assignment to low cost plans in ACA
• Menu restrictions: CAFE standards, appliance efficiency standards, light bulb ban
The Air Quality Index (AQI)
• Established in 1999, it translates ambient readings of ozone, PM, CO, NO2, and SO2 into an overall index of health risk from outdoor exposure to air pollution.
• States are required to use major media outlets to report daily AQI values to the general public in all metro areas with greater than 350,000 people.
• Covers 65% of the U.S. population
The Air Quality Index (AQI)
The AQI is included in weather reports in the television news, newspapers, radio, internet, and apps for computers, tablets, and smartphones.
Frictions and Nudges
Consumers leave money on the table
• Automobiles: Busse et al. (QJE 2015), Lacetera et al. (AER 2012)
• Home loans: Agarwal & Mazumder (AEJ 2013), Woodward & Hall (AER 2012)
• Health insurance: Ketcham et al. (AER 2012,2014), Handel & Kolstad (AER 2015)
• Retirement savings: Bernheim et al. (AER 2015), Chetty et al. (QJE 2015)
• Energy: Allcott & Taubinsky (AER 2015), Gillingham & Palmer (REEP 2014)
Explanations: costly information, cognitive ability, psychological biases,
latent product quality
Federal policies increasingly manipulate choice architecture
• Decision support: EPA air quality index; CMS Medicare plan finder
• Default options: HHS (2014) proposal for assignment to low cost plans in ACA
• Menu restrictions: CAFE standards, appliance efficiency standards, light bulb ban
Under the guidance of the US Social and Behavioral Sciences Team Federal agencies must:
“…identify programs that offer choices and carefully consider how the presentation and structure of those choices, including the order, number, and arrangement of options, can most effectively promote public welfare, as appropriate, giving particular consideration to the selection of default options.”
—Section 1(b)(iii) of Executive Order #13707
September 15, 2015
EXECUTIVE ORDER 17307
Using Behavioral Science Insights to Better Serve the American People
Challenge for Future Research
Challenge: refine the current revealed preference sorting framework to analyze equity and efficiency of policies that modify choice architecture in markets for differentiated goods.
Catch-22: such a framework is only needed in settings where some people’s choices do not reveal their preferences.
Research question: how?
A Conceptual Approach: Bernheim-Rangel (QJE 2009)
1. Identify choices that we suspect do not reveal consumers’ preferences: “suspect choices”
2. Estimate preference parameters from choices that we believe reveal preferences: “non-suspect choices”
3. Assume that observationally identical people making suspect and non-suspect choices share the same underlying preferences
4. Evaluate welfare for prospective policies targeting choice architecture
• Predict policy’s effect on the choice process
• Step 1: identify non-suspect choices
o info treatment (Allcott-Taubinsky AER 2015)
o knowledge test (Handel-Kolstad AER 2015)
o RP test (Ketcham et al. AER 2016)
o Survey information (Ketcham et al. NBER 2016)
• Step 2: estimate decision parameters for suspect and non-suspect groups
• Step 3: evaluate prospective policies
Estimating the Heterogeneous Welfare Effects of Choice Architecture
DEPARTMENT OF ECONOMICS
Jonathan D. Ketcham* Nicolai V. Kuminoff† Christopher A. Powers‡
* Arizona State University Marketing Department
† Arizona State University Economics Department, and NBER
‡ United States Centers for Medicare and Medicaid Services
March 2017
This research was supported by a grant from the National Institute for Health Care Management (NIHCM) Research and Educational Foundation. All results and products
derived from this research are the responsibility of the research team; the findings do not necessarily represent the views of the NIHCM Research and Educational Foundation.
Our Empirical Setting: Medicare Part D
Prescription drug insurance is a differentiated good
• Average consumer choses among 50 plans offered by 20 insurers
• Plans differ in cost, risk protection, and quality
• Average senior spends about $1,300 per year (6% of income)
• Markets enroll 24 million seniors with annual federal outlays over $65 billion
Proposals to simplify choice architecture
• Restrict insurers to sell no more than two plans (2014 CMS proposal)
• Assign consumers to low-cost default plans (2014 HHS proposal)
• Personalized decision support (Kling et al. QJE 2012)
Prescription Drug Plan (PDP) Choice
• Beliefs about PDPs: 𝑉 𝑐𝑗 , 𝑞𝑗 , 𝐹𝑖; 𝜃𝑖 , 𝑏𝑖 = 𝑉 𝑐 𝑖𝑗 , 𝑞 𝑖𝑗 , 𝐹 𝑖; 𝜃𝑖
• Objective measures of PDPs: 𝑉 𝑐𝑗 , 𝑞𝑗 , 𝐹𝑖; 𝜃𝑖
• Revealed preference analysis: 𝜃 = min𝜃∈Θ
𝑄𝑁 𝑐𝑗 , 𝑞𝑗 , 𝐹𝑖 , 𝜃
• Consistency of an estimator for 𝜃:
As 𝑁 → ∞, 𝜃 → 𝜃 if 𝑐𝑗 = 𝑐 𝑖𝑗 , 𝑞𝑗 = 𝑞 𝑖𝑗, and 𝐹𝑖 = 𝐹 𝑖 ∀ 𝑖
Identifying Suspect Choices
max𝑗∈𝐽
𝑉 𝑐 𝑖𝑗𝑡 , 𝑞 𝑖𝑗𝑡 , 𝐹 𝑖𝑡; 𝜃𝑖
Definition. A choice is suspect if we have reason to believe it does not reveal the consumer’s preferences for PDP attributes because 𝑐 𝑖𝑗𝑡 ≠ 𝑐𝑗𝑡, 𝑞 𝑖𝑗𝑡 ≠ 𝑞𝑗𝑡,
and/or 𝐹 𝑖𝑡 ≠ 𝐹𝑖𝑡.
Nonsuspect choices are assumed to be informed in the sense that the decision maker’s beliefs about plan attributes coincide with the objective measures as we have defined them
Potential indicators grounded in theory
1. Incorrect answer to MCBS knowledge question
2. Choice violates axioms of consumer preference theory
Suspect Choice Indicator #1: Knowledge Test
(true / false) “ Your out-of-pocket costs are the same in all Medicare prescription drug plans ”
• True for people with no drug claims
• False for people with any drug claims
• We code it as correct if the enrollee gave the right answer for year t-1 or t
•1
𝑁𝑇 max
𝑗𝑜𝑜𝑝𝑗𝑡 𝑥𝑖𝑡 − min
𝑗𝑜𝑜𝑝𝑗𝑡 𝑥𝑖𝑡𝑁𝑇 = $1,112
Suspect Choice Indicators, by Year
• We observe if the knowledge question is answered by the beneficiary or a proxy.
• Respondents who answered incorrectly could have saved 16% more by switching, on average, than those who answered correctly.
2006 2007 2008 2009 2010
fails knowledge test 44 37 34 29 28
plan dominated ex post 19 18 18 16 15
fails knowledge test U plan dominated ex post 54 48 45 40 38
Suspect choice indicator
Percent of choices
Suspect Choice Indicator #2: Preference Axiom Test
If preferences satisfy completeness, transitivity, monotonicity, and risk aversion, then a consumer will never choose a plan, j, that is dominated by another, k, in the sense that (i)-(iv) hold simultaneously:
(i) E 𝑐𝑖𝑘𝑡 ≤ E 𝑐𝑖𝑗𝑡
(ii) var 𝑐𝑖𝑘𝑡 ≤ var 𝑐𝑖𝑗𝑡
(iii) 𝑞𝑗𝑡 ≤ 𝑞𝑘𝑡
(iv) At least one inequality is strict
Dominated plan choices are:
• Off the efficiency frontier in attribute space (Lancaster JPE 1966)
• Inconsistent with the axioms of consumer theory; e.g. implying the decision maker must be risk loving or have negative marginal utility of income (Ketcham, Kuminoff and Powers AER 2016)
Suspect Choice Indicators, by Year
• Dominated in terms of mean and variance of ex post cost, CMS vertical quality index, and brand.
2006 2007 2008 2009 2010
fails knowledge test 44 37 34 29 28
plan dominated ex post 19 18 18 16 15
fails knowledge test U plan dominated ex post 54 48 45 40 38
Suspect choice indicator
Percent of choices
Suspect Choice Indicators, by Year
2006 2007 2008 2009 2010
fails knowledge test 44 37 34 29 28
plan dominated ex post 19 18 18 16 15
fails knowledge test U plan dominated ex post 54 48 45 40 38
Suspect choice indicator
Percent of choices
Suspect Choice Definition
Enrollment decisions coded as suspect if any of the following are true:
i. Active choice
a. of a dominated plan
b. while giving wrong answer to MCBS knowledge question
ii. Passive reenrollment in a plan that was
a. dominated when it was actively chosen, or
b. actively chosen in enrollment cycle with wrong answer to knowledge question
• Drop about 4,000 passive reenrollments for which we do not observe the enrollee’s knowledge at the time of their original active enrollment into that plan.
• Drop all enrollment decisions in 2006 due to evidence of learning.
• Final sample has 9,119 PDP choices (42% suspect) made by 3,444 enrollees.
• Reference person: 65-69 year old white
male who does not have a high school
degree but makes decisions on his own,
has not been diagnosed with cognitive
illnesses, and has not searched for CMS
info using internet or 1-800-Medicare
• linear probability model includes
indicators for year and CMS region
• Standard errors are robust and
clustered by enrollee
college graduate -0.058 [0.021]***
income>$25k -0.012 [0.019]
currently working 0.009 [0.026]
married 0.011 [0.020]
has living children -0.064 [0.034]*
uses the internet -0.015 [0.022]
searched for CMS info: internet -0.083 [0.021]***
searched for CMS info: 1-800-Medicare -0.066 [0.020]***
has help making insurance decisions 0.016 [0.018]
number of available plans (standardized) -0.003 [0.016]
female 0.028 [0.019]
nonwhite 0.114 [0.036]***
age: 70-74 0.047 [0.023]**
age: 75-79 0.065 [0.027]**
age: 80-84 0.071 [0.028]**
age: over 84 0.118 [0.030]***
dementia including Alzheimer's 0.040 [0.027]
depression 0.011 [0.023]
number of drug claims (standardized) 0.033 [0.008]***
number of plan choices
number of enrollees
mean of the dependent variable
R-squared
Suspect Choice
9,119
3,444
0.42
0.059
Logit Estimation: Initial Enrollment Decisions
𝑈𝑖𝑗𝑠 = 𝑝𝑗𝑠 + 𝜇𝑖𝑗𝑠 𝛼𝑖𝑠 + 𝜎𝑖𝑗𝑠2 𝛽𝑖𝑠 + 𝑞𝑗𝑠𝛾𝑖𝑠 + 𝜖𝑖𝑗𝑠
𝑝𝑗𝑠 + 𝜇𝑖𝑗𝑠 = plan premium + individual OOP costs
𝜎𝑖𝑗𝑠2 = variance of individual OOP costs
𝑞𝑗𝑠 = CMS quality index, brand dummies
𝜖𝑖𝑗𝑠 = iid type I EV preference shocks
• 𝛼𝑖𝑠 = 𝛼0 + 𝑑𝑖𝑠𝛼1, where 𝑑𝑖𝑠 is a vector of demographics
• Costs of search and enrollment assumed to be constant over plans
• Initial plan choice becomes the default in period 𝑡 = 𝑠 + 1
Logit Estimation: Subsequent Decisions
𝑈𝑖𝑗𝑡 = 𝑝𝑗𝑡 + 𝜇𝑖𝑗𝑡 𝛼𝑖𝑡 + 𝜎𝑖𝑗𝑡2 𝛽𝑖𝑡 + 𝑞𝑗𝑡𝛾𝑖𝑡 + 𝜂𝑖𝑡Δ𝐵𝑖𝑗𝑡 + 𝛿𝑖𝑡Δ𝑃𝑖𝑗𝑡 + 𝜖𝑖𝑗𝑡
𝑝𝑗𝑡 + 𝜇𝑖𝑗𝑡 = plan premium + individual OOP costs
𝜎𝑖𝑗𝑡2 = variance of individual OOP costs
𝑞𝑗𝑡 = CMS quality index, brand dummies
𝜖𝑖𝑗𝑡 = iid type I EV preference shocks
Δ𝐵𝑖𝑗𝑡,Δ𝑃𝑖𝑗𝑡 = dummies for active choice to switch out of
the default brand or plan
Logit Estimation: Subsequent Decisions
𝑈𝑖𝑗𝑡 = 𝑝𝑗𝑡 + 𝜇𝑖𝑗𝑡 𝛼𝑖𝑡 + 𝜎𝑖𝑗𝑡2 𝛽𝑖𝑡 + 𝑞𝑗𝑡𝛾𝑖𝑡 + 𝜂𝑖𝑡Δ𝐵𝑖𝑗𝑡 + 𝛿𝑖𝑡Δ𝑃𝑖𝑗𝑡 + 𝜖𝑖𝑗𝑡
𝑝𝑗𝑡 + 𝜇𝑖𝑗𝑡 = plan premium + individual OOP costs
𝜎𝑖𝑗𝑡2 = variance of individual OOP costs
𝑞𝑗𝑡 = CMS quality index, brand dummies
𝜖𝑖𝑗𝑡 = iid type I EV preference shocks
Δ𝐵𝑖𝑗𝑡,Δ𝑃𝑖𝑗𝑡 = dummies for active choice to switch out of
the default brand or plan
Inertia due to hassle costs, latent preferences, and psychological biases
Multinomial Logit Results: Main Effects
• Reference person: 78 years old, white, no college degree, income below $25k, drug claims in middle tercile
of the distribution, no help making decisions, did not search for CMS info on web or 1-800-Medicare.
• Replicate the prior finding that variance coefficient is small and insignificant
(Abaluck and Gruber AER 2011; Ketcham, Kuminoff and Powers AER 2016)
expected cost -0.283 [0.017]*** -0.377 [0.029]*** -0.197 [0.021]***
variance 0.076 [0.085] -0.433 [0.118]*** 0.621 [0.126]***
quality (CMS index) 0.035 [0.078] 0.056 [0.104] -0.012 [0.124]
within-brand switch -3.307 [0.109]*** -3.239 [0.152]*** -3.396 [0.155]***
between-brand switch -5.181 [0.095]*** -4.923 [0.128]*** -5.591 [0.141]***
pseudo R2
number of enrollment decisions
number of enrollees
All ChoicesNon-Suspect
choicesSuspect choices
3,442 2,175 1,560
0.66 0.64 0.71
9,119 5,248 3,871
Multinomial Logit Results: Main Effects
• Reference person: 78 years old, white, no college degree, income below $25k, drug claims in middle tercile
of the distribution, no help making decisions, did not search for CMS info on web or 1-800-Medicare.
• Implied risk premium for NS group in line with prior literature (50-50 bet {1000,892})
(Cohen and Einav AER 2007, Handel AER 2013, Handel and Kolstad AER 2015)
expected cost -0.283 [0.017]*** -0.377 [0.029]*** -0.197 [0.021]***
variance 0.076 [0.085] -0.433 [0.118]*** 0.621 [0.126]***
quality (CMS index) 0.035 [0.078] 0.056 [0.104] -0.012 [0.124]
within-brand switch -3.307 [0.109]*** -3.239 [0.152]*** -3.396 [0.155]***
between-brand switch -5.181 [0.095]*** -4.923 [0.128]*** -5.591 [0.141]***
pseudo R2
number of enrollment decisions
number of enrollees
All ChoicesNon-Suspect
choicesSuspect choices
3,442 2,175 1,560
0.66 0.64 0.71
9,119 5,248 3,871
Multinomial Logit Results: Main Effects
• Reference person: 78 years old, white, no college degree, income below $25k, drug claims in middle tercile
of the distribution, no help making decisions, did not search for CMS info on web or 1-800-Medicare.
• Implied WTP to avoid switching brands is $2,958 for suspect group and $1,292 for
non-suspect group, consistent with prior evidence on inertia
expected cost -0.283 [0.017]*** -0.377 [0.029]*** -0.197 [0.021]***
variance 0.076 [0.085] -0.433 [0.118]*** 0.621 [0.126]***
quality (CMS index) 0.035 [0.078] 0.056 [0.104] -0.012 [0.124]
within-brand switch -3.307 [0.109]*** -3.239 [0.152]*** -3.396 [0.155]***
between-brand switch -5.181 [0.095]*** -4.923 [0.128]*** -5.591 [0.141]***
pseudo R2
number of enrollment decisions
number of enrollees
All ChoicesNon-Suspect
choicesSuspect choices
3,442 2,175 1,560
0.66 0.64 0.71
9,119 5,248 3,871
Welfare Analysis
Non-suspect choices—infer preferences from observed choices
Suspect choices — infer preferences from non-suspect choices
Let 𝜃 = 𝛼, 𝛽, 𝛾, 𝛿, 𝜂
Assumption: 𝜃𝑠 = 𝜃 𝑛and 𝐹𝑠 𝜖𝑖𝑗𝑡 , 𝐹𝑛 𝜖𝑖𝑗𝑡 ~ type I EV
Welfare Effects of Modifying Choice Architecture
𝐸 𝐶𝑆𝑖 =1
𝛼 𝑛 𝑙𝑛 𝑒𝑥𝑝 𝑉 𝑖𝑗𝑗∈𝐽 + 𝐶 + 𝜓𝑖𝑗 𝑉𝑖𝑗𝑛 − 𝑉 𝑖𝑗𝑗∈𝐽
where, 𝑉 𝑖𝑗 is logit estimate for observable part of utility | 𝑖 ∈ 𝑆 or 𝑖 ∈ 𝑁𝑆
𝑉𝑖𝑗𝑛 is logit estimate for observable part of utility | 𝑑𝑖𝑡 , 𝜃
𝑛𝑠 , and
multinomial logit choice probability: 𝜓𝑖𝑗 =𝑒𝑥𝑝 𝑉 𝑖𝑗
𝑒𝑥𝑝 𝑉 𝑖𝑚𝑚∈𝐽
Welfare Effects of Modifying Choice Architecture
𝐸 𝐶𝑆𝑖 =1
𝛼 𝑛 𝑙𝑛 𝑒𝑥𝑝 𝑉 𝑖𝑗𝑗∈𝐽 + 𝐶 + 𝜓𝑖𝑗 𝑉𝑖𝑗𝑛 − 𝑉 𝑖𝑗𝑗∈𝐽
where, 𝑉 𝑖𝑗 is logit estimate for observable part of utility | 𝑖 ∈ 𝑆 or 𝑖 ∈ 𝑁𝑆
𝑉𝑖𝑗𝑛 is logit estimate for observable part of utility | 𝑑𝑖𝑡 , 𝜃
𝑛𝑠 , and
multinomial logit choice probability: 𝜓𝑖𝑗 =𝑒𝑥𝑝 𝑉 𝑖𝑗
𝑒𝑥𝑝 𝑉 𝑖𝑚𝑚∈𝐽
standard expression for consumer surplus derived by Small and Rosen (ECMA 1981)
Welfare Effects of Modifying Choice Architecture
𝐸 𝐶𝑆𝑖 =1
𝛼 𝑛 𝑙𝑛 𝑒𝑥𝑝 𝑉 𝑖𝑗𝑗∈𝐽 + 𝐶 + 𝜓𝑖𝑗 𝑉𝑖𝑗𝑛 − 𝑉 𝑖𝑗𝑗∈𝐽
where, 𝑉 𝑖𝑗 is logit estimate for observable part of utility | 𝑖 ∈ 𝑆 or 𝑖 ∈ 𝑁𝑆
𝑉𝑖𝑗𝑛 is logit estimate for observable part of utility | 𝑑𝑖𝑡 , 𝜃
𝑛𝑠 , and
multinomial logit choice probability: 𝜓𝑖𝑗 =𝑒𝑥𝑝 𝑉 𝑖𝑗
𝑒𝑥𝑝 𝑉 𝑖𝑚𝑚∈𝐽
standard expression for consumer surplus derived by Small and Rosen (ECMA 1981)
Adjustment derived by Leggett (ERE 2002)
Welfare Effects of Modifying Choice Architecture
∆𝐸 𝐶𝑆𝑖𝑛 =
1
𝛼 𝑛 𝑙𝑛 𝑒𝑥𝑝 𝑉𝑖𝑘
𝑛1𝑘∈𝐾
𝑒𝑥𝑝 𝑉𝑖𝑗𝑛0
𝑗∈𝐽
∆𝐸 𝐶𝑆𝑖𝑠 =
1
𝛼 𝑛 𝑙𝑛 𝑒𝑥𝑝 𝑉𝑖𝑘
𝑠1𝑘∈𝐾
𝑒𝑥𝑝 𝑉𝑖𝑗𝑠0
𝑗∈𝐽
+ 𝜓𝑖𝑘1 𝑉𝑖𝑘
𝑛1 − 𝑉𝑖𝑘𝑠1
𝑘∈𝐾 − 𝜓𝑖𝑗0 𝑉𝑖𝑗
𝑛0 − 𝑉𝑖𝑗𝑠0
𝑗∈𝐽
Welfare Effects of Modifying Choice Architecture
∆𝐸 𝐶𝑆𝑖𝑛 =
1
𝛼 𝑛 𝑙𝑛 𝑒𝑥𝑝 𝑉𝑖𝑘
𝑛1𝑘∈𝐾
𝑒𝑥𝑝 𝑉𝑖𝑗𝑛0
𝑗∈𝐽
∆𝐸 𝐶𝑆𝑖𝑠 =
1
𝛼 𝑛 𝑙𝑛 𝑒𝑥𝑝 𝑉𝑖𝑘
𝑠1𝑘∈𝐾
𝑒𝑥𝑝 𝑉𝑖𝑗𝑠0
𝑗∈𝐽
+ 𝜓𝑖𝑘1 𝑉𝑖𝑘
𝑛1 − 𝑉𝑖𝑘𝑠1
𝑘∈𝐾 − 𝜓𝑖𝑗0 𝑉𝑖𝑗
𝑛0 − 𝑉𝑖𝑗𝑠0
𝑗∈𝐽
standard log sum ratio post-policy adjustment pre-policy adjustment
∆𝐸 𝐶𝑆𝑖𝑛 =
1
𝛼 𝑛 𝑙𝑛 𝑒𝑥𝑝 𝑉𝑖𝑘
𝑛1𝑘∈𝐾
𝑒𝑥𝑝 𝑉𝑖𝑗𝑛0
𝑗∈𝐽
∆𝐸 𝐶𝑆𝑖𝑠 =
1
𝛼 𝑛 𝑙𝑛 𝑒𝑥𝑝 𝑉𝑖𝑘
𝑠1𝑘∈𝐾
𝑒𝑥𝑝 𝑉𝑖𝑗𝑠0
𝑗∈𝐽
+ 𝜓𝑖𝑘1 𝑉𝑖𝑘
𝑛1 − 𝑉𝑖𝑘𝑠1
𝑘∈𝐾 − 𝜓𝑖𝑗0 𝑉𝑖𝑗
𝑛0 − 𝑉𝑖𝑗𝑠0
𝑗∈𝐽
• Consistent with Bernheim & Rangel’s (QJE 2009) proposal to use proxies when ancillary conditions inhibit revealed preference logic.
• Consistent with Kahneman et al.’s (QJE 1997) “hedonic” and “decision” utility.
• Consistent with Camerer et al.’s (UPLR 2003) proposal to evaluate policies based on an “asymmetric paternalism” criterion.
Welfare Effects of Modifying Choice Architecture
standard log sum ratio post-policy adjustment pre-policy adjustment
∆𝐸 𝐶𝑆𝑖𝑛 =
1
𝛼 𝑛 𝑙𝑛 𝑒𝑥𝑝 𝑉𝑖𝑘
𝑛1𝑘∈𝐾
𝑒𝑥𝑝 𝑉𝑖𝑗𝑛0
𝑗∈𝐽
∆𝐸 𝐶𝑆𝑖𝑠 =
1
𝛼 𝑛 𝑙𝑛 𝑒𝑥𝑝 𝑉𝑖𝑘
𝑠1𝑘∈𝐾
𝑒𝑥𝑝 𝑉𝑖𝑗𝑠0
𝑗∈𝐽
+ 𝜓𝑖𝑘1 𝑉𝑖𝑘
𝑛1 − 𝑉𝑖𝑘𝑠1
𝑘∈𝐾 − 𝜓𝑖𝑗0 𝑉𝑖𝑗
𝑛0 − 𝑉𝑖𝑗𝑠0
𝑗∈𝐽
Information needed to evaluate a prospective policy
1. 𝑉𝑖𝑗𝑠0, 𝑉𝑖𝑗
𝑛0 infer from observed behavior
2. 𝑉𝑖𝑘𝑠1, 𝑉𝑖𝑘
𝑛1 predict based on features of the policy
Welfare Effects of Modifying Choice Architecture
standard log sum ratio post-policy adjustment pre-policy adjustment
• Step 1: identify non-suspect choices
o info treatment (Allcott-Taubinsky AER 2015)
o knowledge test (Handel-Kolstad AER 2015)
o RP test (Ketcham et al. AER 2016)
o Survey information (Ketcham et al. NBER 2016)
• Step 2: estimate decision parameters for suspect and non-suspect groups
• Step 3: evaluate prospective policies
1. Spatial dispersion of amenities
2. Strategic regulatory behavior
3. Heterogeneity in information and beliefs
4. Revealed preferences analysis when some choices don’t reveal preferences
5. Can features of spatial equilibria inform the VSL?
TOPICS
• EPA regulations account for 44%-56% of all costs and 61%-80% of all benefits.
• Lee and Taylor (2013) estimate that up to 70% of benefits across all rules are due to
mortality reductions valued by VSL
• Also consider
effects on
housing,
agriculture,
forestry, and
recreation
CAAA ⇒ ΔPM2.5 ⇒ 160,000 fewer deaths x $7.4 million = $1.2 trillion in 2010
“ First, the wages of labor vary with the ease
or hardship, the cleanliness or dirtiness, the
honorableness or dishonorableness of the
employment…A journeyman blacksmith,
though an artificer, seldom earns so much in
twelve hours as a collier, who is only a laborer,
does in eight. His work is not quite so dirty, is
less dangerous, and is carried on in day-light,
and above ground. ”
— Adam Smith, Wealth of Nations (1776)
Chapter X, Part I.
• Thaler and Rosen (NBER 1976) formalized
the theory in terms of state-dependent
utility and an equilibrium hedonic wage
function, and reported some early VSL
estimates, building on earlier work by
Jones-Lee (JPE 1974) and Shelling (1968).
Suppose wage is measured as hourly
earnings, risk is measured as annual deaths
per 1,000 workers, and the hedonic wage
equation is linear and additively separable:
𝑉𝑆𝐿 =𝜕𝑤𝑎𝑔𝑒
𝜕𝑟𝑖𝑠𝑘× 𝑎𝑣𝑔 ℎ𝑜𝑢𝑟𝑠 / 𝑦𝑒𝑎𝑟 × 1,000
Industry, occupation
Mining, accountant
Mining, miner
Education, research professor
Education, extension professor
Track job switchers
Implications for Evaluating the Clean Air Act Amendments
CAAA Benefits and Costs in 2020 ($2006 billion)
• $65 billion cost estimate includes some (but not all) costs from Becker and Henderson (JPE 2000) and Greenstone (JPE 2002)
• Excludes $7 billion in foregone wages from Walker (QJE 2013)
Implications for Evaluating the Clean Air Act Amendments
Implications for Evaluating the Clean Air Act Amendments
Implications for Evaluating the Clean Air Act Amendments
Implications for Evaluating the Clean Air Act Amendments
Implications for Evaluating the Clean Air Act Amendments
Implications for Evaluating the Clean Air Act Amendments
Why aren’t health effects of air pollution fully
capitalized into property values?
Hypothesis 1: they are capitalized into wages
Hypothesis 2: homebuyers are uninformed
Hypothesis 3: homebuyers are discounting the future
Hypothesis 4: VSL is overestimated
Hypothesis 5: heterogeneity in willingness to pay to
reduce mortality risk
• Disentangling these hypotheses is crucial for learning how spatial equilibrium can inform benefit transfer and environmental policy
Implications for Evaluating the Clean Air Act Amendments
0
200
400
600
800
1,000
1,200
1,400
1,600
1,800
2,000
EPA cost EPA cost+ wage loss
(Walker QJE 2013)
SortingBKT (JEEM 2009)
EPA Benefit PM Mortality(VSL = $8.8m)
Housing PCE(NIPA - 2015)
Why aren’t health effects of air pollution fully
capitalized into property values?
Hypothesis 1: they are capitalized into wages
Hypothesis 2: homebuyers are uninformed
Hypothesis 3: homebuyers are discounting the future
Hypothesis 4: VSL is overestimated
Hypothesis 5: heterogeneity in willingness to pay to
reduce mortality risk
• Disentangling these hypotheses is crucial for learning how spatial equilibrium can inform benefit transfer and environmental policy
Spatial Migration Tends to Reduce PM2.5 Exposure…
-0.60
-0.55
-0.50
-0.45
-0.40
-0.35
-0.30
-0.25
-0.20
66 71 76 81 86 91 96
ann
ual
ch
ange
in m
ean
ho
url
y P
M2
.5(μ
/m3)
age
non-movers
movers
Implications for Evaluating the Clean Air Act Amendments
0
200
400
600
800
1,000
1,200
1,400
1,600
1,800
2,000
EPA cost EPA cost+ wage loss
(Walker QJE 2013)
SortingBKT (JEEM 2009)
EPA Benefit PM Mortality(VSL = $8.8m)
Housing PCE(NIPA - 2015)
Why aren’t health effects of air pollution fully
capitalized into property values?
Hypothesis 1: they are capitalized into wages
Hypothesis 2: homebuyers are uninformed
Hypothesis 3: homebuyers are discounting the future
Hypothesis 4: VSL is overestimated
Hypothesis 5: heterogeneity in willingness to pay to
reduce mortality risk
• Disentangling these hypotheses is crucial for learning how spatial equilibrium can inform benefit transfer and environmental policy
Whose VSL should we be trying to measure?
Thank you!
Further reading:
1. Papers cited on the website and slides
2. References in Kuminoff, Smith and Timmins (JEL 2013)
Questions: [email protected]