---

Energy Efficiency

[Econ 8852](http://www.richard-sweeney.com/ee_topics) [Prof. Richard L. Sweeney](http://www.richard-sweeney.com/)

[to print slides, append "?print-pdf" to the url]

---

Intro

Lecture draws on:

Allcott and Greenstone (JEP 2012) Allcott (Annual Review 2014) Gerarden, Newell and Stavins (JEL 2018)

---

Energy services require capital

Source: Allcott and Greenstone (2012)

---

Long standing assertion that energy efficient capital suboptimally deployed

• Serious interest began during oil crisis in 1970s

• Engineers concluded that many energy saving technologies were slow to diffuse

  • for example Amory Lovins and RMI

Over time interest has centered on apparent “win-win” from environmental perspective

• Energy consumption associated with many externalities

---

40 years later this view still has a large audience

---

There are many “rational” reasons for EE gap

• agency issues
  • ie landlord tenant
• resale issues
  • home improvements
• engineering models wrong
  • Fowlie, Greenstone and Wolfram (QJE 2018)

---

Research focussed on “behavioral” explanations

• [present bias]
  • tradeoff is over cash flows not consumption
    • car or mortgage payments also flows
  • firm mispricing problem less clear
    • car companies don’t sell you gasoline
• inattention/ salience
  • Chetty et al (2009); Hossain \& Morgan (2006)
• cognitive issues /confusion
  • Abaluck \& Gruber (2011); Allcott (2013)

---

Setup

[from Allcott (2014)]

Consumers have indirect utility \(u_{j} = \eta (y - p_j - \gamma g_j) + \nu_j\)

• $p_j$ is upfront cost
• $\nu_j$ is the usage utility
• $g_j$ is the lifetime energy cost

Assuming $g$ is calculated and discounted appropriately, a natural test is to estimate $\gamma$ and test if its equal to 1.

---

Lifetime energy costs

\[g_{j} = \sum\limits_{t=0}^{T} \delta^t m(r_j,e_t)r_j e_t\]

• $m$ is utilization
• $r_j$ is energy requirement for capital $j$
• $e$ is the energy price

Note

• $\delta, m, T, e$ could all have $i$ subscripts
• if $m$ is endogenous, can lead to a “rebound effect”

---

Early literature tried to estimate $\delta_i$

Hausman (1979) estimates a discrete choice model for AC’s

\[u_{ij} = \eta (y_i - p_j - \delta_i \bar m_i r_j e_t) + \alpha X_j + \epsilon_{ij}\]

• Has a small survey of households with submetered ACs

---

Literature then turned to within product variation

• In the cross section, concern that $E[r_j \epsilon_j] \ne 0$

However, $g_{ij}$ varies for many other reasons within product

• energy prices vary
  • across regions, time
• lifetime varies at time of sale
  • Sallee et al. look at mileage

---

Allcott and Wozny

ReStat (2014)

• What’s the research question?
• What’s the empirical strategy?
• What data do they have?

---

AW cost to own

\(p_{j} + \sum\limits_{t=0}^{T} \delta^t m(r_j,e_t)r_j e_t\)

• used car auction vehicle prices from Manhiem
  • what are the assumptions using this?
• assumed real discount rate of 6%
  • vehicle loan rate 6.9%
  • S\&P avg return 5.9%
• (exogenous) VMT ($m$) and $T$ from NHTSA
• fuel economy from EPA
• national avg gasoline prices / oil futures

---

Controlling for product quality clearly matters

---

Identifying variation

---

AW use hedonic regression

Typical logit share identity (Berry, 1994): \(\ln s_{jat} - \ln s_{ot} = -\eta (p_{jat} - \gamma g_{jat}) + \psi_{ja} + \tilde \xi_{jat}\)

AW actually estimate: \(p_{jat} = - \gamma g_{jat} + \tau_t + \psi_{ja} + \epsilon_{jat}\)

Why do they do this?

---

Berry equation rearranged is \(p_{jat} = - \gamma g_{jat} - \frac{1}{\eta}(\ln s_{jat} - \ln s_{ot} ) + \psi_{ja} + \tilde \xi_{jat}\)

• The don’t have shares.
• Absorb $s_{ot}$ with time FEs.
• Put product FEs in $\psi_{ja}$
• $\xi$ and the remaining share difference in the $\epsilon$

\(p_{jat} = - \gamma g_{jat} + \tau_t + \psi_{ja} + \epsilon_{jat}\)

• So this tests it tests whether relative vehicle prices move one-for-one with changes in the relative gasoline costs.
• Coeff of interest $\gamma$ recovered with OLS.

---

AW also using a grouping estimator

• group all cars by month, above / below median

  • $Z_{jat}^u = 1(f_{ja} > f^{50}) \times 1(t=u$)

• instrument for $G$ with $Z$

• why do they do this?

---

Jerry Hausman’s (2001) “iron law of econometrics”: due to measurement error, the magnitude of a parameter estimate is usually smaller in absolute value than expected

What is the null hypothesis that AW want to test?

\[p_{jat} = - \gamma g_{jat} + \tau_t + \psi_{ja} + \epsilon_{jat}\]

---

Visual results

---

AW Results - Sensitivity

---

What do people takeaway from this paper?

• tradeoff implies discount rate of about 15%
  • for new cars, close to 1
  • old cars, very low

• results sensitive to how G is constructed

• are people making mistakes?

• what questions does this leave open?

---

Allcott and Knittel

AEJ: Policy 2019

---

Recent emphasis has been on experiments

• Many studies similar to AW

• Although panel data helps, interpretation still requires econometrician to construct $g$

• If hypothesis is that $\gamma < 1$ due to inattention, imperfect information or bounded rationality, an alternative is to experimentally vary exactly those margins

---

Alternative: AK 2019 conduct two large field experiments

Experiment #1:

• Hang out at car dealers and intercept potential buyers
• Randomly explain fuel economy savings to some

Experiment #2:

• Conduct and online survey of people how say they are in the market for a new car
• In addition to following up on what people actually bought, AK elicit WTP for fuel economy.

---

---

---

Some advantages of surveys

• can ask people about consideration sets

• can ask about beliefs

• can elicit WTP

---

Consideration sets reveal important differences

---

Beliefs about fuel costs noisy, but not biased

---

Results

---

Policy Implications

---

Some conclusions from core EE gap lit

• assertion that consumers making large mistakes on average appears incorrect

• once you account for product unobservables and use exogenous fuel cost variation, tradeoffs seem close to rational

• experimental studies directly educating or directing consumer attention to energy costs have found very small impacts

---

Future directions

• Heterogeneity in both values and bias
  • policy implications
  • targeting
• What role do firms play here?
  • Are they offering the “right” set of products?
• How do consumer’s form beliefs?
  • we know the true calculation is hard
    • rational inattention (Sallee JLE 2014)

---

Allcott and Taubinsky

AER 2015

---

---

Can policymakers improve welfare in this market?

• develop a theoretical framework in presence of heterogenous bias and tastes

• implement two experiments informing consumers about CFLs

• evaluate welfare effects:

  • optimal subsidy
  • [ban on incandescents]

---

Corrective taxation with damage heterogeneity

Pigou (1920):
\(\tau = D'(e)\)

If damages are heterogenous, first best achieved with polluter specific tax

\[\tau_i = D_i'(e_i)\]

Diamond (BJE 1973): If damages are heterogeneous, but you can only set one tax rate, it should be equal to the average damages at the margin when the tax is implemented.

\[\tau_h = E_i[ D'_i(e_i(\tau_h))]\]

---

AT show this also applies to correcting biases

Setup

• unit demand: $j \in {E,I}$
  • no outside option
• utility: $u_j = v_j + z - P_j$
  • where $z$ is the consumer’s budget
• choose $E$ if $v - b > p$
  • $v = v_E - v_I$
  • $b$ is bias

---

Demand with and without bias

---

Welfare effect of subsidy

\[W(s) = Z(s) + v_I - p_I + \int_{v-b \ge p}(v - p) dF dG\]

• perfect competition: $p = c - s$

Proposition 1:

\[W'(s) = (s - B(p)) D'_B(p)\]

where $B = E(b

v - b = p)$ is the average marginal bias – ie the average bias for consumers on the margin the (subsidized) price

Optimal subsidy: \(s^* = B(c - s^*)\)

---

Average Marginal Bias

---

AMB is the key object of policy interest

Need not be constant

• Average marginal bias $\ne$ average bias

How can we estimate the average marginal bias?

Option 1:

• use within consumer variation in informedness

Option 2:

• estimate demand elasticity wrt tax and information
• calibrate difference

---

TESS experiment

• Artefactual field experiment

• Give consumers $10

• Elicit WTP ($v$) for CFLs

• Inform some consumers about cost to own to recover $b$

---

Many papers have used TESS

---

---

---

---

Informed vs uninformed demand

---

Average marginal bias not constant

---

Welfare effects

---

Welfare effects

---

Allcott, Knittel and Taubinsky

(AER P\&P 2015)

---

Extend AT model to allow correlation in bias and valuation

• consumers of type $j$ have distortions $d$ (nests bias from AT)
  • value $v - d$

• population average $\bar d = \sum \alpha_j d_j$

• define targeting as: \(\tau (s) \equiv cov ( d_j, -Q'_j(c-s) )\)
  • so a high $\tau(s)$ is well “targetted”

---

Welfare and optimal subsidy

Result 1: Poor targeting reduces welfare gains

\[W'(s) = (s - \bar d) \cdot D'(c-s) + \tau(s)\]

Result 2: Optimal poorly targeted subsidy could be small, even if average bias is large

\[s^* = \bar d - \frac{\tau(s)}{D'(c - s)}\]

[ Optimal subsidy increasing in $\tau(s)$ ]

---

Bias is correlated with observables

---

As is policy takeup

---

Where does this leave us?

• Optimal subsidy depends on average marginal bias

• Not sufficient to know bias and responsiveness separately: need to show biased consumers are actually the onces affected by the policy.

---

Allcott and Sweeney

(MS 2017)

---

What we did

• partner with major appliance retailer
• field experiment at call center
  • customer information and rebates
  • sales agent incentives
• audit phone calls to check compliance
• survey consumers

---

---

Some thoughts on running a field experiment

• getting institutional buy in was tough
  • in end, partner lost interest
• compliance a big issue
  • has implications for feasible policy too

---

---

Some agents much better than others

---

Survey reveals widespread confusion, but small bias

---

Main Results

---

Agents Target Scripts

---

What role do firms play in this?

• Allcott & Sweeney: Sales agents can target. Can we leverage this somehow?

• Houde (2014): firms bunch product characteristics around subsidy / label cutoffs

• Houde (2018): labeling may further reduce average purchased quality

---

Endogenous Attention?

---

Endogenous Attention?