Elsevier

Energy Economics

Volume 104, December 2021, 105656
Energy Economics

Split incentives and endogenous inattention in home retrofits uptake: a story of selection on unobservables?

https://doi.org/10.1016/j.eneco.2021.105656Get rights and content

Highlights

  • Rented properties and multi-unit dwellings are less likely to be insulated.

  • The gaps in insulation adoptions risk to be over-estimated due to selection.

  • Advanced sensitivity analysis is necessary for cross-sectional models.

  • Equal selection on observables and unobservables reduces sensibly split incentives gap.

  • Dwelling allocations depend less on observables, but selection risk is still high.

Abstract

Imperfect information in rental sector and housing-induced returns heterogeneity among occupiers are typically estimated with cross-sectional single-equation models. This approach leads to the estimation of conspicuous wedges in insulation investment propensity between tenants vs owner–occupiers (20 percentage points) and low-return vs high-return dwellings households (0.10–0.20pp) in the UK. I analyse their sensitivity to assumptions on unobservables à la Oster (2019) and Cinelli and Hazlett (2020). According to the former’s parametrization, under equally strong observables and unobservables, the effect of split incentives on loft/wall insulation investment can be up to 40%/26% lower, while the effect of housing choices is unaltered. The latter’s strategy suggests that an equal selection scenario would reduce by at least 60% the split incentives estimates, whereas non-random housing would just cause the estimates to drop by less than one third. Hence, I quantify a certain degree of selection on research conclusions and offer some convenient tools to integrate in the assessment of the sources of under-retrofitting with cross-sectional data.

Introduction

The concept of Energy Efficiency Gap is hardly new, and yet central in the energy demand management agenda for many developed economies. In the UK, special attention has been devolved over the years to residential energy efficiency (henceforth, EE), given that the amounts of residential emissions have never been trivial — in 2018 only, residential sector emissions were estimated to be 18% of the total CO2 emissions (BEIS, 2021). Moreover, the housing stock in the UK is reportedly one of the oldest and worst insulated in Europe (Committee on Climate Change, 2019). The UK was also the first country in Europe to introduce obligations on suppliers to save energy at the customer end in 1994 (Rosenow, 2012) with zero out-of-pocket home retrofits, but the results have not been up to expectations (Rosenow and Eyre, 2016).

What is really driving the wedge between the cost-minimizing level of EE and the level actually realized has then become a popular subject of study. The determinants of household propensity to retrofit have been vastly researched, with an eye toward possible sources of underinvestment. Broadly speaking, investment inefficiencies into EE fall within three categories (Gerarden et al., 2017): (i) market failures, (ii) behavioural explanations, and (iii) modelling flaws. While the third set of reasons leads to be cautious with engineering predictions potentially overstating the profitability of EE measures, here I focus on two possible causes of under-retrofitting that typically reflect the existence of market failures and behavioural biases. As a first contribution, I explore the extent to which some of these inefficiencies weigh on the EE Gap in the British residential sector during the period 2012–2018. Specifically, with data from the UK Public Attitude Tracker (PAT) survey covering the aforementioned years, I gauge the magnitude of split incentives in the rental market (a market failure) and endogenous inattention (a behavioural explanation) among owner–occupiers on insulation choices.

Imperfect information in the rental market is a prominent example of investment inefficiency that might cause underinvestment in EE (Allcott and Greenstone, 2012): because imperfectly informed renters will not be willing to pay more for energy efficient dwellings, landlords have reduced incentive to invest in EE. Owner–occupiers, instead, do capture the benefits of improved EE — either through energy savings or increased house value. Such a “landlord–tenant” agency problem (also known as “split incentives” issue) implies that rental properties are less energy efficient than would be socially optimal because of the non-appropriability of energy savings.

Residential retrofits are characterized by the heterogeneity of their effect on energy use (Metcalf and Hassett, 1999). The estimated returns to loft and cavity wall insulation, for instance, differ greatly across the distribution of dwelling types — available from the Energy Saving Trust2 in the UK. The projected annual savings in terms of bill expenditure and carbon dioxide emissions, as well as the costs are provided in Table 1. The reported savings and investment costs, assuming constant energy prices and annual heating consumption over a 25-year lifetime of the investments, translate into IRRs (without carbon savings) as high as 90%, 100% and 172% for loft insulation installation in terraced, semi-detached and single unit houses. Cavity wall insulation yields slightly lower IRRs: 33%, 37%, 53%, 85% for flats, terraced, semi-detached and detached houses respectively. Attention might be endogenous to the stakes of the decision: the higher the potential gross utility gain from EE, the more consumers will pay attention to energy cost savings — hence, consumers may present “endogenous” inattention (Allcott et al., 2014). Varying attention levels can lead to a differential in uptake rates among households, depending on their house potential savings.

In many papers on retrofit decision outcomes, researchers have tested for market equilibria consistent with imperfect information in rental sector and endogenous inattention (although I am the first one to formally call it so in the home retrofit decision-making case) among occupiers by estimating single-equation models exploiting cross-sectional variation.3 Their findings show that renters are substantially less likely to live in insulated dwellings compared to owner–occupiers, and households in low-return dwellings (i.e., flats or terraced houses, or also smaller properties) are less prone to invest in insulation than households in high-return dwellings (such as single-unit buildings). At a first glance to Fig. 1, these patterns are confirmed in the PAT data; it is noticeable how no matter the type of landlord, tenants tend to live in under-insulated houses, while there is virtually no difference in uptake rates between outright owners and mortgagors. The latter have an incidence of investment of more than 80%, while local authority (LA) and private renters uptake rate is lower by more than 25% (over 20 percentage points less). Split incentives thus appear strong in British residential sector. There is also some evidence of attention to savings differentials, given that flat owners are conspicuously less likely to invest in either insulation measures (only 1 in 5 flats has loft/wall insulation); contrarily, uptake rates once moving from terraced to detached dwelling households become as high as 80%.

However, the empirical results in the literature may well be biased — renters’ preferences may be divergent enough from homeowners to explain the difference in EE investments without any market failure (Myers, 2020). Similarly, unobserved household preferences could simultaneously affect both dwelling type allocation and EE outcomes (Bhat, 2015): an environmentally conscious household will locate itself in a multi-family household setting with fewer bathrooms and bedrooms, in a one storey home for energy efficiency. Consequently, the observed differentials may not be causal evidence of endogenous attention but rather reflecting sorting into less or more energy intensive housing according to own energy use habits (i.e., households’ attention is not really driven by heterogeneous returns, rather by unobservable traits).

Hence, I proceed by questioning the robustness of the results when using non-randomized (survey) data, highlighting how the estimated relationships rely on unreasonably strong assumptions on unobservables. In this regard, for my second contribution I propose useful tools to incorporate when assessing the magnitude of the estimates — a practice that is still very rare. The challenge of applied researchers in the identification of actual determinants of EE choices is isolating causal effects, i.e. finding exogenous variation that affects the variable of interest (tenure or dwelling type) but not outcomes. Hence, I recur to estimation strategies that may be helpful when strong prior information is unavailable regarding the exogeneity of either the variable of interest or instruments for that variable.

After a traditional analysis of the split incentive issue and the dwelling-dependent differentials, I bound the estimated coefficients following the proportional selection on observables bounding argument advanced by Oster (2019), which sets a degree of selection into treatment (e.g., assignment to rental properties or lower-return properties) on unobservables relative to the observables and obtains a selection-corrected bound. The reliability of this bound, however, depends on the explanatory power assigned to a full model (with observables and unobservables) in terms of R2: I also report how the degree of proportional selection capable of nullifying the estimates (obtaining a bound 100% lower that the findings) varies with the assumed R2. As a competing strategy, I also report the robustness values formulated by Cinelli and Hazlett (henceforth CH (2020)): these describe the minimum strength of association that unobservables would need to have, both with the independent and the dependent variables, to change the research conclusions. An analogous bounding exercise can be performed by observing coefficient stability to varying strength of unobservables compared to observables in explaining treatment assignment.

By considering tenure choice as completely exogenous, the estimated gap in both loft and wall insulation incidence created by split incentives amounts to more than 20pp in British residential sector between 2012 and 2018. A conspicuous amount of this wedge may be imputed to selection bias: the estimates can be severely reduced when accounting for unobservables. An equal degree of selection on observables and unobservables into rental property generates an Oster bound to the estimates which is lower by 40% and 26% for loft and wall insulation respectively. CH’s parametrization obtains consistent results in terms of selection bias: under equal selection into tenancy on observed (sociodemographics) and unobserved factors, the estimated split incentives gap tends to shrink by at least 60%.

The evidence concerning investment propensity sensitivity towards returns differentials is less fragile: the share of owner–occupiers in low-return dwellings with loft (wall) insulation is lower by around 12pp (21pp), and the estimates are relatively less sensitive, compared to split incentives. Unobservables as strong as observable characteristics bound the estimated effects in the loft and wall at the same magnitude as the point estimates. The higher robustness to confounders is confirmed by CH’s criteria to some extent, since unobservables would need to be three times as strong as socioeconomic conditions in dwelling type prediction to make the estimates reduce sensibly (by 100%/60% for loft/wall insulation). Nevertheless, this may be due to the poor predictive power of observables on housing allocation — the level of association with outcomes and variable of interest (a.k.a., the robustness value) required for unobservables to explain away the gap due to dwelling differences is indeed lower than the one to nullify the gap due to split incentives. Furthermore, I apply CH’s toolkit on literature findings from a group of papers who pursue a traditional analysis with single-equation adoption propensity models, and observe that unobservables with even very low explanatory power on tenure type and dwelling type choices can explain away the uptake differentials identified.

Overall, the traditional analysis on possible sources of the EE Gap in home retrofitting is revealed as incomplete when using cross-sectional variation to infer causal relationship (or assess “determinants”). The robustness of the estimates has been weighed in two different ways, with comparable results — hence, the (in)stability of the estimates does not really depend on the tool choice: new microeconometric techniques to perform sensitivity analysis could be integrated to better corroborate research findings.

The paper is thus developed as outlined: I review prior results and estimation frameworks in Section 2; the datasets and the estimation subsamples are introduced in Section 3. The preliminary strategy for estimation of split incentives and endogenous inattention is pursued in Section 4, which also includes the results. I report the sensitivity analysis tools by Oster (2019) and CH (2020) in Section 5, along with selection bias quantification and the application of the latter author’s methodology to the literature. Concluding remarks are found in Section 6.

Section snippets

Prior findings

Theoretical predictions on the existence of barriers to investment in residential retrofits have found empirical support since the 80’s and researchers have delved into the assessment of different determinants for decades. One of the most adduced reasons for investment decision is the type of occupier: a principal–agent issue arises, for example, in rented properties, where the landlord may be in charge for the installation of any structural improvements that would make the property more energy

Data

The data come from the PAT, a Department of Business, Energy and Industry Strategy (BEIS) survey, which collects cross-sectional data on public opinions and behaviour in various policy areas and has been running quarterly since 2012. It was gathered through face-to-face in-home interviews with around 2000 different households each wave. The survey waves of interest are yearly ones from 2012 to 2018, as questions on insulation investments are only repeated on annual frequency, and the data will

Preliminary analysis: Traditional estimation

The main framework for analysis of these sources of underinvestment is a single-equation model of the type: Y=τD+Xγ+ε,where Y is a latent variable indicating home retrofit valuation, D may be the variable in question (in my case, tenure condition or dwelling type) and X is the usual set of controls. I will estimate a probit model to produce the first set of estimates of the effect of tenure mode and housing allocation on investment probability.

In Table 2, Table 3, a typical sensitivity

Bounded estimates of split incentives and endogenous inattention

A main takeaway of this article is the role of the selection bias in the context of single-equation models to estimate sources of under-retrofitting. Assume that the full model that an investigator would like to estimate is the following: Y=τfullD+XΓ+θZ+εfull,where for simplicity Z is a single unobserved covariate. When the available information only allows the investigator to estimate a model as in (1), then ττfull because of the bias introduced by unobservable selection. Hence, the actual

Conclusions

This paper gauges the magnitude of split incentives and possible endogenous inattention in the residential sector when evaluating home retrofits. Split incentives derives from tenants’ lack of information, and subsequent landlords’ under-investment in energy efficiency. Endogenous inattention manifests as consumers’ increasing attention towards higher potential savings — corresponding to households increasingly likely to forego returns when they are too small as in the case of flat and terraced

CRediT authorship contribution statement

Stefano Cellini: Conceptualization, Design, Data curation, Formal analysis, Interpretation of results, Writing – original draft.

References (29)

Cited by (3)

I am grateful to Mona Chitnis, Martin Foureaux-Koppensteiner, Matthias Parey, Giuseppe Moscelli, Sahil Ravgotra and Francisco Nobre for helpful feedback and support.

1

Address: Elizabeth Fry Building (AD), University of Surrey, Guildford GU2 7XH.

View full text