Using narratives to infer preferences in understanding the energy efficiency gap

The study addresses the persistent energy efficiency gap in the residential building sector: despite seemingly cost-effective technologies and generous subsidies, households underinvest in energy efficiency. The key question is whether systematic barriers and household preferences impede adoption. To uncover decision-makers’ thought processes beyond preset survey options, the authors elicit open-ended narratives from single-family homeowners in Zurich, Switzerland, and quantify them using natural language processing. The purpose is to identify nuanced barriers/determinants of retrofits and to compare insights from narratives with traditional closed-ended questions, informing more effective, targeted policy. The study is important because it overcomes limitations of prior research that relied on predetermined variables and may have missed key behavioural and institutional factors driving or impeding retrofits.

The paper situates itself within the energy efficiency gap literature that distinguishes economic, behavioural, and organizational perspectives on barriers to adoption. It refines this into four operational categories: behavioural, financial, non-market (co-benefits and hassle/hidden costs), and market (heterogeneity in buildings, technologies, preferences). The authors note that most empirical work uses closed-ended surveys or revealed-preference designs with limited ability to capture decision processes, leading to researcher bias and poor comparability across studies. Two recent reviews summarize 26 empirical studies, identifying limited robust findings, such as positive associations of income, younger age, energy/financial literacy, and comfort with investments, and policy associations with higher take-up. The paper contributes to emerging economics research using narratives/open-ended questions and NLP to elicit first-order concerns and policy preferences, addressing shortcomings of unsupervised topic models (e.g., LDA) for short survey texts and proposing a scalable dictionary-based approach supported by validation against human coding and embedding-based coherence measures.

Design: The authors conducted an online survey (SurveyMonkey) of single-family homeowners in the Canton of Zurich, Switzerland, in Feb–Mar 2020. Personalized letters were mailed to a stratified random sample (n=16,700) targeting: homes built before 1990; 50% with recent renovation permits (2014–2019); stratification by tenant age and household size; and a subsample of Minergie-certified (energy-efficient) new buildings. Response and sample: 3,471 respondents started the survey (20.8% response rate) with 82% completion and ~30 minutes average duration. After excluding tenants/apartments and setting aside Minergie-certified homeowners for a separate study, the analysis sample includes 2,187 households with open-ended responses. Households were classified as: non-takers (no retrofits in past 5 years and no plans in next 5 years; n=461, 21%) and takers (retrofits in past 5 years or plans in next 5 years; n=1,726, 79%). Elicitation: For both non-takers and takers, the survey first asked closed-ended (multiple-choice) questions on barriers (17 options) or determinants (8 options) of retrofits, followed by an open-ended question asking respondents to describe reasons for not doing/doing energy-efficient retrofits (~4 sentences). A separate open-ended question elicited policy preferences for promoting energy-efficient construction/renovation. The barriers/determinants open-ended questions were mandatory for non-Minergie respondents; the policy question was optional and presented to all. Text processing and topic classification: The authors implemented a semi-automated dictionary approach supported by NLP. Steps: (1) Pre-processing: tokenize unigrams, lemmatize (spaCy), retain nouns/adjectives/verbs/adverbs (POS tagging), drop very short tokens, and map words to German fastText embeddings to enable cosine-similarity distances. (2) Clustering: k-means clustering of unique keywords within each POS using embeddings to group semantically similar words. (3) Topic extraction: manually align clustered keywords to predefined topics (from closed-ended lists) and add new topics emerging from narratives (e.g., old age). Build keyword dictionaries (including reintroduction of inflected forms from lemmas) and label each response if any topic keywords appear. Dictionaries cover ~15–20% of pre-processed words; non-lemmatized forms increase coverage by 7–21%. Validation: Two validations are reported: (a) Comparison to human coders on a subsample showed comparable agreement between coders and the dictionary-based method; coder-coder agreement was only marginally higher than coder-dictionary. (b) Embedding-based quality metric showed intra-topic coherence exceeds inter-topic similarity (quality >1 for all topics). Inconsistencies between closed- and open-ended responses for major topics were manually checked; initial classifications were largely confirmed. Quantitative analysis: Topic frequencies from narratives were compared to closed-ended selections for ranking barriers/determinants. Heterogeneity analyses used linear probability models (LPMs) with binary outcomes indicating whether a respondent mentioned a given barrier (non-takers) or determinant (takers) in the open-ended responses. Covariates include demographics (income, age, gender, children, education), building characteristics (age, floor size, log rental value), psychographics (energy literacy, studied economics, math proficiency, energy-saving score, donations to environmental organizations, happiness), and policy indices/preferences (awareness, usage, market/behavioural/non-market preferences). Standard errors and significance levels are reported; explained variance (R²) is low, in line with limited predictability from observables.

Sample characteristics: Among 2,187 homeowners (non-Minergie), 21% are non-takers (n=461) and 79% takers (n=1,726). Takers and non-takers are similar in building/demographic features; non-takers are slightly older (mean age 61.36 vs 58.13). Heating type and construction year distributions are similar across groups (Table 1). Barriers for non-takers (Table 2, open-ended vs closed-ended shares):

• Building already energy efficient (Market): 49.5% (open) vs 38.4% (closed) – the most prevalent barrier.
• Too expensive (Financial): 26.2% (open) vs 21.9% (closed).
• Old age of respondent (Market): 8.9% (open) vs 0.0% (closed) – emerged only in narratives.
• Too complicated/hassle (Non-market): 7.2% (open) vs 10.0% (closed).
• Aesthetics (Market): 3.0% (open) vs 8.0% (closed); Historic building difficulties (Market): 2.8% vs 6.1%; Lack of information (Behavioural): 0.0% vs 8.7%; Planning to move (Market): 0.0% vs 6.5%; Leaving the house during renovation (Non-market): 0.0% vs 5.9%; Difficult to find experts/materials (Non-market): 0.0% vs 4.6%; Financing difficulties (Financial): 0.7% vs 4.1%. Determinants for takers (Table 3, open vs closed):
• Replace broken elements (Market): 45.5% (open) vs 57.6% (closed) – top narrative determinant, highlighting opportunistic retrofits.
• Reduce ecological footprint (Non-market): 30.0% vs 69.0%.
• Save money (Financial): 28.8% vs 36.7%.
• Increase comfort (Non-market): 24.6% vs 68.5%.
• Increase resale value (Financial): 4.8% vs 25.2%.
• Regulatory (Non-market) appears in narratives (4.3%). Policy preferences from narratives (Table 4, shares non-takers vs takers): more subsidy (40.6% vs 44.4%), more information (20.8% vs 19.8%), less bureaucracy (16.3% vs 19.5%), focus on PV (12.6% vs 17.9%), focus on heating (11.3% vs 13.8%), standards (10.0% vs 10.7%), tax deduction (9.1% vs 8.9%), pollution tax (8.5% vs 8.6%), focus on insulation (2.2% vs 4.6%). Subsidies are popular, but over half mention other measures, notably information and bureaucracy reduction. Heterogeneity among non-takers (Table 5, LPM on major barriers):
• “Already efficient”: few significant predictors among observables; difficult to target based on observables.
• “Too expensive”: less likely female (−0.135*, P<0.1) and university degree (−0.118*, P<0.1); higher log rental value (+0.140**, P<0.05); lower policy usage (−0.086***, P<0.01).
• “Old age”: older age (+0.004**, P<0.05), female (+0.082*, P<0.1), donations (+0.060*, P<0.1), lower log rental value (−0.070*, P<0.1). R² values are low (0.056–0.095), indicating limited explanatory power from observables. Heterogeneity among takers (Table 6, LPM on determinants from narratives):
• Replacement driven by younger age (−0.005***, P<0.01) and older buildings (+0.002**, P<0.05); negatively associated with policy awareness (−0.037**, P<0.05); positively associated with market policy preferences (+0.083***, P<0.01) and non-market policy preferences (+0.100***, P<0.01).
• Save money: less likely with university degree (−0.062*, P<0.1) and donations (−0.051*, P<0.1); positively associated with policy usage (+0.022*, P<0.1) and market policy preferences (+0.075***, P<0.01).
• Comfort: positively associated with building age (+0.002***, P<0.01) and policy usage (+0.025**, P<0.05).
• Environmental: positively associated with income (+0.139***, P<0.01) and donations (+0.082***, P<0.01); positively with policy usage (+0.020*, P<0.1) and behavioural policy preferences (+0.111***, P<0.01); weak negative association with non-market policy preferences (−0.057*, P<0.1). R²≈0.024–0.047. Synthesis: Energy-efficiency retrofits are highly opportunistic; many takers act when equipment fails rather than through planned efficiency upgrades. Financial barriers are salient for non-takers, yet financial motives are not the primary drivers among takers. Non-market co-benefits (comfort, environmental concerns) are equally or more important than monetary savings. Awareness and use of existing policies are modest; respondents favour reducing bureaucracy and improving information. Subsidies are popular but risk mistargeting and free riding as many retrofits would occur upon failure regardless of incentives.

The narrative-based approach directly reveals households’ perceived barriers/determinants and clarifies the energy efficiency gap. Non-takers often believe their homes have limited efficiency potential or face financial or age-related constraints. Among takers, retrofit decisions are primarily triggered by replacement needs, with non-market co-benefits (environment and comfort) playing key roles, while direct financial motivations are relatively less prominent in narratives. The discrepancies between closed- and open-ended responses highlight intention–behaviour gaps and potential cheap talk in closed-ended selections. Policy implications include limited potential for targeting non-takers via standard observables (except age) and the risk of free riding from broad subsidies. Targeting could focus on “replacers” to encourage earlier planning and to reduce institutional frictions (bureaucracy) and information barriers. Aligning interventions with behavioural and institutional factors—information provision, standards, and streamlined permitting/subsidy processes—may improve effectiveness and cost-efficiency relative to purely financial incentives.

The study demonstrates that open-ended narratives combined with NLP uncover nuanced barriers and determinants behind homeowners’ retrofit decisions and provide a scalable, replicable method. It finds that energy-efficiency investments are largely opportunistic, triggered by equipment failure, and that non-market co-benefits (comfort, environmental concerns) are as important as, or more important than, financial savings among takers. Many non-takers perceive limited efficiency opportunities or face financial constraints, and policy awareness/use remains modest. Broad subsidies risk mistargeting and free riding; effective policies should address institutional barriers (bureaucracy) and information accessibility, and consider targeted strategies for replacers to foster earlier, planned retrofits. Future research can leverage advances in AI and text analytics to compare narrative elicitation across contexts, study discrepancies between closed- and open-ended responses, and test policy designs that reduce institutional frictions and improve targeting.

Findings rely on cross-sectional survey data from single-family homeowners in one Swiss canton, which may limit generalizability to other regions or housing types. Narratives are short texts and topic assignment depends on a semi-manual dictionary approach; despite validation, some misclassification is possible. Differences between closed- and open-ended responses may reflect intention–behaviour gaps or survey mode effects rather than true preference differences. Observational heterogeneity analyses (linear probability models) identify associations, not causal effects, and R² values are low, indicating limited predictive power from observables. Policy awareness/use measures are self-reported. The study period (early 2020) and local policy landscape may affect external validity.