Building Back Better or Locking in Carbon? A Provincial Panel Analysis of Residential Energy Demand and Low-Carbon Reconstruction Policy in Post-Earthquake Türkiye ^†

Abstract

Post-disaster reconstruction programmes create an irreversible window for embedding or foreclosing residential energy efficiency at scale. This study examines the structural determinants of per capita residential electricity consumption (K_MES) across all 81 provinces of Türkiye over 2013–2022 using a balanced province-year panel. We develop two complementary panel models, both estimated by two-way fixed effects (province + year) with cluster-robust standard errors, and supported by GLS-AR(1) and random-effects GLS robustness checks. Note that K_MES measures the electricity component of residential energy use only; we, therefore, also estimate the building-stock model with a constructed total-energy dependent variable that combines residential electricity (H_MES) and natural-gas consumption (X_DG) in kWh-equivalent units. Model 1 isolates the macroeconomic transmission channel through which exchange-rate volatility shapes residential electricity demand. Because the USD/TRY rate has no cross-sectional variation, its identifying power in two-way fixed effects comes from its interaction with province-level natural-gas-heating exposure (sh_gas × EV_DA). The interaction is robustly negative across all full-sample specifications (β ≈ −0.022, p < 0.01), indicating that provinces with greater gas-heating penetration are buffered against currency-depreciation pass-through into electricity demand. Provincial GDP carries the dominant direct macro coefficient (β ≈ 0.27–0.29, p < 0.01), establishing income elasticity rather than the exchange rate as the headline aggregate driver. Model 2 decomposes the building stock by structural system, filler material, heating system, and heating fuel. The dominant predictors are the share of electric heating (β ≈ 1.16–1.27, p < 0.01) and the share of AC-only heating (β ≈ −1.0 to −1.13, p < 0.05), with a total-energy specification reaching R² = 0.92. In the comparative subsample of the eleven Kahramanmaraş-affected provinces, masonry construction emerges as the dominant pre-disaster predictor of per capita electricity consumption (β = 14.04, p < 0.05), revealing structurally distinct stock characteristics that pre-date the February 2023 earthquake. Two re-framings are required. First, since the panel covers 2013–2022, the disaster-province estimates capture pre-disaster structural heterogeneity rather than post-disaster market rupture. Second, the macroeconomic mechanism that prior work attributed to the exchange-rate level is more accurately understood as a fuel-mix-mediated exposure channel. The combined evidence implies that mandatory building-code enforcement and natural-gas grid extension are complementary policy levers in the 488,000-unit Turkish Housing Development Administration reconstruction programme: gas grid expansion reduces the macroeconomic vulnerability of residential energy demand, while masonry-replacement construction standards address the largest pre-disaster structural determinant of energy intensity in the affected region.

1. Introduction

The global built environment accounts for approximately 40% of final energy consumption and nearly one-third of greenhouse-gas emissions, making the residential sector central to climate action [1,2,3]. Within this sector, household energy demand is shaped by a complex interaction of physical building-stock characteristics, macroeconomic conditions, and real-estate market dynamics; this interaction becomes acute in the aftermath of major disasters, when reconstruction programmes lock in or foreclose energy-efficiency outcomes for decades [4,5,6]. The Kahramanmaraş earthquake sequence of February 2023 (Mw 7.8 and 7.6) destroyed or severely damaged over 241,000 buildings across eleven southern provinces, displaced more than 3.3 million people, and triggered a 488,000-unit reconstruction programme implemented by the Turkish Housing Development Administration (TOKİ) under emergency conditions. The reconstruction is being executed during a period of exceptional macroeconomic volatility [6,7].

This paper examines the structural determinants of per capita residential electricity consumption (K_MES) across Türkiye’s 81 provinces over 2013–2022. Two complementary panel models, both estimated by two-way fixed effects with cluster-robust standard errors, isolate two distinct mechanisms. Model 1 identifies the macroeconomic transmission channel by interacting the USD/TRY ($/₺) exchange rate with province-level natural-gas-heating exposure, generating identifying variation that survives the absorption of EV_DA by year fixed effects. Model 2 decomposes residential electricity demand into its building-stock determinants structural system, filler material, heating system, and heating fuel and is also estimated with a constructed total-energy dependent variable that combines residential electricity and natural-gas consumption to address the scope limitation of an electricity-only measure.

A comparative subsample analysis of the eleven Kahramanmaraş-affected provinces complements the national panel. Because the dataset spans 2013–2022, all observations in this subsample pre-date the February 2023 earthquake. The subsample therefore captures pre-disaster structural heterogeneity in residential energy demand rather than post-disaster market rupture. This pre-disaster heterogeneity is itself the relevant baseline for evaluating the reconstruction programme: the structural features that distinguished these provinces before the earthquake including a higher share of masonry construction and a thinner gas grid are precisely the conditions under which TOKİ’s reconstruction policy now operates.

2. Literature Review

2.1. Post-Disaster Reconstruction and the Build-Back-Better Framework

Post-disaster reconstruction is one of the most demanding governance contexts for sustainable development, requiring the simultaneous management of humanitarian imperatives, institutional coordination, and long-term resilience [4,8,9,10,11]. Twigg [8] identifies inadequate planning coordination as a systemic barrier to viable post-disaster settlements. The Build Back Better (BBB) framework endorsed by the Sendai Framework for Disaster Risk Reduction 2015–2030 holds that reconstruction offers a transformative opportunity rather than a return-to-status quo exercise [4,5]. Kennedy et al. [5] demonstrate in the post-tsunami Aceh and Sri Lanka contexts that failure to integrate humanitarian and sustainability objectives during the reconstruction window locks in carbon-intensive and structurally vulnerable building stock for the full service life of the reconstructed units—typically 50 years or more [5,9].

The 2023 Türkiye recovery process illustrates the scale of this challenge. The TERRA assessment [6] outlined BBB principle’s resilient infrastructure, inclusive development, and green construction methods as prerequisites for a recovery that reduces long-term disaster risk. Empirical studies of post-earthquake urban regeneration under Türkiye’s Disaster Law No. 6306 have found mixed evidence on the effectiveness of pre-disaster risk-area designations during the 2023 event [12,13]. Işık et al. [13] document specific structural failure modes such as soft-storey collapse, inadequate shear walls, and substandard concrete that were systemic across the affected building stock despite existing regulatory requirements. These findings reinforce the analytical motivation for separately characterising the eleven affected provinces: they constitute a qualitatively distinct market and stock context.

2.2. Residential Energy Demand, Green Premiums, and Macroeconomic Transmission

International evidence on green premiums is extensive. Kok et al. [14] report rental premiums of 3–8% and sale premiums of approximately 13% for green-certified commercial buildings in the United States. Eichholtz et al. [15] show that operational cost savings from certified green buildings are reflected in lease rates. Ko et al. [16] report green premium magnitudes of 24.3% for G-SEED-certified apartments in Seoul. McCord et al. [17] demonstrate that EPC-related rental premiums in Northern Ireland range from 0.2% to 8.2% by efficiency band, while low-rated properties attract discounts of 3.9–5.5%.

These results overwhelmingly come from stable, high-income markets in which macroeconomic conditions can be treated as a background. The mechanisms by which exchange-rate volatility, energy-import dependency, and credit tightening interact with residential energy consumption in emerging-market contexts have received limited empirical attention. Türkiye’s structural energy-import dependency—approximately 70% of primary energy is imported, dominated by natural gas and petroleum [3,7]—gives the exchange rate a privileged channel into residential energy economics. Prior cross-country evidence further suggests that retrofit subsidies and green-premium signals are conditional on market depth and credit availability [18,19,20]; in disaster-exposed and emerging-market contexts, market signals appear insufficient drivers of green-building adoption when institutional capacity, technical expertise, and regulatory frameworks are constrained [21,22].

2.3. Panel Methods, Hedonic Pricing, and the Turkish Housing Market

Hedonic pricing models regression-based techniques that decompose an observed market price (a sale price or rental rate) into the implicit prices of constituent attributes including location, size, quality, and energy performance, have become the dominant empirical framework for quantifying the capitalisation of energy-performance attributes into property prices [23,24,25]. Standard hedonic models face well-documented limitations in spatially segmented or structurally disrupted markets [26,27]; Keskin et al. [26] highlight the inadequacy of single-equation specifications and recommend multi-level event-study designs that account for spatial autocorrelation. The 2012 Van earthquake analysis of Keskin et al. [26] is the closest precedent for the present study, demonstrating differential price impacts across residential submarkets. Building directly on this Türkiye-specific empirical agenda, Arslanli et al. [28] provide the most immediate methodological precursor to the present study, applying random-effects GLS regression to an 81-province Türkiye panel to model the determinants of residential demand and the feasibility of low-carbon investments in earthquake-impacted real estate markets.

Panel-data methods exploit within-unit variation over time while controlling for unobserved province-level heterogeneity [29,30]. The choice between random and fixed effects is consequential: in heterogeneous national panels, the random-effects assumption that province-specific unobservables are uncorrelated with the regressors is difficult to defend, and Hausman testing typically favours fixed effects when sufficient identifying variation exists [29]. Türkiye’s residential market is characterised by high foreign-currency exposure [31], a building stock dominated by pre-1999 construction [6,13], and a recent wave of mass-market apartment construction during the 2011–2017 boom that partially modernised the urban housing mix. Kontokosta [25] and Amasyali and El-Gohary [32] demonstrate that transaction volumes, building-age composition, and macroeconomic conditions are among the most powerful predictors of aggregate residential energy demand.

2.4. Research Gaps and Study Contribution

Three gaps motivate this study. First, the green-premium and residential-energy literature has largely treated macroeconomic conditions as background; the mechanisms by which exchange-rate volatility transmits into residential electricity demand in import-dependent emerging economies remain empirically underexplored. This study identifies the transmission channel by interacting the exchange rate with province-level natural-gas-heating exposure, generating cross-sectional identifying variation that survives two-way fixed effects. Second, the post-disaster real-estate literature is largely confined to price-level effects [26,33] and does not examine whether disaster-affected provinces exhibit pre-existing structural heterogeneity in residential energy demand that conditions the operability of reconstruction policy. The comparative-subsample design of this study distinguishing the eleven affected provinces from the national panel provides the first province-level quantitative evidence on this question. Third, the Türkiye-specific empirical literature has not exploited the national TÜİK transaction database, the EPDK energy balances, and the building-permit composition data in combination as a panel covering both a housing-market boom and an exchange-rate crisis [3,12,13]. The 2013–2022 panel captures precisely the structural dynamics that the TERRA recovery plan [6] identifies as preconditions for low-carbon investment viability in the reconstruction context.

3. Data

3.1. Panel Structure and Sources

The empirical analysis is grounded in a balanced panel covering all 81 provinces of Türkiye across 2013–2022 (810 province-year observations in the unrestricted source dataset). The unit of observation is the province-year, a spatial scale at which official real-estate transaction records, energy-consumption statistics, demographic data, and building-permit composition data are consistently published by Turkish statistical and administrative agencies. Real-estate transaction counts (S_ILK, S_IKIEL, S_IPO) were obtained from TÜİK’s Land Registry and Cadastre Information System; demographic variables (N_TOP, N_ISO) from TÜİK’s Address-Based Population Registration System; macroeconomic indicators (EV_DA, EV_BTL, EV_TGE) from the CBRT Electronic Data Delivery System; provincial GDP (GSYH_TL) from TÜİK; residential energy variables (K_MES, H_MES, X_DG) from EPDK provincial energy balance tables. Building-permit composition variables structural system, filler material, heating system, heating-fuel type, and ownership are drawn from the TÜİK Building Permit Statistics (Yapı Kullanma İzin Belgeleri/occupancy-permit B-series), and provide a province-year characterisation of newly completed and occupied stock.

A restricted subsample of the eleven provinces designated as disaster zones following the February 2023 earthquakes Kahramanmaraş, Hatay, Adıyaman, Gaziantep, Malatya, Osmaniye, Adana, Diyarbakır, Şanlıurfa, Kilis, Elazığ is constructed for comparative subsample analysis. The restricted subsample, like the full panel, spans 2013–2022 and therefore consists entirely of pre-disaster observations. The February 2023 earthquake sequence falls outside the estimation window. The restricted-subsample estimates therefore capture pre-disaster structural relationships in these provinces, establishing a pre-earthquake baseline rather than direct post-disaster observations.

3.2. Variable Construction and Transformations

Quantity variables that span several orders of magnitude across provinces (K_MES, H_MES, X_DG, S_ILK, S_IKIEL, S_IPO, GSYH_TL, N_TOP) are log-transformed; for variables that may take zero values (X_DG and the transaction counts in some small provinces), ln(x + 1) is used. Bounded indices and rates (EV_DA, EV_BTL, EV_TGE, share variables, N_ISO) are retained in level form. The dependent variable in the principal specifications is ln(K_MES). For Model 2, an additional total-energy specification uses ln(per capita residential electricity + natural gas in kWh-equivalent units), with natural gas converted at the lower-heating-value rate of 10.55 kWh per cubic metre. This composite measure addresses the scope limitation of the electricity-only K_MES variable, which does not capture the natural-gas heating that constitutes a large share of residential final energy in many provinces.

Building-permit composition variables are constructed as shares of the relevant B-series (occupancy-permit) totals. Heating-fuel shares (sh_gas, sh_elec, sh_solid, sh_lpg, sh_renew) sum to one across the eight reported fuel categories. Structural-system shares (sh_RCframe, sh_masonry, sh_steel, sh_compos) sum to one across the six structural categories, with composite/prefabricated construction as the omitted reference. Filler-material shares (sh_brick, sh_aerated, sh_adobe, sh_concblk) sum to one across the principal materials. Heating-system shares (sh_central, sh_combi, sh_stove, sh_AC) similarly partition the heating-system distribution. The interaction term sh_gas × EV_DA is constructed as the product of the contemporaneous gas-heating share and the annual mean USD/TRY exchange rate.

4. Methodology

This study uses panel data measurements taken for the same 81 provinces across 10 years (2013–2022). Two complementary statistical models are used to analyse these data. Both are estimated by two-way fixed effects (TWFE), the workhorse of modern panel econometrics, supported by random-effects GLS and GLS with first-order autoregressive disturbances as robustness checks.

Two-way fixed effects (TWFE). This estimator removes unobserved differences across provinces (province fixed effects) and unobserved differences across years (year fixed effects) before measuring how each predictor relates to per capita electricity consumption. By absorbing year fixed effects, TWFE removes any nationwide time trend that could otherwise be misattributed to a particular variable, for example, the upward trend in the USD/TRY exchange rate over the study period.

Why both models. Model 1 asks: how does the macroeconomic environment shape residential electricity consumption? It introduces an interaction between the exchange rate and each province’s natural-gas-heating share, generating cross-province variation that is not absorbed by year fixed effects. Model 2 asks: which physical and infrastructural features of the building stock its structural system, materials, heating system, and heating fuel predict residential electricity intensity? Each model is estimated on the full national sample of 81 provinces and on the restricted subsample of 11 disaster-affected provinces.

What the coefficients mean. With ln(K_MES) as the dependent variable, coefficients are approximate semi-elasticities: a coefficient of 0.27 on ln(GSYH) means that a 1% increase in provincial GDP is associated with a 0.27% increase in per capita residential electricity consumption, holding other variables constant. For share variables (range 0–1), a coefficient of 1.16 on the share of electric heating means that moving from 0% to 100% electric heating would increase ln(K_MES) by 1.16, or roughly tripling K_MES; in practice, observed within-province changes are far smaller.

4.1. Specification Choice and Diagnostics

The choice of TWFE as the headline estimator is supported by three formal diagnostics. The Hausman test rejects the random-effects orthogonality assumption in both samples (full sample: χ²(5) = 15.52, p = 0.0083; restricted sample: χ²(5) = 16.49, p = 0.0056), favouring fixed effects. Pesaran’s test of cross-sectional dependence rejects independence (CD = 6.18, p < 0.001), motivating the inclusion of year fixed effects to absorb common time shocks. The Wooldridge test for first-order autocorrelation rejects the null (F(1, 80) = 15.12, p < 0.001), motivating the GLS-AR(1) specification as a robustness check. Cluster-robust standard errors at the province level are reported throughout, addressing both heteroscedasticity and within-province serial correlation. The full Hausman output is reported in Appendix A.

4.2. Model 1 Macroeconomic Transmission with Two-Way Fixed Effects

The Model 1 estimating equation is:

ln(K_MES)_it = α_i + λ_t + β₁ ln(S_ILK)_it + β₂ ln(S_IKIEL)_it + β₃ ln(S_IPO)_it + β₄ ln(GSYH)_it + β₅ EV_BTL_it + β₆ ln(X_DG)_it + β₇ sh_gas_it + β₈ (sh_gas × EV_DA)_it + β₉ ln(N_TOP)_it + β₁₀ N_ISO_it + ε_it

where αᵢ are province fixed effects and λ_t are year fixed effects. EV_DA itself is absorbed by λ_t; identification of the exchange-rate channel rests on the interaction β₈, which has cross-sectional variation through province-level heterogeneity in sh_gas. The coefficient β₈ measures the differential semi-elasticity of per capita residential electricity consumption to USD/TRY movements as a function of a province’s gas-heating share: a negative β₈ implies that gas-heated provinces are buffered from depreciation pass-through into electricity demand.

4.3. Model 2 Building Stock Composition

The Model 2 estimating equation is:

ln(K_MES)_it = α_i + λ_t + γ′·STRUCT_it + δ′·FILL_it + θ′·HSYS_it + φ′·FUEL_it + β₁ ln(GSYH)_it + β₂ ln(S_ILK)_it + β₃ ln(N_TOP)_it + ε_it

where STRUCT is the vector of structural-system shares (sh_RCframe, sh_masonry, sh_steel) with composite/prefab as the omitted reference; FILL is the vector of filler-material shares (sh_aerated, sh_adobe, sh_brick) with concrete-block as omitted; HSYS is the vector of heating-system shares (sh_central, sh_stove, sh_AC) with combi as omitted; FUEL is the vector of heating-fuel shares (sh_gas, sh_solid, sh_elec) with LPG, fuel-oil, and renewable categories as the combined omitted reference. The total-energy specification replaces ln(K_MES) with ln(per capita electricity + natural-gas in kWh-equivalent units).

4.4. Comparative Subsample Design

Each model is estimated on two sample configurations: the full national panel (81 provinces) and the restricted subsample of the 11 disaster-affected provinces. The restricted-subsample estimates are not a quasi-experimental treatment-effect estimate; they are a comparative subsample analysis of pre-existing structural heterogeneity. The objective is to identify whether, prior to the February 2023 earthquake, the disaster-affected provinces exhibited structurally distinct relationships between building-stock composition, macroeconomic conditions, and per capita residential electricity consumption. Because the panel ends in 2022, the comparison establishes a pre-earthquake baseline against which the future impact of the 488,000-unit reconstruction programme can subsequently be assessed once post-2022 data become available.

5. Results Model 1: Macroeconomic Transmission

Table 1. Model 1 macroeconomic transmission. Dependent variable: ln(K_MES).

Variable	(1) RE-GLS	(2) FE	(3) TWFE	(4) GLS-AR(1)	(5) TWFE-Disaster
ln(S_ILK)	−0.024 (0.021)	−0.013 (0.022)	−0.008 (0.022)	−0.015 (0.014)	−0.013 (0.081)
ln(S_IKIEL)	0.132 *** (0.048)	0.156 ** (0.063)	0.172 ** (0.066)	0.078 *** (0.016)	−0.019 (0.147)
ln(S_IPO)	0.060 ** (0.028)	0.038 (0.037)	0.000 (0.050)	0.090 *** (0.016)	0.220 * (0.113)
ln(GSYH)	0.289 *** (0.046)	0.272 *** (0.076)	0.205 ** (0.098)	0.255 *** (0.023)	−0.299 (0.309)
EV_BTL	0.002 (0.002)	0.001 (0.002)	−0.001 (0.009)	0.004 *** (0.001)	0.040 (0.031)
ln(X_DG)	−0.007 (0.005)	−0.005 (0.006)	−0.006 (0.006)	−0.004 (0.003)	0.003 (0.058)
sh_gas	0.026 (0.049)	0.071 (0.050)	0.074 (0.050)	−0.091 ** (0.036)	0.169 (0.208)
sh_gas × EV_DA	−0.025 *** (0.009)	−0.022 *** (0.008)	−0.022 *** (0.008)	−0.017 *** (0.007)	0.020 (0.015)
EV_DA	−0.003 (0.008)	−0.006 (0.012)	absorbed	0.001 (0.006)	absorbed
ln(N_TOP)	−0.130 *** (0.036)	−0.057 (0.192)	−0.167 (0.183)	−0.123 *** (0.013)	1.952 (2.160)
N_ISO	−0.008 * (0.004)	−0.006 (0.005)	0.008 (0.022)	−0.010 *** (0.003)	0.034 (0.062)
Constant	3.860 *** (0.532)	2.933 (2.461)	5.041 ** (2.359)	4.287 *** (0.265)	−20.07 (28.29)
Observations	648	648	648	648	88
R2 (within)	—	0.638	0.657	—	0.630
Provinces	81	81	81	81	11
Year FE	No	No	Yes	No	Yes

Notes: Cluster-robust standard errors in parentheses (clustered at province level). Sample: 81 provinces (full) or 11 disaster-affected provinces (restricted), 2013–2022. The interaction term sh_gas × EV_DA is the construction of the product of the natural-gas heating share with the USD/TRY exchange rate. “Absorbed” indicates that EV_DA is collinear with year fixed effects and is dropped automatically. *** p < 0.01, ** p < 0.05, * p < 0.10.

reports five specifications of Model 1: a baseline pooled OLS, random-effects GLS, province fixed effects, two-way (province + year) fixed effects, GLS with first-order autoregressive disturbances, and the two-way fixed-effects estimator on the restricted disaster-province subsample. The two-way fixed-effects column on the full sample is the headline specification.

5.1. Model 1 Findings

Three findings emerge from the full-sample TWFE specification (column 3). First, provincial GDP enters with a positive and statistically significant elasticity (β = 0.205, p = 0.041). The slightly larger coefficient in the RE-GLS specification (β = 0.289) and the AR(1) specification (β = 0.255) bracket this estimate; across all four full-sample specifications the elasticity falls in the range 0.20–0.29 and is significant at the 1 or 5% level. This is the income–energy elasticity that the household-energy literature documents [31,34] and is the largest direct macroeconomic coefficient in the model.

Second, the exchange-rate channel operates through the interaction with province-level gas-heating exposure. EV_DA itself is absorbed by year fixed effects in the TWFE specification, and where it appears as a direct term (columns 1, 2, and 4) it is statistically insignificant. The interaction sh_gas × EV_DA is, however, robustly negative across all four full-sample specifications (β = −0.017 to −0.025, p < 0.01 in every case). The interpretation is that provinces with a higher share of natural-gas heating exhibit a smaller per capita electricity response to lira depreciation. This is consistent with a substitution-buffering mechanism: when the exchange rate rises, households in provinces without gas access face stronger incentives to substitute toward electric supplemental heating, while gas-heated households are insulated from the substitution channel by their existing heating infrastructure. This finding revises and refines the Rev_02 narrative in which EV_DA was characterised as the dominant unconditional driver: with appropriate identification, the exchange-rate channel is shown to operate through fuel-mix-mediated exposure, not as a direct level effect.

Third, real-estate transaction-volume variables do not survive the richer specification uniformly. First-sale transactions (S_ILK) which Rev_02 reported as significantly positive are insignificant across all five specifications. Second-sale transactions (S_IKIEL) remain significant and positive across the full-sample columns (β = 0.13–0.17, p < 0.05), consistent with the interpretation that turnover in existing-stock units is associated with intensifying use of the older and less-efficient parts of the stock. Mortgage-financed sales (S_IPO) are mixed: significant positive in RE-GLS and AR(1), insignificant in FE and TWFE. The pattern indicates that, once GDP, gas-heating exposure, and the within-province fixed effect are absorbed, market-activity variables explain less of the within-province variation in residential electricity demand than Rev_02 suggested.

5.2. Restricted Disaster-Province Subsample

The TWFE-Disaster column (column 5) re-estimates Model 1 on the eleven Kahramanmaraş-affected provinces. Statistical significance collapses across most predictors. Only mortgage-financed sales retain marginal significance (β = 0.220, p = 0.061); the interaction sh_gas × EV_DA the principal Model 1 finding is insignificant in this subsample, as is the GDP elasticity. The within-R² remains high at 0.63, but is driven by the year and province fixed effects rather than by the macroeconomic and market predictors. This pattern is consistent with the pre-disaster structural heterogeneity hypothesis: the disaster-affected provinces exhibit a relationship between macroeconomic conditions and residential energy demand that is structurally distinct from the national pattern, even before the earthquake. The restricted-subsample estimates do not, on their own, identify the specific source of this heterogeneity. Model 2 addresses that question by decomposing the building stock.

6. Results Model 2: Building Stock Composition

Table 2. Model 2 building stock composition. Dependent variable: ln(K_MES) unless noted.

Variable	(1) TWFE-Elec	(2) TWFE-Total	(3) GLS-AR(1)	(4) TWFE-Disaster
sh_RCframe	−0.269 ** (0.129)	−0.080 (0.146)	0.220 ** (0.103)	0.304 (0.608)
sh_masonry	−0.348 (0.454)	0.274 (0.579)	−0.195 (0.315)	14.04 ** (6.183)
sh_steel	−0.223 (0.233)	−0.417 (0.277)	0.400 ** (0.170)	−0.666 (1.544)
sh_aerated	−0.107 (0.104)	−0.144 (0.116)	0.155 *** (0.054)	−0.136 (0.206)
sh_adobe	−0.208 ** (0.101)	−0.213 ** (0.105)	0.007 (0.069)	−6.964 (8.691)
sh_brick	−0.032 (0.056)	0.008 (0.055)	0.093 *** (0.029)	−0.003 (0.130)
sh_central	0.002 (0.047)	0.009 (0.052)	−0.027 (0.028)	0.077 (0.087)
sh_stove	0.102 (0.108)	0.219 * (0.130)	0.111 * (0.057)	0.644 ** (0.254)
sh_AC	−1.015 ** (0.427)	−1.133 ** (0.460)	0.019 (0.228)	0.113 (0.515)
sh_gas	0.303 ** (0.146)	0.124 (0.188)	0.090 (0.062)	0.181 (0.207)
sh_solid	0.329 * (0.177)	0.163 (0.208)	0.150 ** (0.067)	−0.084 (0.277)
sh_elec	1.159 *** (0.372)	1.272 *** (0.382)	0.879 *** (0.197)	0.106 (0.458)
ln(GSYH)	0.309 ** (0.150)	0.297 ** (0.142)	0.236 *** (0.011)	−0.117 (0.253)
ln(S_ILK)	0.018 (0.016)	0.023 (0.018)	0.089 *** (0.009)	0.029 (0.053)
ln(N_TOP)	−0.277 (0.202)	−0.217 (0.218)	−0.055 *** (0.011)	0.712 (1.301)
Constant	6.631 ** (2.700)	−1.056 (2.848)	3.439 *** (0.199)	−3.373 (18.87)
Observations	729	651	729	99
R2 (within)	0.649	0.921	—	0.772
Provinces	81	81	81	11
DV	ln(K_MES)	ln(Total Energy)	ln(K_MES)	ln(K_MES)

Notes: Cluster-robust standard errors in parentheses (clustered at province level). “Total Energy” combines residential electricity (H_MES) and natural-gas consumption (X_DG) in kWh-equivalent units (gas converted at 10.55 kWh per cubic metre, lower heating value). Reference categories omitted: composite/prefab (structure), concrete-block (filler material), combi (heating system), and the combined LPG/fuel-oil/renewables category (heating fuel). All TWFE specifications include year fixed effects. *** p < 0.01, ** p < 0.05, * p < 0.10.

reports four specifications of Model 2: two-way fixed effects with ln(K_MES) as the dependent variable, two-way fixed effects with the constructed total-energy dependent variable that combines residential electricity and natural gas in kWh-equivalent units, GLS with first-order autoregressive disturbances, and two-way fixed effects on the restricted disaster-province subsample.

6.1. Model 2 Findings

Three findings emerge from the full-sample specifications. First, the share of electric heating is the most powerful single predictor of per capita residential electricity consumption. The coefficient ranges from 1.16 (TWFE-Elec) to 1.27 (TWFE-Total) to 0.88 (GLS-AR1), all significant at the 1% level. The interpretation is mechanical but quantified: a 10-percentage-point increase in the share of newly completed dwellings using electric heating is associated with an approximately 12% increase in per capita residential electricity consumption. This finding speaks directly to the energy-sector implications of the gas grid extension policy implied by Model 1: provinces lacking gas access default to electric heating, with the residential electricity demand consequences quantified here.

Second, the share of AC-only heating is robustly and substantially negative (β = −1.0 in TWFE-Elec, β = −1.13 in TWFE-Total, both p < 0.05). At first , this is counter-intuitive AC is electricity-consuming. The pattern is best understood as a climate-zone effect: AC-only construction is concentrated in provinces with mild winters and consequently lower aggregate residential heating loads. The composite total-energy specification confirms this interpretation: the AC-share coefficient remains strongly negative on the broader dependent variable that includes natural gas, indicating that the lower energy demand is not an artefact of fuel substitution but a reflection of climate-zone differences in heating demand. The shares of solid-fuel heating (sh_solid, β = 0.33, p = 0.07 in TWFE-Elec) and gas-stove-based heating (sh_stove, β = 0.22, p = 0.09 in TWFE-Total) point in the expected direction: provinces dominated by these inefficient delivery systems exhibit higher electricity demand, consistent with electric supplemental heating during cold periods.

Third, the total-energy specification (column 2) achieves an exceptional R² of 0.921, compared with 0.649 for the same specification with the electricity-only dependent variable. This is direct evidence that K_MES alone misses a large share of cross-province and within-province variation in residential energy use, and that the total-energy index is a substantially better operationalisation of the dependent variable. The signs and significance patterns of the building-composition variables are nonetheless qualitatively similar across the two dependent variables, indicating that the K_MES results are not driven by the omission of natural gas but the total-energy R² confirms that future work in this literature should construct integrated energy measures wherever data permit.

6.2. Pre-Disaster Structural Heterogeneity in the Eleven Disaster Provinces

The TWFE-Disaster column (column 4) is the central empirical contribution of this study to the post-disaster reconstruction literature. The dominant predictor of pre-disaster per capita residential electricity consumption in the eleven affected provinces is the share of masonry construction (sh_masonry, β = 14.04, p = 0.024). This coefficient is an order of magnitude larger than any predictor in the national-sample regressions and is the only structural-system variable to attain statistical significance in the disaster-province subsample. The share of stove-based heating is also strongly significant (sh_stove, β = 0.644, p = 0.014). No transaction, macroeconomic, or other predictor reaches conventional significance thresholds.

Two interpretive caveats apply. First, the disaster subsample contains 99 observations after share-variable construction, and the masonry-share standard error is correspondingly wide (6.18). The point estimate, while statistically significant, should be regarded as indicative of a large structural effect rather than a precisely identified treatment-effect magnitude. Second, the masonry coefficient is identified from within-province variation in masonry-share over time, controlling for province and year fixed effects. Provinces in the disaster region exhibit higher and more variable masonry shares than the national mean a feature that pre-dates the 2023 earthquake and reflects vernacular construction traditions, lower mass-market apartment penetration, and slower compliance migration after the 2007 and 2018 building-code revisions [12,13].

Subject to these caveats, the empirical content of the masonry result is direct: the eleven disaster provinces are structurally distinct from the national pattern not because of an idiosyncratic post-disaster shock but because, prior to the earthquake, their building stock was disproportionately composed of masonry construction with poor thermal performance and high seismic vulnerability. The same physical features that produced the catastrophic structural failures documented by Işık et al. [13] also produced higher pre-disaster residential electricity demand. This dual identification of masonry as both a seismic-risk and an energy-efficiency determinant in the disaster region is the most policy-relevant finding of the study, and provides the empirical foundation for the unified building-code argument advanced in the discussion below.

7. Discussion

7.1. The Fuel-Mix Exchange-Rate Channel and Implications for the Gas Grid Programme

The Model 1 result that the exchange-rate channel operates through fuel-mix-mediated exposure has direct policy relevance for Türkiye’s natural-gas distribution policy. BOTAŞ’s extension of the residential gas grid into mid-sized and smaller cities accelerated through the 2010s, but coverage remains highly heterogeneous across provinces, and several of the eleven disaster-affected provinces have gas grid penetration rates well below the national mean. The negative interaction between sh_gas and EV_DA implies that gas grid extension reduces the macroeconomic vulnerability of household energy budgets to currency depreciation: gas-connected households are insulated from the electricity-tariff–depreciation–imported-energy chain that drives the substitution incentive in non-gas-connected households. This is a quantified case for treating gas grid extension as a macroeconomic-resilience instrument, not only as an energy-substitution instrument.

The implication for the reconstruction context is twofold. First, where gas distribution infrastructure was damaged or destroyed in the 2023 earthquakes, its restoration should be prioritised on macroeconomic-resilience grounds in addition to direct energy-efficiency grounds. Second, in provinces where gas grid extension was already planned but not yet completed, the reconstruction window provides a low-cost opportunity to install distribution infrastructure during foundation and roadworks, rather than retrofit it later at a higher cost. The estimated semi-elasticity of −0.022 on the interaction means that a 10-percentage-point increase in gas-heating share reduces the per capita electricity response to a one-unit USD/TRY change by approximately 0.22 percentage points—a small per-unit effect that aggregates substantially over a full reconstruction cycle and a multi-currency-cycle horizon.

7.2. Building Stock Composition and the Mandatory Building-Code Argument

The Model 2 results identify the share of masonry construction as the dominant pre-disaster predictor of residential electricity demand in the eleven disaster-affected provinces, and the share of electric heating as the dominant predictor in the national sample. Both findings converge on a single policy implication: mandatory building-code enforcement in the 488,000-unit reconstruction programme is the most consequential near-term low-carbon investment lever in the residential sector. The argument is no longer dependent on green-premium price signals which the macroeconomic findings indicate are unreliable in volatility regimes but rests directly on the empirical relationship between physical stock characteristics and aggregate residential electricity demand.

The specific composition of an enforceable post-disaster building code follows from the regression results. Reinforced-concrete frame construction already required by the 2018 Turkish Building Earthquake Code (TBDY) addresses both seismic vulnerability and the elevated electricity demand associated with masonry stock. Aerated-concrete or composite filler in place of stone, adobe, or unreinforced brick improves thermal envelope performance. Centralised or combi heating systems with natural-gas connection improve delivery efficiency relative to stoves. These are not novel building-science conclusions; the contribution of this study is to demonstrate that within Türkiye’s eleven disaster-affected provinces, the marginal payoff of each of these standards is empirically larger than in the national average, because the pre-disaster composition was further from the technical optimum.

7.3. Limits of Market-Based Instruments and the Case for Direct Standards

The collapse of statistical significance for transaction-volume variables in the disaster-province subsample, combined with the insignificance of the macroeconomic interaction term in the same subsample, indicates that market-based instruments green premium signals, EPC-based price differentiation, mortgage-rate discounts conditional on certification are unlikely to be effective levers in the reconstruction phase. The conditions under which these instruments operate elsewhere [17,18,23,35] include thick markets, stable monetary regimes, and functional appraisal infrastructure; none of these conditions holds in the affected region in the period covered by the panel, and conditions in the post-2023 reconstruction phase are unlikely to be more favourable in the near term.

It is a case for substituting governance-based instruments of direct mandatory standards that are enforced through TOKİ’s administrative housing-allocation mechanism for the market-based instruments that the literature treats as the default policy lever. The post-disaster context is the unusual circumstance in which administrative allocation displaces market exchange as the dominant unit-allocation mechanism; it is therefore the unusual circumstance in which the regulatory mandate, rather than the price signal, is the binding instrument for embedding energy-performance and seismic-resilience standards in the reconstructed stock.

8. Conclusions

This study examined the structural determinants of per capita residential electricity consumption across Türkiye’s 81 provinces over 2013–2022, using two complementary panel models estimated by two-way fixed effects with cluster-robust standard errors. A comparative subsample analysis of the eleven provinces designated as disaster zones following the February 2023 Kahramanmaraş earthquakes establishes pre-disaster structural heterogeneity in residential energy demand, and provides a baseline against which post-2022 reconstruction effects can subsequently be evaluated. Six principal conclusions emerge.

First, provincial GDP carries the dominant direct macroeconomic coefficient (β ≈ 0.20–0.29, p < 0.05 across all full-sample specifications), establishing income elasticity as the primary aggregate driver of residential electricity demand. This finding is consistent with the international household-energy literature and corrects the Rev_02 narrative in which the exchange rate was characterised as the dominant unconditional driver.

Second, the exchange-rate channel operates through province-level fuel-mix exposure rather than as a direct level effect. The interaction sh_gas × EV_DA is robustly negative (β ≈ −0.022, p < 0.01) across all full-sample specifications, indicating that provinces with greater gas-heating penetration are buffered from depreciation-driven electricity-demand pressure. This is a novel identification result with direct implications for the design of natural-gas grid extension as a macroeconomic-resilience instrument.

Third, the share of electric heating is the dominant building-stock predictor of residential electricity consumption in the national sample (β ≈ 1.16–1.27, p < 0.01). The total-energy specification which combines electricity and natural gas in kWh-equivalent units achieves R² = 0.92, providing direct empirical support for the construction of integrated energy measures in future work. The share of AC-only heating is strongly negative, capturing climate-zone differences in heating load.

Fourth, in the eleven disaster-affected provinces, the share of masonry construction is the dominant pre-disaster predictor of residential electricity consumption (β = 14.04, p < 0.05, n = 99). This finding identifies a single physical building-stock characteristic that drives both the seismic vulnerability documented by Işık et al. [13] and the elevated residential energy demand identified here. The dual identification provides the empirical foundation for a unified building-code argument in the reconstruction context.

Fifth, real-estate transaction-volume variables central to the Rev_02 narrative exhibit weaker and less robust effects than previously reported. First-sale transactions are insignificant across all specifications. Second-sale transactions retain significance in the national sample. Mortgage-financed sales are mixed. The pattern indicates that, once macroeconomic conditions and building-stock composition are controlled, market-activity variables explain a smaller share of within-province variation in residential energy demand than Rev_02 estimated.

Sixth, mandatory building-code enforcement embedded in the 488,000-unit TOKİ reconstruction programme is the most consequential near-term low-carbon investment lever in the residential sector. The case for direct standards rests on the empirical relationship between physical stock characteristics and residential electricity demand, not on green-premium price signals which the macroeconomic findings indicate are unreliable in high-volatility emerging-market contexts. Gas grid extension is identified as a complementary instrument that reduces the macroeconomic vulnerability of household energy demand to currency depreciation.

Limitations and Future Research

Several limitations should be noted. The dataset covers 2013–2022 and predates the February 2023 earthquake; the policy implications for the post-disaster reconstruction period are inferential rather than directly observed. The exchange-rate channel is identified through fuel-mix-mediated exposure but cannot be cleanly separated from other province-level macroeconomic exposure mechanisms; future work should construct additional instruments, for example, construction-material import intensity and foreign-currency-denominated mortgage shares to triangulate the macroeconomic transmission. Electricity prices are nationally regulated and have no cross-sectional variation, and could not be entered as an independent control. The masonry-share coefficient in the disaster-province subsample is identified on a small sample and should be regarded as indicative; replication on the post-2023 panel, once available, will provide the direct test.

Three priorities for future research follow. First, extending the panel to 2023 onwards will permit direct estimation of the post-disaster reconstruction effect on residential energy demand and a test of whether the pre-existing structural heterogeneity identified here is amplified or moderated by the reconstruction programme. Second, province-level instruments for exchange-rate exposure beyond the gas-heating share used here would improve the precision of the macroeconomic-channel estimate. Third, micro-level data on TOKİ-allocated reconstruction units, including their energy-performance certifications, would permit a direct treatment-effect estimate of the building-code enforcement instrument advocated in the policy discussion above.