Who cuts emissions, who turns up the heat? causal machine learning estimates of energy efficiency interventions

TL;DR

The study tackles the uneven impact of domestic energy efficiency upgrades by estimating causal effects of external wall insulation on gas consumption using a causal graphical model applied to the English Housing Survey Fuel Poverty data. It derives both the population-wide average treatment effect (ATE ≈ $-2980$ kWh/year, ≈ $-19\%$) and covariate-specific effects (CATE) that vary strongly with energy burden, showing large savings for low-burden households (≈ $-3720$ kWh/year, ≈ $-26\%$) but modest gains for the most burdened (≈ $-500$ kWh/year, ≈ $-3\%$). The results reveal a behaviourally driven mechanism where high energy costs drive households to reallocate savings toward thermal comfort, reducing climate gains for vulnerable groups but delivering health and well-being co-benefits. The authors advocate integrating equity and welfare indicators into carbon accounting and energy policy design, supported by a national-scale causal inference framework demonstrated on representative data. These insights have practical implications for tailoring energy efficiency programs to energy-burden profiles and for broader low-carbon transitions that balance emissions reductions with social equity.

Abstract

Reducing domestic energy demand is central to climate mitigation and fuel poverty strategies, yet the impact of energy efficiency interventions is highly heterogeneous. Using a causal machine learning model trained on nationally representative data of the English housing stock, we estimate average and conditional treatment effects of wall insulation on gas consumption, focusing on distributional effects across energy burden subgroups. While interventions reduce gas demand on average (by as much as 19 percent), low energy burden groups achieve substantial savings, whereas those experiencing high energy burdens see little to no reduction. This pattern reflects a behaviourally-driven mechanism: households constrained by high costs-to-income ratios (e.g. more than 0.1) reallocate savings toward improved thermal comfort rather than lowering consumption. Far from wasteful, such responses represent rational adjustments in contexts of prior deprivation, with potential co-benefits for health and well-being. These findings call for a broader evaluation framework that accounts for both climate impacts and the equity implications of domestic energy policy.

Figures (5)

• Figure 1: Causal graphs specification representing the structure of assumed causal relationships. a: full directed acyclic graph (DAG) $\mathcal{G}$, incorporating the intervention variable, $X$, the outcome, $Y_0$, and additional covariates representing potential confounders and mediators such as energy burden, $W$. b: the core causal mechanism is first modelled as a cyclic temporal graph, thus capturing the dynamic feedback loop initiated by the intervention $X$ and mediated through gas consumption $Y_0$ at time $t$, then gas cost ($V_1$), and household energy burden ($W$), eventually influencing future gas consumption at time $t+\Delta t$. c: modified acyclic version of the core causal graph, derived by unfolding the feedback loop into two temporally undifferentiated causal pathways, i.e. a direct physical effect ($X \rightarrow Y_0$) and an indirect behavioural effect ($X \rightarrow V_1 \rightarrow W \rightarrow Y_0$), as formalised in Eq. (\ref{['two_chains']}). This transformation from cyclic to acyclic ensures compatibility with the requirement of DAG-based causal inference modelling.
• Figure 2: Estimated causal effect of walls insulation intervention, $X$, on annual gas consumption for space heating, $Y_0$. a: post-intervention distributions of gas consumption under the two interventional scenarios: $P(Y_0 \mid do(X=true))$ i.e. treatment group, and $P(Y_0 \mid do(X=false))$ i.e. control group. The downward shift in expectation $\mathbb{E}(Y_0 \mid do(X))$ under treatment $X=true$ indicates a clear average reduction in energy use ($\approx-2980$ kWh/year). b: probability ratio, $PR(Y_0)$, plotted across the range of annual gas consumption levels. $PR(Y_0) > 1$ suggests an increased likelihood of remaining in a low-consumption regime post-insulation, while $PR(Y_0) < 1$ indicates a reduced likelihood of persisting in high-consumption states. A clear pattern emerges, with higher-consuming households ($>13,000$ kWh/yr) consistently more likely to reduce their energy use compared to lower-consuming households.
• Figure 3: Conditional average treatment effects (CATE) of external wall insulation on annual gas consumption by energy burden subgroups. a: estimated post-intervention outcome distributions $P(Y_0 \mid do(X), W)$ by energy burden level $W$, defined as the share of income spent on energy. Confidence intervals (40–60%), medians, and conditional expectations $\mathbb{E}(Y_0 \mid do(X), W)$ are reported for both treated and untreated scenarios. b: plots of the Conditional Average Treatment Effect (CATE) as a function of $W$, revealing a declining trend in treatment efficacy with increasing energy burden: households spending 2% ($\pm 0.5\%$) of income on energy see reductions of $\approx 3720$ kWh/year ($\approx –26\%$), whereas the most burdened ($W>11\%$) only achieve $\approx$500 kWh/year ($\approx –3\%$). c: probability ratios $PR(Y_0)$ of population subgroups experiencing different levels of energy burden, showing that low-burden households with medium-to-high baseline gas consumption are consistently more likely to reduce demand post-intervention ($PR(Y_0) < 1$). This pattern disappears at higher burden levels (e.g. $W=0.11\pm 0.5\%$), where all households show limited post-treatment change, regardless of initial gas consumption.
• Figure 4: Sensitivity analysis of the estimated average treatment effect ($ATE_{\mathcal{G}}$) to a hypothetical unmeasured confounder $U$. The heat-map shows the adjusted treatment effect, $ATE_{\text{adj}} = ATE_{\mathcal{G}} - \pi \cdot \gamma$, where $\pi$ denotes the confounder–treatment association (increase in probability of dwelling insulation attributable to the confounder), and $\gamma$ denotes the confounder–outcome association (change in annual gas consumption attributable to the confounder, independent of insulation). Shading represents the magnitude of the adjusted ATE. The solid black contour represents the tipping line, i.e. the locus of ($\pi, \gamma$) combinations that would exactly explain away the estimated effect of insulation ($ATE_{\mathcal{G}} = -2980.1$ kWh/year).
• Figure S1: The directed acyclic graph $\mathcal{G}$ of the implemented semi-Markovian causal model is shown at the top-left. Subgraphs resulting from selectively removing edges from $\mathcal{G}$ are shown at the top-right ($\mathcal{G_{\underline{X}}}$) and bottom-left ($\mathcal{G_{\overline{X}}}$). Specifically: $\mathcal{G_{\underline{X}}}$ is obtained by removing from $\mathcal{G}$ all directed edges originating from node $X$, whereas $\mathcal{G_{\overline{X}}}$ is obtained by removing all edges in $\mathcal{G}$ pointing to node $X$. The figure on the bottom-right side shows the augmented graph $\mathcal{G_{S}}$ after removing all edges pointing to $X$.

Theorems & Definitions (1)

• proof : Proof of Eq. (\ref{['Eq: covariate_y']}).