Speaker
Description
Urban planning is increasingly expected to contribute to climate mitigation through decisions concerning urban form, housing, infrastructure and the spatial organisation of everyday life. This expectation has simultaneously increased reliance on quantitative models and data-intensive methods to identify where and how planning interventions can reduce energy demand. The relationship between urban development and energy consumption, however, is neither simple nor stable. Household energy consumption emerges through interactions between the built environment, dwelling characteristics, household composition, socioeconomic position and behaviour. Consequently, apparently straightforward planning questions—such as whether compactness, income, dwelling characteristics or particular socioeconomic conditions increase or decrease energy consumption—frequently produce divergent empirical answers.
This problem is already visible in the urban-energy literature. The relationship between urban form and energy use has long been recognised, particularly through debates concerning compact versus dispersed urban development (Newman & Kenworthy, 1989; Owens, 1986). Low-density and spatially dispersed development can increase energy demand through both mobility patterns and characteristics of the housing stock, whereas compact development may provide opportunities for reduced energy consumption (Shammin et al., 2010). Yet these relationships are not universally transferable. Urban morphology, building characteristics, socioeconomic conditions and household behaviour interact in ways that can modify aggregate relationships (Chen et al., 2020; Namazkhan et al., 2020; Tian et al., 2016).
Socioeconomic determinants introduce further ambiguity. Income, for instance, can increase household energy consumption through larger dwellings, higher appliance ownership and consumption-intensive lifestyles, while simultaneously improving households’ capacity to invest in energy efficiency (Cayla et al., 2011). Similar complexity exists for education, tenure and household characteristics. Consequently, the effects of socioeconomic variables cannot necessarily be interpreted as stable, linear relationships independent of their spatial and household contexts. The literature reviewed in the underlying research consequently indicates that spatial and socioeconomic determinants are frequently analysed separately and that contradictory findings are often attributed broadly to “context”, without adequately identifying under which conditions, at which scales and beyond which thresholds relationships change.
This paper argues that this is not only an empirical problem but also a planning uncertainty problem. When planning increasingly relies on models to translate complex urban processes into actionable evidence, analytical choices influence what becomes visible as a policy-relevant relationship. A conventional regression model may identify an average association; a non-linear specification may reveal a threshold; quantile regression may show that the association changes across the consumption distribution; and an explainable machine-learning model may identify another variable as more influential for prediction. These findings are not necessarily mutually contradictory. Rather, each model interrogates a different dimension of the same underlying urban process.
The paper conceptualises this condition as analytical uncertainty: uncertainty in the interpretation of an urban problem generated by choices concerning model specification, analytical method, scale and the population or spatial context being examined. It asks:
How does the interpretation of household energy determinants change across analytical methods, consumption regimes, spatial scales and urban contexts, and how should planners navigate this variation when translating model-mediated evidence into intervention?
From Average Effects to Heterogeneous Urban Energy Relationships
A central limitation of conventional modelling is its tendency to represent an “average” household or an average relationship between explanatory variables and energy consumption. Yet residential energy demand is intrinsically heterogeneous. Building characteristics, socioeconomic circumstances and behavioural factors operate simultaneously (Namazkhan et al., 2020), while the relative importance of these factors can vary considerably within urban environments (Tian et al., 2016).
Quantile regression provides one way of exposing this heterogeneity because, unlike conventional conditional-mean regression, it estimates relationships at different points of the dependent-variable distribution (Koenker & Bassett, 1978; Koenker, 2005). This is particularly relevant for household energy research because determinants operating among low-consuming households need not operate identically among median or high-consuming households. Kaza (2010), for example, demonstrates the analytical value of examining the spectrum of residential energy consumption rather than relying solely on mean effects, while Kostakis (2020) similarly employs quantile regression to identify heterogeneous socioeconomic determinants of household electricity use. The methodological rationale adopted in the underlying research follows precisely this distinction: quantile analysis makes it possible to investigate which factors become dominant within different consumption regimes rather than assuming one common relationship across the distribution.
The increasing use of machine learning introduces a related challenge. Predictive models may capture complex and non-linear relationships but can distance model performance from planning interpretation. Explainable Artificial Intelligence approaches such as SHAP provide mechanisms for identifying the relative contribution of variables to model predictions (Lundberg & Lee, 2017). Applications to urban electricity consumption already demonstrate the potential of SHAP-based approaches for understanding heterogeneous determinants (Wang et al., 2024). Yet variable importance should not automatically be equated with policy importance. A factor can be highly influential for prediction without constituting a directly modifiable planning variable.
The issue therefore moves beyond deciding which model is statistically “best”. For planning, the more consequential question is whether findings remain stable when the analytical lens changes—and, when they do not, what that instability tells us about the urban process being modelled.
Methodology: Comparing Evidence Across Methods, Scales and Contexts
The empirical analysis uses Ankara and Dublin as deliberately contrasting urban contexts. The comparison is not intended to treat the two cities as equivalent cases or to search for identical coefficients. Rather, their differences provide an analytical test of the contextual robustness of model-derived evidence. Ankara is characterised by a comparatively expansive and heterogeneous urban structure, while Dublin displays different development dynamics, housing typologies and socioeconomic spatial patterns. The original research therefore selected the two cases specifically to examine whether relationships between socioeconomic and built-environment characteristics and energy consumption persist or change across distinct urban settings.
A multi-scalar design links three analytical levels. At the micro scale, household surveys capture residential energy consumption alongside socioeconomic, demographic, dwelling and behavioural characteristics. At the meso scale, neighbourhoods in Ankara and Electoral Divisions in Dublin are differentiated through socioeconomic and spatial clustering. At the macro-comparative scale, the two cities provide contrasting contexts through which the transferability of relationships is evaluated. The thesis therefore deliberately integrates household-level mechanisms with neighbourhood context rather than treating energy consumption solely as an individual or building-level phenomenon.
Rather than relying on a single analytical technique, the study combines linear OLS regression, polynomial specifications, interaction models, Johnson–Neyman analysis, quantile regression, clustering and SHAP-based explainable machine learning. These approaches perform different analytical functions. OLS identifies average relationships; polynomial models test non-linearity; interaction and Johnson–Neyman analyses identify conditional relationships and ranges within which effects change; quantile regression examines heterogeneity across consumption levels; and SHAP identifies the relative contribution and dominance of predictors within more complex model structures. In the integrated analysis, household-level SHAP contributions, consumption regimes and meso-scale urban clusters are brought together to examine which mechanisms dominate, for whom and within which urban context.
Results: When the Analytical Lens Changes the Planning Story
The comparison demonstrates that residential energy consumption in both cities cannot be adequately explained through isolated variables or average effects alone. When linear OLS, polynomial and interaction models, Johnson–Neyman analysis, quantile regression and SHAP are considered together, energy consumption emerges as a heterogeneous and context-dependent process.
In Ankara, conventional OLS models identify dwelling size and household size as relatively strong and consistent determinants. Taken alone, this could support a relatively straightforward planning interpretation centred on housing scale and household composition. However, non-linear and interaction analyses alter this interpretation. Income and dwelling size exhibit threshold behaviour, while relationships between spatial and socioeconomic conditions become particularly relevant for winter energy consumption. The threshold analysis provides an especially clear example: below the identified income threshold the estimated income coefficient for summer energy consumption is positive, whereas above it the coefficient becomes negative. Thus, describing “income” as having one uniform effect would conceal a structurally different relationship between socioeconomic resources and energy demand across household conditions.
Quantile analysis adds another dimension by showing that determinants need not operate identically across low-, median- and high-consumption households. This is analytically important because a policy based on the estimated average household may fail to address either constrained low-consuming households or high-consuming households whose energy use is produced through different combinations of dwelling size, household characteristics and socioeconomic resources.
In Dublin, the analytical story differs. Household size emerges as one of the more stable determinants in the linear models, while insulation becomes particularly relevant to winter consumption and therefore presents a more direct built-environment intervention point. In contrast to Ankara, income-related threshold and interaction mechanisms are weaker and less continuous; demographic characteristics and the physical characteristics of the housing stock become comparatively more important.
The cross-city comparison therefore illustrates two forms of uncertainty simultaneously. First, methodological variation changes the evidence visible within each city: moving from average to non-linear, distributional and explainable-machine-learning analyses reveals mechanisms hidden by simpler specifications. Second, contextual variation changes the interpretation of nominally similar variables across cities. The same variable label does not necessarily represent the same planning mechanism.
The meso scale further complicates the picture productively. Combining neighbourhood clusters, quantile consumption regimes and SHAP dominance reveals that similar household characteristics can operate differently depending on their socio-spatial setting. In Dublin, for example, central, higher-density, apartment- and rental-oriented areas are dominated by low and medium consumption regimes, with dwelling size, income and demographic structure emerging as important explanatory mechanisms. This suggests that spatial context is not simply another independent variable; it structures the conditions within which household-level relationships are produced.
Navigating Analytical Uncertainty in Planning
These findings suggest that disagreement or variation between models should not automatically be interpreted as analytical failure. In complex urban systems, it may instead reveal heterogeneity that is itself planning-relevant. A mean-effect model answers a different question from a quantile model, while predictive importance identified by SHAP is conceptually different from a regression coefficient. The danger emerges when outputs from these different analytical logics are translated directly into universal policy prescriptions.
The paper therefore proposes a robustness-oriented framework for model-mediated planning evidence, structured around four sequential questions:
Consistency: Which relationships remain visible across different analytical methods?
Heterogeneity: Which apparently average relationships change across consumption groups, household profiles or thresholds?
Context: Which relationships persist across neighbourhood and city contexts, and which are context-specific?
Decision relevance: Which of the identified determinants correspond to variables that planning institutions can legitimately and effectively influence?
This framework reframes robustness. Robust evidence does not necessarily mean finding one coefficient that remains identical everywhere. Instead, robustness can mean understanding the conditions under which a relationship holds and the boundaries beyond which it should not be transferred.
This distinction has important implications for evidence-informed planning. Uniform energy-efficiency interventions risk assuming that similar households or neighbourhoods respond similarly to the same intervention. The multi-scalar results instead indicate that planning decisions should consider spatial structure, socioeconomic conditions and household behaviour jointly, and that energy-efficiency objectives should extend beyond isolated building interventions toward neighbourhood- and city-scale interpretation. This is also the central planning implication identified by the underlying doctoral research.
Conclusion
Data-intensive planning does not remove uncertainty simply by increasing the number or sophistication of models available to planners. Models actively structure how urban problems become visible. Average-effect models, non-linear specifications, distribution-sensitive approaches and explainable machine-learning techniques can produce different yet simultaneously valid representations of household energy consumption.
The Ankara–Dublin comparison demonstrates that these differences are neither statistical noise nor merely methodological inconvenience. They reveal that energy relationships are conditional on household characteristics, consumption regimes, spatial scale and urban context. The planning challenge is therefore not to eliminate analytical uncertainty by searching indefinitely for a single definitive model, but to make uncertainty interpretable.
For urban energy planning, this requires shifting the core question from “Which variable determines energy consumption?” toward a more conditional formulation:
Which mechanisms matter, for whom, under what socio-spatial conditions, at what scale, and with what degree of robustness across analytical perspectives?
Such an approach transforms analytical plurality from a source of confusion into a planning resource. By explicitly examining consistency, heterogeneity, context and decision relevance, model-mediated evidence can support interventions that are not only technically informed but also more context-sensitive and less dependent on assumptions about an “average” urban household.
References
Cayla, J.-M., Maïzi, N., & Marchand, C. (2011). The role of income in energy consumption behaviour: Evidence from French households data. Energy Policy, 39(12), 7874–7883.
Chen, H. C., Han, Q., & De Vries, B. (2020). Urban morphology indicator analyzes for urban energy modeling. Sustainable Cities and Society, 52, 101863.
Kaza, N. (2010). Understanding the spectrum of residential energy consumption: A quantile regression approach. Energy Policy, 38(11), 6574–6585.
Koenker, R. (2005). Quantile regression. Cambridge University Press.
Koenker, R., & Bassett, G. (1978). Regression quantiles. Econometrica, 46(1), 33–50.
Kostakis, I. (2020). Socio-demographic determinants of household electricity consumption: Evidence from Greece using quantile regression analysis. Current Research in Environmental Sustainability, 1, 23–30.
Lundberg, S. M., & Lee, S.-I. (2017). A unified approach to interpreting model predictions. Advances in Neural Information Processing Systems, 30.
Namazkhan, M., Albers, C., & Steg, L. (2020). A decision tree method for explaining household gas consumption: The role of building characteristics, socio-demographic variables, psychological factors and household behaviour. Renewable and Sustainable Energy Reviews, 119, 109542.
Newman, P. G., & Kenworthy, J. R. (1989). Cities and automobile dependence: An international sourcebook. Gower Publishing.
Owens, S. E. (1986). Energy, planning and urban form. Pion.
Robinson, J. (2016). Comparative urbanism: New geographies and cultures of theorizing the urban. International Journal of Urban and Regional Research, 40(1), 187–199.
Shammin, M. R., Herendeen, R. A., Hanson, M. J., & Wilson, E. J. (2010). A multivariate analysis of the energy intensity of sprawl versus compact living in the US for 2003. Ecological Economics, 69(12), 2363–2373.
Tian, W., Liu, Y., Heo, Y., Yan, D., Li, Z., An, J., & Yang, S. (2016). Relative importance of factors influencing building energy in urban environment. Energy, 111, 237–250.
Wang, Y., Hu, L., Hou, L., Wang, L., Chen, J., He, Y., & Su, X. (2024). A SHAP machine learning-based study of factors influencing urban residents' electricity consumption—Evidence from Chinese provincial data. Environment, Development and Sustainability, 26(12), 30445–30476.