Transcript Slide 1
Small Area Estimates of Fuel Poverty in Scotland Phil Clarke (ONS), Ganka Mueller (Scottish Government) Fuel Poverty Indicator • A fuel poor household is one that would need to spend more than 10% of its income on adequate energy use • Drivers: Thermal properties of the dwelling Price (type of fuel) Household income Fuel Poverty Indicator: data sources • Detailed property survey: building physics model • Household interview: income • Scottish House Condition Survey (SHCS): annual ~ 3000 households 32 Local Authority – 3 year combined data Fuel Poverty Indicator: policy demand • Section 88 Housing (Scotland) Act 2001: establishes a duty to set out how the following objective will be achieved (by 2016): “… ensuring, so far as reasonably practicable, that persons do not live in fuel poverty.” • Policies and programmes: financial assistance and energy efficiency upgrades Targeting Interventions Design and delivery: • Area-based approach Home Energy Efficiency Programmes for Scotland (HEEPS): around £60 m annually; Energy Company Obligation (ECO): Carbon Saving Communities Obligation (CSCO) • Cost effectiveness Local area estimation – Good timing for investigation • Scottish House Conditions Survey can be used to directly estimate fuel poverty at local authority level • Below this level direct estimates too imprecise so there is a need to encompass model-based techniques • Recent availability of census local area statistics means that timing is good for such investigation. Small area estimation models • Framework is to associate sample survey indicators with publicly available administrative/census data for small areas. • Fit a statistical model to describe this association. • Use this model to derive more precise estimates for small areas. • Here we have : indicator = household fuel poverty status small areas = Scotland Intermediate Zones (~1230) Survey data available SHCS individual household records Two years of data giving 9121 respondents in 1231 of the total 1235 intermediate zones. Records coded by IZ. Records have binary indicator variable for household fuel poor or not Distribution of number of responding households by IZ Auxiliary data available at IZ level • For successful model based estimation auxiliary data variables should correlate well with quantity of interest. • For this investigation a set of variables from three main sources were considered -: Census 2011 : indicators relating to household social, employment, health and housing status; Scottish Neighbourhood Statistics : DWP benefit claimant rates and property council tax bandings; Dept. of Energy and Climate Change : energy consumption The underlying model • Fuel poverty status is a binary variable • So in modelling it we build a probability linking model • Let pid be the probability that a household i living in area d is fuel poor • The equation linking pid with auxiliary data xd is a multilevel logistic regression model: pid logit pid ln 1 pid T xd β ud ud ~ N 0, u2 As auxiliary data is all at area level the individual probability can be reinterpreted as the proportion of households in fuel poverty in the area and denoted pd Rationale for modelling • SHCS data is aggregated to IZ area giving sample proportions in fuel poverty. • Then merged with appropriate IZ auxiliary data variables. • First a null model is fitted with no auxiliary data : logit pd ud ud ~ N 0, 2 u • u is a measure of unexplained between area variability • Modelling then proceeds to fit a best set of auxiliary variables as explanatory variables 2 • As variables are fitted the value of u reduces. This value is a major determinant of precision of estimates. 2 Final fitted model Variables in final fitted model are : • Proportion of people aged 16 to 64 claiming income support • Proportion of people in households who are living as couples • Proportion of persons aged 16-74 whose NS-SEC is ‘managerial or professional’ • Average number of rooms per household • Proportion of properties built before 1919 • Average Economy7 domestic electricity consumption • Consumption of ordinary domestic electricity as a proportion of total domestic energy consumption • Interaction of “Consumption of ordinary domestic electricity as a proportion of total domestic energy consumption” with “Average number of rooms per household” Using model to determine estimates • Fitting a model determines parameter estimates ˆ , βˆ and ˆu2. Also estimates are made of the random effects uˆd . • These are then applied to each area’s auxiliary data to determine an estimate on the logistic scale, ˆ uˆ IZ estimate ˆ xTd β d 2 T ˆ )x • Mean squared error of estimate = ˆu xdVar (β d • Estimates on the logistic scale can then be back transformed to probability scale using the function : exp( x) /(1 exp( x)) Final fitted model – was it a success? • Model fitted the data well and passed technical diagnostics. • Reduction of u2 from null model = 79.7%. • Estimate precision has some drawbacks : Coefficients of variation can exceed 20%. Final fitted model – was it a success? Estimates and confidence intervals of proportion of households in fuel poverty. 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 Chart shows that vast majority of estimates between 20% and 35% with confidence intervals +/- 10 percentage points. Only about 140 IZs at bottom of range (11.4% of total) can be distinguished from same number at top of range. Estimates Estimates for central belt Achievements and conclusions • A methodology and a fully documented set of SAS programs have been written permitting further development. • A set of estimates of fuel poverty for Intermediate Zones with useful precision have been determined. • Due to estimates being alike over a large number of Intermediate Zones though, the precision measures are not sufficiently good for high discrimination. • The estimates though are sufficiently good for general scale categorisation of areas.