Small Area Estimation of Unemployment for Cross Cutting
Total Page:16
File Type:pdf, Size:1020Kb
Small Area Estimation of Unemployment for Cross Cutting Geographies in Great Britain Denise Silva1; Philip Clarke1; Bob Watson1; Nicholas Misoulis2 1ONS – Office for National Statistics, Segensworth Road, Titchfield, PO15 5RR, UK. 2National Housing and Planning Advice Unit- NHPAU - UK. E-mail: [email protected] Sessions CPMs 1. Introduction Since 1999 ONS has been continuously devoting efforts to develop and improve a small area estimation methodology to provide local area estimates of unemployment based on annual Labour Force Survey(LFS)1 data. The survey is a key source of labour market statistics but it is not able to deliver direct estimates of 1 The Labour Force Survey is a continuous survey, with a sample of around 60,000 households in each three-month period, and measures unemployment according to the International Labour Organisation (ILO) definition. unemployment with adequate precision for every local authority (LA) 2 ,or lower level geographies as parliamentary constituencies (PCs)3, in Great Britain due to small sample size in many areas. The model based approach relies on determining a strong relationship between unemployment (as measured by the LFS according to ILO definition) and auxiliary information. The main source of this auxiliary information is the number of beneficiaries of job seekers allowance known as claimant count (CC). The small area estimation model at LA level is an area level logistic mixed model that relates the probability of being unemployed for an individual of a particular sex and age group within a LA with the corresponding CC information, incorporating additional explanatory variables accounting for age, sex and regional differentials. The LA random effects capture unexplained sources of variation and area heterogeneity that Sessions CPMs may not be explained by the auxiliary data. The model formulation is given by: ⎧ p ⎫ exp()X T β + u logit()p = ln di =X T β +usuch that= di d ( 1) di ⎨ − ⎬ di d di + ()T + ⎩1 pdi ⎭ 1 expX di β ud where is the probability that an individual in (age-sex) group = from LA d is unemployed, pdi i 1 ,K , 6 Xdi 2 Local Authorities are the main tier of local government in UK. There are 434 LAs with an average population of 140,000. However they vary widely in size from around 60,000 to over 1 million people. 3 The UK is currently divided into 646 PCAs, each of which is represented by one MP in the House of Commons. is the column vector of group indicators and covariates for age-sex group i in LA d, β is the vector of fixed ∼ φ effect coefficients and uNd (0, ) is the random area effect. The model covariates are indicators of age- sex groups (male/female for age groups: 16 to 24; 25 to 49; 50 and over), of the 12 government office regions4 of GB and of the ONS socioeconomic family classification for LAs plus the logit of the CC proportion in each age-sex group within each LA and the logit of the CC in the LA. An estimate of the unemployment level in LA d is given by where ysdi is the LFS sample count of number of unemployed, 6 6 θˆ = ˆ θ =y{}+() N − nˆp Ndi and ndi are d∑i=1 di ∑i=1 sdi di di di the population and sample size for group i in LA d ()− and Ndi di n ˆp diis the predicted value for yrdi - the LFS non-sample count of number of unemployed obtained from the modelling procedure fitted just on the sample data. To ensure that the model based Sessions estimates are consistent with the LFS published estimates at high geographical levels, they are constrained to CPMs the corresponding direct LFS estimates of unemployment. 2. Unemployment Estimates for Non-Nested Small Areas ONS publishes annual LA model based estimates every quarter for unemployment and unemployment rates5 4 The nine Government Office Regions (GORs) are the primary statistical subdivisions of England and each GOR covers a number of local authorities (http://www.statistics.gov.uk/geography/glossary/g.asp). 5 Local area labour markets: statistical indicators – (www.statistics.gov.uk/STATBASE/Product.asp?vlnk=14160) and in addition there is a need for reliable labour market unemployment data for PCs. There are 35 parliamentary constituencies which are identical areas to local authorities and there are also LA areas which are coterminous to a number of PCs. This is a case in which estimates are requested for two non-nested geographies. The original model developed for LA estimation was taken as the basis for PC model which in turn is supplied with the same sort of survey and administrative data (social benefits) for the constituencies. The results show good performance but with less explanatory power than the local authority model. Issues arise from having alternative estimates for the same areas from the LA and the PC models where areas coincide and also because PC model estimates may not add up to LA model based published figures for cases Sessions where several PCs exactly make up one LA (this affects 48 LAs and 150 PCs). Where a PC is the same area CPMs as a LA, the estimation procedure constrains the model based PC estimate to be the same as the model based estimate for the LA. Moreover, PC estimates are also calibrated to LA figures for all other LA that are conterminous with multiple PCs. In addition, PC estimates have to be constrained to direct survey estimates of regional totals. 3. Results The new PC estimates will be published in 2009 and results for the period April 2006 to March 2007 are presented in the graph below. The PC estimates align with the LA model based estimates and are calibrated at regional level to the LFS direct estimates by age-sex groups. Unemployment Rates Estimates and 95% Confidence Intervals for Parliamentary Constituencies in GB - April06 to March07 25 Sessions 20 CPMs 15 10 Unemployment Rate (%) 5 0 0 50 100 150 200 250 300 350 400 450 500 550 600 650 Parliamentary Constituencies Ordered by Unemployment Rate The development of this project exposed the duality between producing estimates based on statistical procedures that are optimised for a given geography and the need for calibrating/constraining the estimates for presentation purposes. Consistency is an important feature of any statistical system and the establishment of a small area estimation framework certainly has to address the issue of consistency between model based estimates produced for different geographies besides the consistency of model based estimates and direct surveys estimates published for higher/broader geographies. REFERENCES [1] Ambler, R., Caplan, D., Chambers, R., Kovacevic, M., Wang, S. (2001) Combining unemployment benefits data and LFS data to estimate ILO unemployment for small areas: An application of a Sessions modified Fay-Herriot method. Proceedings of the 53th ISI Session, Seoul. CPMs [2] Clarke, P., McGrath, K., Chandra, H. and Tzavidis, N. (2007) Developments in Small Area Estimation in UK with Focus on Current Research Activities. Paper presented at IASS Satellite Conference on Small Area Estimation. 56th ISI Session. [3] Hastings, D., Maine, N., Brown, G. Cruddas, M. (2003). Development of improved estimation methods for local area unemployment levels and rates. Labour Market Trends, vol. 111, no 1. ONS publication. [4] Saei, S., Chambers, R. (2003) Small Area Estimation Under Linear and Generalized Linear Mixed Models with Time and Area Effects. S3RI Methodology Working Paper M03/15, University of Southampton. Sessions CPMs .