ONS report: New approach for producing population estimates by ethnic group

PGRG Blog #2, September 2017

The Office for National Statistics (ONS) publishes detailed population estimates by ethnic group for areas in England and Wales following each census. However, there are currently no reliable population estimates by ethnic group available at the local authority level for the years since the 2011 census. Given the user interest in regular, local authority ethnicity data, ONS has been investigating alternative methods to meet this need.

Ludi Simpson (University of Manchester) discusses the ONS report –

Population estimates with ethnic group dimension

Two methods to estimate sub-national ethnic group populations were released by ONS 25th August 2017. For further information see here . The article requests feedback to pop.info@ons.gov.uk. The two are a ‘simple method based on APS’ and a ‘GSPREE modelling using APS and other data sources’.

Which methods should be considered?

The utility of the outputs and the accuracy of preliminary outputs is poor for both the proposed methods (see below). Is it wise to restrict the discussion to only these methods? Choice of method for the 2010s and beyond should include consideration of:

  • Cohort component methods, as attempted in the 2000s by ONS, and by Leeds University in the 2010s, to reflect each group’s age composition: its inherent ‘demographic potential’ from the balance of child-bearing and mortality.
  • Modelling with multiple datasets but with more simple assumptions than GSPREE, including IPF or SPREE.
  • Use of further datasets, including name-analysis of patient register or other administrative records that contribute to the SPD.

The 2011 Census output provides a benchmark that promising methods should approximate closely. What other way of validating results is possible?

An evaluation of methods requires aims

Is the aim to approximately replicate the experience of the Census so far? This would seem to be a sensible criterion and is used in other evaluations of new methods.

  • The GSPREE method is assessed against the Census. It would be helpful if the simple APS method was similarly assessed.
  • The detail of output of the two methods released is much broader than the information required by providers of services, and by policy analysts of discrimination. Consultation during past decades suggests these users require (a) age breakdown (eg children; populations of young adults and of working age, elderly, …), (b) detailed ethnic group (because Pakistani not equal to Indian, African not equal to Black Caribbean), and (c) estimates of small populations (since equalities policies are as relevant to Wrexham as to Westminster).
  • What is the target accuracy? Accuracy when compared to the 2011 Census outturn is one outcome discussed in the GSPREE report. How accurate is good enough?

ONS Method 1 – Population estimates of ethnic groups for LADs for 2016, based on APS and MYE

The new estimates provide estimates of the total population of 6 broad ethnic groups (White British, All other White, Mixed, Asian, Black, Other, to the nearest 1000, for the 348 Local Authority Districts (LAD) in England and Wales. Where the APS sample is not big enough the result is ‘na’; this occurs for 2016 in 32% of the minority ethnic group-district combinations (551 of the 348*5= 1740 possible estimates).

Concerns with the method:

  • To complete the missing estimates, why not substitute an estimate based on a sample of several years, using more years until the sample reaches a minimum acceptable level? These time-averaged estimates for each ethnic group could be scaled so that their total makes the national total of LADs for the ethnic group consistent with a nationally-estimated total. This would recognise the utility of estimates of small populations.
  • The report on GSPREE estimates using administrative data and the APS concludes that “To obtain a valid input to the model going forward, either the APS data would need to be optimised for the collection of ethnicity data or an alternative source of data would be required.” Doesn’t the same conclusion hold for these estimates of the 2016 population?
  • The population not in households is maintained as the 2011 Census proportion of the total population. Official household projections assume a constant 2011 Census number up to age 74, and a constant proportion from age 75. This introduces an inconsistency between official statistics series, which could be considered and avoided by adjusting one or the other.
  • The proportion of the population not in households for minority groups is highly influenced by the number of overseas students. The assumption that the ethnic group population will move in the same way in households and not in households is unlikely to hold for any group, because the factors affecting numbers of overseas students only marginally affect the household population, which has a different dynamic of change. University Districts could be dealt with separately, keeping these institutional populations constant rather than changing over time.

Method 2 –  Ethnic group population estimates from administrative records using GSPREE modelling

The GSPREE method is an extension of SPREE, itself a probability model based on the more commonly known Iterative Proportional Fitting. In this case GSPREE has been used to combine ethnic group population estimates from the APS survey data with the English School Census and 2001 Census, to estimate the total population of six broad ethnic groups in each LAD of England. The research outputs are a 2011 population total for each ethnic group in each LAD, which have been compared to the 2011 Census equivalents.

Concerns with the method:

  • The comparison against the 2011 Census shows errors frequently more than 10%, often much greater than 25%. This is a level of error in the total population which suggests failure in the aim to achieve usable outputs.
  • The account suggests that only the White population is well estimated.
  • It is difficult to judge GSPREE as promising without comparison with other methods, given the poor accuracy of its results in this case.

Conclusion

Social policy requires ethnic group population estimates with a degree of detail and accuracy that go beyond the current estimates and proposals from ONS.

Ludi Simson
University of Manchester
Sept 5th 2017