Presentation on theme: "In search of a population sampling frame for UK postal surveys Gerry Nicolaas and Sarah Tipping."— Presentation transcript:
In search of a population sampling frame for UK postal surveys Gerry Nicolaas and Sarah Tipping
Acknowledgements Our sponsors: The Socio-economic Inequalities Branch (SEIB), Office for National Statistics Our collaborators at the Office for National Statistics: Amanda Wilmot, Abigail Dewar, Zeeshan Rahman, Colin Hand
Postcode Address File (PAF) List of all addresses to which mail is sent Small user file (less than 50 mail items per day) 10% of addresses are non-residential Almost complete coverage Most commonly used sampling frame for major surveys Sample of addresses, NOT individuals No names!
Electoral Register (ER) List of eligible adults aged 18+ registered to vote Can be used to sample individuals Includes names as well as addresses Coverage problems!
Coverage problems of ER: Does not include people who are not eligible to vote Does not include people who fail to register Does not include people who opt out of the publicly available version of the ER (the edited ER)
Commercial alternatives: CACIs Consumer Register Experians National Canvasse Equifax ConnectSelect EuroDirects Data Exchange NB. These databases were not developed for probability sampling of the general population.
CACIs Consumer Register: Edited Electoral Register Supplemented with data from Claritas, DataWorks, Bountys market leading mother and baby file, large mail order databases Updated quarterly Includes names and addresses Also includes information which could possibly be used for sampling, e.g. modelled age Claims to be the most comprehensive database Almost 40 million names and addresses of UK adults
Design of the study: Sample of 6,030 PAF addresses in GB selected for the ONS Omnibus Survey (April - June 2005) Names from CACIs Consumer Register matched to sampled PAF addresses Omnibus interviewers collected household grid information for responding cases, including first name, sex and age Manual matching of Consumer Register names with names from the household grid Omnibus household grid treated as correct
Not on CR On CR Incorrect name & address - 1,050 Individuals at responding addresses Correct name & address 2,350 4,436
Not on CR On CR Incorrect name & address - 1,050 Individuals at responding addresses Correct name & address 2,350 4,436 6,784 (35%) (65%)
Not on CR On CR Incorrect name & address - 1,050 (19%) Individuals at responding addresses Correct name & address 2,350 4,436 (81%) 5,486
Characteristics of adults aged 18+ not included on the CR: Sex distribution is not significantly different More likely to be aged More likely to be renting their accommodation More likely to have a degree
If sampling adults aged 18+ from Consumer Register: 35% of all adults aged 18+ in GB are not included on the CR Under-representation of specific groups may bias survey results 19% of adults sampled from CR will not be found at address
BUT…….. Is the Consumer Register a better sampling frame for general population surveys than the edited Electoral Register? NB. 86% of adults listed on the Consumer Register were also listed on the edited Electoral Register
Not on edited ER On edited ER Incorrect name & address Individuals at responding addresses Correct name & address 2,929 3,857
Not on edited ER On edited ER Incorrect name & address Individuals at responding addresses Correct name & address 2,929 3,857 6,786 (43%) (57%)
Not on edited ER On edited ER Incorrect name & address (18%) Individuals at responding addresses Correct name & address 2,929 3,857 (82%) 4,718
Not found at address Coverage of adults aged 18+ in GB Edited ER % Sample of individuals Consumer Register % 19 65
Conclusion: CR has greater coverage of GB adults than edited ER BUT not much greater (57% on edited ER and 65% on CR) no improvement in sample representativeness
So the search continues……. Other commercial databases? Combine commercial databases? Mixed mode? Personalised and unpersonalised mailings? ……….
Limitation of study: We have assumed that the matching rate at the individual level was the same for responding and non- responding households BUT this doesnt have impact on comparison between PAF, edited ER and CR estimated coverage rate is likely to be an overestimate
For further information: Gerry Nicolaas tel Amanda Wilmot tel