Survey data quality in panel surveys: trade-offs between nonresponse and measurement errors
Longitudinal survey data are the source for many social scientists, economists, and policy makers who study or take policy decisions regarding poverty, income and well-being.
In the UK, research based on longitudinal surveys such as the British Household Panel Survey (BHPS), the Birth Cohort Studies and the English Longitudinal Study of Aging (ELSA) has over many years had a huge impact on policy debates and policy decisions. The accuracy of the data is, therefore, extremely important.
Whilst surveys are designed to exclude bias, two error sources are not easily controlled which in turn threaten to make panel data invalid and unreliable.
- Non-response; when respondents “drop out” of a longitudinal survey, this is particularly likely to lead to biased data.
- Measurement error; this might occur for specific groups of people. For example, people with lower cognitive abilities have more difficulty understanding questions, and remembering any information required to give a correct answer.
Answers to questions which ask people to estimate or guess, or where they are asked about desirable and non- desirable attitudes and behaviours will also be subject to inaccuracy and bias.
The project aims to:
improve data quality in panel surveys and lower survey costs
- enlarge knowledge about trade-offs between nonresponse and measurement error in panel survey data
- describe the extent of error for a range of variables in the BHPS
- develop practical guidelines for researchers to study the trade-off between non response and measurement errors
Data sources and methods
The first part of this project is devoted to developing models to separate the two types of error; there is normally no information on measurement error for those people that do not participate, so statistical models are used to overcome this problem.
Non-respondents in any wave can be compared to respondents in that wave. Using recent innovations in Structural Equation Modelling and Latent Class techniques, distinct non-response patterns can be related to reasons for dropout. The second part studies common causes of the two error sources, and identifies which respondent characteristics are responsible for them.
The final part studies the trade-offs or common causes between measurement error and nonresponse.
Associate Professor - University of Utrecht