Symposium: Analysing Cohort Data in Practice: The CLS User Guides
Handling Missing Data (Martina Narayanan, UCL Centre for Longitudinal Studies)
Non-response is common in longitudinal surveys. Missing values due to non-response mean less efficient estimates because of the reduced size of the analysis sample but also introduce the potential for bias since respondents are often systematically different from non-respondents. This talk will cover the CLS user guide on handling missing data. Content will include: background on the problem of missing data; description and illustration of the main statistical methods for dealing with missing data (multiple imputation, inverse probability weighting, and full information maximum likelihood); and the results of work to assess the performance of these methods in CLS data, along with advice on implementing the methods in practice.
Causal Inference (Liam Wright, UCL Centre for Longitudinal Studies)
Correlations between variables in cohort and other observational studies do not necessarily indicate causal effects and may instead reflect bias due to confounding, sample selection processes, or measurement issues. Thankfully, multiple statistical methods, frameworks, and research designs have been developed that (when used judiciously) can aid making causal inferences with observational data. This talk will cover the recent CLS user guide on causal inference. Content will include: background on the problem of causal inference in observational data; discussion of the attributes of CLS’ cohorts that support causal inference; outline of methods, research designs and a relatively novel epistemic approach (referred to as “Elaborate Theories”) that can help making causal inferences with CLS’ cohort data.
Handling Mode Effects (Richard Silverwood, UCL Centre for Longitudinal Studies)
Each of CLS’ cohorts has collected data using a mixture of survey ‘modes’ – e.g., surveying cohort members via face-to-face, telephone or video interview or with a web questionnaire. Mixing modes, either within or between sweeps of data collection, has a number of advantages, but can lead to mode effects – systematic differences in how people respond to survey items between modes. For instance, cohort members may more accurately report sensitive information in anonymous (e.g., web) vs. non-anonymous (e.g., face-to-face) modes. Mode effects can generate bias in all types of analysis of mixed-mode survey data and need accounting for in statistical analyses. This talk will cover the CLS user guide on handling mode effects. Content will include: background on the factors that lead to mode effects; discussion of the bias from mode effects using a simple-to-follow Causal Directed Acyclic Graph (DAG) framework; and outline of methods and recommendations for handling mode effects in analyses of CLS mixed mode data.
