Applied statistical methods

All users of longitudinal data need to handle the issue of missing data in their research since some non-response is inevitable. Strategies for how to deal with missing data depend on the nature of non-response.

We know different types of people tend to drop out of our studies at different times, depending on their individual circumstances and characteristics, and so we can take account of that in the methods we use. We have developed approaches to deal with ‘missingness’ which capitalise on the rich data cohort members provided over the years, before they left the study, in order to deal with missing data and reduce bias.

We use well known methods such as multiple imputation, inverse probability weighting and full information maximum likelihood. These rely on the assumption that the data are missing at random (MAR), implying that systematic differences between the missing values and the observed values can be explained by observed data. Most studies employing MAR methods rely on a largely arbitrary selection of variables used as predictors of ‘missingness’, and the extent to which the plausibility of the MAR assumption is maximised for a given dataset is not known. We have implemented a systematic data-driven approach to predict non-response and have published a user guide which provides detailed guidance on how to adopt this approach in your own research using NCDS.

Causal inference in observational data is far from straightforward. However, the wealth of information we have collected about cohort members about the whole of their lives, and even about the circumstances of their birth, gives data users the opportunity to select rich controls for multivariable adjustment.

There are also circumstances in which causal identification may also be achieved with approaches such as instrumental variable modelling/Mendelian randomisation, regression discontinuity, and fixed effects/correlated random effects methods.

We are developing a programme of work on causal inference in our cohorts, which will use a range of methods. We will use directed acyclic graphs to represent assumptions about potential confounders, and apply techniques such as negative controls and simulations of unmeasured confounders to test the degree to which omitted variables might lead to bias.

Data from self-reported measures can be biased due to processes driven by cohort members’ personalities and circumstances. On the other hand, data from objective measures may also be affected by instrumental errors, for example the precision of the blood pressure device used by the nurse or laboratory variations in blood analysis. Additional sources of error arise when comparing data from multiple studies as there can be variation in how different groups interpret the same question and in response tendencies.

In our work on measurement error, we use the latest extensions of the generalised latent variable modelling framework to specify complex error structures. This allows us to investigate the properties of some key areas of measurement of our cohort data, including on physical health, mental health and cognition, and to establish within and between cohort equivalent measures.

Two ongoing projects funded by CLOSER investigate the measurement properties of mental health and cognitive ability measures in British birth cohorts.