In an analysis of longitudinal data, you find variables which change over time (e.g. marital status) and those that do not (e.g. date of birth). I was forced to think about the place of time-varying covariates in longitudinal data analysis yesterday by a seminar presented by Nick Parr about recent fertility trends in Australia. He used a longitudinal dataset called HILDA, which tracks women over time.
His work used marital status (among many other things) as a predictor for childbearing: married women are more likely to experience a birth than single, divorced, separated or widowed women. So far, so understandable. However, even though changes in marital status were recorded, the change itself was not taken into account in the model. It’s plausible to suggest that a change from single to married increases the likelihood of a woman experiencing a birth in the period (say) 2-3 years after the marriage. Equally, a change from married to separated might be associated with a greatly reduced likelihood of experiencing a birth in the two year period after the separation (all other things being equal) – this was the point raised by another academic, Maire Ni Bhrolchain, in discussion after the seminar.
In longitudinal data analysis, a status change could be more important than the status itself.