Rasch modelling to deal with changes in the questionnaires used during long-term follow-up of cohort studies: a simulation study

BACKGROUND A specific measurement issue often occurs in cohort studies with long-term follow-up: the substitution of the classic instruments used to assess one or several factors or outcomes studied by new, more reliable, more accurate or more convenient instruments. This study aimed to compare three techniques to deal with this issue when the substituted instrument is a questionnaire measuring a subjective phenomenon: one using only the items shared by the different questionnaires over time, i.e. computation of the raw score; the two others using every item, i.e. computation of the standardised score or estimation of the latent variable score using the Rasch model. METHODS Two hundred databases were simulated, corresponding to longitudinal 10-item questionnaire data from three trajectory groups of subjects for the subjective phenomenon of interest ("increasing", "stable-low" or "stable-high" mean trajectory over time). Three copies of these databases were generated and the subjects' responses to some items were removed at some collection times leading to a number of shared items over time varying from 4 to 10 in the 800 datasets. The performances of Latent Class Growth Analysis (LCGA) applied to the raw score, the standardised score or the latent variable score were studied on these databases according to the number of shared items over time. RESULTS Surprisingly, LCGA applied to the latent variable score estimate did not perform as well as LCGA applied to the standardised score, where it was the most efficient whatever the number of shared items. However, the proportions of correctly classified subjects by LCGA applied to the latent variable score were more balanced across trajectory groups. CONCLUSIONS The use of the standardised score to deal with questionnaire changes over time was more efficient than the raw score and also, surprisingly, than the latent variable score. LCGA applied to the raw score was the least efficient and exhibited the most unbalanced misclassifications across trajectory groups. As prospective longitudinal studies with long-term follow-up are more and more common, researchers should be aware of this phenomenon and should reconsider the use of the raw score when changes in the questionnaires used occurred during follow-up.