Introduction Although there is evidence for distinct behavioural sub-phenotypes in Alzheimer's disease (AD), their inter-relationships and the effect of clinical variables on their expression have been little investigated. Methods We have analysed a sample of 1850 probable AD patients from the UK and Greece with 10 item Neuropsychiatric Inventory (NPI) data. We applied a Multiple Indicators Multiple Causes (MIMIC) approach to investigate the effect of MMSE, disease duration, gender, age and age of onset on the structure of a four-factor model consisting of “psychosis”, “moods”, “agitation” and “behavioural dyscontrol”. Results Specific clinical variables predicted the expression of individual factors. When the inter-relationship of factors is modelled, some previously significant associations are lost. For example, lower MMSE scores predict psychosis, agitation and behavioural dyscontrol factors, but psychosis and mood predict the agitation factor. Taking these associations into account MMSE scores did not predict agitation. Conclusions The complexity of the inter-relations between symptoms, factors and clinical variables is efficiently captured by this MIMIC model. ; PUBLISHED ; We are grateful for funding from the Alzheimer's Research Trust, the MRC Centre for Neurodegeneration Research, the NIHR BRC Centre for Mental Health at the South London and Maudsley NHS Foundation Trust and Institute of Psychiatry, KCL, the Alzheimer's Society, Stewart Bequest and Ulster Garden Villages.