OBJECTIVES: To use baseline results from a prospective observational study in Alzheimer’s disease (AD) to evaluate methods for dealing with missing longitudinal AD cost data. METHODS: GERAS is an 18-month observational study of costs associated with AD. Total societal costs included patient healthcare costs (including hospitalizations, outpatient visits, and medication) and social care costs (including home-care and day-center sessions), and caregiver informal care costs (from time spend on informal care). Missing longitudinal cost data due to patient death/institutionalization was classified as not missing at random (NMAR). Cost data missing for other reasons was classified as missing at random (MAR) or missing completely at random (MCAR). To assess the impact of imputing missing longitudinal cost data, patterns of missing data during follow-up were simulated based on baseline GERAS data to generate 10%, 20%, 30% and 40% missing data for MCAR, MAR and NMAR classifications. Naïve methods (including complete case analysis, mean imputation and regression models), multiple imputation (MI) and a fixed cost were applied to each dataset and %bias assessed using (estimated-actual)/actual cost*100. RESULTS: Total baseline societal costs were available for 1488 (99.4%) of enrolled patients, with a mean monthly cost of €2101(95% CI: €1980-€2222). For MCAR datasets, naïve methods performed as well as MI (20% missing data:  0.6-10.9% bias naïve methods vs 0.2-6.1% MI). For MAR data, MI methods performed better (-3.2% to -14.3% bias) than naïve methods (6.6%-18.0% bias). All approaches were consistently poor with NMAR data (bias range -31.4% to -38.6%); the best performing approach was to impute a fixed value (monthly cost of institutionalisation) with -22.6% bias. For all approaches %bias increased with missing data volume. CONCLUSIONS: Methods used to impute missing cost data in AD should be tailored depending on the type of missing data, using sensitivity analysis to assess the impact of any assumptions.