OBJECTIVES: Multimorbid individuals consume a disproportionally large share of health care resources. Usually standard (1-component) regression techniques are applied to analyse costs in samples of patients with multiple chronic conditions. However, the patient specific number and combination of co-occurring single diseases results in inhomogeneous data leading to biased estimates when using traditional regression techniques. In this study we analyse health care costs in a sample of patients suffering from multimorbidity using a more elaborate approach to address this heterogeneity. METHODS: We used a subsample of N=1050 patients from  a multicentre prospective cohort study of multimorbid primary care patients aged 65 to 85 years in Germany who completed a questionnaire on healthcare utilization covering a 6-month-period. We applied a finite mixture of generalized linear models, which belongs to the group of statistical learning algorithms, in order to control for unobserved heterogeneity of patient level health care costs focussing on the identification of multimorbidity patterns. RESULTS: We detected four different groups of patients with regard to total costs. The effect of the presence of an additional disease on costs differs between these groups. Two diametrically opposed cost trends were detected with respect to the number of co-occurring diseases. While in one group containing hypertension, joint arthrosis, diabetes, gout, anxiety and lower limb varicosis cost increased with the number of co-occurring diseases, in a second group including severe hearing loss, asthma/COPD, osteoporosis, neuropathies, Parkinson’s disease and chronic ischemic heart disease cost decreased. Diversities between groups were also found in the results indicated by diametrically opposed influence of single diseases. CONCLUSIONS: Our results indicate existing unobserved heterogeneity in costs among patients suffering from multimorbidity with different combinations of single diseases which would remain unconsidered using standard regression techniques. Especially different costs trends were detected with regard to the number and nature of co-existing diseases.