Objectives: We developed and validated claims-based prediction models for transitions from the community to nursing home in an elderly population without dementia. We sought to compare three models: model 1 included prescription drug class and disease conditions variables, model 2 excluded prescription drug class variables, and model 3 excluded disease condition variables. Methods: The study sample was a retrospective cohort of 454,656 elderly Medicare beneficiaries with employer-sponsored supplemental health insurance. We developed models for predicting the probability of nursing home admission within six months after a baseline year of no nursing home admissions or diagnosis of dementia, using a combination of literature-based risk factors for transitions, stepwise logistic regression, and Akaike's information criteria. A split-sample approach was used to assess reliability of final models. Model discrimination was evaluated using the C-statistic. Model calibration was measured by using the Hosmer and Lemeshow test to assess Chi-square goodness-of-fit of the model and then inspecting residuals and checking the existence of influential data points. Results: In addition to age, sex, geographic region, insurance type, prior hospitalization and number of prescriptions, the final prediction model 1 for beneficiaries without dementia contained 36 co-morbidities and 16 drug categories. The C-statistics of model 1, model 2, and model 3 were 0.83, 0.81, and 0.82, respectively. The Hosmer and Lemeshow goodness-of-fit tests for the models were not significant except for the model 3. In each case, less than 5% of standardized residuals had a value outside the range [-1.96; 1.96]. No influential points were found in any of the models. Conclusion: Prediction models using administrative claims can be a valuable screening tool for identifying beneficiaries who are at high risk of nursing home admission. Reliable prediction models for nursing home admission can be based on data that include or exclude drugs/disease information.