Approximately half of all emergency department (ED) visits are primary-care sensitive (PCS) – meaning that they could potentially be avoided with timely, effective primary care. Reducing undesirable types of healthcare utilization (including PCS ED use) requires the ability to define, measure, and predict such use in a population. In this retrospective, observational study, we quantified ED use in 2 privately insured populations and developed ED risk prediction models. One dataset, obtained from a Massachusetts managed-care network (MCN), included data from 2009-11. The second was the MarketScan database, with data from 2007-08. The MCN study included 64,623 individuals enrolled for at least 1 base-year month and 1 prediction-year month in Massachusetts whose primary care provider (PCP) participated in the MCN. The MarketScan study included 15,136,261 individuals enrolled for at least 1 base-year month and 1 prediction-year month in the 50 US states plus DC, Puerto Rico, and the US Virgin Islands. We used medical claims to identify principal diagnosis codes for ED visits, and scored each according to the New York University Emergency Department algorithm. We defined primary-care sensitive (PCS) ED visits as those in 3 subcategories: nonemergent, emergent but primary-care treatable, and emergent but preventable/avoidable. We then: 1) defined and described the distributions of 3 ED outcomes: any ED use; number of ED visits; and a new outcome, based on the NYU algorithm, that we call PCS ED use; 2) built and validated predictive models for these outcomes using administrative claims data; 3) compared the performance of models predicting any ED use, number of ED visits, and PCS ED use; 4) enhanced these models by adding enrollee characteristics from electronic medical records, neighborhood characteristics, and payor/provider characteristics, and explored differences in performance between the original and enhanced models. In the MarketScan sample, 10.6% of enrollees had at least 1 ED visit, with about half of utilization scored as PCS. For the top risk group (those in the 99.5th percentile), the model’s sensitivity was 3.1%, specificity was 99.7%, and positive predictive value (PPV) was 49.7%. The model predicting PCS visits yielded sensitivity of 3.8%, specificity of 99.7%, and PPV of 40.5% for the top risk group. In the MCN sample, 14.6% (±0.1%) had at least 1 ED visit during the prediction period, with an overall rate of 18.8 (±0.2) visits per 100 persons and 7.6 (±0.1) PCS ED visits per 100 persons. Measuring PCS ED use with a threshold-based approach resulted in many fewer visits counted as PCS, discarding information unnecessarily. Out of 45 practices, 5 to 11 (11-24%) had observed values that were statistically significantly different from their expected values. Models predicting ED utilization using age, sex, race, morbidity, and prior use only (claims-based models) had lower R2 (ranging from 2.9% to 3.7%) and poorer predictive ability than the enhanced models that also included payor, PCP type and quality, problem list conditions, and covariates from the EMR, Census tract, and MCN provider data (enhanced model R2 ranged from 4.17% to 5.14%). In adjusted analyses, age, claims-based morbidity score, any ED visit in the base year, asthma, congestive heart failure, depression, tobacco use, and neighborhood poverty were strongly associated with increased risk for all 3 measures (all P<.001).