Using Hurdle Models to Analyze Zero-Inflated Count Data

Many outcomes in medical research are counts of some event. For example, heart rate is expressed as a count of beats per minute, or a patient may have more than one infection during an ICU stay, or a patient may be re-hospitalized a number of times following their index hospitalization. When the counts are large, like heart rate, they are analyzed as continuous. Some, like re-hospitalizations, are zero for most patients, and the counts are small for patients that are re-hospitalized. In this case, the data are said to be zero-inflated and do not follow a Poisson distribution. Hurdle models provide a simple way of addressing the issue. Dr. Kolm is Director of Biostatistics at Christiana Care Health System, Research Professor of Medicine at Thomas Jefferson University and Director of the Biostatistical Core of the Bridging Advanced Developments for Exceptional Rehabilitation (BADER) Consortium funded by the Department of Defense. He has over 30 years of experience in consulting with investigators in the design and analysis of clinical trials, retrospective and observational studies, and large patient registries. He has considerable experience in the application of general and generalized linear and hierarchical models, classification and tree regression, time-to-event analysis, multivariate analysis, cost-effectiveness analyses and multiple imputation methods for missing data.