Zero-Inflated Poisson Regression to Analyze Lengths of Hospital Stays Adjusting for Intra-Center Correlation

• Published in: Communications in statistics. Simulation and computation Vol. 34; no. 1; pp. 235 - 241
• Main Author: SONG, JAMES X.
• Format: Journal Article
• Language: English
• Published: Colchester Taylor & Francis Group 23.02.2005; Taylor & Francis
• Subjects: Applications; Correlated length of hospital stays; Exact sciences and technology; Generalized estimating equations; Insurance, economics, finance; Mathematics; Medical sciences; Multi-center clinical trial; Numerical analysis; Numerical analysis. Scientific computation; Numerical methods in probability and statistics; Probability and statistics; Probability theory and stochastic processes; Sciences and techniques of general use; Special processes (renewal theory, markov renewal processes, semi-markov processes, statistical mechanics type models, applications); Statistics; Zero-inflated model; Economics; Correlation; Statistical analysis; Error estimation; Correlated length of hospital stays; Multi-center clinical trial; ZIP model; Poisson process; Error type I; Statistical method; Statistical parameter; Correlation analysis; Generalized equation; Zero-inflated model; Clinical trial; Estimating equation; Generalized estimating equations
• ISSN: 0361-0918; 1532-4141
• DOI: 10.1081/SAC-200047118

In the anti-infective clinical trials, in addition to the efficacy and safety outcomes, health economic outcomes such as length of hospital stays (LOS) and number of hours missed from work are usually compared between treatment groups. Since excess zeros are often exhibited in the LOS data, a zero-inflated Poisson (ZIP) model is adopted to model such data. In a multi-center trial, as patients from the same center are often treated by the same doctor and have similar socioeconomic backgrounds, correlation among subjects' LOS within centers may exist, and is commonly termed the intracluster correlation (ICC). Ignoring such intracluster correlation in statistical analysis leads to erroneous parameter estimates and usually inflated Type I error. To adjust for the intracluster variations, the generalized estimating equations (GEE) method is introduced to the ZIP model. The GEEs have consistent solutions even when the dependence is misspecified. The proposed model in this study provides an extension to the regular ZIP model when analyzing correlated LOS data.