Dealing with Excess Zeros and Highly Skewed Distributions in Nursing Research

• Activity Number: 677
• Type: Topic Contributed
• Date/Time: Thursday, August 8, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Consulting
• Author(s): Melinda Higgins*+
• Companies: Emory University
• Keywords: joint distributions ; mixture models ; zero-inflated ; ZIP ; ZINB ; Hurdle

Abstract

In nursing research, highly skewed data distributions are often encountered. An excess of zeros are readily seen in hospital/clinic frequencies and durations, as well as in symptom assessment surveys relative to the presence/absence of symptoms followed by frequency/severity ratings. Excessive high percentages (many 100 %'s) are also seen in adherence studies resulting in highly left-skewed distributions. Statistical models utilizing t-tests, chi-square, and logistic regression approaches will be compared with models using the Poisson, Negative Binomial, zero-inflated, hurdle and zero-altered forms of these distributions. Availability and "ease of use" of these models will be compared across multiple softwares. Utilizing these zero-inflated distributions enhances the power and ability to detect smaller changes in the longer but sparse tails of these highly skewed distributions. This presentation will illustrate the availability and utility of these powerful methods in nursing research with examples from recent nursing and related health research literature.