Modeling Count Data Is Used For A Normal Distribution

710 Words3 Pages
Modeling count data Count data are frequently collected by social scientists. The number of drinks a student consumes, the number of pens an employee steals, and the number of trips to an emergency room are all examples of count data that are collected by psychologists. Researchers typically rely on ordinary least squares regression (OLS) to analyze these data. Unfortunately, OLS regression is usually inappropriate as count data are typically non-normal and heteroskedastic (Atkins & Gallop, 2007). In other words, the frequencies of these occurrences rarely exemplify the bell curve representing a normal distribution, often positively skewed with most frequencies stacked at or near zero, and the variances are unequal across groups. Most students do not drink, most employees do not steal pens, and most people do not visit the emergency room. Attempting to model associations of this sort violates fundamental assumptions of OLS regression. Poisson regression is uniquely equipped to handle count data, and zero-inflated models allow researchers to simultaneously model excess zeros as well as associations among key variables. Model testing One assumption of Poisson regression is that the dependent variable’s conditional mean should equal the variance. Overdispersion is a common concern regarding Poisson regression, and occurs when the conditional variance exceeds the conditional mean (Cameron & Trivedi, 2013). Failure to address overdispersion can result in inflated standard errors and t statistics. This means researchers and clinicians may obtain spurious results. Zero-inflated models should be employed in these circumstances as they examine the excess zeros within the logistic portions of the models, while simultaneously allowing rese... ... middle of paper ... ...alyses to account for variance in game day drinking attributed to date of attendance. Complex contrast codes were used to represent intervention condition in order to assess whether the gender-specific condition was more effective at reducing alcohol use than the gender-neutral condition. The first contrast represented the control group versus the gender-neutral and gender-specific conditions, and the second contrast represented the gender-neutral versus gender-specific conditions. Next, two contrast codes were specified in order to evaluate the efficacy of the gender-specific condition, relative to the control condition. The first contrast code examined the effects of the gender-specific and control conditions, relative to the gender-neutral condition. The second contrast code examined the effects of the gender-specific condition, compared to the control condition.
Open Document