A cautionary note regarding count models of alcohol consumption in randomized controlled trials

Abstract

Background: Alcohol consumption is commonly used as a primary outcome in randomized alcohol treatment studies. The distribution of alcohol consumption is highly skewed, particularly in subjects with alcohol dependence.

Predictors of Depressive Symptoms Among Israeli Jews and Arabs During the Al Aqsa Intifada: A Population-Based Cohort Study

“…Because the distribution of the total depression scale score from the five items of the PHQ-9 was highly skewed and similar to that of a count variable, the Poisson distribution was considered for modeling the outcome, rather than the normal distribution. The depression scale score also exhibited overdispersion (i.e., variance greater than the mean, contrary to the assumption of equal mean and variance inherent in the Poisson distribution), as well as more zero values than would be expected under the Poisson distribution (39% of the depression scale scores were zero), so we also considered using a negative binomial model and zero-inflated models (19)(20)(21). Likelihood ratio and Vuong tests were used to compare models and select the most appropriate model, which turned out to be a zero-inflated negative binomial model (22).…”

“…Zero-inflated negative binomial models incorporate 2 separate models: a logistic model evaluating the association between covariates and having a non-zero versus a zero value for the outcome of interest, and a negative binomial model evaluating the association between covariates and a non-zero outcome following the negative binomial distribution (23). We present results as incidence rate ratios, which are calculated by exponentiating the coefficient estimates from the negative binomial model (20) and are interpreted as the risk associated with a particular characteristic relative to a reference group.…”

A Predictive Microsimulation Model to Estimate the Clinical Relevance of Reducing Alcohol Consumption in Alcohol Dependence

“…The model used English hospitalisation data to derive the incidence of alcohol-attributable diseases and injuries in the general population. The relative risks of such clinical events for each patient in the model, based on their simulated alcohol consumption and the incidence of events in the general population, were derived from published international epidemiological studies measuring the association between level of alcohol consumption and the risk of alcohol-attributable injuries or diseases [11,40,41,42,43,44,45,46,47,48,49,50,51]. A summary of the main inputs to the microsimulation model is shown in figure 1.…”

Analysing count data of Butterflies communities in Jasin, Melaka: A Poisson regression analysis

“…In addition, previous researcher [9] proposed a cautionary note regarding count models of alcohol consumption in randomized controlled trials. In this paper, the researchers considered Poisson, overdispersed Poisson, NB, ZIP and ZINB.…”

“…Zeroinflated models estimate two equations simultaneously; one for the count model and the other one are the excess zeros which are caused by the real ecological effect of interest. The analysis data with accessing high zero comprised the model of Poisson [3][4][5], Negative Binomial (NB) regression, zeroInflated Poisson (ZIP) [6][7][8][9], and Zero-Inflated Negative Binomial (ZINB) is widely used [10].…”